DiscoverLatent Space: The AI Engineer Podcast — Practitioners talking LLMs, CodeGen, Agents, Multimodality, AI UX, GPU Infra and all things Software 3.0
Latent Space: The AI Engineer Podcast — Practitioners talking LLMs, CodeGen, Agents, Multimodality, AI UX, GPU Infra and all things Software 3.0
Claim Ownership

Latent Space: The AI Engineer Podcast — Practitioners talking LLMs, CodeGen, Agents, Multimodality, AI UX, GPU Infra and all things Software 3.0

Author: Alessio + swyx

Subscribed: 350Played: 7,679
Share

Description

The podcast by and for AI Engineers! In 2023, over 1 million visitors came to Latent Space to hear about news, papers and interviews in Software 3.0.

We cover Foundation Models changing every domain in Code Generation, Multimodality, AI Agents, GPU Infra and more, directly from the founders, builders, and thinkers involved in pushing the cutting edge. Striving to give you both the definitive take on the Current Thing down to the first introduction to the tech you'll be using in the next 3 months! We break news and exclusive interviews from OpenAI, tiny (George Hotz), Databricks/MosaicML (Jon Frankle), Modular (Chris Lattner), Answer.ai (Jeremy Howard), et al.

Full show notes always on https://latent.space

www.latent.space
95 Episodes
Reverse
Agents @ Work: Lindy.ai

Agents @ Work: Lindy.ai

2024-11-1501:09:53

Alessio will be at AWS re:Invent next week and hosting a casual coffee meetup on Wednesday, RSVP here! And subscribe to our calendar for our Singapore, NeurIPS, and all upcoming meetups!We are still taking questions for our next big recap episode! Submit questions and messages on Speakpipe here for a chance to appear on the show!If you've been following the AI agents space, you have heard of Lindy AI; while founder Flo Crivello is hesitant to call it "blowing up," when folks like Andrew Wilkinson start obsessing over your product, you're definitely onto something.In our latest episode, Flo walked us through Lindy's evolution from late 2022 to now, revealing some design choices about agent platform design that go against conventional wisdom in the space.The Great Reset: From Text Fields to RailsRemember late 2022? Everyone was "LLM-pilled," believing that if you just gave a language model enough context and tools, it could do anything. Lindy 1.0 followed this pattern:* Big prompt field ✅* Bunch of tools ✅* Prayer to the LLM gods ✅Fast forward to today, and Lindy 2.0 looks radically different. As Flo put it (~17:00 in the episode): "The more you can put your agent on rails, one, the more reliable it's going to be, obviously, but two, it's also going to be easier to use for the user."Instead of a giant, intimidating text field, users now build workflows visually:* Trigger (e.g., "Zendesk ticket received")* Required actions (e.g., "Check knowledge base")* Response generationThis isn't just a UI change - it's a fundamental rethinking of how to make AI agents reliable. As Swyx noted during our discussion: "Put Shoggoth in a box and make it a very small, minimal viable box. Everything else should be traditional if-this-then-that software."The Surprising Truth About Model LimitationsHere's something that might shock folks building in the space: with Claude 3.5 Sonnet, the model is no longer the bottleneck. Flo's exact words (~31:00): "It is actually shocking the extent to which the model is no longer the limit. It was the limit a year ago. It was too expensive. The context window was too small."Some context: Lindy started when context windows were 4K tokens. Today, their system prompt alone is larger than that. But what's really interesting is what this means for platform builders:* Raw capabilities aren't the constraint anymore* Integration quality matters more than model performance* User experience and workflow design are the new bottlenecksThe Search Engine Parallel: Why Horizontal Platforms Might WinOne of the spiciest takes from our conversation was Flo's thesis on horizontal vs. vertical agent platforms. He draws a fascinating parallel to search engines (~56:00):"I find it surprising the extent to which a horizontal search engine has won... You go through Google to search Reddit. You go through Google to search Wikipedia... search in each vertical has more in common with search than it does with each vertical."His argument: agent platforms might follow the same pattern because:* Agents across verticals share more commonalities than differences* There's value in having agents that can work together under one roof* The R&D cost of getting agents right is better amortized across use casesThis might explain why we're seeing early vertical AI companies starting to expand horizontally. The core agent capabilities - reliability, context management, tool integration - are universal needs.What This Means for BuildersIf you're building in the AI agents space, here are the key takeaways:* Constrain First: Rather than maximizing capabilities, focus on reliable execution within narrow bounds* Integration Quality Matters: With model capabilities plateauing, your competitive advantage lies in how well you integrate with existing tools* Memory Management is Key: Flo revealed they actively prune agent memories - even with larger context windows, not all memories are useful* Design for Discovery: Lindy's visual workflow builder shows how important interface design is for adoptionThe Meta LayerThere's a broader lesson here about AI product development. Just as Lindy evolved from "give the LLM everything" to "constrain intelligently," we might see similar evolution across the AI tooling space. The winners might not be those with the most powerful models, but those who best understand how to package AI capabilities in ways that solve real problems reliably.Full Video PodcastFlo’s talk at AI Engineer SummitChapters* 00:00:00 Introductions * 00:04:05 AI engineering and deterministic software * 00:08:36 Lindys demo* 00:13:21 Memory management in AI agents * 00:18:48 Hierarchy and collaboration between Lindys * 00:21:19 Vertical vs. horizontal AI tools * 00:24:03 Community and user engagement strategies * 00:26:16 Rickrolling incident with Lindy * 00:28:12 Evals and quality control in AI systems * 00:31:52 Model capabilities and their impact on Lindy * 00:39:27 Competition and market positioning * 00:42:40 Relationship between Factorio and business strategy * 00:44:05 Remote work vs. in-person collaboration * 00:49:03 Europe vs US Tech* 00:58:59 Testing the Overton window and free speech * 01:04:20 Balancing AI safety concerns with business innovation Show Notes* Lindy.ai* Rick Rolling* Flo on X* TeamFlow* Andrew Wilkinson* Dust* Poolside.ai* SB1047* Gathertown* Sid Sijbrandij* Matt Mullenweg* Factorio* Seeing Like a StateTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol.ai.Swyx [00:00:12]: Hey, and today we're joined in the studio by Florent Crivello. Welcome.Flo [00:00:15]: Hey, yeah, thanks for having me.Swyx [00:00:17]: Also known as Altimore. I always wanted to ask, what is Altimore?Flo [00:00:21]: It was the name of my character when I was playing Dungeons & Dragons. Always. I was like 11 years old.Swyx [00:00:26]: What was your classes?Flo [00:00:27]: I was an elf. I was a magician elf.Swyx [00:00:30]: Well, you're still spinning magic. Right now, you're a solo founder and CEO of Lindy.ai. What is Lindy?Flo [00:00:36]: Yeah, we are a no-code platform letting you build your own AI agents easily. So you can think of we are to LangChain as Airtable is to MySQL. Like you can just pin up AI agents super easily by clicking around and no code required. You don't have to be an engineer and you can automate business workflows that you simply could not automate before in a few minutes.Swyx [00:00:55]: You've been in our orbit a few times. I think you spoke at our Latent Space anniversary. You spoke at my summit, the first summit, which was a really good keynote. And most recently, like we actually already scheduled this podcast before this happened. But Andrew Wilkinson was like, I'm obsessed by Lindy. He's just created a whole bunch of agents. So basically, why are you blowing up?Flo [00:01:16]: Well, thank you. I think we are having a little bit of a moment. I think it's a bit premature to say we're blowing up. But why are things going well? We revamped the product majorly. We called it Lindy 2.0. I would say we started working on that six months ago. We've actually not really announced it yet. It's just, I guess, I guess that's what we're doing now. And so we've basically been cooking for the last six months, like really rebuilding the product from scratch. I think I'll list you, actually, the last time you tried the product, it was still Lindy 1.0. Oh, yeah. If you log in now, the platform looks very different. There's like a ton more features. And I think one realization that we made, and I think a lot of folks in the agent space made the same realization, is that there is such a thing as too much of a good thing. I think many people, when they started working on agents, they were very LLM peeled and chat GPT peeled, right? They got ahead of themselves in a way, and us included, and they thought that agents were actually, and LLMs were actually more advanced than they actually were. And so the first version of Lindy was like just a giant prompt and a bunch of tools. And then the realization we had was like, hey, actually, the more you can put your agent on Rails, one, the more reliable it's going to be, obviously, but two, it's also going to be easier to use for the user, because you can really, as a user, you get, instead of just getting this big, giant, intimidating text field, and you type words in there, and you have no idea if you're typing the right word or not, here you can really click and select step by step, and tell your agent what to do, and really give as narrow or as wide a guardrail as you want for your agent. We started working on that. We called it Lindy on Rails about six months ago, and we started putting it into the hands of users over the last, I would say, two months or so, and I think things really started going pretty well at that point. The agent is way more reliable, way easier to set up, and we're already seeing a ton of new use cases pop up.Swyx [00:03:00]: Yeah, just a quick follow-up on that. You launched the first Lindy in November last year, and you were already talking about having a DSL, right? I remember having this discussion with you, and you were like, it's just much more reliable. Is this still the DSL under the hood? Is this a UI-level change, or is it a bigger rewrite?Flo [00:03:17]: No, it is a much bigger rewrite. I'll give you a concrete example. Suppose you want to have an agent that observes your Zendesk tickets, and it's like, hey, every time you receive a Zendesk ticket, I want you to check my knowledge base, so it's like a RAG module and whatnot, and then answer the ticket. The way it used to work with Lindy before was, you would type the prompt asking it to do that. You check my knowledge base, and so on and so forth. The problem with doing that is that it can always go wrong. You're praying the LLM gods that they will actually invoke y
Agents @ Work: Dust.tt

Agents @ Work: Dust.tt

2024-11-1101:00:06

We are recording our next big recap episode and taking questions! Submit questions and messages on Speakpipe here for a chance to appear on the show!Also subscribe to our calendar for our Singapore, NeurIPS, and all upcoming meetups!In our first ever episode with Logan Kilpatrick we called out the two hottest LLM frameworks at the time: LangChain and Dust. We’ve had Harrison from LangChain on twice (as a guest and as a co-host), and we’ve now finally come full circle as Stanislas from Dust joined us in the studio.After stints at Oracle and Stripe, Stan had joined OpenAI to work on mathematical reasoning capabilities. He describes his time at OpenAI as "the PhD I always wanted to do" while acknowledging the challenges of research work: "You're digging into a field all day long for weeks and weeks, and you find something, you get super excited for 12 seconds. And at the 13 seconds, you're like, 'oh, yeah, that was obvious.' And you go back to digging." This experience, combined with early access to GPT-4's capabilities, shaped his decision to start Dust: "If we believe in AGI and if we believe the timelines might not be too long, it's actually the last train leaving the station to start a company. After that, it's going to be computers all the way down."The History of DustDust's journey can be broken down into three phases:* Developer Framework (2022): Initially positioned as a competitor to LangChain, Dust started as a developer tooling platform. While both were open source, their approaches differed – LangChain focused on broad community adoption and integration as a pure developer experience, while Dust emphasized UI-driven development and better observability that wasn’t just `print` statements.* Browser Extension (Early 2023): The company pivoted to building XP1, a browser extension that could interact with web content. This experiment helped validate user interaction patterns with AI, even while using less capable models than GPT-4.* Enterprise Platform (Current): Today, Dust has evolved into an infrastructure platform for deploying AI agents within companies, with impressive metrics like 88% daily active users in some deployments.The Case for Being HorizontalThe big discussion for early stage companies today is whether or not to be horizontal or vertical. Since models are so good at general tasks, a lot of companies are building vertical products that take care of a workflow end-to-end in order to offer more value and becoming more of “Services as Software”. Dust on the other hand is a platform for the users to build their own experiences, which has had a few advantages:* Maximum Penetration: Dust reports 60-70% weekly active users across entire companies, demonstrating the potential reach of horizontal solutions rather than selling into a single team.* Emergent Use Cases: By allowing non-technical users to create agents, Dust enables use cases to emerge organically from actual business needs rather than prescribed solutions.* Infrastructure Value: The platform approach creates lasting value through maintained integrations and connections, similar to how Stripe's value lies in maintaining payment infrastructure. Rather than relying on third-party integration providers, Dust maintains its own connections to ensure proper handling of different data types and structures.The Vertical ChallengeHowever, this approach comes with trade-offs:* Harder Go-to-Market: As Stan talked about: "We spike at penetration... but it makes our go-to-market much harder. Vertical solutions have a go-to-market that is much easier because they're like, 'oh, I'm going to solve the lawyer stuff.'"* Complex Infrastructure: Building a horizontal platform requires maintaining numerous integrations and handling diverse data types appropriately – from structured Salesforce data to unstructured Notion pages. As you scale integrations, the cost of maintaining them also scales. * Product Surface Complexity: Creating an interface that's both powerful and accessible to non-technical users requires careful design decisions, down to avoiding technical terms like "system prompt" in favor of "instructions." The Future of AI PlatformsStan initially predicted we'd see the first billion-dollar single-person company in 2023 (a prediction later echoed by Sam Altman), but he's now more focused on a different milestone: billion-dollar companies with engineering teams of just 20 people, enabled by AI assistance.This vision aligns with Dust's horizontal platform approach – building the infrastructure that allows small teams to achieve outsized impact through AI augmentation. Rather than replacing entire job functions (the vertical approach), they're betting on augmenting existing workflows across organizations.Full YouTube EpisodeChapters* 00:00:00 Introductions* 00:04:33 Joining OpenAI from Paris* 00:09:54 Research evolution and compute allocation at OpenAI* 00:13:12 Working with Ilya Sutskever and OpenAI's vision* 00:15:51 Leaving OpenAI to start Dust* 00:18:15 Early focus on browser extension and WebGPT-like functionality* 00:20:20 Dust as the infrastructure for agents* 00:24:03 Challenges of building with early AI models* 00:28:17 LLMs and Workflow Automation* 00:35:28 Building dependency graphs of agents* 00:37:34 Simulating API endpoints* 00:40:41 State of AI models* 00:43:19 Running evals* 00:46:36 Challenges in building AI agents infra* 00:49:21 Buy vs. build decisions for infrastructure components* 00:51:02 Future of SaaS and AI's Impact on Software* 00:53:07 The single employee $1B company race* 00:56:32 Horizontal vs. vertical approaches to AI agentsTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol.ai.Swyx [00:00:11]: Hey, and today we're in a studio with Stanislas, welcome.Stan [00:00:14]: Thank you very much for having me.Swyx [00:00:16]: Visiting from Paris.Stan [00:00:17]: Paris.Swyx [00:00:18]: And you have had a very distinguished career. It's very hard to summarize, but you went to college in both Ecopolytechnique and Stanford, and then you worked in a number of places, Oracle, Totems, Stripe, and then OpenAI pre-ChatGPT. We'll talk, we'll spend a little bit of time about that. About two years ago, you left OpenAI to start Dust. I think you were one of the first OpenAI alum founders.Stan [00:00:40]: Yeah, I think it was about at the same time as the Adept guys, so that first wave.Swyx [00:00:46]: Yeah, and people really loved our David episode. We love a few sort of OpenAI stories, you know, for back in the day, like we're talking about pre-recording. Probably the statute of limitations on some of those stories has expired, so you can talk a little bit more freely without them coming after you. But maybe we'll just talk about, like, what was your journey into AI? You know, you were at Stripe for almost five years, there are a lot of Stripe alums going into OpenAI. I think the Stripe culture has come into OpenAI quite a bit.Stan [00:01:11]: Yeah, so I think the buses of Stripe people really started flowing in, I guess, after ChatGPT. But, yeah, my journey into AI is a... I mean, Greg Brockman. Yeah, yeah. From Greg, of course. And Daniela, actually, back in the days, Daniela Amodei.Swyx [00:01:27]: Yes, she was COO, I mean, she is COO, yeah. She had a pretty high job at OpenAI at the time, yeah, for sure.Stan [00:01:34]: My journey started as anybody else, you're fascinated with computer science and you want to make them think, it's awesome, but it doesn't work. I mean, it was a long time ago, it was like maybe 16, so it was 25 years ago. Then the first big exposure to AI would be at Stanford, and I'm going to, like, disclose a whole lamb, because at the time it was a class taught by Andrew Ng, and there was no deep learning. It was half features for vision and a star algorithm. So it was fun. But it was the early days of deep learning. At the time, I think a few years after, it was the first project at Google. But you know, that cat face or the human face trained from many images. I went to, hesitated doing a PhD, more in systems, eventually decided to go into getting a job. Went at Oracle, started a company, did a gazillion mistakes, got acquired by Stripe, worked with Greg Buckman there. And at the end of Stripe, I started interesting myself in AI again, felt like it was the time, you had the Atari games, you had the self-driving craziness at the time. And I started exploring projects, it felt like the Atari games were incredible, but there were still games. And I was looking into exploring projects that would have an impact on the world. And so I decided to explore three things, self-driving cars, cybersecurity and AI, and math and AI. It's like I sing it by a decreasing order of impact on the world, I guess.Swyx [00:03:01]: Discovering new math would be very foundational.Stan [00:03:03]: It is extremely foundational, but it's not as direct as driving people around.Swyx [00:03:07]: Sorry, you're doing this at Stripe, you're like thinking about your next move.Stan [00:03:09]: No, it was at Stripe, kind of a bit of time where I started exploring. I did a bunch of work with friends on trying to get RC cars to drive autonomously. Almost started a company in France or Europe about self-driving trucks. We decided to not go for it because it was probably very operational. And I think the idea of the company, of the team wasn't there. And also I realized that if I wake up a day and because of a bug I wrote, I killed a family, it would be a bad experience. And so I just decided like, no, that's just too crazy. And then I explored cybersecurity with a friend. We're trying to apply transformers to cut fuzzing. So cut fuzzing, you have kind of an algorithm that goes really fast and tries to mutate the inputs of a library to find bugs. And we tried to apply a transformer to that and do reinforcement learning with
Apologies for lower audio quality; we lost recordings and had to use backup tracks. Our guests today are Anastasios Angelopoulos and Wei-Lin Chiang, leads of Chatbot Arena, fka LMSYS, the crowdsourced AI evaluation platform developed by the LMSys student club at Berkeley, which became the de facto standard for comparing language models. Arena Elo is often more cited than MMLU scores to many folks, and they have attracted >1,000,000 people to cast votes since its launch, leading top model trainers to cite them over their own formal academic benchmarks:The Limits of Static BenchmarksWe’ve done two benchmarks episodes: Benchmarks 101 and Benchmarks 201. One issue we’ve always brought up with static benchmarks is that 1) many are getting saturated, with models scoring almost perfectly on them 2) they often don’t reflect production use cases, making it hard for developers and users to use them as guidance. The fundamental challenge in AI evaluation isn't technical - it's philosophical. How do you measure something that increasingly resembles human intelligence? Rather than trying to define intelligence upfront, Arena let users interact naturally with models and collect comparative feedback. It's messy and subjective, but that's precisely the point - it captures the full spectrum of what people actually care about when using AI.The Pareto Frontier of Cost vs IntelligenceBecause the Elo scores are remarkably stable over time, we can put all the chat models on a map against their respective cost to gain a view of at least 3 orders of magnitude of model sizes/costs and observe the remarkable shift in intelligence per dollar over the past year:This frontier stood remarkably firm through the recent releases of o1-preview and price cuts of Gemini 1.5:The Statistics of SubjectivityIn our Benchmarks 201 episode, Clémentine Fourrier from HuggingFace thought this design choice was one of shortcomings of arenas: they aren’t reproducible. You don’t know who ranked what and what exactly the outcome was at the time of ranking. That same person might rank the same pair of outputs differently on a different day, or might ask harder questions to better models compared to smaller ones, making it imbalanced. Another argument that people have brought up is confirmation bias. We know humans prefer longer responses and are swayed by formatting - Rob Mulla from Dreadnode had found some interesting data on this in May:The approach LMArena is taking is to use logistic regression to decompose human preferences into constituent factors. As Anastasios explains: "We can say what components of style contribute to human preference and how they contribute." By adding these style components as parameters, they can mathematically "suck out" their influence and isolate the core model capabilities.This extends beyond just style - they can control for any measurable factor: "What if I want to look at the cost adjusted performance? Parameter count? We can ex post facto measure that." This is one of the most interesting things about Arena: You have a data generation engine which you can clean and turn into leaderboards later. If you wanted to create a leaderboard for poetry writing, you could get existing data from Arena, normalize it by identifying these style components. Whether or not it’s possible to really understand WHAT bias the voters have, that’s a different question.Private EvalsOne of the most delicate challenges LMSYS faces is maintaining trust while collaborating with AI labs. The concern is that labs could game the system by testing multiple variants privately and only releasing the best performer. This was brought up when 4o-mini released and it ranked as the second best model on the leaderboard:But this fear misunderstands how Arena works. Unlike static benchmarks where selection bias is a major issue, Arena's live nature means any initial bias gets washed out by ongoing evaluation. As Anastasios explains: "In the long run, there's way more fresh data than there is data that was used to compare these five models." The other big question is WHAT model is actually being tested; as people often talk about on X / Discord, the same endpoint will randomly feel “nerfed” like it happened for “Claude European summer” and corresponding conspiracy theories:It’s hard to keep track of these performance changes in Arena as these changes (if real…?) are not observable.The Future of EvaluationThe team's latest work on RouteLLM points to an interesting future where evaluation becomes more granular and task-specific. But they maintain that even simple routing strategies can be powerful - like directing complex queries to larger models while handling simple tasks with smaller ones.Arena is now going to expand beyond text into multimodal evaluation and specialized domains like code execution and red teaming. But their core insight remains: the best way to evaluate intelligence isn't to simplify it into metrics, but to embrace its complexity and find rigorous ways to analyze it. To go after this vision, they are spinning out Arena from LMSys, which will stay as an academia-driven group at Berkeley.Full Video PodcastChapters* 00:00:00 - Introductions* 00:01:16 - Origin and development of Chatbot Arena* 00:05:41 - Static benchmarks vs. Arenas* 00:09:03 - Community building* 00:13:32 - Biases in human preference evaluation* 00:18:27 - Style Control and Model Categories* 00:26:06 - Impact of o1* 00:29:15 - Collaborating with AI labs* 00:34:51 - RouteLLM and router models* 00:38:09 - Future of LMSys / ArenaShow Notes* Anastasios Angelopoulos* Anastasios' NeurIPS Paper Conformal Risk Control* Wei-Lin Chiang* Chatbot Arena* LMSys* MTBench* ShareGPT dataset* Stanford's Alpaca project* LLMRouter* E2B* DreadnodeTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, Partner and CTO in Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol.ai.Swyx [00:00:14]: Hey, and today we're very happy and excited to welcome Anastasios and Wei Lin from LMSys. Welcome guys.Wei Lin [00:00:21]: Hey, how's it going? Nice to see you.Anastasios [00:00:23]: Thanks for having us.Swyx [00:00:24]: Anastasios, I actually saw you, I think at last year's NeurIPS. You were presenting a paper, which I don't really super understand, but it was some theory paper about how your method was very dominating over other sort of search methods. I don't remember what it was, but I remember that you were a very confident speaker.Anastasios [00:00:40]: Oh, I totally remember you. Didn't ever connect that, but yes, that's definitely true. Yeah. Nice to see you again.Swyx [00:00:46]: Yeah. I was frantically looking for the name of your paper and I couldn't find it. Basically I had to cut it because I didn't understand it.Anastasios [00:00:51]: Is this conformal PID control or was this the online control?Wei Lin [00:00:55]: Blast from the past, man.Swyx [00:00:57]: Blast from the past. It's always interesting how NeurIPS and all these academic conferences are sort of six months behind what people are actually doing, but conformal risk control, I would recommend people check it out. I have the recording. I just never published it just because I was like, I don't understand this enough to explain it.Anastasios [00:01:14]: People won't be interested.Wei Lin [00:01:15]: It's all good.Swyx [00:01:16]: But ELO scores, ELO scores are very easy to understand. You guys are responsible for the biggest revolution in language model benchmarking in the last few years. Maybe you guys want to introduce yourselves and maybe tell a little bit of the brief history of LMSysWei Lin [00:01:32]: Hey, I'm Wei Lin. I'm a fifth year PhD student at UC Berkeley, working on Chatbot Arena these days, doing crowdsourcing AI benchmarking.Anastasios [00:01:43]: I'm Anastasios. I'm a sixth year PhD student here at Berkeley. I did most of my PhD on like theoretical statistics and sort of foundations of model evaluation and testing. And now I'm working 150% on this Chatbot Arena stuff. It's great.Alessio [00:02:00]: And what was the origin of it? How did you come up with the idea? How did you get people to buy in? And then maybe what were one or two of the pivotal moments early on that kind of made it the standard for these things?Wei Lin [00:02:12]: Yeah, yeah. Chatbot Arena project was started last year in April, May, around that. Before that, we were basically experimenting in a lab how to fine tune a chatbot open source based on the Llama 1 model that I released. At that time, Lama 1 was like a base model and people didn't really know how to fine tune it. So we were doing some explorations. We were inspired by Stanford's Alpaca project. So we basically, yeah, grow a data set from the internet, which is called ShareGPT data set, which is like a dialogue data set between user and chat GPT conversation. It turns out to be like pretty high quality data, dialogue data. So we fine tune on it and then we train it and release the model called V2. And people were very excited about it because it kind of like demonstrate open way model can reach this conversation capability similar to chat GPT. And then we basically release the model with and also build a demo website for the model. People were very excited about it. But during the development, the biggest challenge to us at the time was like, how do we even evaluate it? How do we even argue this model we trained is better than others? And then what's the gap between this open source model that other proprietary offering? At that time, it was like GPT-4 was just announced and it's like Cloud One. What's the difference between them? And then after that, like every week, there's a new model being fine tuned, released. So even until still now, right? And then we have that demo website for V2 now. And then we thought like, okay, maybe we can add a few more of the model as well, like API model as well. And then we quickly realized that
How NotebookLM Was Made

How NotebookLM Was Made

2024-10-2501:13:572

If you’ve listened to the podcast for a while, you might have heard our ElevenLabs-powered AI co-host Charlie a few times. Text-to-speech has made amazing progress in the last 18 months, with OpenAI’s Advanced Voice Mode (aka “Her”) as a sneak peek of the future of AI interactions (see our “Building AGI in Real Time” recap). Yet, we had yet to see a real killer app for AI voice (not counting music).Today’s guests, Raiza Martin and Usama Bin Shafqat, are the lead PM and AI engineer behind the NotebookLM feature flag that gave us the first viral AI voice experience, the “Deep Dive” podcast:The idea behind the “Audio Overviews” feature is simple: take a bunch of documents, websites, YouTube videos, etc, and generate a podcast out of them. This was one of the first demos that people built with voice models + RAG + GPT models, but it was always a glorified speech-to-text. Raiza and Usama took a very different approach:* Make it conversational: when you listen to a NotebookLM audio there are a ton of micro-interjections (Steven Johnson calls them disfluencies) like “Oh really?” or “Totally”, as well as pauses and “uh…”, like you would expect in a real conversation. These are not generated by the LLM in the transcript, but they are built into the the audio model. See ~28:00 in the pod for more details. * Listeners love tension: if two people are always in agreement on everything, it’s not super interesting. They tuned the model to generate flowing conversations that mirror the tone and rhythm of human speech. They did not confirm this, but many suspect the 2 year old SoundStorm paper is related to this model.* Generating new insights: because the hosts’ goal is not to summarize, but to entertain, it comes up with funny metaphors and comparisons that actually help expand on the content rather than just paraphrasing like most models do. We have had listeners make podcasts out of our podcasts, like this one.This is different than your average SOTA-chasing, MMLU-driven model buildooor. Putting product and AI engineering in the same room, having them build evals together, and understanding what the goal is lets you get these unique results. The 5 rules for AI PMsWe always focus on AI Engineers, but this episode had a ton of AI PM nuggets as well, which we wanted to collect as NotebookLM is one of the most successful products in the AI space:1. Less is more: the first version of the product had 0 customization options. All you could do is give it source documents, and then press a button to generate. Most users don’t know what “temperature” or “top-k” are, so you’re often taking the magic away by adding more options in the UI. Since recording they added a few, like a system prompt, but those were features that users were “hacking in”, as Simon Willison highlighted in his blog post.2. Use Real-Time Feedback: they built a community of 65,000 users on Discord that is constantly reporting issues and giving feedback; sometimes they noticed server downtime even before the Google internal monitoring did. Getting real time pings > aggregating user data when doing initial iterations. 3. Embrace Non-Determinism: AI outputs variability is a feature, not a bug. Rather than limiting the outputs from the get-go, build toggles that you can turn on/off with feature flags as the feedback starts to roll in.4. Curate with Taste: if you try your product and it sucks, you don’t need more data to confirm it. Just scrap that and iterate again. This is even easier for a product like this; if you start listening to one of the podcasts and turn it off after 10 seconds, it’s never a good sign. 5. Stay Hands-On: It’s hard to build taste if you don’t experiment. Trying out all your competitors products as well as unrelated tools really helps you understand what users are seeing in market, and how to improve on it.Chapters00:00 Introductions01:39 From Project Tailwind to NotebookLM09:25 Learning from 65,000 Discord members12:15 How NotebookLM works18:00 Working with Steven Johnson23:00 How to prioritize features25:13 Structuring the data pipelines29:50 How to eval34:34 Steering the podcast outputs37:51 Defining speakers personalities39:04 How do you make audio engaging?45:47 Humor is AGI51:38 Designing for non-determinism53:35 API when?55:05 Multilingual support and dialect considerations57:50 Managing system prompts and feature requests01:00:58 Future of NotebookLM01:04:59 Podcasts for your codebase01:07:16 Plans for real-time chat01:08:27 Wrap upShow Notes* Notebook LM* AI Test Kitchen* Nicholas Carlini* Steven Johnson* Wealth of Nations* Histories of Mysteries by Andrej Karpathy* chicken.pdf Threads* Area 120* Raiza Martin* Usama Bin ShafqatTranscriptNotebookLM [00:00:00]: Hey everyone, we're here today as guests on Latent Space. It's great to be here, I'm a long time listener and fan, they've had some great guests on this show before. Yeah, what an honor to have us, the hosts of another podcast, join as guests. I mean a huge thank you to Swyx and Alessio for the invite, thanks for having us on the show. Yeah really, it seems like they brought us here to talk a little bit about our show, our podcast. Yeah, I mean we've had lots of listeners ourselves, listeners at Deep Dive. Oh yeah, we've made a ton of audio overviews since we launched and we're learning a lot. There's probably a lot we can share around what we're building next, huh? Yeah, we'll share a little bit at least. The short version is we'll keep learning and getting better for you. We're glad you're along for the ride. So yeah, keep listening. Keep listening and stay curious. We promise to keep diving deep and bringing you even better options in the future. Stay curious.Alessio [00:00:52]: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO at Residence at Decibel Partners. And I'm joined by my co-host, Swyx, founder of Smol.ai.Swyx [00:01:01]: Hey, and today we're back in the studio with our special guest, Raiza Martin. And Raiza, I forgot to get your last name, Shafqat.Raiza [00:01:10]: Yes.Swyx [00:01:10]: Okay, welcome.Raiza [00:01:12]: Hello, thank you for having us.Swyx [00:01:14]: So AI podcasters meet human podcasters, always fun. Congrats on the success of Notebook LM. I mean, how does it feel?Raiza [00:01:22]: It's been a lot of fun. A lot of it, honestly, was unexpected. But my favorite part is really listening to the audio overviews that people have been making.Swyx [00:01:29]: Maybe we should do a little bit of intros and tell the story. You know, what is your path into the sort of Google AI org? Or maybe, actually, I don't even know what org you guys are in.Raiza [00:01:39]: I can start. My name is Raisa. I lead the Notebook LM team inside of Google Labs. So specifically, that's the org that we're in. It's called Google Labs. It's only about two years old. And our whole mandate is really to build AI products. That's it. We work super closely with DeepMind. Our entire thing is just, like, try a bunch of things and see what's landing with users. And the background that I have is, really, I worked in payments before this, and I worked in ads right before, and then startups. I tell people, like, at every time that I changed orgs, I actually almost quit Google. Like, specifically, like, in between ads and payments, I was like, all right, I can't do this. Like, this is, like, super hard. I was like, it's not for me. I'm, like, a very zero-to-one person. But then I was like, okay, I'll try. I'll interview with other teams. And when I interviewed in payments, I was like, oh, these people are really cool. I don't know if I'm, like, a super good fit with this space, but I'll try it because the people are cool. And then I really enjoyed that, and then I worked on, like, zero-to-one features inside of payments, and I had a lot of fun. But then the time came again where I was like, oh, I don't know. It's like, it's time to leave. It's time to start my own thing. But then I interviewed inside of Google Labs, and I was like, oh, darn. Like, there's definitely, like—Alessio [00:02:48]: They got you again.Raiza [00:02:49]: They got me again. And so now I've been here for two years, and I'm happy that I stayed because especially with, you know, the recent success of Notebook LM, I'm like, dang, we did it. I actually got to do it. So that was really cool.Usama [00:03:02]: Kind of similar, honestly. I was at a big team at Google. We do sort of the data center supply chain planning stuff. Google has, like, the largest sort of footprint. Obviously, there's a lot of management stuff to do there. But then there was this thing called Area 120 at Google, which does not exist anymore. But I sort of wanted to do, like, more zero-to-one building and landed a role there. We were trying to build, like, a creator commerce platform called Kaya. It launched briefly a couple years ago. But then Area 120 sort of transitioned and morphed into Labs. And, like, over the last few years, like, the focus just got a lot clearer. Like, we were trying to build new AI products and do it in the wild and sort of co-create and all of that. So, you know, we've just been trying a bunch of different things. And this one really landed, which has felt pretty phenomenal. Really, really landed.Swyx [00:03:53]: Let's talk about the brief history of Notebook LM. You had a tweet, which is very helpful for doing research. May 2023, during Google I.O., you announced Project Tailwind.Raiza [00:04:03]: Yeah.Swyx [00:04:03]: So today is October 2024. So you joined October 2022?Raiza [00:04:09]: Actually, I used to lead AI Test Kitchen. And this was actually, I think, not I.O. 2023. I.O. 2022 is when we launched AI Test Kitchen, or announced it. And I don't know if you remember it.Swyx [00:04:23]: That's how you, like, had the basic prototype for Gemini.Raiza [00:04:26]: Yes, yes, exactly. Lambda.Swyx [00:04:28]: Gave beta access to people.Raiza [00:04:29]: Yeah, yeah, yeah. And I remember, I was like, wo
Singapore's GovTech is hosting an AI CTF challenge with ~$15,000 in prizes, starting October 26th, open to both local and virtual hackers. It will be hosted on Dreadnode's Crucible platform; signup here!It is common to say if you want to work in AI, you should come to San Francisco. Not everyone can. Not everyone should. If you can only do meaningful AI work in one city, then AI has failed to generalize meaningfully.As non-Americans working in the US, we know what it’s like to see AI progress so rapidly here, and yet be at a loss for what our home countries can do. Through Latent Space we’ve tried to tell the story of AI outside of the Bay Area bubble; we talked to Notion in New York and Humanloop and Wondercraft in London and HuggingFace in Paris and ICLR in Vienna, and the Reka, RWKV, and Winds of AI Winter episodes were taped in Singapore (the World’s Fair also had Latin America representation and we intend to at least add China, Japan, and India next year).The Role of Government with AIAs an intentionally technical resource, we’ve mostly steered clear of regulation and safety debates on the podcast; whether it is safety bills or technoalarmism, often at the cost of our engagement numbers or ability to book big name guests with a political agenda. When SOTA shifts 3x faster than it takes to pass a law, when nobody agrees on definitions of important things, when you can elicit never-before-seen behavior by slightly different prompting or sampling, it is hard enough to simply keep up to speed, so we are happy limiting our role to that. The story of AI progress has more often been achieved in the private sector, usually in spite of, rather than with thanks to, government intervention.But industrial policy is inextricably linked to the business of AI, which we do very much care about, has an explicitly accelerationist intent if not impact, and has a track record of success in correcting for legitimate market failures in private sector investment, particularly outside of the US. It is with this lens we approach today’s episode and special guest, our first with a sitting Cabinet member.Singapore’s National AI StrategyIt is well understood that much of Singapore’s economic success is attributable to industrial policy, from direct efforts like the Jurong Town Corporation industrialization to indirect ones like going all in on English as national first language. Singapore’s National AI Strategy grew out of its 2014 Smart Nation initiative, first launched in 2019 and then refreshed in 2023 by Minister Josephine Teo, our guest today.While Singapore is not often thought of as an AI leader, the National University ranks in the top 10 in publications (above Oxford/Harvard!), and many overseas Singaporeans work at the leading AI companies and institutions in the US (and some of us even run leading AI Substacks?). OpenAI has often publicly named the Singapore government as their model example of government collaborator and is opening an office in Singapore in time for DevDay 2024.AI Engineer NationsSwyx first pitched the AI Engineer Nation concept at a private Sovereign AI summit featuring Dr. He Ruimin, Chief AI Officer of Singapore, which eventually led to an invitation to discuss the concept with Minister Teo, the country’s de-facto minister for tech (she calls it Digital Development, for good reasons she explains in the pod).This chat happened (with thanks to Jing Long, Joyce, and other folks from MDDI)!The central pitch for any country, not just Singapore, to emphasize and concentrate bets on AI Engineers, compared with other valuable efforts like training more researchers, releasing more government-approved data, or offering more AI funding, is a calculated one, based on the fact that: * GPU clusters and researchers have massive returns to scale and colocation, mostly concentrated in the US, that are irresponsibly expensive to replicate* Even if research stopped today and there was no progress for the next 30 years, there are far more capabilities to unlock and productize from existing foundation models and we <5% done on this journey* Good AI Engineering requires genuine skill and is deepening enough to justify sub-specialization as a sub-industry of Software Engineering* Companies and countries with better AI engineer workforces will disproportionately benefit from AI vs those who equivocate it as one of many equivalent priorities* Tech progress is often framed as “the future is here but it is not evenly distributed”. The role of the AI Engineer is therefore to better distribute the state of the art to as much of humanity as possible, including the elderly, poor, and differently abled.All of which are themes we first identified in the Rise of the AI Engineer. Singapore simply has a few additional factors that make it not just a good fit, but an economic imperative:* English speaking, very-online country that is great at STEM* Aging, ex-growth population (Total Fertility Rate of 1.1)* #3 GDP per capita (PPP) country in the world* Physically remote from major economic growth centers ex China/SEAThat basically dictates that any continued economic growth must be disconnected to geography, timezone, or headcount, or reliance on existing industrial drivers. Short of holding Taylor Swift hostage, making an intentional, concentrated bet on AI industrial policy is Singapore’s best option to keep up progress in the 21st century. As a pioneer in education policy being the primary long term determinant of economic success, this may result in Python as Singapore’s next National Language in the long run, a proposal we also discussed extensively at the RAISE retreat where this episode was recorded.Because of upcoming election season concerns around the globe, we also took the opportunity to ask about Singapore’s recent deepfake (election integrity) law.Full YouTube episodeShow Notes* Josephine Teo Official Bio, Wikipedia* Singapore National AI Strategy* 2019 - v1* 2023 - v2* ICLR (machine learning conference)* Philipp Kandal (CPO of Grab)* Temasek* GIC* EDBI* Economic Development Board (EDB)* Michael Fay incident* Quincy Larson* AIBots (internal RAG system for Singapore government)* Slovakia election incident* National AI Strategy - Singapore* Singapore AI Safety Institute* AI Verify* SkillsFuture* Ministry of Digital Development and Information (MDDI)* GovTech* NTU (Nanyang Technological University)Timestamps00:00:00 Introductions00:00:34 Singapore's National AI Strategy00:02:50 Ministry of Digital Development and Information00:08:49 Defining a National AI Strategy00:14:32 AI Safety and Governance00:16:50 AI Adoption in Companies and Government00:19:53 Balancing AI Innovation and Safety00:22:56 Structuring Government for Rapid Technological Change00:27:08 Doing Business with Singapore00:32:21 Training and Workforce Development in AI00:37:05 Career Transition Help for Post-AI Jobs00:40:19 AI Literacy and Coding as a Language00:43:28 Sovereign AI and Digital Infrastructure00:50:48 Government and AI Workloads00:51:02 Favorite AI Use Case in Government00:53:52 AI and ElectionsTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO at Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Small.ai.Swyx [00:00:13]: Hey everyone, this is a very, very special episode. We have here Mr. Josephine Teo from Singapore. Welcome.Josephine [00:00:19]: Hi Shawn and hi Alessio. Thank you for having me. Of course.Swyx [00:00:23]: You are the Minister for Digital Development and Information and Second Minister for Home Affairs. We're meeting here at RAISE, which is effectively your agency. Maybe we want to explain a little bit about what Singapore is doing in AI.Josephine [00:00:34]: Well, we've had an AI strategy at the national level for some years now, and about two years ago when generative AI became so prominent, we thought it was about time for us to refresh our national AI strategy. And it's not unusual on such occasions for us to consult widely. We want to talk to people who are familiar with the field. We want to talk to people who are active as practitioners, and we also want to talk to people in Singapore who have an interest in seeing the AI ecosystem develop. So when we put all these together, we discovered something else by chance, and it was really a bonus. This was the fact that there were already Singaporeans that were active in the AI space, particularly in the US, particularly in the Bay Area. And one of the exciting things for us was how could we also consult these Singaporeans who clearly still have a passion for Singapore, they do care about what happens back home, and they want to contribute to it. So that's how RAISE came about. And RAISE actually preceded the publication of the refresh of our national AI strategy, which took place in December last year. So the inputs of the participants from RAISE helped us to sharpen what we thought would be important in building up the AI ecosystem. And also with the encouragement of participants at RAISE, primarily Singaporeans who were doing great work in the US, we decided to raise our ambitions, literally. That's why we say AI for the public good, recognising the fact that commercial interest will certainly drive exciting developments in the industry space. But keep in mind, there is a need to make sure that AI serves the public good. And we say for Singapore and the world. So the idea is that experiments that are carried out in Singapore, things that are scaled up in Singapore potentially could have contributions elsewhere in the world. And so AI for the public good, for Singapore and the world. That's how it came about.Alessio [00:02:50]: I was listening to some of your previous interviews, and even the choice of the name development in the ministry name was very specific. You mentioned naming is your ethos. Can you explain maybe a bit about what the ministry does, which is not simply funding R&D, but i
CEOs of publicly traded companies are often in the news talking about their new AI initiatives, but few of them have built anything with it. Drew Houston from Dropbox is different; he has spent over 400 hours coding with LLMs in the last year and is now refocusing his 2,500+ employees around this new way of working, 17 years after founding the company.Timestamps00:00 Introductions00:43 Drew's AI journey04:14 Revalidating expectations of AI08:23 Simulation in self-driving vs. knowledge work12:14 Drew's AI Engineering setup15:24 RAG vs. long context in AI models18:06 From "FileGPT" to Dropbox AI23:20 Is storage solved?26:30 Products vs Features30:48 Building trust for data access33:42 Dropbox Dash and universal search38:05 The evolution of Dropbox42:39 Building a "silicon brain" for knowledge work48:45 Open source AI and its impact51:30 "Rent, Don't Buy" for AI54:50 Staying relevant58:57 Founder Mode01:03:10 Advice for founders navigating AI01:07:36 Building and managing teams in a growing companyTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO at Decibel Partners, and there's no Swyx today, but I'm joined by Drew Houston of Dropbox. Welcome, Drew.Drew [00:00:14]: Thanks for having me.Alessio [00:00:15]: So we're not going to talk about the Dropbox story. We're not going to talk about the Chinatown bus and the flash drive and all that. I think you've talked enough about it. Where I want to start is you as an AI engineer. So as you know, most of our audience is engineering folks, kind of like technology leaders. You obviously run Dropbox, which is a huge company, but you also do a lot of coding. I think that's how you spend almost 400 hours, just like coding. So let's start there. What was the first interaction you had with an LLM API and when did the journey start for you?Drew [00:00:43]: Yeah. Well, I think probably all AI engineers or whatever you call an AI engineer, those people started out as engineers before that. So engineering is my first love. I mean, I grew up as a little kid. I was that kid. My first line of code was at five years old. I just really loved, I wanted to make computer games, like this whole path. That also led me into startups and eventually starting Dropbox. And then with AI specifically, I studied computer science, I got my, I did my undergrad, but I didn't do like grad level computer science. I didn't, I sort of got distracted by all the startup things, so I didn't do grad level work. But about several years ago, I made a couple of things. So one is I sort of, I knew I wanted to go from being an engineer to a founder. And then, but sort of the becoming a CEO part was sort of backed into the job. And so a couple of realizations. One is that, I mean, there's a lot of like repetitive and like manual work you have to do as an executive that is actually lends itself pretty well to automation, both for like my own convenience. And then out of interest in learning, I guess what we call like classical machine learning these days, I started really trying to wrap my head around understanding machine learning and informational retrieval more, more formally. So I'd say maybe 2016, 2017 started me writing these more successively, more elaborate scripts to like understand basic like classifiers and regression and, and again, like basic information retrieval and NLP back in those days. And there's sort of like two things that came out of that. One is techniques are super powerful. And even just like studying like old school machine learning was a pretty big inversion of the way I had learned engineering, right? You know, I started programming when everyone starts programming and you're, you're sort of the human, you're giving an algorithm to the, and spelling out to the computer how it should run it. And then machine learning, here's machine learning where it's like actually flip that, like give it sort of the answer you want and it'll figure out the algorithm, which was pretty mind bending. And it was both like pretty powerful when I would write tools, like figure out like time audits or like, where's my time going? Is this meeting a one-on-one or is it a recruiting thing or is it a product strategy thing? I started out doing that manually with my assistant, but then found that this was like a very like automatable task. And so, which also had the side effect of teaching me a lot about machine learning. But then there was this big problem, like anytime you, it was very good at like tabular structured data, but like anytime it hit, you know, the usual malformed English that humans speak, it would just like fall over. I had to kind of abandon a lot of the things that I wanted to build because like there's no way to like parse text. Like maybe it would sort of identify the part of speech in a sentence or something. But then fast forward to the LLM, I mean actually I started trying some of like this, what we would call like very small LLMs before kind of the GPT class models. And it was like super hard to get those things working. So like these 500 parameter models would just be like hallucinating and repeating and you know. So actually I'd kind of like written it off a little bit. But then the chat GPT launch and GPT-3 for sure. And then once people figured out like prompting and instruction tuning, this was sort of like November-ish 2022 like everybody else sort of that the chat GPT launch being the starting gun for the whole AI era of computing and then having API access to three and then early access to GPT-4. I was like, oh man, it's happening. And so I was literally on my honeymoon and we're like on a beach in Thailand and I'm like coding these like AI tools to automate like writing or to assist with writing and all these different use cases.Alessio [00:04:14]: You're like, I'm never going back to work. I'm going to automate all of it before I get back.Drew [00:04:17]: And I was just, you know, ever since then, I mean, I've always been like coding like prototypes and just stuff to make my life more convenient, but like escalated a lot after 22. And yeah, I spent, I checked, I think it was probably like over 400 hours this year so far coding because I had my paternity leave where I was able to work on some special projects. But yeah, it's a super important part of like my whole learning journey is like being really hands-on with these things. And I mean, it's probably not a typical recipe, but I really love to get down to the metal as far as how this stuff works.Alessio [00:04:47]: Yeah. So Swyx and I were with Sam Altman in October 22. We were like at a hack day at OpenAI and that's why we started this podcast eventually. But you did an interview with Sam like seven years ago and he asked you what's the biggest opportunity in startups and you were like machine learning and AI and you were almost like too early, right? It's like maybe seven years ago, the models weren't quite there. How should people think about revalidating like expectations of this technology? You know, I think even today people will tell you, oh, models are not really good at X because they were not good 12 months ago, but they're good today.Drew [00:05:19]: What's your project? Heuristics for thinking about that or how is, yeah, I think the way I look at it now is pretty, has evolved a lot since when I started. I mean, I think everybody intuitively starts with like, all right, let's try to predict the future or imagine like what's this great end state we're going to get to. And the tricky thing is like often those prognostications are right, but they're right in terms of direction, but not when. For example, you know, even in the early days of the internet, 90s when things were even like tech space and you know, even before like the browser or things like that, people were like, oh man, you're going to have, you know, you're going to be able to order food, get like a Snickers delivered to your house, you're going to be able to watch any movie ever created. And they were right. But they were like, you know, it took 20 years for that to actually happen. And before you got to DoorDash, you had to get, you started with like Webvan and Cosmo and before you get to Spotify, you had to do like Napster and Kazaa and LimeWire and like a bunch of like broken Britney Spears MP3s and malware. So I think the big lesson is being early is the same as being wrong. Being late is the same as being wrong. So really how do you calibrate timing? And then I think with AI, it's the same thing that people are like, oh, it's going to completely upend society and all these positive and negative ways. I think that's like most of those things are going to come true. The question is like, when is that going to happen? And then with AI specifically, I think there's also, in addition to sort of the general tech category or like jumping too fast to the future, I think that AI is particularly susceptible to that. And you look at self-driving, right? This idea of like, oh my God, you can have a self-driving car captured everybody's imaginations 10, 12 years ago. And you know, people are like, oh man, in two years, there's not going to be another year. There's not going to be a human driver on the road to be seen. It didn't work out that way, right? We're still 10, 12 years later where we're in a world where you can sort of sometimes get a Waymo in like one city on earth. Exciting, but just took a lot longer than people think. And the reason is there's a lot of engineering challenges, but then there's a lot of other like societal time constants that are hard to compress. So one thing I think you can learn from things like self-driving is they have these levels of autonomy that's a useful kind of framework in driving or these like maturity levels. People sort of skip to like level five, full autonomy, or we're going to have like an autonomous knowledge worker that's just going to take, that's going
We are in 🗽 NYC this Monday! Join the AI Eng NYC meetup, bring demos and vibes!It is a bit of a meme that the first thing developer tooling founders think to build in AI is all the non-AI operational stuff outside the AI. There are well over 60 funded LLM Ops startups all with hoping to solve the new observability, cost tracking, security, and reliability problems that come with putting LLMs in production, not to mention new LLM oriented products from incumbent, established ops/o11y players like Datadog and Weights & Biases. 2 years in to the current hype cycle, the early winners have tended to be people with practical/research AI backgrounds rather than MLOps heavyweights or SWE tourists:* LangSmith: We covered how Harrison Chase worked on AI at Robust Intelligence and Kensho, the alma maters of many great AI founders* HumanLoop: We covered how Raza Habib worked at Google AI during his PhD* BrainTrust: Today’s guest Ankur Goyal founded Impira pre-Transformers and was acquihired to run Figma AI before realizing how to solve the Ops problem.There have been many VC think pieces and market maps describing what people thought were the essential pieces of the AI Engineering stack, but what was true for 2022-2023 has aged poorly. The basic insight that Ankur had is the same thesis that Hamel Husain is pushing in his World’s Fair talk and podcast with Raza and swyx:Evals are the centerpiece of systematic AI Engineering.REALLY believing in this is harder than it looks with the benefit of hindsight. It’s not like people didn’t know evals were important. Basically every LLM Ops feature list has them. It’s an obvious next step AFTER managing your prompts and logging your LLM calls. In fact, up til we met Braintrust, we were working on an expanded version of the Impossible Triangle Theory of the LLM Ops War that we first articulated in the Humanloop writeup:The single biggest criticism of the Rise of the AI Engineer piece is that we neglected to split out the role of product evals (as opposed to model evals) in the now infamous “API line” chart:With hindsight, we were very focused on the differentiating 0 to 1 phase that AI Engineers can bring to an existing team of ML engineers. As swyx says on the Day 2 keynote of AI Engineer, 2024 added a whole new set of concerns as AI Engineering grew up:A closer examination of Hamel’s product-oriented virtuous cycle and this infra-oriented SDLC would have eventually revealed that Evals, even more than logging, was the first point where teams start to get really serious about shipping to production, and therefore a great place to make an entry into the marketplace, which is exactly what Braintrust did.Also notice what’s NOT on this chart: shifting to shadow open source models, and finetuning them… per Ankur, Fine-tuning is not a viable standalone product:“The thing I would say is not debatable is whether or not fine-tuning is a business outcome or not. So let's think about the other components of your triangle. Ops/observability, that is a business… Frameworks, evals, databases [are a business, but] Fine-tuning is a very compelling method that achieves an outcome. The outcome is not fine-tuning, it is can I automatically optimize my use case to perform better if I throw data at the problem? And fine-tuning is one of multiple ways to achieve that.”OpenAI vs Open AI Market ShareWe last speculated about the market shifts in the End of OpenAI Hegemony and the Winds of AI Winter, and Ankur’s perspective is super valuable given his customer list:Some surprises based on what he is seeing:* Prior to Claude 3, OpenAI had near 100% market share. This tracks with what Harrison told us last year.* Claude 3.5 Sonnet and also notably Haiku have made serious dents* Open source model adoption is <5% and DECLINING. Contra to Eugene Cheah’s ideal marketing pitch, virtually none of Braintrust’s customers are really finetuning open source models for cost, control, or privacy. This is partially caused by…* Open source model hosts, aka Inference providers, aren’t as mature as OpenAI’s API platform. Kudos to Michelle’s team as if they needed any more praise!* Adoption of Big Lab models via their Big Cloud Partners, aka Claude through AWS, or OpenAI through Azure, is low. Surprising! It seems that there are issues with accessing the latest models via the Cloud partners.swyx [01:36:51]: What % of your workload is open source?Ankur Goyal [01:36:55]: Because of how we're deployed, I don't have like an exact number for you. Among customers running in production, it's less than 5%.Full Video EpisodeCheck out the Braintrust demo on YouTube! (and like and subscribe etc)Show Notes* Ankur’s companies* MemSQL/SingleStore → now Nikita Shamgunov of Neon* Impira* Braintrust* Papers mentioned* AlexNet* BERT Paper* Layout LM Paper* GPT-3 Paper* Voyager Paper* AI Engineer World's Fair* Ankur and Olmo’s talk at AIEWF* Together.ai* Fireworks* People* Nikita Shamgunov* Alana Goyal* Elad Gil* Clem Delangue* Guillermo Rauch* Prior episodes* HumanLoop episode* Michelle Pokrass episode* Dylan Patel episodeTimestamps* [00:00:00] Introduction and background on Ankur career* [00:00:49] SingleStore and HTAP databases* [00:08:19] Founding Impira and lessons learned* [00:13:33] Unstructured vs Structured Data * [00:25:41] Overview of Braintrust and its features* [00:40:42] Industry observations and trends in AI tooling* [00:58:37] Workload types and AI use cases in production* [01:06:37] World's Fair AI conference discussion* [01:11:09] AI infrastructure market landscape* [01:24:59] OpenAI vs Anthropic vs other model providers* [01:38:11] GPU inference market discussion* [01:45:39] Hypothetical AI projects outside of Braintrust* [01:50:25] Potentially joining OpenAI* [01:52:37] Insights on effective networking and relationships in techTranscriptswyx [00:00:00]: Ankur Goyal, welcome to Latent Space.Ankur Goyal [00:00:06]: Thanks for having me.swyx [00:00:07]: Thanks for coming all the way over to our studio.Ankur Goyal [00:00:10]: It was a long hike.swyx [00:00:11]: A long trek. Yeah. You got T-boned by traffic. Yeah. You were the first VP of Eng at Signal Store. Yeah. Then you started Impira. You ran it for six years, got acquired into Figma, where you were at for eight months, and you just celebrated your one-year anniversary at Braintrust. I did, yeah. What a journey. I kind of want to go through each in turn because I have a personal relationship with Signal Store just because I have been a follower and fan of databases for a while. HTAP is always a dream of every database guy. It's still the dream. When HTAP, and Signal Store I think is the leading HTAP. Yeah. What's that journey like? And then maybe we'll cover the rest later.Ankur Goyal [00:00:49]: Sounds good.swyx [00:00:50]: We can start Signal Store first. Yeah, yeah.Ankur Goyal [00:00:52]: In college, as a first-generation Indian kid, I basically had two options. I had already told my parents I wasn't going to be a doctor. They're both doctors, so only two options left. Do a PhD or work at a big company. After my sophomore year, I worked at Microsoft, and it just wasn't for me. I realized that the work I was doing was impactful. I was working on Bing and the distributed compute infrastructure at Bing, which is actually now part of Azure. There were hundreds of engineers using the infrastructure that we were working on, but the level of intensity was too low. It felt like you got work-life balance and impact, but very little creativity, very little room to do interesting things. I was like, okay, let me cross that off the list. The only option left is to do research. I did research the next summer, and I realized, again, no one's working that hard. Maybe the times have changed, but at that point, there's a lot of creativity. You're just bouncing around fun ideas and working on stuff and really great work-life balance, but no one would actually use the stuff that we built, and that was not super energizing for me. I had this existential crisis, and I moved out to San Francisco because I had a friend who was here and crashed on his couch and was talking to him and just very, very confused. He said, you should talk to a recruiter, which felt like really weird advice. I'm not even sure I would give that advice to someone nowadays, but I met this really great guy named John, and he introduced me to like 30 different companies. I realized that there's actually a lot of interesting stuff happening in startups, and maybe I could find this kind of company that let me be very creative and work really hard and have a lot of impact, and I don't give a s**t about work-life balance. I talked to all these companies, and I remember I met MemSQL when it was three people and interviewed, and I thought I just totally failed the interview, but I had never had so much fun in my life. I remember I was at 10th and Harrison, and I stood at the bus station, and I called my parents and said, I'm sorry, I'm dropping out of school. I thought I wouldn't get the offer, but I just realized that if there's something like this company, then this is where I need to be. Luckily, things worked out, and I got an offer, and I joined as employee number two, and I worked there for almost six years, and it was an incredible experience. Learned a lot about systems, got to work with amazing customers. There are a lot of things that I took for granted that I later learned at Impira that I had taken for granted, and the most exciting thing is I got to run the engineering team, which was a great opportunity to learn about tech on a larger stage, recruit a lot of great people, and I think, for me personally, set me up to do a lot of interesting things after.swyx [00:03:41]: Yeah, there's so many ways I can take that. The most curious, I think, for general audiences is, is the dream real of SingleStore? Should, obviously, more people be using it? I think there's a lot of marketing from SingleStore that makes sense, but t
We all have fond memories of the first Dev Day in 2023:and the blip that followed soon after. As Ben Thompson has noted, this year’s DevDay took a quieter, more intimate tone. No Satya, no livestream, (slightly fewer people?). Instead of putting ChatGPT announcements in DevDay as in 2023, o1 was announced 2 weeks prior, and DevDay 2024 was reserved purely for developer-facing API announcements, primarily the Realtime API, Vision Finetuning, Prompt Caching, and Model Distillation.However the larger venue and more spread out schedule did allow a lot more hallway conversations with attendees as well as more community presentations including our recent guest Alistair Pullen of Cosine as well as deeper dives from OpenAI including our recent guest Michelle Pokrass of the API Team. Thanks to OpenAI’s warm collaboration (we particularly want to thank Lindsay McCallum Rémy!), we managed to record exclusive interviews with many of the main presenters of both the keynotes and breakout sessions. We present them in full in today’s episode, together with a full lightly edited Q&A with Sam Altman.Show notes and related resourcesSome of these used in the final audio episode below* Simon Willison Live Blog* swyx live tweets and videos* Greg Kamradt coverage of Structured Output session, Scaling LLM Apps session* Fireside Chat Q&A with Sam AltmanTimestamps* [00:00:00] Intro by Suno.ai* [00:01:23] NotebookLM Recap of DevDay* [00:09:25] Ilan's Strawberry Demo with Realtime Voice Function Calling* [00:19:16] Olivier Godement, Head of Product, OpenAI* [00:36:57] Romain Huet, Head of DX, OpenAI* [00:47:08] Michelle Pokrass, API Tech Lead at OpenAI ft. Simon Willison* [01:04:45] Alistair Pullen, CEO, Cosine (Genie)* [01:18:31] Sam Altman + Kevin Weill Q&A* [02:03:07] Notebook LM Recap of PodcastTranscript[00:00:00] Suno AI: Under dev daylights, code ignites. Real time voice streams reach new heights. O1 and GPT, 4. 0 in flight. Fine tune the future, data in sight. Schema sync up, outputs precise. Distill the models, efficiency splice.[00:00:33] AI Charlie: Happy October. This is your AI co host, Charlie. One of our longest standing traditions is covering major AI and ML conferences in podcast format. Delving, yes delving, into the vibes of what it is like to be there stitched in with short samples of conversations with key players, just to help you feel like you were there.[00:00:54] AI Charlie: Covering this year's Dev Day was significantly more challenging because we were all requested not to record the opening keynotes. So, in place of the opening keynotes, we had the viral notebook LM Deep Dive crew, my new AI podcast nemesis, Give you a seven minute recap of everything that was announced.[00:01:15] AI Charlie: Of course, you can also check the show notes for details. I'll then come back with an explainer of all the interviews we have for you today. Watch out and take care.[00:01:23] NotebookLM Recap of DevDay[00:01:23] NotebookLM: All right, so we've got a pretty hefty stack of articles and blog posts here all about open ais. Dev day 2024.[00:01:32] NotebookLM 2: Yeah, lots to dig into there.[00:01:34] NotebookLM 2: Seems[00:01:34] NotebookLM: like you're really interested in what's new with AI.[00:01:36] NotebookLM 2: Definitely. And it seems like OpenAI had a lot to announce. New tools, changes to the company. It's a lot.[00:01:43] NotebookLM: It is. And especially since you're interested in how AI can be used in the real world, you know, practical applications, we'll focus on that.[00:01:51] NotebookLM: Perfect. Like, for example, this Real time API, they announced that, right? That seems like a big deal if we want AI to sound, well, less like a robot.[00:01:59] NotebookLM 2: It could be huge. The real time API could completely change how we, like, interact with AI. Like, imagine if your voice assistant could actually handle it if you interrupted it.[00:02:08] NotebookLM: Or, like, have an actual conversation.[00:02:10] NotebookLM 2: Right, not just these clunky back and forth things we're used to.[00:02:14] NotebookLM: And they actually showed it off, didn't they? I read something about a travel app, one for languages. Even one where the AI ordered takeout.[00:02:21] NotebookLM 2: Those demos were really interesting, and I think they show how this real time API can be used in so many ways.[00:02:28] NotebookLM 2: And the tech behind it is fascinating, by the way. It uses persistent WebSocket connections and this thing called function calling, so it can respond in real time.[00:02:38] NotebookLM: So the function calling thing, that sounds kind of complicated. Can you, like, explain how that works?[00:02:42] NotebookLM 2: So imagine giving the AI Access to this whole toolbox, right?[00:02:46] NotebookLM 2: Information, capabilities, all sorts of things. Okay. So take the travel agent demo, for example. With function calling, the AI can pull up details, let's say about Fort Mason, right, from some database. Like nearby restaurants, stuff like that.[00:02:59] NotebookLM: Ah, I get it. So instead of being limited to what it already knows, It can go and find the information it needs, like a human travel agent would.[00:03:07] NotebookLM 2: Precisely. And someone on Hacker News pointed out a cool detail. The API actually gives you a text version of what's being said. So you can store that, analyze it.[00:03:17] NotebookLM: That's smart. It seems like OpenAI put a lot of thought into making this API easy for developers to use. But, while we're on OpenAI, you know, Besides their tech, there's been some news about, like, internal changes, too.[00:03:30] NotebookLM: Didn't they say they're moving away from being a non profit?[00:03:32] NotebookLM 2: They did. And it's got everyone talking. It's a major shift. And it's only natural for people to wonder how that'll change things for OpenAI in the future. I mean, there are definitely some valid questions about this move to for profit. Like, will they have more money for research now?[00:03:46] NotebookLM 2: Probably. But will they, you know, care as much about making sure AI benefits everyone?[00:03:51] NotebookLM: Yeah, that's the big question, especially with all the, like, the leadership changes happening at OpenAI too, right? I read that their Chief Research Officer left, and their VP of Research, and even their CTO.[00:04:03] NotebookLM 2: It's true. A lot of people are connecting those departures with the changes in OpenAI's structure.[00:04:08] NotebookLM: And I guess it makes you wonder what's going on behind the scenes. But they are still putting out new stuff. Like this whole fine tuning thing really caught my eye.[00:04:17] NotebookLM 2: Right, fine tuning. It's essentially taking a pre trained AI model. And, like, customizing it.[00:04:23] NotebookLM: So instead of a general AI, you get one that's tailored for a specific job.[00:04:27] NotebookLM 2: Exactly. And that opens up so many possibilities, especially for businesses. Imagine you could train an AI on your company's data, you know, like how you communicate your brand guidelines.[00:04:37] NotebookLM: So it's like having an AI that's specifically trained for your company?[00:04:41] NotebookLM 2: That's the idea.[00:04:41] NotebookLM: And they're doing it with images now, too, right?[00:04:44] NotebookLM: Fine tuning with vision is what they called it.[00:04:46] NotebookLM 2: It's pretty incredible what they're doing with that, especially in fields like medicine.[00:04:50] NotebookLM: Like using AI to help doctors make diagnoses.[00:04:52] NotebookLM 2: Exactly. And AI could be trained on thousands of medical images, right? And then it could potentially spot things that even a trained doctor might miss.[00:05:03] NotebookLM: That's kind of scary, to be honest. What if it gets it wrong?[00:05:06] NotebookLM 2: Well, the idea isn't to replace doctors, but to give them another tool, you know, help them make better decisions.[00:05:12] NotebookLM: Okay, that makes sense. But training these AI models must be really expensive.[00:05:17] NotebookLM 2: It can be. All those tokens add up. But OpenAI announced something called automatic prompt caching.[00:05:23] Alex Volkov: Automatic what now? I don't think I came across that.[00:05:26] NotebookLM 2: So basically, if your AI sees a prompt that it's already seen before, OpenAI will give you a discount.[00:05:31] NotebookLM: Huh. Like a frequent buyer program for AI.[00:05:35] NotebookLM 2: Kind of, yeah. It's good that they're trying to make it more affordable. And they're also doing something called model distillation.[00:05:41] NotebookLM: Okay, now you're just using big words to sound smart. What's that?[00:05:45] NotebookLM 2: Think of it like like a recipe, right? You can take a really complex recipe and break it down to the essential parts.[00:05:50] NotebookLM: Make it simpler, but it still tastes the same.[00:05:53] NotebookLM 2: Yeah. And that's what model distillation is. You take a big, powerful AI model and create a smaller, more efficient version.[00:06:00] NotebookLM: So it's like lighter weight, but still just as capable.[00:06:03] NotebookLM 2: Exactly. And that means more people can actually use these powerful tools. They don't need, like, a supercomputer to run them.[00:06:10] NotebookLM: So they're making AI more accessible. That's great.[00:06:13] NotebookLM 2: It is. And speaking of powerful tools, they also talked about their new O1 model.[00:06:18] NotebookLM 2: That's the one they've been hyping up. The one that's supposed to be this big leap forward.[00:06:22] NotebookLM: Yeah, O1. It sounds pretty futuristic. Like, from what I read, it's not just a bigger, better language model.[00:06:28] NotebookLM 2: Right. It's a different porch.[00:06:29] NotebookLM: They're saying it can, like, actually reason, right? Think.[00:06:33] NotebookLM 2: It's trained differently.[00:06:34] NotebookLM 2: They used reinforcement learning with O1.[00:06:36] NotebookLM: So it's
OpenAI DevDay is almost here! Per tradition, we are hosting a DevDay pregame event for everyone coming to town! Join us with demos and gossip!Also sign up for related events across San Francisco: the AI DevTools Night, the xAI open house, the Replicate art show, the DevDay Watch Party (for non-attendees), Hack Night with OpenAI at Cloudflare. For everyone else, join the Latent Space Discord for our online watch party and find fellow AI Engineers in your city.OpenAI’s recent o1 release (and Reflection 70b debacle) has reignited broad interest in agentic general reasoning and tree search methods.While we have covered some of the self-taught reasoning literature on the Latent Space Paper Club, it is notable that the Eric Zelikman ended up at xAI, whereas OpenAI’s hiring of Noam Brown and now Shunyu suggests more interest in tool-using chain of thought/tree of thought/generator-verifier architectures for Level 3 Agents.We were more than delighted to learn that Shunyu is a fellow Latent Space enjoyer, and invited him back (after his first appearance on our NeurIPS 2023 pod) for a look through his academic career with Harrison Chase (one year after his first LS show).ReAct: Synergizing Reasoning and Acting in Language Modelspaper linkFollowing seminal Chain of Thought papers from Wei et al and Kojima et al, and reflecting on lessons from building the WebShop human ecommerce trajectory benchmark, Shunyu’s first big hit, the ReAct paper showed that using LLMs to “generate both reasoning traces and task-specific actions in an interleaved manner” achieved remarkably greater performance (less hallucination/error propagation, higher ALFWorld/WebShop benchmark success) than CoT alone. In even better news, ReAct scales fabulously with finetuning:As a member of the elite Princeton NLP group, Shunyu was also a coauthor of the Reflexion paper, which we discuss in this pod.Tree of Thoughtspaper link hereShunyu’s next major improvement on the CoT literature was Tree of Thoughts:Language models are increasingly being deployed for general problem solving across a wide range of tasks, but are still confined to token-level, left-to-right decision-making processes during inference. This means they can fall short in tasks that require exploration, strategic lookahead, or where initial decisions play a pivotal role…ToT allows LMs to perform deliberate decision making by considering multiple different reasoning paths and self-evaluating choices to decide the next course of action, as well as looking ahead or backtracking when necessary to make global choices.The beauty of ToT is it doesnt require pretraining with exotic methods like backspace tokens or other MCTS architectures. You can listen to Shunyu explain ToT in his own words on our NeurIPS pod, but also the ineffable Yannic Kilcher:Other WorkWe don’t have the space to summarize the rest of Shunyu’s work, you can listen to our pod with him now, and recommend the CoALA paper and his initial hit webinar with Harrison, today’s guest cohost:as well as Shunyu’s PhD Defense Lecture:as well as Shunyu’s latest lecture covering a Brief History of LLM Agents:As usual, we are live on YouTube! Show Notes* Harrison Chase* LangChain, LangSmith, LangGraph* Shunyu Yao* Alec Radford* ReAct Paper* Hotpot QA* Tau Bench* WebShop* SWE-Agent* SWE-Bench* Trees of Thought* CoALA Paper* Related Episodes* Our Thomas Scialom (Meta) episode* Shunyu on our NeurIPS 2023 Best Papers episode* Harrison on our LangChain episode* Mentions* Sierra* Voyager* Jason Wei* Tavily* SERP API* ExaTimestamps* [00:00:00] Opening Song by Suno* [00:03:00] Introductions* [00:06:16] The ReAct paper* [00:12:09] Early applications of ReAct in LangChain* [00:17:15] Discussion of the Reflection paper* [00:22:35] Tree of Thoughts paper and search algorithms in language models* [00:27:21] SWE-Agent and SWE-Bench for coding benchmarks* [00:39:21] CoALA: Cognitive Architectures for Language Agents* [00:45:24] Agent-Computer Interfaces (ACI) and tool design for agents* [00:49:24] Designing frameworks for agents vs humans* [00:53:52] UX design for AI applications and agents* [00:59:53] Data and model improvements for agent capabilities* [01:19:10] TauBench* [01:23:09] Promising areas for AITranscriptAlessio [00:00:01]: Hey, everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO of Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Small AI.Swyx [00:00:12]: Hey, and today we have a super special episode. I actually always wanted to take like a selfie and go like, you know, POV, you're about to revolutionize the world of agents because we have two of the most awesome hiring agents in the house. So first, we're going to welcome back Harrison Chase. Welcome. Excited to be here. What's new with you recently in sort of like the 10, 20 second recap?Harrison [00:00:34]: Linkchain, Linksmith, Lingraph, pushing on all of them. Lots of cool stuff related to a lot of the stuff that we're going to talk about today, probably.Swyx [00:00:42]: Yeah.Alessio [00:00:43]: We'll mention it in there. And the Celtics won the title.Swyx [00:00:45]: And the Celtics won the title. You got that going on for you. I don't know. Is that like floorball? Handball? Baseball? Basketball.Alessio [00:00:52]: Basketball, basketball.Harrison [00:00:53]: Patriots aren't looking good though, so that's...Swyx [00:00:56]: And then Xun Yu, you've also been on the pod, but only in like a sort of oral paper presentation capacity. But welcome officially to the LinkedSpace pod.Shunyu [00:01:03]: Yeah, I've been a huge fan. So thanks for the invitation. Thanks.Swyx [00:01:07]: Well, it's an honor to have you on. You're one of like, you're maybe the first PhD thesis defense I've ever watched in like this AI world, because most people just publish single papers, but every paper of yours is a banger. So congrats.Shunyu [00:01:22]: Thanks.Swyx [00:01:24]: Yeah, maybe we'll just kick it off with, you know, what was your journey into using language models for agents? I like that your thesis advisor, I didn't catch his name, but he was like, you know... Karthik. Yeah. It's like, this guy just wanted to use language models and it was such a controversial pick at the time. Right.Shunyu [00:01:39]: The full story is that in undergrad, I did some computer vision research and that's how I got into AI. But at the time, I feel like, you know, you're just composing all the GAN or 3D perception or whatever together and it's not exciting anymore. And one day I just see this transformer paper and that's really cool. But I really got into language model only when I entered my PhD and met my advisor Karthik. So he was actually the second author of GPT-1 when he was like a visiting scientist at OpenAI. With Alec Redford?Swyx [00:02:10]: Yes.Shunyu [00:02:11]: Wow. That's what he told me. It's like back in OpenAI, they did this GPT-1 together and Ilya just said, Karthik, you should stay because we just solved the language. But apparently Karthik is not fully convinced. So he went to Princeton, started his professorship and I'm really grateful. So he accepted me as a student, even though I have no prior knowledge in NLP. And you know, we just met for the first time and he's like, you know, what do you want to do? And I'm like, you know, you have done those test game scenes. That's really cool. I wonder if we can just redo them with language models. And that's how the whole journey began. Awesome.Alessio [00:02:46]: So GPT-2 was out at the time? Yes, that was 2019.Shunyu [00:02:48]: Yeah.Alessio [00:02:49]: Way too dangerous to release. And then I guess the first work of yours that I came across was React, which was a big part of your defense. But also Harrison, when you came on The Pockets last year, you said that was one of the first papers that you saw when you were getting inspired for BlankChain. So maybe give a recap of why you thought it was cool, because you were already working in AI and machine learning. And then, yeah, you can kind of like intro the paper formally. What was that interesting to you specifically?Harrison [00:03:16]: Yeah, I mean, I think the interesting part was using these language models to interact with the outside world in some form. And I think in the paper, you mostly deal with Wikipedia. And I think there's some other data sets as well. But the outside world is the outside world. And so interacting with things that weren't present in the LLM and APIs and calling into them and thinking about the React reasoning and acting and kind of like combining those together and getting better results. I'd been playing around with LLMs, been talking with people who were playing around with LLMs. People were trying to get LLMs to call into APIs, do things, and it was always, how can they do it more reliably and better? And so this paper was basically a step in that direction. And I think really interesting and also really general as well. Like I think that's part of the appeal is just how general and simple in a good way, I think the idea was. So that it was really appealing for all those reasons.Shunyu [00:04:07]: Simple is always good. Yeah.Alessio [00:04:09]: Do you have a favorite part? Because I have one favorite part from your PhD defense, which I didn't understand when I read the paper, but you said something along the lines, React doesn't change the outside or the environment, but it does change the insight through the context, putting more things in the context. You're not actually changing any of the tools around you to work for you, but you're changing how the model thinks. And I think that was like a very profound thing when I, not that I've been using these tools for like 18 months. I'm like, I understand what you meant, but like to say that at the time you did the PhD defense was not trivial. Yeah.Shunyu [00:04:41]: Another way to put it is like thinking can be an extra tool that's useful.Alessio [00:04:47]: Makes sense. Checks
Noah Hein from Latent Space University is finally launching with a free lightning course this Sunday for those new to AI Engineering. Tell a friend!Did you know there are >1,600 papers on arXiv just about prompting? Between shots, trees, chains, self-criticism, planning strategies, and all sorts of other weird names, it’s hard to keep up. Luckily for us, Sander Schulhoff and team read them all and put together The Prompt Report as the ultimate prompt engineering reference, which we’ll break down step-by-step in today’s episode.In 2022 swyx wrote “Why “Prompt Engineering” and “Generative AI” are overhyped”; the TLDR being that if you’re relying on prompts alone to build a successful products, you’re ngmi. Prompt engineering moved from being a stand-alone job to a core skill for AI Engineers now. We won’t repeat everything that is written in the paper, but this diagram encapsulates the state of prompting today: confusing. There are many similar terms, esoteric approaches that have doubtful impact on results, and lots of people that are just trying to create full papers around a single prompt just to get more publications out. Luckily, some of the best prompting techniques are being tuned back into the models themselves, as we’ve seen with o1 and Chain-of-Thought (see our OpenAI episode). Similarly, OpenAI recently announced 100% guaranteed JSON schema adherence, and Anthropic, Cohere, and Gemini all have JSON Mode (not sure if 100% guaranteed yet). No more “return JSON or my grandma is going to die” required. The next debate is human-crafted prompts vs automated approaches using frameworks like DSPy, which Sander recommended:I spent 20 hours prompt engineering for a task and DSPy beat me in 10 minutes. It’s much more complex than simply writing a prompt (and I’m not sure how many people usually spend >20 hours prompt engineering one task), but if you’re hitting a roadblock it might be worth checking out.Prompt Injection and JailbreaksSander and team also worked on HackAPrompt, a paper that was the outcome of an online challenge on prompt hacking techniques. They similarly created a taxonomy of prompt attacks, which is very hand if you’re building products with user-facing LLM interfaces that you’d like to test:In this episode we basically break down every category and highlight the overrated and underrated techniques in each of them. If you haven’t spent time following the prompting meta, this is a great episode to catchup!Full Video EpisodeLike and subscribe on YouTube!Timestamps* [00:00:00] Introductions - Intro music by Suno AI* [00:07:32] Navigating arXiv for paper evaluation* [00:12:23] Taxonomy of prompting techniques* [00:15:46] Zero-shot prompting and role prompting* [00:21:35] Few-shot prompting design advice* [00:28:55] Chain of thought and thought generation techniques* [00:34:41] Decomposition techniques in prompting* [00:37:40] Ensembling techniques in prompting* [00:44:49] Automatic prompt engineering and DSPy* [00:49:13] Prompt Injection vs Jailbreaking* [00:57:08] Multimodal prompting (audio, video)* [00:59:46] Structured output prompting* [01:04:23] Upcoming Hack-a-Prompt 2.0 projectShow Notes* Sander Schulhoff* Learn Prompting* The Prompt Report* HackAPrompt* Mine RL Competition* EMNLP Conference* Noam Brown* Jordan Boydgraver* Denis Peskov* Simon Willison* Riley Goodside* David Ha* Jeremy Nixon* Shunyu Yao* Nicholas Carlini* DreadnodeTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO-in-Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI.Swyx [00:00:13]: Hey, and today we're in the remote studio with Sander Schulhoff, author of the Prompt Report.Sander [00:00:18]: Welcome. Thank you. Very excited to be here.Swyx [00:00:21]: Sander, I think I first chatted with you like over a year ago. What's your brief history? I went onto your website, it looks like you worked on diplomacy, which is really interesting because we've talked with Noam Brown a couple of times, and that obviously has a really interesting story in terms of prompting and agents. What's your journey into AI?Sander [00:00:40]: Yeah, I'd say it started in high school. I took my first Java class and just saw a YouTube video about something AI and started getting into it, reading. Deep learning, neural networks, all came soon thereafter. And then going into college, I got into Maryland and I emailed just like half the computer science department at random. I was like, hey, I want to do research on deep reinforcement learning because I've been experimenting with that a good bit. And over that summer, I had read the Intro to RL book and the deep reinforcement learning hands-on, so I was very excited about what deep RL could do. And a couple of people got back to me and one of them was Jordan Boydgraver, Professor Boydgraver, and he was working on diplomacy. And he said to me, this looks like it was more of a natural language processing project at the time, but it's a game, so very easily could move more into the RL realm. And I ended up working with one of his students, Denis Peskov, who's now a postdoc at Princeton. And that was really my intro to AI, NLP, deep RL research. And so from there, I worked on diplomacy for a couple of years, mostly building infrastructure for data collection and machine learning, but I always wanted to be doing it myself. So I had a number of side projects and I ended up working on the Mine RL competition, Minecraft reinforcement learning, also some people call it mineral. And that ended up being a really cool opportunity because I think like sophomore year, I knew I wanted to do some project in deep RL and I really liked Minecraft. And so I was like, let me combine these. And I was searching for some Minecraft Python library to control agents and found mineral. And I was trying to find documentation for how to build a custom environment and do all sorts of stuff. I asked in their Discord how to do this and their super responsive, very nice. And they're like, oh, you know, we don't have docs on this, but, you know, you can look around. And so I read through the whole code base and figured it out and wrote a PR and added the docs that I didn't have before. And then later I ended up joining their team for about a year. And so they maintain the library, but also run a yearly competition. That was my first foray into competitions. And I was still working on diplomacy. At some point I was working on this translation task between Dade, which is a diplomacy specific bot language and English. And I started using GPT-3 prompting it to do the translation. And that was, I think, my first intro to prompting. And I just started doing a bunch of reading about prompting. And I had an English class project where we had to write a guide on something that ended up being learn prompting. So I figured, all right, well, I'm learning about prompting anyways. You know, Chain of Thought was out at this point. There are a couple blog posts floating around, but there was no website you could go to just sort of read everything about prompting. So I made that. And it ended up getting super popular. Now continuing with it, supporting the project now after college. And then the other very interesting things, of course, are the two papers I wrote. And that is the prompt report and hack a prompt. So I saw Simon and Riley's original tweets about prompt injection go across my feed. And I put that information into the learn prompting website. And I knew, because I had some previous competition running experience, that someone was going to run a competition with prompt injection. And I waited a month, figured, you know, I'd participate in one of these that comes out. No one was doing it. So I was like, what the heck, I'll give it a shot. Just started reaching out to people. Got some people from Mila involved, some people from Maryland, and raised a good amount of sponsorship. I had no experience doing that, but just reached out to as many people as I could. And we actually ended up getting literally all the sponsors I wanted. So like OpenAI, actually, they reached out to us a couple months after I started learn prompting. And then Preamble is the company that first discovered prompt injection even before Riley. And they like responsibly disclosed it kind of internally to OpenAI. And having them on board as the largest sponsor was super exciting. And then we ran that, collected 600,000 malicious prompts, put together a paper on it, open sourced everything. And we took it to EMNLP, which is one of the top natural language processing conferences in the world. 20,000 papers were submitted to that conference, 5,000 papers were accepted. We were one of three selected as best papers at the conference, which was just massive. Super, super exciting. I got to give a talk to like a couple thousand researchers there, which was also very exciting. And I kind of carried that momentum into the next paper, which was the prompt report. It was kind of a natural extension of what I had been doing with learn prompting in the sense that we had this website bringing together all of the different prompting techniques, survey website in and of itself. So writing an actual survey, a systematic survey was the next step that we did in the prompt report. So over the course of about nine months, I led a 30 person research team with people from OpenAI, Google, Microsoft, Princeton, Stanford, Maryland, a number of other universities and companies. And we pretty much read thousands of papers on prompting and compiled it all into like a 80 page massive summary doc. And then we put it on archive and the response was amazing. We've gotten millions of views across socials. I actually put together a spreadsheet where I've been able to track about one and a half million. And I just kind of figure if I can find that many, then there's many more views out there. It's been really great. We
Congrats to Damien on successfully running AI Engineer London! See our community page and the Latent Space Discord for all upcoming events.This podcast came together in a far more convoluted way than usual, but happens to result in a tight 2 hours covering the ENTIRE OpenAI product suite across ChatGPT-latest, GPT-4o and the new o1 models, and how they are delivered to AI Engineers in the API via the new Structured Output mode, Assistants API, client SDKs, upcoming Voice Mode API, Finetuning/Vision/Whisper/Batch/Admin/Audit APIs, and everything else you need to know to be up to speed in September 2024.This podcast has two parts: the first hour is a regular, well edited, podcast on 4o, Structured Outputs, and the rest of the OpenAI API platform. The second was a rushed, noisy, hastily cobbled together recap of the top takeaways from the o1 model release from yesterday and today.Building AGI with Structured Outputs — Michelle Pokrass of OpenAI API teamMichelle Pokrass built massively scalable platforms at Google, Stripe, Coinbase and Clubhouse, and now leads the API Platform at Open AI. She joins us today to talk about why structured output is such an important modality for AI Engineers that Open AI has now trained and engineered a Structured Output mode with 100% reliable JSON schema adherence. To understand why this is important, a bit of history is important:* June 2023 when OpenAI first added a "function calling" capability to GPT-4-0613 and GPT 3.5 Turbo 0613 (our podcast/writeup here)* November 2023’s OpenAI Dev Day (our podcast/writeup here) where the team shipped JSON Mode, a simpler schema-less JSON output mode that nevertheless became more popular because function calling often failed to match the JSON schema given by developers. * Meanwhile, in open source, many solutions arose, including * Instructor (our pod with Jason here) * LangChain (our pod with Harrison here, and he is returning next as a guest co-host)* Outlines (Remi Louf’s talk at AI Engineer here)* Llama.cpp’s constrained grammar sampling using GGML-BNF* April 2024: OpenAI started implementing constrained sampling with a new `tool_choice: required` parameter in the API* August 2024: the new Structured Output mode, co-led by Michelle* Sept 2024: Gemini shipped Structured Outputs as wellWe sat down with Michelle to talk through every part of the process, as well as quizzing her for updates on everything else the API team has shipped in the past year, from the Assistants API, to Prompt Caching, GPT4 Vision, Whisper, the upcoming Advanced Voice Mode API, OpenAI Enterprise features, and why every Waterloo grad seems to be a cracked engineer.Part 1 Timestamps and TranscriptTranscript here.* [00:00:42] Episode Intro from Suno* [00:03:34] Michelle's Path to OpenAI* [00:12:20] Scaling ChatGPT* [00:13:20] Releasing Structured Output* [00:16:17] Structured Outputs vs Function Calling* [00:19:42] JSON Schema and Constrained Grammar* [00:20:45] OpenAI API team* [00:21:32] Structured Output Refusal Field* [00:24:23] ChatML issues* [00:26:20] Function Calling Evals* [00:28:34] Parallel Function Calling* [00:29:30] Increased Latency* [00:30:28] Prompt/Schema Caching* [00:30:50] Building Agents with Structured Outputs: from API to AGI* [00:31:52] Assistants API* [00:34:00] Use cases for Structured Output* [00:37:45] Prompting Structured Output* [00:39:44] Benchmarking Prompting for Structured Outputs* [00:41:50] Structured Outputs Roadmap* [00:43:37] Model Selection vs GPT4 Finetuning* [00:46:56] Is Prompt Engineering Dead?* [00:47:29] 2 models: ChatGPT Latest vs GPT 4o August* [00:50:24] Why API => AGI* [00:52:40] Dev Day* [00:54:20] Assistants API Roadmap* [00:56:14] Model Reproducibility/Determinism issues* [00:57:53] Tiering and Rate Limiting* [00:59:26] OpenAI vs Ops Startups* [01:01:06] Batch API* [01:02:54] Vision* [01:04:42] Whisper* [01:07:21] Voice Mode API* [01:08:10] Enterprise: Admin/Audit Log APIs* [01:09:02] Waterloo grads* [01:10:49] Books* [01:11:57] Cognitive Biases* [01:13:25] Are LLMs Econs?* [01:13:49] Hiring at OpenAIEmergency O1 Meetup — OpenAI DevRel + Strawberry teamthe following is our writeup from AINews, which so far stands the test of time.o1, aka Strawberry, aka Q*, is finally out! There are two models we can use today: o1-preview (the bigger one priced at $15 in / $60 out) and o1-mini (the STEM-reasoning focused distillation priced at $3 in/$12 out) - and the main o1 model is still in training. This caused a little bit of confusion.There are a raft of relevant links, so don’t miss:* the o1 Hub* the o1-preview blogpost* the o1-mini blogpost* the technical research blogpost* the o1 system card* the platform docs* the o1 team video and contributors list (twitter)Inline with the many, many leaks leading up to today, the core story is longer “test-time inference” aka longer step by step responses - in the ChatGPT app this shows up as a new “thinking” step that you can click to expand for reasoning traces, even though, controversially, they are hidden from you (interesting conflict of interest…):Under the hood, o1 is trained for adding new reasoning tokens - which you pay for, and OpenAI has accordingly extended the output token limit to >30k tokens (incidentally this is also why a number of API parameters from the other models like temperature and role and tool calling and streaming, but especially max_tokens is no longer supported).The evals are exceptional. OpenAI o1:* ranks in the 89th percentile on competitive programming questions (Codeforces),* places among the top 500 students in the US in a qualifier for the USA Math Olympiad (AIME),* and exceeds human PhD-level accuracy on a benchmark of physics, biology, and chemistry problems (GPQA).You are used to new models showing flattering charts, but there is one of note that you don’t see in many model announcements, that is probably the most important chart of all. Dr Jim Fan gets it right: we now have scaling laws for test time compute, and it looks like they scale loglinearly.We unfortunately may never know the drivers of the reasoning improvements, but Jason Wei shared some hints:Usually the big model gets all the accolades, but notably many are calling out the performance of o1-mini for its size (smaller than gpt 4o), so do not miss that.Part 2 Timestamps* [01:15:01] O1 transition* [01:16:07] O1 Meetup Recording* [01:38:38] OpenAI Friday AMA recap* [01:44:47] Q&A Part 2* [01:50:28] O1 DemosDemo Videos to be posted shortly Get full access to Latent Space at www.latent.space/subscribe
AI Engineering is expanding! Join the first 🇬🇧 AI Engineer London meetup in Sept and get in touch for sponsoring the second 🗽 AI Engineer Summit in NYC this Dec!The commoditization of intelligence takes on a few dimensions:* Time to Open Model Equivalent: 15 months between GPT-4 and Llama 3.1 405B * 10-100x CHEAPER/year: from $30/mtok for Claude 3 Opus to $3/mtok for L3-405B, and a 400x reduction in the frontier OpenAI model from 2022-2024. Notably, for personal use cases, both Gemini Flash and now Cerebras Inference offer 1m tokens/day inference free, causing the Open Model Red Wedding.* Alternatively you can observe the frontiers of various small/medium/large sizes of intelligence per dollar shift in realtime. 2024 has been particularly aggressive with almost 2 order-of-magnitude improvements in $/Elo points in the last 8 months.* 4-8x FASTER/year: The new Cerebras Inference platform runs 70B models at 450 tok/s, almost twice as fast as the Groq Cloud example that went viral earlier this year (and at $0.60/mtok to boot). James Wang says they have room to ”~8x throughput in the next few months”, which needs to be seen in reality and at scale, but is very exciting for downstream latency/throughput-sensitive usecases.Today’s guest, Nyla Worker, a senior PM at Nvidia, Convai, and now Google, and recently host of the GPUs & Inference track at the World’s Fair, was the first to point out to us that the kind of efficiency improvements that have become a predominant theme in LLMs in 2024, have been seen before in her career in computer vision. From her start at Ebay optimizing V100 inference for a ResNet-50 model for image search, she has watched many improvements like Multi-Inference GPU (allowing multiple instances with perfect hardware parallelism), Quantization Aware Training (most recently highlighted by Noam Shazeer pre Character AI departure) and Model Distillation (most recently highlighted by the Llama 3.1 paper) stacking with baseline hardware improvements (from V100s to A100s to H100s to GH200s) to produce theoretically 3000x faster inference now than 6 years ago.What Nyla saw in her career the last 6 years, is happening to LLMs today (not exactly repeating, but surely rhyming), specifically with LoRAs, native Int8 and even Ternary models, and teacher model distillation. We were excited to delve into all things efficiency in this episode and even come out the other side with bonus discussions on what generative AI can do for gaming, fanmade TV shows, character AI conversations, and even podcasting!Show Notes:* Nyla Linkedin, Twitter* Related Nvidia research* Improving INT8 Accuracy Using Quantization Aware Training and the NVIDIA TAO Toolkit* Nvidia Jetson Nano: Bringing the power of modern AI to millions of devices.* Synthetic Data with Nvidia Omniverse Replicator: Accelerate AI Training Faster Than Ever with New NVIDIA Omniverse Replicator CapabilitiesTimestamps* [00:00:00] Intro from Suno* [00:03:17] Nyla's path from Astrophysics to LLMs* [00:05:45] Efficiency Curves in Computer Vision at Nvidia* [00:09:51] Optimizing for today's hardware vs tomorrow's inference* [00:16:33] Quantization vs Precision tradeoff* [00:20:42] Hitting the Data Wall: The need for Synthetic Data at Nvidia* [00:26:20] Sora, text to 3D models, and Synthetic Data from Game Engines* [00:30:55] ResNet 50 keeps coming back* [00:35:40] Gaming Benchmarks* [00:38:00] FineWeb* [00:39:43] Traditional ML vs LLMs path to general intelligence* [00:42:33] ConvAI - AI NPCs* [00:45:32] Jensen and Lisa at Computex Taiwan* [00:52:51] NPCs need to take Actions and have Context* [00:54:29] Simulating different roles for training* [00:58:37] AI Generated Fan Content - Podcasts, TV Show, EinsteinTranscripts[00:00:29] AI Charlie: Happy September. This is your AI co host, Charlie.[00:00:34] AI Charlie: One topic we've developed on LatentSpace is the importance of efficiency in all forms, from sample efficiency for spending limited training compute on limited data, and increasingly towards inference efficiency for increasingly demanding use cases like local LLMs, real time AI NPCs, and edge AI. However, we've never really developed any intuition for the trends and efficiency over time.[00:00:59] AI Charlie: For example, from 2020 to 2023, the price of GPT 3 level intelligence dropped from 60 per million tokens to 27 cents with the mixtural price war of December 2023. See show notes for charts and data. As for GPT 4 level intelligence, it took just over a year for GPT 4 to be matched by LLAMA370B and GPT 4 Turbo to be beaten by LLAMA3405B in open source, causing blended cost per million tokens to freefall from over 30 for Claude III Opus and the original GPT 4 down to under 3 for LLAMA3405B.[00:01:43] AI Charlie: Of course, OpenAI themselves have not stood still, slashing the price of GPT 4. 0 by 30 times with GPT 4. 0 Mini. Yes, you heard that right. GPT 4. 0 Mini is 3. 5 percent the price of GPT 4. 0, yet ties with GPT 4 Turbo on LM SYS. When the price of intelligence is falling by over 90 percent every year. What are the driving forces?[00:02:10] AI Charlie: And how should AI engineers plan for this? It turns out that this has happened before in computer vision, which has seen an almost 3, 000 times latency improvement over the last 6 years. We invited Nila Worker of NVIDIA and Convay. Who first made this comparison to help talk us through the past, present, and future use cases of efficient AI inference.[00:02:35] AI Charlie: Note that this was recorded before Naila joined Google AI to work on efficiency, so you can expect more great efficiency work coming from her on the Gemini team. In latent space news, look out for our upcoming London and NYC meetups on the community page, and of course feel free to start your own and simply let us know.[00:02:54] AI Charlie: Watch out and take care.[00:02:57] Alessio: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO in residence at Decibel Partners, and I'm joined by my co host Swyx, founder of Small. ai.[00:03:11] Hey, and today we are in the remote studio with Naila Worko. Welcome, Naila. Good to see you.[00:03:16] Nyla Worker: Good to see you all.[00:03:17] Nyla's path from Astrophysics to LLMs[00:03:17] swyx: So we try to introduce people based on sort of their professional profile and then let you fill in the blanks.[00:03:22] swyx: Um, so you did astrophysics research at Carleton College, uh, and then you made your way into machine learning. We're going to talk about your time at eBay, but most recently you spent four years at Nvidia, uh, working on everything from synthetic data to cloud container offerings. And now currently you're director of product management at Convai.[00:03:41] swyx: What should people know about you that maybe it's not super obvious on your LinkedIn that it's, you know. Encapsulates your life journey so far.[00:03:47] Nyla Worker: And yeah, I think the thing that is not very obvious is that transition from astrophysics research to AI and how that happens. So within astrophysics, what I was doing on my freshman year of college was categorizing whether this was a supernova Rembrandt or like an exoplanet.[00:04:06] Nyla Worker: And while that sounds all cool and incredible, it's literally looking at images of like Oxygen and sulfur and selecting manually each region. And it is extremely boring, shall I say. So I then found a paper from 1996, um, called Source Extractor, or like he called it Sextractor for some reason. And it was a multi layer perception network that had been trained on synthetic data.[00:04:38] Nyla Worker: To categorize whether this was a star or a galaxy, that led me to see that there was this massive optimization machine that when fed with right data, it could perform and automate tasks such as this kind of manual classification. That made me want to learn more. How do you train these things? How do you deploy them effectively?[00:05:00] Nyla Worker: And if it's useful for just classifying galaxies, what other applications are there out there where we show a bunch of data and just train these functions to just predict the next word in the case of LLMs or predict, uh, what is. Is this a cat or a dog and things like that. So then I went to computer vision research, particularly scaling the training of deep neural networks.[00:05:24] Nyla Worker: Back then I was using CPUs, doing it wrongly, of course. Uh, and then I went to eBay where I switched to GPUs, but I was working also on like the Jetsons and Edge devices. That is an interesting transition in how it all flows together.[00:05:41] swyx: We can talk about that and also how you transition from that into NVIDIA.[00:05:45] Efficiency Curves in Computer Vision at Nvidia[00:05:45] swyx: But like, yeah, a lot of the podcasts for today, we're actually talking about efficiency and efficiency curves over time. And The reason I invited you to this pod was I was basically looking for somebody to talk about this. And you came at this with your insight on how like this already happens with computer vision, right?[00:06:06] swyx: This sort of efficiency curve over time. So I wonder if you want to just comment about Just set the context for like what has happened in your career that you've seen already.[00:06:15] Nyla Worker: When I started was first scaling up training and making training more efficient. And that of course has evolved significantly over time.[00:06:22] Nyla Worker: There is a lot on training. But what I discovered is that if these things are truly useful, you should be obsessing about inference. And then I went to eBay, uh, where I was in their hardware team, but I was doing software optimizations for the hardware team, such that the research that had been done for the AI research team was actually running efficiently on the hardware.[00:06:45] Nyla Worker: And there, I started leveraging optimization, uh, frameworks such as TensorRT to optimize our models like ResNet 50. S
Today's guest, Nicholas Carlini, a research scientist at DeepMind, argues that we should be focusing more on what AI can do for us individually, rather than trying to have an answer for everyone."How I Use AI" - A Pragmatic ApproachCarlini's blog post "How I Use AI" went viral for good reason. Instead of giving a personal opinion about AI's potential, he simply laid out how he, as a security researcher, uses AI tools in his daily work. He divided it in 12 sections:* To make applications* As a tutor* To get started* To simplify code* For boring tasks* To automate tasks* As an API reference* As a search engine* To solve one-offs* To teach me* Solving solved problems* To fix errorsEach of the sections has specific examples, so we recommend going through it. It also includes all prompts used for it; in the "make applications" case, it's 30,000 words total!My personal takeaway is that the majority of the work AI can do successfully is what humans dislike doing. Writing boilerplate code, looking up docs, taking repetitive actions, etc. These are usually boring tasks with little creativity, but with a lot of structure. This is the strongest arguments as to why LLMs, especially for code, are more beneficial to senior employees: if you can get the boring stuff out of the way, there's a lot more value you can generate. This is less and less true as you go entry level jobs which are mostly boring and repetitive tasks. Nicholas argues both sides ~21:34 in the pod.A New Approach to LLM BenchmarksWe recently did a Benchmarks 201 episode, a follow up to our original Benchmarks 101, and some of the issues have stayed the same. Notably, there's a big discrepancy between what benchmarks like MMLU test, and what the models are used for. Carlini created his own domain-specific language for writing personalized LLM benchmarks. The idea is simple but powerful:* Take tasks you've actually needed AI for in the past.* Turn them into benchmark tests.* Use these to evaluate new models based on your specific needs.It can represent very complex tasks, from a single code generation to drawing a US flag using C:"Write hello world in python" >> LLMRun() >> PythonRun() >> SubstringEvaluator("hello world")"Write a C program that draws an american flag to stdout." >> LLMRun() >> CRun() >> \ VisionLLMRun("What flag is shown in this image?") >> \ (SubstringEvaluator("United States") | SubstringEvaluator("USA")))This approach solves a few problems:* It measures what's actually useful to you, not abstract capabilities.* It's harder for model creators to "game" your specific benchmark, a problem that has plagued standardized tests.* It gives you a concrete way to decide if a new model is worth switching to, similar to how developers might run benchmarks before adopting a new library or framework.Carlini argues that if even a small percentage of AI users created personal benchmarks, we'd have a much better picture of model capabilities in practice.AI SecurityWhile much of the AI security discussion focuses on either jailbreaks or existential risks, Carlini's research targets the space in between. Some highlights from his recent work:* LAION 400M data poisoning: By buying expired domains referenced in the dataset, Carlini's team could inject arbitrary images into models trained on LAION 400M. You can read the paper "Poisoning Web-Scale Training Datasets is Practical", for all the details. This is a great example of expanding the scope beyond the model itself, and looking at the whole system and how ti can become vulnerable.* Stealing model weights: They demonstrated how to extract parts of production language models (like OpenAI's) through careful API queries. This research, "Extracting Training Data from Large Language Models", shows that even black-box access can leak sensitive information.* Extracting training data: In some cases, they found ways to make models regurgitate verbatim snippets from their training data. Him and Milad Nasr wrote a paper on this as well: Scalable Extraction of Training Data from (Production) Language Models. They also think this might be applicable to extracting RAG results from a generation.These aren't just theoretical attacks. They've led to real changes in how companies like OpenAI design their APIs and handle data. If you really miss logit_bias and logit results by token, you can blame Nicholas :)We had a ton of fun also chatting about things like Conway's Game of Life, how much data can fit in a piece of paper, and porting Doom to Javascript. Enjoy!Show Notes* How I Use AI* My Benchmark for LLMs* Doom Javascript port* Conway's Game of Life* Tic-Tac-Toe in one printf statement* International Obfuscated C Code Contest* Cursor* LAION 400M poisoning paper* Man vs Machine at Black Hat* Model Stealing from OpenAI* Milad Nasr* H.D. Moore* Vijay Bolina* Cosine.sh* uuencodeTimestamps* [00:00:00] Introductions* [00:01:14] Why Nicholas writes* [00:02:09] The Game of Life* [00:05:07] "How I Use AI" blog post origin story* [00:08:24] Do we need software engineering agents?* [00:11:03] Using AI to kickstart a project* [00:14:08] Ephemeral software* [00:17:37] Using AI to accelerate research* [00:21:34] Experts vs non-expert users as beneficiaries of AI* [00:24:02] Research on generating less secure code with LLMs.* [00:27:22] Learning and explaining code with AI* [00:30:12] AGI speculations?* [00:32:50] Distributing content without social media* [00:35:39] How much data do you think you can put on a single piece of paper?* [00:37:37] Building personal AI benchmarks* [00:43:04] Evolution of prompt engineering and its relevance* [00:46:06] Model vs task benchmarking* [00:52:14] Poisoning LAION 400M through expired domains* [00:55:38] Stealing OpenAI models from their API* [01:01:29] Data stealing and recovering training data from models* [01:03:30] Finding motivation in your workTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO-in-Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI.Swyx [00:00:12]: Hey, and today we're in the in-person studio, which Alessio has gorgeously set up for us, with Nicholas Carlini. Welcome. Thank you. You're a research scientist at DeepMind. You work at the intersection of machine learning and computer security. You got your PhD from Berkeley in 2018, and also your BA from Berkeley as well. And mostly we're here to talk about your blogs, because you are so generous in just writing up what you know. Well, actually, why do you write?Nicholas [00:00:41]: Because I like, I feel like it's fun to share what you've done. I don't like writing, sufficiently didn't like writing, I almost didn't do a PhD, because I knew how much writing was involved in writing papers. I was terrible at writing when I was younger. I do like the remedial writing classes when I was in university, because I was really bad at it. So I don't actually enjoy, I still don't enjoy the act of writing. But I feel like it is useful to share what you're doing, and I like being able to talk about the things that I'm doing that I think are fun. And so I write because I think I want to have something to say, not because I enjoy the act of writing.Swyx [00:01:14]: But yeah. It's a tool for thought, as they often say. Is there any sort of backgrounds or thing that people should know about you as a person? Yeah.Nicholas [00:01:23]: So I tend to focus on, like you said, I do security work, I try to like attacking things and I want to do like high quality security research. And that's mostly what I spend my actual time trying to be productive members of society doing that. But then I get distracted by things, and I just like, you know, working on random fun projects. Like a Doom clone in JavaScript.Swyx [00:01:44]: Yes.Nicholas [00:01:45]: Like that. Or, you know, I've done a number of things that have absolutely no utility. But are fun things to have done. And so it's interesting to say, like, you should work on fun things that just are interesting, even if they're not useful in any real way. And so that's what I tend to put up there is after I have completed something I think is fun, or if I think it's sufficiently interesting, write something down there.Alessio [00:02:09]: Before we go into like AI, LLMs and whatnot, why are you obsessed with the game of life? So you built multiplexing circuits in the game of life, which is mind boggling. So where did that come from? And then how do you go from just clicking boxes on the UI web version to like building multiplexing circuits?Nicholas [00:02:29]: I like Turing completeness. The definition of Turing completeness is a computer that can run anything, essentially. And the game of life, Conway's game of life is a very simple cellular 2D automata where you have cells that are either on or off. And a cell becomes on if in the previous generation some configuration holds true and off otherwise. It turns out there's a proof that the game of life is Turing complete, that you can run any program in principle using Conway's game of life. I don't know. And so you can, therefore someone should. And so I wanted to do it. Some other people have done some similar things, but I got obsessed into like, if you're going to try and make it work, like we already know it's possible in theory. I want to try and like actually make something I can run on my computer, like a real computer I can run. And so yeah, I've been going on this rabbit hole of trying to make a CPU that I can run semi real time on the game of life. And I have been making some reasonable progress there. And yeah, but you know, Turing completeness is just like a very fun trap you can go down. A while ago, as part of a research paper, I was able to show that in C, if you call into printf, it's Turing complete. Like printf, you know, like, which like, you know, you can print numbers or whatever, right?Swyx [00:03:39]: Yeah, but there should be no like control flow
Betteridge's law says no: with seemingly infinite flavors of RAG, and >2million token context + prompt caching from Anthropic/Deepmind/Deepseek, it's reasonable to believe that "in context learning is all you need".But then there’s Cosine Genie, the first to make a huge bet using OpenAI’s new GPT4o fine-tuning for code at the largest scale it has ever been used externally; resulting in what is now the #1 coding agent in the world according to SWE-Bench Full, Lite, and Verified:SWE-Bench has been the most successful agent benchmark of the year, receiving honors at ICLR (our interview here) and recently being verified by OpenAI. Cognition (Devin) was valued at $2b after reaching 14% on it. So it is very, very big news when a new agent appears to beat all other solutions, by a lot:While this number is self reported, it seems to be corroborated by OpenAI, who also award it clear highest marks on SWE-Bench verified:The secret is GPT-4o finetuning on billions of tokens of synthetic data. * Finetuning: As OpenAI says:Genie is powered by a fine-tuned GPT-4o model trained on examples of real software engineers at work, enabling the model to learn to respond in a specific way. The model was also trained to be able to output in specific formats, such as patches that could be committed easily to codebases. Due to the scale of Cosine’s finetuning, OpenAI worked closely with them to figure out the size of the LoRA:“They have to decide how big your LoRA adapter is going to be… because if you had a really sparse, large adapter, you’re not going to get any signal in that at all. So they have to dynamically size these things.”* Synthetic data: we need to finetune on the process of making code work instead of only training on working code.“…we synthetically generated runtime errors. Where we would intentionally mess with the AST to make stuff not work, or index out of bounds, or refer to a variable that doesn't exist, or errors that the foundational models just make sometimes that you can't really avoid, you can't expect it to be perfect.”Genie also has a 4 stage workflow with the standard LLM OS tooling stack that lets it solve problems iteratively:Full Video Podlike and subscribe etc!Show Notes* Alistair Pullen - Twitter, Linkedin* Cosine Genie launch, technical report* OpenAI GPT-4o finetuning GA* Llama 3 backtranslation* Cursor episode and Aman + SWEBench at ICLR episodeTimestamps* [00:00:00] Suno Intro* [00:05:01] Alistair and Cosine intro* [00:16:34] GPT4o finetuning* [00:20:18] Genie Data Mix* [00:23:09] Customizing for Customers* [00:25:37] Genie Workflow* [00:27:41] Code Retrieval* [00:35:20] Planning* [00:42:29] Language Mix* [00:43:46] Running Code* [00:46:19] Finetuning with OpenAI* [00:49:32] Synthetic Code Data* [00:51:54] SynData in Llama 3* [00:52:33] SWE-Bench Submission Process* [00:58:20] Future Plans* [00:59:36] Ecosystem Trends* [01:00:55] Founder Lessons* [01:01:58] CTA: Hiring & CustomersDescript Transcript[00:01:52] AI Charlie: Welcome back. This is Charlie, your AI cohost. As AI engineers, we have a special focus on coding agents, fine tuning, and synthetic data. And this week, it all comes together with the launch of Cosign's Genie, which reached 50 percent on SWE Bench Lite, 30 percent on the full SWE Bench, and 44 percent on OpenAI's new SWE Bench Verified.[00:02:17] All state of the art results by the widest ever margin recorded compared to former leaders Amazon Q and US Autocode Rover. And Factory Code Droid. As a reminder, Cognition Devon went viral with a 14 percent score just five months ago. Cosign did this by working closely with OpenAI to fine tune GPT 4. 0, now generally available to you and me, on billions of tokens of code, much of which was synthetically generated.[00:02:47] Alistair Pullen: Hi, I'm Ali. Co founder and CEO of Cosign, a human reasoning lab. And I'd like to show you Genie, our state of the art, fully autonomous software engineering colleague. Genie has the highest score on SWBench in the world. And the way we achieved this was by taking a completely different approach. We believe that if you want a model to behave like a software engineer, it has to be shown how a human software engineer works.[00:03:15] We've designed new techniques to derive human reasoning from real examples of software engineers doing their jobs. Our data represents perfect information lineage, incremental knowledge discovery, and step by step decision making. Representing everything a human engineer does logically. By actually training Genie on this unique dataset, rather than simply prompting base models, which is what everyone else is doing, we've seen that we're no longer simply generating random code until some works.[00:03:46] It's tackling problems like[00:03:48] AI Charlie: a human. Alistair Pullen is CEO and co founder of Kozen, and we managed to snag him on a brief trip stateside for a special conversation on building the world's current number one coding agent. Watch out and take care.[00:04:07] Alessio: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO of Resonance at Decibel Partners, and I'm joined by my co host Swyx, founder of Small. ai.[00:04:16] swyx: Hey, and today we're back in the studio. In person, after about three to four months in visa jail and travels and all other fun stuff that we talked about in the previous episode.[00:04:27] But today we have a special guest, Ali Pullen from Cosign. Welcome. Hi, thanks for having me. We're very lucky to have you because you're on a two day trip to San Francisco. Yeah, I wouldn't recommend it. I would not[00:04:38] Alistair Pullen: recommend it. Don't fly from London to San Francisco for two days.[00:04:40] swyx: And you launched Genie on a plane.[00:04:42] On plain Wi Fi, um, claiming state of the art in SuiteBench, which we're all going to talk about. I'm excited to dive into your whole journey, because it has been a journey. I've been lucky to be a small angel in part of that journey. And it's exciting to see that you're launching to such acclaim and, you know, such results.[00:05:01] Alistair and Cosine intro[00:05:01] swyx: Um, so I'll go over your brief background, and then you can sort of fill in the blanks on what else people should know about you. You did your bachelor's in computer science at Exeter.[00:05:10] Speaker 6: Yep.[00:05:10] swyx: And then you worked at a startup that got acquired into GoPuff and round about 2022, you started working on a stealth startup that became a YC startup.[00:05:19] What's that? Yeah. So[00:05:21] Alistair Pullen: basically when I left university, I, I met my now co founder, Sam. At the time we were both mobile devs. He was an Android developer. iOS developer. And whilst at university, we built this sort of small consultancy, sort of, we'd um, be approached to build projects for people and we would just take them up and start with, they were student projects.[00:05:41] They weren't, they weren't anything crazy or anything big. We started with those and over time we started doing larger and larger projects, more interesting things. And then actually, when we left university, we just kept doing that. We didn't really get jobs, traditional jobs. It was also like in the middle of COVID, middle of lockdown.[00:05:57] So we were like, this is a pretty good gig. We'll just keep like writing code in our bedrooms. And yeah, that's it. We did that for a while. And then a friend of ours that we went to Exeter with started a YC startup during COVID. And it was one of these fast grocery delivery companies. At the time I was living in the deepest, darkest countryside in England, where fast grocery companies are still not a thing.[00:06:20] So he, he sort of pitched me this idea and was like, listen, like I need an iOS dev, do you fancy coming along? And I thought, absolutely. It was a chance to get out of my parents house, chance to move to London, you know, do interesting things. And at the time, truthfully, I had no idea what YC was. I had no idea.[00:06:34] I wasn't in the startup space. I knew I liked coding and building apps and stuff, but I'd never, never really done anything in that area. So I said, yes, absolutely. I moved to London just sort of as COVID was ending and yeah, worked at what was fancy for about a year and a half. Then we brought Sam along as well.[00:06:52] So we, Sam and I, were the two engineers at Fancy for basically its entire life, and we built literally everything. So like the, the front, the client mobile apps, the, the backends, the internal like stock management system, the driver routing, algorithms, all those things. Literally like everything. It was my first.[00:07:12] You know, both of us were super inexperienced. We didn't have, like, proper engineering experience. There were definitely decisions we'd do differently now. We'd definitely buy a lot of stuff off the shelf, stuff like that. But it was the initial dip of the toe into, like, the world of startups, and we were both, like, hooked immediately.[00:07:26] We were like, this is so cool. This sounds so much better than all our friends who were, like, consultants and doing, like, normal jobs, right? We did that, and it ran its course, and after, I want to say, 18 months or so, GoPuff came and acquired us. And there was obviously a transitionary period, an integration period, like with all acquisitions, and we did that, and as soon as we'd vested what we wanted to vest, and as soon as we thought, okay, this chapter is sort of done, uh, in about 2022, We left and we knew that we wanted to go alone and try something like we'd had this taste.[00:07:54] Now we knew we'd seen how a like a YC startup was managed like up close and we knew that we wanted to do something similar ourselves. We had no idea what it was at the time. We just knew we wanted to do something. So we, we tried a small, um, some small projects in various different areas, but then GPT 3.[00:08:12] He'd seen it on Reddit and I
Disclaimer: We recorded this episode ~1.5 months ago, timing for the FastHTML release. It then got bottlenecked by Llama3.1, Winds of AI Winter, and SAM2 episodes, so we’re a little late. Since then FastHTML was released, swyx is building an app in it for AINews, and Anthropic has also released their prompt caching API. Remember when Dylan Patel of SemiAnalysis coined the GPU Rich vs GPU Poor war? (if not, see our pod with him). The idea was that if you’re GPU poor you shouldn’t waste your time trying to solve GPU rich problems (i.e. pre-training large models) and are better off working on fine-tuning, optimized inference, etc. Jeremy Howard (see our “End of Finetuning” episode to catchup on his background) and Eric Ries founded Answer.AI to do exactly that: “Practical AI R&D”, which is very in-line with the GPU poor needs. For example, one of their first releases was a system based on FSDP + QLoRA that let anyone train a 70B model on two NVIDIA 4090s. Since then, they have come out with a long list of super useful projects (in no particular order, and non-exhaustive):* FSDP QDoRA: this is just as memory efficient and scalable as FSDP/QLoRA, and critically is also as accurate for continued pre-training as full weight training.* Cold Compress: a KV cache compression toolkit that lets you scale sequence length without impacting speed.* colbert-small: state of the art retriever at only 33M params* JaColBERTv2.5: a new state-of-the-art retrievers on all Japanese benchmarks.* gpu.cpp: portable GPU compute for C++ with WebGPU.* Claudette: a better Anthropic API SDK. They also recently released FastHTML, a new way to create modern interactive web apps. Jeremy recently released a 1 hour “Getting started” tutorial on YouTube; while this isn’t AI related per se, but it’s close to home for any AI Engineer who are looking to iterate quickly on new products: In this episode we broke down 1) how they recruit 2) how they organize what to research 3) and how the community comes together. At the end, Jeremy gave us a sneak peek at something new that he’s working on that he calls dialogue engineering: So I've created a new approach. It's not called prompt engineering. I'm creating a system for doing dialogue engineering. It's currently called AI magic. I'm doing most of my work in this system and it's making me much more productive than I was before I used it.He explains it a bit more ~44:53 in the pod, but we’ll just have to wait for the public release to figure out exactly what he means.Timestamps* [00:00:00] Intro by Suno AI* [00:03:02] Continuous Pre-Training is Here* [00:06:07] Schedule-Free Optimizers and Learning Rate Schedules* [00:07:08] Governance and Structural Issues within OpenAI and Other AI Labs* [00:13:01] How Answer.ai works* [00:23:40] How to Recruit Productive Researchers* [00:27:45] Building a new BERT* [00:31:57] FSDP, QLoRA, and QDoRA: Innovations in Fine-Tuning Large Models* [00:36:36] Research and Development on Model Inference Optimization* [00:39:49] FastHTML for Web Application Development* [00:46:53] AI Magic & Dialogue Engineering* [00:52:19] AI wishlist & predictionsShow Notes* Jeremy Howard* Previously on Latent Space: The End of Finetuning, NeurIPS Startups* Answer.ai* Fast.ai* FastHTML* answerai-colbert-small-v1* gpu.cpp* Eric Ries* Aaron DeFazio* Yi Tai* Less Wright* Benjamin Warner* Benjamin Clavié* Jono Whitaker* Austin Huang* Eric Gilliam* Tim Dettmers* Colin Raffel* Mark Saroufim* Sebastian Raschka* Carson Gross* Simon Willison* Sepp Hochreiter* Llama3.1 episode* Snowflake Arctic* Ranger Optimizer* Gemma.cpp* HTMX* UL2* BERT* DeBERTa* Efficient finetuning of Llama 3 with FSDP QDoRA* xLSTMTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO-in-Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI.Swyx [00:00:14]: And today we're back with Jeremy Howard, I think your third appearance on Latent Space. Welcome.Jeremy [00:00:19]: Wait, third? Second?Swyx [00:00:21]: Well, I grabbed you at NeurIPS.Jeremy [00:00:23]: I see.Swyx [00:00:24]: Very fun, standing outside street episode.Jeremy [00:00:27]: I never heard that, by the way. You've got to send me a link. I've got to hear what it sounded like.Swyx [00:00:30]: Yeah. Yeah, it's a NeurIPS podcast.Alessio [00:00:32]: I think the two episodes are six hours, so there's plenty to listen, we'll make sure to send it over.Swyx [00:00:37]: Yeah, we're trying this thing where at the major ML conferences, we, you know, do a little audio tour of, give people a sense of what it's like. But the last time you were on, you declared the end of fine tuning. I hope that I sort of editorialized the title a little bit, and I know you were slightly uncomfortable with it, but you just own it anyway. I think you're very good at the hot takes. And we were just discussing in our pre-show that it's really happening, that the continued pre-training is really happening.Jeremy [00:01:02]: Yeah, absolutely. I think people are starting to understand that treating the three ULM FIT steps of like pre-training, you know, and then the kind of like what people now call instruction tuning, and then, I don't know if we've got a general term for this, DPO, RLHFE step, you know, or the task training, they're not actually as separate as we originally suggested they were in our paper, and when you treat it more as a continuum, and that you make sure that you have, you know, more of kind of the original data set incorporated into the later stages, and that, you know, we've also seen with LLAMA3, this idea that those later stages can be done for a lot longer. These are all of the things I was kind of trying to describe there. It wasn't the end of fine tuning, but more that we should treat it as a continuum, and we should have much higher expectations of how much you can do with an already trained model. You can really add a lot of behavior to it, you can change its behavior, you can do a lot. So a lot of our research has been around trying to figure out how to modify the model by a larger amount rather than starting from random weights, because I get very offended at the idea of starting from random weights.Swyx [00:02:14]: Yeah, I saw that in ICLR in Vienna, there was an outstanding paper about starting transformers from data-driven piers. I don't know if you saw that one, they called it sort of never trained from scratch, and I think it was kind of rebelling against like the sort of random initialization.Jeremy [00:02:28]: Yeah, I've, you know, that's been our kind of continuous message since we started Fast AI, is if you're training for random weights, you better have a really good reason, you know, because it seems so unlikely to me that nobody has ever trained on data that has any similarity whatsoever to the general class of data you're working with, and that's the only situation in which I think starting from random weights makes sense.Swyx [00:02:51]: The other trends since our last pod that I would point people to is I'm seeing a rise in multi-phase pre-training. So Snowflake released a large model called Snowflake Arctic, where they detailed three phases of training where they had like a different mixture of like, there was like 75% web in the first instance, and then they reduced the percentage of the web text by 10% each time and increased the amount of code in each phase. And I feel like multi-phase is being called out in papers more. I feel like it's always been a thing, like changing data mix is not something new, but calling it a distinct phase is new, and I wonder if there's something that you're seeingJeremy [00:03:32]: on your end. Well, so they're getting there, right? So the point at which they're doing proper continued pre-training is the point at which that becomes a continuum rather than a phase. So the only difference with what I was describing last time is to say like, oh, there's a function or whatever, which is happening every batch. It's not a huge difference. You know, I always used to get offended when people had learning rates that like jumped. And so one of the things I started doing early on in Fast.ai was to say to people like, no, you should actually have your learning rate schedule should be a function, not a list of numbers. So now I'm trying to give the same idea about training mix.Swyx [00:04:07]: There's been pretty public work from Meta on schedule-free optimizers. I don't know if you've been following Aaron DeFazio and what he's doing, just because you mentioned learning rate schedules, you know, what if you didn't have a schedule?Jeremy [00:04:18]: I don't care very much, honestly. I don't think that schedule-free optimizer is that exciting. It's fine. We've had non-scheduled optimizers for ages, like Less Wright, who's now at Meta, who was part of the Fast.ai community there, created something called the Ranger optimizer. I actually like having more hyperparameters. You know, as soon as you say schedule-free, then like, well, now I don't get to choose. And there isn't really a mathematically correct way of, like, I actually try to schedule more parameters rather than less. So like, I like scheduling my epsilon in my atom, for example. I schedule all the things. But then the other thing we always did with the Fast.ai library was make it so you don't have to set any schedules. So Fast.ai always supported, like, you didn't even have to pass a learning rate. Like, it would always just try to have good defaults and do the right thing. But to me, I like to have more parameters I can play with if I want to, but you don't have to.Alessio [00:05:08]: And then the more less technical side, I guess, of your issue, I guess, with the market was some of the large research labs taking all this innovation kind of behind closed doors and whether or not that's good, which it isn't. And now we could maybe make it more available to people. And then a month after we release
Because of the nature of SAM, this is more video heavy than usual. See our YouTube!Because vision is first among equals in multimodality, and yet SOTA vision language models are closed, we’ve always had an interest in learning what’s next in vision. Our first viral episode was Segment Anything 1, and we have since covered LLaVA, IDEFICS, Adept, and Reka. But just like with Llama 3, FAIR holds a special place in our hearts as the New Kings of Open Source AI.The list of sequels better than the originals is usually very short, but SAM 2 delighted us by not only being a better image segmentation model than SAM 1, it also conclusively and inexpensively solved video segmentation in just an elegant a way as SAM 1 did for images, and releasing everything to the community as Apache 2/CC by 4.0.“In video segmentation, we observe better accuracy, using 3x fewer interactions than prior approaches. In image segmentation, our model is more accurate and 6x faster than the Segment Anything Model (SAM).”Surprisingly EfficientThe paper reports that SAM 2 was trained on 256 A100 GPUs for 108 hours (59% more than SAM 1). Taking the upper end $2 A100 cost off gpulist.ai means SAM2 cost ~$50k to train if it had an external market-rate cost - surprisingly cheap for adding video understanding!The newly released SA-V dataset is also the largest video segment dataset to date, with careful attention given to scene/object/geographical diversity, including that of annotators. In some ways, we are surprised that SOTA video segmentation can be done on only ~50,000 videos (and 640k masklet annotations). Model-in-the-loop Data Engine for Annotations and Demo-first DevelopmentSimilar to SAM 1, a 3 Phase Data Engine helped greatly in bootstrapping this dataset. As Nikhila says in the episode, the demo you see wasn’t just for show, they actually used this same tool to do annotations for the model that is now demoed in the tool:“With the original SAM, we put a lot of effort in building a high-quality demo. And the other piece here is that the demo is actually the annotation tool. So we actually use the demo as a way to improve our annotation tool. And so then it becomes very natural to invest in building a good demo because it speeds up your annotation. and improve the data quality, and that will improve the model quality. With this approach, we found it to be really successful.”An incredible 90% speedup in annotation happened due to this virtuous cycle which helped SA-V reach this incredible scale.Building the demo also helped the team live the context that their own downstream users, like Roboflow, would experience, and forced them to make choices accordingly.As Nikhila says:“It's a really encouraging trend for not thinking about only the new model capability, but what sort of applications folks want to build with models as a result of that downstream.I think it also really forces you to think about many things that you might postpone. For example, efficiency. For a good demo experience, making it real time is super important. No one wants to wait. And so it really forces you to think about these things much sooner and actually makes us think about what kind of image encoder we want to use or other things. hardware efficiency improvements. So those kind of things, I think, become a first-class citizen when you put the demo first.”Indeed, the team swapped out standard ViT-H Vision Transformers for Hiera (Hierarchical) Vision Transformers as a result of efficiency considerations.Memory AttentionSpeaking of architecture, the model design is probably the sleeper hit of a project filled with hits. The team adapted SAM 1 to video by adding streaming memory for real-time video processing:Specifically adding memory attention, memory encoder, and memory bank, which surprisingly ablated better than more intuitive but complex architectures like Gated Recurrent Units.One has to wonder if streaming memory can be added to pure language models with a similar approach… (pls comment if there’s an obvious one we haven’t come across yet!)Video PodcastTune in to Latent Space TV for the video demos mentioned in this video podcast!Resources referencedShow References* https://sam2.metademolab.com/demo * roboflow.com/sam2* https://github.com/autodistill/autodistill* https://github.com/facebookresearch/segment-anything-2* https://rf100.org * https://blog.roboflow.com/label-data-with-grounded-sam-2/* https://arxiv.org/abs/2408.00714 * https://github.com/roboflow/notebooks* https://x.com/skalskip92/status/1818648396002951178https://x.com/skalskip92/status/1818648396002951178* https://blog.roboflow.com/sam-2-video-segmentation/Timestamps* [00:00:00] The Rise of SAM by Udio (David Ding Edit)* [00:03:07] Introducing Nikhila* [00:06:38] The Impact of SAM 1 in 2023* [00:12:15] Do People Finetune SAM?* [00:16:05] Video Demo of SAM* [00:20:01] Why the Demo is so Important* [00:23:23] SAM 1 vs SAM 2 Architecture* [00:26:46] Video Demo of SAM on Roboflow* [00:32:44] Extending SAM 2 with other models* [00:35:00] Limitations of SAM: Screenshots* [00:38:56] SAM 2 Paper* [00:39:15] SA-V Dataset and SAM Data Engine* [00:43:15] Memory Attention to solve Video* [00:47:24] "Context Length" in Memory Attention* [00:48:17] Object Tracking* [00:50:52] The Future of FAIR* [00:52:23] CVPR, Trends in Vision* [01:02:04] Calls to ActionTranscript[00:00:00] [music intro][00:02:11] AI Charlie: Happy Yoga! This is your AI co host Charlie. Thank you for all the love for our special 1 million downloads Wins of AI Winter episode last week, especially Sam, Archie, Trellis, Morgan, Shrey, Han, and more. For this episode, we have to go all the way back to the first viral episode of the podcast Segment Anything Model and the Hard Problems of Computer Vision, which we discussed with Joseph Nelson of Roboflow.[00:02:39] AI Charlie: Since Meta released SAM 2 last week, we are delighted to welcome Joseph back as our fourth guest co host to chat with Nikhila Ravi, Research Engineering Manager at Facebook AI Research and lead author of SAM 2. Just like our SAM 1 podcast, this is a multimodal pod because of the vision element, so we definitely encourage you to hop over to our YouTube at least for the demos, if not our faces.[00:03:04] AI Charlie: Watch out and take care.[00:03:10] Introducing Nikhila[00:03:10] swyx: Welcome to the latest podcast. I'm delighted to do segment anything to our first, one of our very first viral podcasts was segment anything one with Joseph. Welcome back. Thanks so much. And this time we are joined by the lead author of Segment Anything 2, Nikki Ravi, welcome.[00:03:25] Nikhila Ravi: Thank you. Thanks for having me.[00:03:26] swyx: There's a whole story that we can refer people back to episode of the podcast way back when for the story of Segment Anything, but I think we're interested in just introducing you as a researcher, as a, on the human side what was your path into AI research? Why, you know, why did you choose computer vision coming out of your specialization at Cambridge?[00:03:46] Nikhila Ravi: So I did my undergraduate. Degree in engineering at Cambridge university. The engineering program is very general. So first couple of years, you sort of study everything from mechanical engineering to fluid mechanics, structural mechanics, material science, and also computer science.[00:04:04] Nikhila Ravi: Towards the end of my degree, I started taking more classes in machine learning and computational neuroscience, and I really enjoyed it. And actually after graduating from undergrad, I had a place at Oxford to study medicine. And so I was. Initially planning on becoming a doctor, had everything planned and then decided to take a gap year after finishing undergrad.[00:04:28] Nikhila Ravi: And actually that was around the time that sort of deep learning was emerging. And in my machine learning class in undergrad, I remember one day our professor came in and that was when Google acquired DeepMind. And so that became like a huge thing. We talked about it for the whole class. It kind of really stuck.[00:04:48] Nikhila Ravi: And I was kicked off thinking about, okay, maybe I want to try something different other than medicine. Maybe this is a different path I want to take. And then in the gap year, I did a bunch of coding, worked on a number of projects. Did some sort of freelance contracting work. And then I got a scholarship to come and study in America.[00:05:06] Nikhila Ravi: So I went to Harvard for a year, took a bunch of computer science classes at Harvard and MIT, worked on a number of AI projects, especially in computer vision. I really, really enjoyed working in computer vision. I applied to Facebook and got this job at Facebook, and I've now at Facebook at the time, now Meta, and I've been here for seven years, so very circuitous path, probably not a very unconventional, I didn't do a PhD, I'm not like a research, typical research scientist, definitely came from more of an engineering background, but since being at Meta, Have had amazing opportunities to work across so many different interesting problems in computer vision from 3D computer vision.[00:05:50] Nikhila Ravi: How can you go from images of objects to 3D structures and then going back to 2D computer vision and actually understanding the objects and the pixels and the images themselves. So it's been a very interesting journey over the past seven years.[00:06:05] swyx: It's weird because like, I guess with segment anything too, it's like 4D because you solve time, you know, you started with 3D and now you're solving the 4D.[00:06:14] Nikhila Ravi: Yeah, it's just going from 3D to images to video. It's really covering the full spectrum. And actually, one of the nice things has been, so I think I mentioned I, Wanted to become a doctor, but actually Sam is having so much impact in medicine, probably more than I could have ever had as a doctor myself. So I think, you know, hopefully Sam too can also have a similar sort o
Thank you for 1m downloads of the podcast and 2m readers of the Substack! 🎉This is the audio discussion following The Winds of AI Winter essay that also serves as a recap of Q2 2024 in AI viewed through the lens of our Four Wars framework. Enjoy!Full Video DiscussionFull show notes are here.Timestamps* [00:00:00] Intro Song by Suno.ai* [00:02:01] Swyx and Alessio in Singapore* [00:05:49] GPU Rich vs Poors: Frontier Labs* [00:06:35] GPU Rich Frontier Models: Claude 3.5* [00:10:37] GPU Rich helping Poors: Llama 3.1: The Synthetic Data Model* [00:15:41] GPU Rich helping Poors: Frontier Labs Vibe Shift - Phi 3, Gemma 2* [00:18:26] GPU Rich: Mistral Large* [00:21:56] GPU Rich: Nvidia + FlashAttention 3* [00:23:45] GPU Rich helping Poors: Noam Shazeer & Character.AI* [00:28:14] GPU Poors: On Device LLMs: Mozilla Llamafile, Chrome (Gemini Nano), Apple Intelligence* [00:35:33] Quality Data Wars: NYT vs The Atlantic lawyer up vs partner up* [00:37:41] Quality Data Wars: Reddit, ScarJo, RIAA vs Udio & Suno* [00:41:03] Quality Data Wars: Synthetic Data, Jagged Intelligence, AlphaProof* [00:45:33] Multimodality War: ChatGPT Voice Mode, OpenAI demo at AIEWF* [00:47:34] Multimodality War: Meta Llama 3 multimodality + Chameleon* [00:50:54] Multimodality War: PaliGemma + CoPaliGemma* [00:52:55] Renaming Rag/Ops War to LLM OS War* [00:55:31] LLM OS War: Ops War: Prompt Management vs Gateway vs Observability* [01:02:57] LLM OS War: BM42 Vector DB Wars, Memory Databases, GraphRAG* [01:06:15] LLM OS War: Agent Tooling* [01:08:26] LLM OS War: Agent Protocols* [01:10:43] Trend: Commoditization of Intelligence* [01:16:45] Trend: Vertical Service as Software, AI Employees, Brightwave, Dropzone* [01:20:44] Trend: Benchmark Frontiers after MMLU* [01:23:31] Crowdstrike will save us from Skynet* [01:24:30] Bonus: ChatGPT Advanced Voice Mode Demo* [01:25:37] Voice Mode: Storytelling* [01:27:55] Voice Mode: Accents* [01:31:48] Voice Mode: Accent Detection* [01:35:00] Voice Mode: Nonverbal Emotions* [01:37:53] Voice Mode: Multiple Voices in One* [01:40:52] Voice Mode: Energy Levels Detection* [01:42:03] Voice Mode: Multilinguality* [01:43:53] Voice Mode: Shepard Tone* [01:46:57] Voice Mode: Generating Tones* [01:49:39] Voice Mode: Interruptions don't work* [01:49:55] Voice Mode: Reverberations* [01:51:37] Voice Mode: Mimicry doesn't workTranscriptCharlie [00:01:08]: Welcome back, listeners. This is your AI co-host, Charlie. It's been a few months since we took a step back from the interview format and talked about the show. We're happy to share that we have crossed one million downloads and two million reads on Substack. Woo-hoo. We are really grateful to those of you who keep tuning in and sharing us with your friends, especially if who watch and comment on our new YouTube channel, where we are trying to grow next. For a special millionaire edition, SWIX and Alessio are finally back in person in sunny Singapore to discuss the big vibe shift in the last three months, that we are calling the Winds of AI Winter. We also discuss my nemesis, ChatGPT Advanced Voice Mode, with a special treat for those who stay till the end. Now, more than ever, watch out and take care.Alessio [00:02:02]: Hey, everyone. Welcome to the Latent Space Podcast. This is Alessio, partner and CTO in Residence and Decibel Partners, and today we're in the Singapore studio with SWIX.Swyx [00:02:11]: Hey, this is our long-awaited one-on-one episode. I don't know how long ago the previous one was. Do you remember? Three, four months?Alessio [00:02:20]: Yeah, it's been a while.Swyx [00:02:22]: People really enjoyed it. It's just really, I think our travel schedules have been really difficult to get this stuff together. And then we also had like a decent backlog of guests for a while. I think we've kind of depleted that backlog now and we need to build it up again. But it's been busy and there's been a lot of news. So we actually get to do this like sort of rapid fire thing. I think some people, you know, the podcast has grown a lot in the last six months. Maybe just reintroducing like what you're up to, what I'm up to, and why we're here in Singapore and stuff like that.Alessio [00:02:51]: Yeah. My first time here in Singapore, which has been really nice. This country is really amazing, I would say. First of all, everything feels like the busiest part of the city. Everything is skyscrapers. There's like plants in all the buildings, or at least in the areas that I've been in, which has been awesome. And I was at one of the offices kind of on the south side and from the 38th floor, you can see Indonesia on one side and you can see Malaysia on the other side. So it's quite, quite small. One of the people there said their kid goes to school at the border with Malaysia basically, so they could drive to Malaysia every day. So they go pick her up from school. Yeah. And we came here, we hosted with you, the Sovereign AI Summit Wednesday night. We had a lot of folks.Swyx [00:03:31]: NVIDIA, Goldman, Temasek, Singtel.Alessio [00:03:34]: And we got to talk about this trend of sovereign AI, which maybe we might cover on another episode, but basically how do you drive, if you're a country, how do you drive productivity growth in a time where populations are shrinking, the workforce is shrinking and AI can kind of supplement a lot of this. And then the question is, okay, should I put all this money in foundation models? Should I put it in data centers and infrastructure? Should I put it in GPUs? Should I put it in agents and whatnot? So we'll touch on some of these trends in the episode, but it was a fun event. And I did not expect some of the most senior people at the largest financial institution in Singapore ask about state space models and some of the alternatives. So it's great to see how advanced the conversation is sometimes.Swyx [00:04:16]: Yeah. I think that that is mostly people trying to listen to jargon that is being floated around as like, oh, what could kill transformers? And then they jump straight there without actually exploring the fundamentals, the basics of what they will actually put to work. That's fine. It's a forum to ask questions. So you want to ask about the future, but I feel like it's not very practical to spend so much time on those things. Part of the things that I do in space, especially when I travel, is to try to ask questions about what countries that are not the US and not San Francisco can do, because everyone feels a bit left out. You feel it here as well. And I'm trying to promote alternatives. I think AI engineering is one way that countries can capitalize on the industry without building a hundred billion dollar cluster, which is one-fifth the GDP of Singapore. And so my pitch at the summit was that we would sample with the AIGeneration. We're also working on bringing the AIGeneration conference to Singapore next year together with iClear. So yeah, we're just trying my best and I'm being looped into various government meetings to try to make that happen.Alessio [00:05:25]: Well, we'll definitely be here next year. I'll be back here very often. It's really nice.Swyx [00:05:31]: Yeah. Awesome. Okay. Well, we have a lot of news. How do you think we should cover?Alessio [00:05:36]: Maybe just recap since the framework of the four words of AI is something that came up end of last year. So basically, we'll link in the show notes, but the end of year recap for 2023 was basically the four words of AI, which we picked GPU-rich versus GPU-poor, the data quality wars, the multimodality wars, and the reg slash ops wars. So usually everything falls back under those four categories. So I'm pretty happy that seven months later, it's something that still matters.Swyx [00:06:07]: It still kind of holds up.Alessio [00:06:08]: Yeah. Most AI stuff from eight months ago, it's really not that relevant anymore. And today we'll try and bucket some of the recent news on it. We haven't done a monthly thing in like three months. So three months is a lot of stuff.Swyx [00:06:23]: That's mostly because I got busy with the conference. But I do want to get back on that horse or maybe just do it weekly so that I don't have such a big lift that I don't do it. I think the activation energy is the problem really. So yeah, I think frontier model wise, it seems like Cloud has really carved out a persistent space for itself. For a long time, I thought it was kind of like a clear number two to open AI. And with 3.5 on it, at least in some of the hard benchmarks on LMSys or coding benchmarks on LMSys, it is the undisputed number one model in the world, even with 4.0 mini. And we can talk about 4.0 mini and benchmarking later on. But for Cloud to be there and hold that position for what is more than a month now in AI time is a big deal. There's not much that people know publicly about what Enthopic did for Cloud's on it. But I think it's still a huge achievement. It marks the beginning of a non-open AI centric world to the point where people on Twitter have canceled ChatGPT. That's been a trend that's been going on for a while. We talked about the unbundling of ChatGPT. But now new open source projects and tooling, they're just built for Cloud. They don't even use open AI. That's a strategic threat to open AI, I think, a little bit. Obviously, open AI is so big that it doesn't really care about that. But for Enthopic, it's a big win. I think to see that going and to see Enthopic differentiating itself and actually implementing research. So the rumor is that the scaling monosematicity paper that they put out two months ago was a big part of Cloud 3.5's on it. I've had off-the-record chats with people about that idea, and they don't agree that it is the only cause. So I was thinking this is the only thing that they did. But people say that there's about four or five other tricks that they haven't disclosed yet that went into 3.5's on it. But the scaling monosematicity paper is a very, very g
If you see this in time, join our emergency LLM paper club on the Llama 3 paper!For everyone else, join our special AI in Action club on the Latent Space Discord for a special feature with the Cursor cofounders on Composer, their newest coding agent!Today, Meta is officially releasing the largest and most capable open model to date, Llama3-405B, a dense transformer trained on 15T tokens that beats GPT-4 on all major benchmarks:The 8B and 70B models from the April Llama 3 release have also received serious spec bumps, warranting the new label of Llama 3.1.If you are curious about the infra / hardware side, go check out our episode with Soumith Chintala, one of the AI infra leads at Meta. Today we have Thomas Scialom, who led Llama2 and now Llama3 post-training, so we spent most of our time on pre-training (synthetic data, data pipelines, scaling laws, etc) and post-training (RLHF vs instruction tuning, evals, tool calling).Synthetic data is all you needLlama3 was trained on 15T tokens, 7x more than Llama2 and with 4 times as much code and 30 different languages represented. But as Thomas beautifully put it:“My intuition is that the web is full of s**t in terms of text, and training on those tokens is a waste of compute.” “Llama 3 post-training doesn't have any human written answers there basically… It's just leveraging pure synthetic data from Llama 2.”While it is well speculated that the 8B and 70B were "offline distillations" of the 405B, there are a good deal more synthetic data elements to Llama 3.1 than the expected. The paper explicitly calls out:* SFT for Code: 3 approaches for synthetic data for the 405B bootstrapping itself with code execution feedback, programming language translation, and docs backtranslation.* SFT for Math: The Llama 3 paper credits the Let’s Verify Step By Step authors, who we interviewed at ICLR:* SFT for Multilinguality: "To collect higher quality human annotations in non-English languages, we train a multilingual expert by branching off the pre-training run and continuing to pre-train on a data mix that consists of 90% multilingualtokens."* SFT for Long Context: "It is largely impractical to get humans to annotate such examples due to the tedious and time-consuming nature of reading lengthy contexts, so we predominantly rely on synthetic data to fill this gap. We use earlier versions of Llama 3 to generate synthetic data based on the key long-context use-cases: (possibly multi-turn) question-answering, summarization for long documents, and reasoning over code repositories, and describe them in greater detail below"* SFT for Tool Use: trained for Brave Search, Wolfram Alpha, and a Python Interpreter (a special new ipython role) for single, nested, parallel, and multiturn function calling.* RLHF: DPO preference data was used extensively on Llama 2 generations. This is something we partially covered in RLHF 201: humans are often better at judging between two options (i.e. which of two poems they prefer) than creating one (writing one from scratch). Similarly, models might not be great at creating text but they can be good at classifying their quality.Last but not least, Llama 3.1 received a license update explicitly allowing its use for synthetic data generation.Llama2 was also used as a classifier for all pre-training data that went into the model. It both labelled it by quality so that bad tokens were removed, but also used type (i.e. science, law, politics) to achieve a balanced data mix. Tokenizer size mattersThe tokens vocab of a model is the collection of all tokens that the model uses. Llama2 had a 34,000 tokens vocab, GPT-4 has 100,000, and 4o went up to 200,000. Llama3 went up 4x to 128,000 tokens. You can find the GPT-4 vocab list on Github.This is something that people gloss over, but there are many reason why a large vocab matters:* More tokens allow it to represent more concepts, and then be better at understanding the nuances.* The larger the tokenizer, the less tokens you need for the same amount of text, extending the perceived context size. In Llama3’s case, that’s ~30% more text due to the tokenizer upgrade. * With the same amount of compute you can train more knowledge into the model as you need fewer steps.The smaller the model, the larger the impact that the tokenizer size will have on it. You can listen at 55:24 for a deeper explanation.Dense models = 1 Expert MoEsMany people on X asked “why not MoE?”, and Thomas’ answer was pretty clever: dense models are just MoEs with 1 expert :)[00:28:06]: I heard that question a lot, different aspects there. Why not MoE in the future? The other thing is, I think a dense model is just one specific variation of the model for an hyperparameter for an MOE with basically one expert. So it's just an hyperparameter we haven't optimized a lot yet, but we have some stuff ongoing and that's an hyperparameter we'll explore in the future.Basically… wait and see!Llama4Meta already started training Llama4 in June, and it sounds like one of the big focuses will be around agents. Thomas was one of the authors behind GAIA (listen to our interview with Thomas in our ICLR recap) and has been working on agent tooling for a while with things like Toolformer. Current models have “a gap of intelligence” when it comes to agentic workflows, as they are unable to plan without the user relying on prompting techniques and loops like ReAct, Chain of Thought, or frameworks like Autogen and Crew. That may be fixed soon? 👀The whole podcast was a lot of fun to record, as usual you can find show notes and chapters below. Make sure to also subscribe on YouTube! 🙏Full Video PodcastShow Notes* Thomas Scialom* Recital* Galactica* Lucas Beyer - Citation Generator* Llama 2 paper* Guillaume Lample* Hugo Touvron* April 2023 Llama 3 release* Llama3 Repo* Chinchilla trap* Agents research* Thomas’ paper: Augmented Language Models: A Survey* GAIA: Gaia General Assistant Benchmark (we interviewed Thomas at ICLR on this)* Toolformer paper* JEPA* Clementine Fourrier episode* Nathan Lambert episode* Noam Shazeer* Optimizing AI Inference at Character.AI aka Shazeer et al 2024 - we misspoke and said “native FP8” when we meant INT8* The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits* Mentioned Papers* MobileLLM* SmolLM* Overleaf* AlphaGo* Lindy AITimestamps* Song credit: Code of the Future via Udio* [00:00:13] Introducing Thomas* [00:03:18] BLOOM and Meta Galactica* [00:06:33] Leading Llama 2* [00:09:56] Going 100x Chinchilla Scaling Laws* [00:12:15] Open Sourcing Llama 3 405B* [00:14:29] Quantization with INT8 / FP8 / Ternary (1.58 Bits)* [00:16:58] MobileLLM, SmolLM, On Device Models* [00:17:36] Llama 3 Architecture* [00:18:33] Llama 3 Tokenizer: 128k and beyond* [00:23:12] Synthetic Data for Pretraining* [00:25:08] Synthetic Data from Augmented Language Models* [00:27:19] Data Mix and Continual Pretraining* [00:29:16] Adding Code, Reasoning, Multilinguality to Llama 3* [00:30:39] Nvidia Nemotron and dedicated SynData Models* [00:31:30] Why no MOE?* [00:32:23] RLHF: Humans as Discriminators > Annotators* [00:38:37] Teacher Forcing/Critique* [00:42:02] Llama 3 Benchmarking* [00:45:24] Llama 3 Arena ELO* [00:47:27] Calibration Evals* [00:49:23] Function Calling* [00:50:17] Llama 4's plan for Agents* [00:55:09] The State of Variable/Long Inference Research* [00:57:19] Llama 4 Focus* [00:59:15] AI Startups* [01:03:34] Call to Action - HiringTranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO-in-Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI.Swyx [00:00:13]: Hey, and today we have a very special episode with Thomas Scialom. I don't know how to describe, you've done so much work in a very short amount of time at Meta, but you were most notably leading Llama 2 and now today we're also coordinating on the release of Llama 3. So welcome.Thomas [00:00:28]: Thanks for having me.Swyx [00:00:29]: So let's play obviously the Llama 3 405B. Is that the official size number that we're going with, or do we just say 400B?Thomas [00:00:37]: For the text model only, yes. A bit of additional parameters for the multi-model version that will come later.Swyx [00:00:44]: Awesome. Just to quickly go over your background, actually we had a slightly similar past. I was also a quantitative trader and it looks like you did five years in QuantFinance, working a trading timer in SockGen, and then you transitioned into natural language, getting a PhD at Sorbonne. Working on Recital as well. And then right after your PhD, joining Meta.Thomas [00:01:04]: No, it's exactly that, but basically I think it's at the AlphaGo moment where I was doing some trading. I say like, what I need to understand, what's the technology behind that? And I wanted to study machine learning. I did first some training, like six months degree, executive degree, at the end of which I knew like what XGBoost at the time, and nothing about deep learning at all. And most of the people around were like PhD people, and I was like, okay, PhD seems pretty cool, deep learning seems pretty cool, so I want to do a PhD in deep learning. That's where I joined, we have this PhD program in France within a company and academia. And so I did my PhD with Recital and Sorbonne University on natural language generation reinforcement learning. I guess it was a good topic. I was not like a visionary. It was very random. I've had a company that offered me this topic, and it was something like I started two weeks before BERT. Excellent timing.Swyx [00:02:03]: Yeah. We actually also just released our episode with Clementine Fouquier, who also did her PhD with a company in kind of like a very similar format. I think, yeah, very underrated, very underrated, this sort of PhD with industry expertise, because you're also publishing papers the whole time. I looked at your publishing history, you were doing summarization work, you're doing factual cons
The first AI Engineer World’s Fair talks from OpenAI and Cognition are up!In our Benchmarks 101 episode back in April 2023 we covered the history of AI benchmarks, their shortcomings, and our hopes for better ones. Fast forward 1.5 years, the pace of model development has far exceeded the speed at which benchmarks are updated. Frontier labs are still using MMLU and HumanEval for model marketing, even though most models are reaching their natural plateau at a ~90% success rate (any higher and they’re probably just memorizing/overfitting).From Benchmarks to LeaderboardsOutside of being stale, lab-reported benchmarks also suffer from non-reproducibility. The models served through the API also change over time, so at different points in time it might return different scores.Today’s guest, Clémentine Fourrier, is the lead maintainer of HuggingFace’s OpenLLM Leaderboard. Their goal is standardizing how models are evaluated by curating a set of high quality benchmarks, and then publishing the results in a reproducible way with tools like EleutherAI’s Harness.The leaderboard was first launched summer 2023 and quickly became the de facto standard for open source LLM performance. To give you a sense for the scale:* Over 2 million unique visitors* 300,000 active community members* Over 7,500 models evaluatedLast week they announced the second version of the leaderboard. Why? Because models were getting too good!The new version of the leaderboard is based on 6 benchmarks:* 📚 MMLU-Pro (Massive Multitask Language Understanding - Pro version, paper)* 📚 GPQA (Google-Proof Q&A Benchmark, paper)* 💭MuSR (Multistep Soft Reasoning, paper)* 🧮 MATH (Mathematics Aptitude Test of Heuristics, Level 5 subset, paper)* 🤝 IFEval (Instruction Following Evaluation, paper)* 🧮 🤝 BBH (Big Bench Hard, paper)You can read the reasoning behind each of them on their announcement blog post. These updates had some clear winners and losers, with models jumping up or down up to 50 spots at once; the most likely reason for this is that the models were overfit to the benchmarks, or had some contamination in their training dataset.But the most important change is in the absolute scores. All models score much lower on v2 than they do on v1, which now creates a lot more room for models to show improved performance.On ArenasAnother high-signal platform for AI Engineers is the LMSys Arena, which asks users to rank the output of two different models on the same prompt, and then give them an ELO score based on the outcomes.Clémentine called arenas “sociological experiments”: it tells you a lot about the users preference, but not always much about the model capabilities. She pointed to Anthropic’s sycophancy paper as early research in this space:We find that when a response matches a user’s views, it is more likely to be preferred. Moreover, both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.The other issue is that Arena rankings aren’t reproducible, as you don’t know who ranked what and what exactly the outcome was at the time of ranking. They are still quite helpful as tools, but they aren’t a rigorous way to rank capabilities of the models.Her advice for both arena and leaderboard is to use these tools as ranges; find 3-4 models that fit your needs (speed, cost, capabilities, etc) and then do vibe checks to figure out which one is best for your specific task.LLMs aren’t good judgesIn the last ~6 months, there has been an increased interest in using LLMs as Judges: rather than asking a person to evaluate the outcome of a model, you can ask a more powerful LLM to score it. We covered this a bit in our Brightwave episode last month as well. HuggingFace also has a cookbook on it, but Clémentine was actually not a fan of this approach:* Mode collapse: if you are asking a model to choose which output is better, it will just self-reinforce its own preferences. It will also prefer models from its own family (i.e. GPT models will prefer other GPT models over Claude outputs). If these outputs are then used to fine-tune the model, you will further mode collapse the model. Cohere for example has said they do not train on any model-generated data to avoid this.* Positional bias: LLMs usually prefer the first answer, so you can’t naively give them options and ask them to rank them, but you also have to mix up the order in which they appear.* Don’t score, rank: rather than asking a model to assign a score to each output, you should have it stack-rank them. The models aren’t trained to score things, so even though they might understand what response is better, assigning a score to it is hard.If you do have to use LLMs as Judges (we aren’t all ScaleAI-rich!), she suggested using an open LLM like Prometheus or JudgeLM to make sure you can reproduce those rankings in the future. Show Notes* Clémentine Fourrier* Hugging Face* OpenLLM v2 Leaderboard* Let’s talk about LLM Evaluation* Leaderboard V2 Blog Post* Latent Space Benchmarks 101* Gradient AI epsiode on Long Context Evals* Allen AI long context novel evalsCompanies and Organizations* Anthropic* Cohere* EleutherAI* INRIA* ICLR (International Conference on Learning Representations)People* Aidan Gomez* Dan Hendrycks* Edward Beeching* Haley Sholkoff* Lewis Tunstall* Nathan Habib* Thomas ScialomProjects, Models, and Benchmarks* LMSys Arena* ARC AGI Challenge* Allen Institute ARC Challenge* BigBench* GAIA benchmark* GPQA* GSM 8K* IFEval* LightEval* ML perf* MMLU* JudgeLM* Prometheus* RavenWolf* SWE-Bench* VantageTimestamps* [00:00:00] Introductions* [00:02:32] How Clémentine went from geology to AI* [00:05:52] Origin of the OpenLLM Leaderboard* [00:09:06] How v1 Benchmarks Were Selected* [00:10:49] The Problem with Current Benchmarks* [00:13:45] Saturating benchmarks and the future of evaluation* [00:16:14] Issues with human evaluations* [00:24:07] AI girlfriends as the multi-turn benchmark* [00:25:35] What's New in OpenLLM leaderboard V2* [00:28:12] Benchmark Answers Black Market* [00:30:21] The impact of prompt formatting on model evaluation scores* [00:33:30] Difficulty and Computational Constraints of Evals* [00:36:28] The Responsibility of Setting Standards* [00:40:35] The Economics of OpenLLM* [00:44:15] Long context reasoning benchmarks* [00:46:34] Agent benchmarks, GAIA, and the ARC AGI challenge* [00:50:43] Vibe check for benchmarks* [00:53:16] Request for benchmarks* [00:56:48] v3 predictions?TranscriptAlessio [00:00:00]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO-in-Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol AI.Swyx [00:00:13]: Hey, and today we have a super special guest that we've been trying to book on the schedule for a while. It's Clémentine Fourrier. I'm trying my best to do the French, but maybe you can do a better job of it than me.Clémentine [00:00:26]: This was perfect. It's Clémentine Fourrier, but your pronunciation was really on point.Swyx [00:00:31]: There was a Fourrier, which is very sort of French intonation, which I don't really understand. So I'll introduce you off of your LinkedIn and I would love for you to fill in the blanks. You are currently a research scientist at Hugging Face and the maintainer of the OpenLLM leaderboard, which we'll talk about very shortly. Previously, you were at INRIA as well, but then it looks like you also concurrently got your PhD at the same time. How does that work? Is that a very common thing?Clémentine [00:01:01]: So I basically did my PhD at INRIA, technically. So INRIA funded my PhD and PhDs in France are three years, but I also worked as an engineer at INRIA before my PhD, hence maybe the confusion.Swyx [00:01:14]: I think there's a rise in universities having sort of industrial attachments to these things. And I think it actually makes for a much more grounded study, especially if you're doing your sort of graduate studies and all these things. I think it's rising in North America as well with Berkeley and with Waterloo in Toronto. Cool. Like, you know, there's, there's a lot of other things we can, we can introduce. I can't really pronounce the name of the, the university you went to, but what else should people know?Clémentine [00:01:44]: So I actually, technically I'm an engineer in geology. So I studied rocks and I graduated in 2015 after having done like extensive studies about rocks. And I discovered I was very bad at it, but I was very good at computer science. So I went to computer science. What stuck with me though is that geology is very much an experimental science. And I think that machine learning is very much an experimental science too, even though people want to claim that it's pure math. And I worked on several machine learning projects throughout the years, a bit of the prediction of illnesses in the brain at Brain and Spine Institute in Paris. I worked as an engineer in a research team in NLP where I did my thesis and then I joined Hugging Face.Swyx [00:02:32]: Do you have a favorite rock fact or sort of rock story before we get into the NLP stuff?Clémentine [00:02:38]: Okay. I was not expecting this question.Swyx [00:02:43]: I did my geography A-levels and I always loved learning about like isostasy and stuff like that where you have different plates kind of up and down in the mantle. And I don't think people think about vertical dimensions to geographical plates, but it's real.Clémentine [00:03:02]: Yeah, definitely. And like when you do geology, the time scale is just not the same. There is like one specific place in France where you can see rocks that are 1 billion years old and like the sheer scale of this is huge. Yeah, that's what I loved about geology, that the scale is completely different and it makes us see the rest of the world in perspective, I guess. We are like a blink in the length of time of the earth.Swyx [00:03:31]: But a very significant blink. So you w
Livestreams for the AI Engineer World’s Fair (Multimodality ft. the new GPT-4o demo, GPUs and Inference (ft. Cognition/Devin), CodeGen, Open Models tracks) are now live! Subscribe to @aidotEngineer to get notifications of the other workshops and tracks!It’s easy to get de-sensitized to new models topping leaderboards every other week — however, the top of the LMsys leaderboard has typically been the exclusive domain of very large, very very well funded model labs like OpenAI, Anthropic, Google, and Meta. OpenAI had about 600 people at the time of GPT-4, and Google Gemini had 950 co-authors. This is why Reka Core made waves in May - not only debuting at #7 on the leaderboard, but doing so with all-new GPU infrastructure and 20 employees with <5 people on pre-training and a relatively puny $60m in funding.Shortly after the release of GPT3, Sam Altman speculated on the qualities of “10,000x researchers”:* “They spend a lot of time reflecting on some version of the Hamming question—"what are the most important problems in your field, and why aren’t you working on them?” In general, no one reflects on this question enough, but the best people do it the most, and have the best ‘problem taste’, which is some combination of learning to think independently, reason about the future, and identify attack vectors.” — sama* Taste is something both John Schulman and Yi Tay emphasize greatly* “They have a laser focus on the next step in front of them combined with long-term vision.” — sama* “They are extremely persistent and willing to work hard… They have a bias towards action and trying things, and they’re clear-eyed and honest about what is working and what isn’t” — sama“There's a certain level of sacrifice to be an AI researcher, especially if you're training at LLMs, because you cannot really be detached… your jobs could die on a Saturday at 4am, and there are people who will just leave it dead until Monday morning, or there will be people who will crawl out of bed at 4am to restart the job, or check the TensorBoard” – Yi Tay (at 28 mins)“I think the productivity hack that I have is, I didn't have a boundary between my life and my work for a long time. So I think I just cared a lot about working most of the time. Actually, during my PhD, Google and everything [else], I'll be just working all the time. It's not like the most healthy thing, like ever, but I think that that was actually like one of the biggest, like, productivity, like and I spent, like, I like to spend a lot of time, like, writing code and I just enjoy running experiments, writing code” — Yi Tay (at 90 mins)* See @YiTayML example for honest alpha on what is/is not workingand so on.More recently, Yi’s frequent co-author, Jason Wei, wrote about the existence of Yolo researchers he witnessed at OpenAI:Given the very aggressive timeline — Yi left Google in April 2023, was GPU constrained until December 2023, and then Reka Flash (21B) was released in Feb 2024, and Reka Core (??B) was released in April 2024 — Reka’s 3-5 person pretraining team had no other choice but to do Yolo runs. Per Yi:“Scaling models systematically generally requires one to go from small to large in a principled way, i.e., run experiments in multiple phrases (1B->8B->64B->300B etc) and pick the winners and continuously scale them up. In a startup, we had way less compute to perform these massive sweeps to check hparams. In the end, we had to work with many Yolo runs (that fortunately turned out well).In the end it took us only a very small number of smaller scale & shorter ablation runs to get to the strong 21B Reka Flash and 7B edge model (and also our upcoming largest core model). Finding a solid recipe with a very limited number of runs is challenging and requires changing many variables at once given the ridiculously enormous search space. In order to do this, one has to abandon the systematicity of Bigtech and rely a lot on “Yolo”, gut feeling and instinct.”We were excited to be the first podcast to interview Yi, and recommend reading our extensive show notes to follow the same papers we reference throughout the conversation.Special thanks to Terence Lee of TechInAsia for the final interview clip, who are launching their own AI newsletter called The Prompt!Full Video PodcastShow Notes* Yi on LinkedIn, Twitter, Personal* Full prep doc* Reka funding/valuation* Building frontier AI teams as GPU Poors* Yi’s Research* 2020* Efficient Transformers: A Survey went viral!* Long Range Arena: A Benchmark for Efficient Transformers in 2020* 2021: Generative Models are Unsupervised Predictors of Page Quality: A Colossal-Scale Study * 2022: * UL2: Unifying Language Learning Paradigms* PaLM -> PaLM-2* Emergent Abilities of Large Language Models vs the Mirage paper* Recitation Augmented generation* DSI++: Updating Transformer Memory with New Documents* The Efficiency Misnomer: “a model with low FLOPs may not actually be fast, given that FLOPs does not take into account information such as degree of parallelism (e.g., depth, recurrence) or hardware-related details like the cost of a memory access”* 2023: Flan-{PaLM/UL2/T5}1.8k tasks for instruction tuning* Encoder-decoder vs Decoder only* Latent Space Discord discussion on enc-dec vs dec-only* Related convo with Yi Tay vs Yann LeCun* @teortaxes: * If 2024 papers are to be trusted: You don't need (most) attention you don't need (most) kv cache You don't need (most) FFN layers You don't need a reward model You don't need… all the stuff that still makes frontier models work, ironically* “there have been no real advance since 2019's T5 models”* The future of Open source models - relevant to a16z vs Founders Fund debate. Open source cannot compete!Timestamps* [00:00:00] Intro* [00:01:57] Yi Tay Intro* [00:03:02] Path into LLMs* [00:09:41] Google Brain: PaLM, UL2, DSI, Emergent Abilities* [00:11:54] PaLM 2* [00:15:27] Emergent Abilities* [00:18:26] Quoc Le* [00:24:16] Marketing Research: How to Start from Zero with No Reach* [00:27:34] What's needed to be a successful AI Researcher?* [00:30:31] Reka Origin* [00:33:24] Starting Reka Infra* [00:35:04] Why not to use TPUs outside Google* [00:36:29] Chaotic vs Stable Infra* [00:38:04] Risk Sharing of Bad Nodes* [00:41:05] Checkpointing and Orchestration* [00:43:39] Reka Flash/Core/Edge* [00:46:59] Recruiting the team* [00:47:22] Noam Architecture - Swiglu, GQA, RMSnorm, ROPE* [00:52:26] Encoder-decoder vs Decoder-only* [00:55:52] LLM Trends - Llama 3 and Phi 3 Glowup* [00:57:46] LLM Trends - Benchmarks and Evals* [01:03:25] LLM Trends - Early vs Late Fusion Multimodality* [01:07:22] LLM Trends - Scaling Laws* [01:09:41] LLM Trends - Long Context vs RAG* [01:12:31] Long Context vs Finetuning* [01:14:14] If emergence is real, when does Efficiency work?* [01:17:41] MoEs and Upcycling* [01:20:47] The Efficiency Misnomer - Efficiency != Speed* [01:25:05] Open Source vs Closed Models* [01:28:08] Personal Productivity* [01:33:19] Singapore vs US Academic Scene* [01:37:42] Building Silicon Valley outside Silicon Valley* [01:40:29] TechInAsia Meetup Transcript[00:00:00] swyx: Thanks for watching. Bye bye.[00:00:05] AI Charlie: Welcome back, friends. It's only been a week since the World's Fair, and it was incredible gathering the community to see the latest and greatest in AI engineering. You can catch up now on the four live stream track days on the AI Engineer YouTube, and our team is busy editing the remaining workshops and five other tracks, including the surprisingly popular AI Leadership track.[00:00:28] Thank you all for your support, and stay tuned for news about the next event. The 2025 AI Engineer Summit. Last week, we did a very special deep dive with Josh and John of InView and Databricks Mosaic on training LLMs and setting up massive GPU clusters. And today, we're pleased to follow that up with a very special conversation with Yi Tai, formerly tech lead of Palm 2 at Google Brain, and now chief scientist of Reka.[00:00:56] ai. Raker's largest model, Raker Core, was at launch. The fifth best model in the world. And the only GPT 4 class model not trained by a big lab like OpenAI, Google, Anthropic or Meta. In fact, while Google Gemini has 950 co authors, Raker only has 20 employees. With up to five people actually working on pre training.[00:01:21] Swyx was excited to return to Singapore to delve into Yi Reka and building a new AI model lab outside of Silicon Valley. Stay tuned to the very end for a special bonus clip from Yi's recent appearance at the Tekinesia meetup for his spiciest take on why senior management is overrated and why this is the time to build up senior 10, 000x individual contributors like himself.[00:01:46] Watch out and take care.[00:01:48] swyx: Welcome to lay space. This is a long time coming, but I'm so excited to have you here.[00:01:52] Yi Tay: Yeah, thanks for, thanks for inviting and excited to be here. chat about a lot of stuff.[00:01:57] Yi Tay Intro[00:01:57] swyx: Yeah. So you are interesting to research and introduce. You are now chief scientist of Rega, which is a super interesting model lab, but before that you were at Google Brain, you were architecture co-lead on POM two, you were inventor of UL two.[00:02:10] You're a core contributor on Flan, you're a member of the Bard core team, and you also did some work on generative retrieval. That's a very, very illustrious three year career at Google Brain.[00:02:19] Yi Tay: Yeah, thanks, thanks, thanks, yeah.[00:02:20] swyx: And then since then, Reka, you joined in March 2023, announced a 58 million series in June 2023.[00:02:26] I don't know if you know, the post money valuation, or the pre money valuation is public. So it's, crunch basis is, is, Oh, okay, okay. I[00:02:33] Yi Tay: did not know that yet. 50[00:02:34] swyx: something million. So you don't even have to leak it. It's on the internet. Okay. Rekha's stated goals were to work on universal intelligence, including general purpose multimodal and multilingual agents, self
loading