Into AI Safety

The almost Dr. Igor Krawczuk joins me for what is the equivalent of 4 of my previous episodes. We get into all the classics: eugenics, capitalism, philosophical toads... Need I say more?If you're interested in connecting with Igor, head on over to his website, or check out placeholder for thesis (it isn't published yet).Because the full show notes have a whopping 115 additional links, I'll highlight some that I think are particularly worthwhile here:The best article you'll ever read on Open Source AIThe best article you'll ever read on emergence in MLKate Crawford's Atlas of AI (Wikipedia)On the Measure of IntelligenceThomas Piketty's Capital in the Twenty-First Century (Wikipedia)Yurii Nesterov's Introductory Lectures on Convex OptimizationChapters(02:32) - Introducing Igor (10:11) - Aside on EY, LW, EA, etc., a.k.a. lettersoup (18:30) - Igor on AI alignment (33:06) - "Open Source" in AI (41:20) - The story of infinite riches and suffering (59:11) - On AI threat models (01:09:25) - Representation in AI (01:15:00) - Hazard fishing (01:18:52) - Intelligence and eugenics (01:34:38) - Emergence (01:48:19) - Considering externalities (01:53:33) - The shape of an argument (02:01:39) - More eugenics (02:06:09) - I'm convinced, what now? (02:18:03) - AIxBio (round ??) (02:29:09) - On open release of models (02:40:28) - Data and copyright (02:44:09) - Scientific accessibility and bullshit (02:53:04) - Igor's point of view (02:57:20) - Outro LinksLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance. All references, including those only mentioned in the extended version of this episode, are included.Suspicious Machines Methodology, referred to as the "Rotterdam Lighthouse Report" in the episodeLIONS Lab at EPFLThe meme that Igor referencesOn the Hardness of Learning Under SymmetriesCourse on the concept of equivariant deep learningAside on EY/EA/etc.Sources on Eliezer YudkowskiScholarly Community EncyclopediaTIME100 AIYudkowski's personal websiteEY WikipediaA Very Literary Wiki -TIME article: Pausing AI Developments Isn’t Enough. We Need to Shut it All Down documenting EY's ruminations of bombing datacenters; this comes up later in the episode but is included here because it about EY.LessWrongLW WikipediaMIRICoverage on Nick Bostrom (being a racist)The Guardian article: ‘Eugenics on steroids’: the toxic and contested legacy of Oxford’s Future of Humanity InstituteThe Guardian article: Oxford shuts down institute run by Elon Musk-backed philosopherInvestigative piece on Émile TorresOn the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜NY Times article: We Teach A.I. Systems Everything, Including Our BiasesNY Times article: Google Researcher Says She Was Fired Over Paper Highlighting Bias in A.I.Timnit Gebru's WikipediaThe TESCREAL Bundle: Eugenics and the Promise of Utopia through Artificial General IntelligenceSources on the environmental impact of LLMsThe Environmental Impact of LLMsThe Cost of Inference: Running the ModelsEnergy and Policy Considerations for Deep Learning in NLPThe Carbon Impact of AI vs Search EnginesFilling Gaps in Trustworthy Development of AI (Igor is an author on this one)A Computational Turn in Policy Process Studies: Coevolving Network Dynamics of Policy ChangeThe Smoothed Possibility of Social Choice, an intro in social choice theory and how it overlaps with MLRelating to Dan HendrycksNatural Selection Favors AIs over Humans"One easy-to-digest source to highlight what he gets wrong [is] Social and Biopolitical Dimensions of Evolutionary Thinking" -IgorIntroduction to AI Safety, Ethics, and Society, recently published textbook"Source to the section [of this paper] that makes Dan one of my favs from that crowd." -IgorTwitter post referenced in the episode<...

As always, the best things come in 3s: dimensions, musketeers, pyramids, and... 3 installments of my interview with Dr. Peter Park, an AI Existential Safety Post-doctoral Fellow working with Dr. Max Tegmark at MIT.As you may have ascertained from the previous two segments of the interview, Dr. Park cofounded StakeOut.AI along with Harry Luk and one other cofounder whose name has been removed due to requirements of her current position. The non-profit had a simple but important mission: make the adoption of AI technology go well, for humanity, but unfortunately, StakeOut.AI had to dissolve in late February of 2024 because no granter would fund them. Although it certainly is disappointing that the organization is no longer functioning, all three cofounders continue to contribute positively towards improving our world in their current roles.If you would like to investigate further into Dr. Park's work, view his website, Google Scholar, or follow him on Twitter00:00:54 ❙ Intro00:02:41 ❙ Rapid development00:08:25 ❙ Provable safety, safety factors, & CSAM00:18:50 ❙ Litigation00:23:06 ❙ Open/Closed Source00:38:52 ❙ AIxBio00:47:50 ❙ Scientific rigor in AI00:56:22 ❙ AI deception01:02:45 ❙ No takesies-backsies01:08:22 ❙ StakeOut.AI's start01:12:53 ❙ Sustainability & Agency01:18:21 ❙ "I'm sold, next steps?" -you01:23:53 ❙ Lessons from the amazing Spiderman01:33:15 ❙ "I'm ready to switch careers, next steps?" -you01:40:00 ❙ The most important question01:41:11 ❙ OutroLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.StakeOut.AIPause AIAI Governance Scorecard (go to Pg. 3)CIVITAIArticle on CIVITAI and CSAMSenate Hearing: Protecting Children OnlinePBS Newshour CoverageThe Times Sues OpenAI and Microsoft Over A.I. Use of Copyrighted WorkOpen Source/Weights/Release/InterpretationOpen Source InitiativeHistory of the OSIMeta’s LLaMa 2 license is not Open SourceIs Llama 2 open source? No – and perhaps we need a new definition of open…Apache License, Version 2.03Blue1Brown: Neural NetworksOpening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generatorsThe online tableSignalBloomz model on HuggingFaceMistral websiteNASA TragediesChallenger disaster on WikipediaColumbia disaster on WikipediaAIxBio RiskDual use of artificial-intelligence-powered drug discoveryCan large language models democratize access to dual-use biotechnology?Open-Sourcing Highly Capable Foundation Models (sadly, I can't rename the article...)Propaganda or Science: Open Source AI and Bioterrorism RiskExaggerating the risks (Part 15: Biorisk from LLMs)Will releasing the weights of future large language models grant widespread access to pandemic agents?On the Societal Impact of Open Foundation ModelsPolicy briefApart ResearchScienceCiceroHuman-level play in the game of Diplomacy by combining language models with strategic reasoningCicero webpageAI Deception: A Survey of Examples, Risks, and Potential SolutionsOpen Sourcing the AI Revolution: Framing the debate on open source, artificial intelligence and regulationAI Safety CampInto AI Safety Patreon

Join me for round 2 with Dr. Peter Park, an AI Existential Safety Postdoctoral Fellow working with Dr. Max Tegmark at MIT. Dr. Park was a cofounder of StakeOut.AI, a non-profit focused on making AI go well for humans, along with Harry Luk and one other individual, whose name has been removed due to requirements of her current position.In addition to the normal links, I wanted to include the links to the petitions that Dr. Park mentions during the podcast. Note that the nonprofit which began these petitions, StakeOut.AI, has been dissolved.Right AI Laws, to Right Our Future: Support Artificial Intelligence Safety Regulations NowIs Deepfake Illegal? Not Yet! Ban Deepfakes to Protect Your Family & Demand Deepfake LawsBan Superintelligence: Stop AI-Driven Human Extinction Risk 00:00:54 - Intro00:02:34 - Battleground 1: Copyright00:06:28 - Battleground 2: Moral Critique of AI Collaborationists00:08:15 - Rich Sutton00:20:41 - OpenAI Drama00:34:28 - Battleground 3: Contract Negotiations for AI Ban Clauses00:37:57 - Tesla, Autopilot, and FSD00:40:02 - Recycling00:47:40 - Battleground 4: New Laws and Policies00:50:00 - Battleground 5: Whistleblower Protections00:53:07 - Whistleblowing on Microsoft00:54:43 - Andrej Karpathy & Exercises in Empathy01:05:57 - OutroLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.StakeOut.AIThe Times Sues OpenAI and Microsoft Over A.I. Use of Copyrighted WorkSusman Godfrey LLPRich SuttonReinforcement Learning: An Introduction (textbook)AI Succession (presentation by Rich Sutton)The Alberta Plan for AI ResearchMoore's LawThe Future of Integrated Electronics (original paper)Computer History Museum's entry on Moore's LawStochastic gradient descent (SGD) on WikipediaOpenAI DramaMax Read's Substack postZvi Mowshowitz's Substack series, in order of postingOpenAI: Facts from a WeekendOpenAI: The Battle of the BoardOpenAI: Altman ReturnsOpenAI: Leaks Confirm the Story ← best singular post in the seriesOpenAI: The Board ExpandsOfficial OpenAI announcementWGA on WikipediaSAG-AFTRA on WikipediaTesla's False AdvertisingTesla's response to the DMV's false-advertising allegations: What took so long?Tesla Tells California DMV that FSD Is Not Capable of Autonomous DrivingWhat to Call Full Self-Driving When It Isn't Full Self-Driving?Tesla fired an employee after he posted driverless tech reviews on YouTubeTesla's page on Autopilot and Full Self-DrivingRecyclingBoulder County Recycling Center Stockpiles Accurately Sorted Recyclable MaterialsOut of sight, out of mindBoulder Eco-Cycle Recycling GuidelinesDivide-and-Conquer Dynamics in AI-Driven DisempowermentMicrosoft WhistleblowerWhistleblowers call out AI's flawsShane's LinkedIn postLetters sent by JonesKarpathy announces departure from OpenAI

UPDATE: Contrary to what I say in this episode, I won't be removing any episodes that are already published from the podcast RSS feed.After getting some advice and reflecting more on my own personal goals, I have decided to shift the direction of the podcast towards accessible content regarding "AI" instead of the show's original focus. I will still be releasing what I am calling research ride-along content to my Patreon, but the show's feed will consist only of content that I aim to make as accessible as possible.00:35 - TL;DL01:12 - Advice from Pete03:10 - My personal goal05:39 - Reflection on refining my goal09:08 - Looking forward (logistics

Dr. Peter Park is an AI Existential Safety Postdoctoral Fellow working with Dr. Max Tegmark at MIT. In conjunction with Harry Luk and one other cofounder, he founded ⁠StakeOut.AI, a non-profit focused on making AI go well for humans.00:54 - Intro03:15 - Dr. Park, x-risk, and AGI08:55 - StakeOut.AI12:05 - Governance scorecard19:34 - Hollywood webinar22:02 - Regulations.gov comments23:48 - Open letters 26:15 - EU AI Act35:07 - Effective accelerationism40:50 - Divide and conquer dynamics45:40 - AI "art"53:09 - OutroLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.StakeOut.AIAI Governance Scorecard (go to Pg. 3)Pause AIRegulations.gov USCO StakeOut.AI Comment OMB StakeOut.AI Comment AI Treaty open letterTAISCAlpaca: A Strong, Replicable Instruction-Following ModelReferences on EU AI Act and Cedric O Tweet from Cedric O EU policymakers enter the last mile for Artificial Intelligence rulebook AI Act: EU Parliament’s legal office gives damning opinion on high-risk classification ‘filters’ EU’s AI Act negotiations hit the brakes over foundation models The EU AI Act needs Foundation Model Regulation BigTech’s Efforts to Derail the AI Act Open Sourcing the AI Revolution: Framing the debate on open source, artificial intelligence and regulationDivide-and-Conquer Dynamics in AI-Driven Disempowerment

Take a trip with me through the paper Large Language Models, A Survey, published on February 9th of 2024. All figures and tables mentioned throughout the episode can be found on the Into AI Safety podcast website.00:36 - Intro and authors01:50 - My takes and paper structure04:40 - Getting to LLMs07:27 - Defining LLMs & emergence12:12 - Overview of PLMs15:00 - How LLMs are built18:52 - Limitations if LLMs23:06 - Uses of LLMs25:16 - Evaluations and Benchmarks28:11 - Challenges and future directions29:21 - Recap & outroLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.Large Language Models, A SurveyMeysam's LinkedIn PostClaude E. ShannonA symbolic analysis of relay and switching circuits (Master's Thesis)Communication theory of secrecy systemsA mathematical theory of communicationPrediction and entropy of printed EnglishFuture ML Systems Will Be Qualitatively DifferentMore Is DifferentSleeper Agents: Training Deceptive LLMs that Persist Through Safety TrainingAre Emergent Abilities of Large Language Models a Mirage?Are Emergent Abilities of Large Language Models just In-Context Learning?Attention is all you needDirect Preference Optimization: Your Language Model is Secretly a Reward ModelKTO: Model Alignment as Prospect Theoretic OptimizationOptimization by Simulated AnnealingMemory and new controls for ChatGPTHallucinations and related concepts—their conceptual background

Esben reviews an application that I would soon submit for Open Philanthropy's Career Transitition Funding opportunity. Although I didn't end up receiving the funding, I do think that this episode can be a valuable resource for both others and myself when applying for funding in the future.Head over to Apart Research's website to check out their work, or the Alignment Jam website for information on upcoming hackathons.A doc-capsule of the application at the time of this recording can be found at this link.01:38 - Interview starts05:41 - Proposal11:00 - Personal statement14:00 - Budget21:12 - CV22:45 - Application questions34:06 - Funding questions44:25 - OutroLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.AI governance talent profiles we’d like to seeThe AI Governance Research SprintReasoning TransparencyPlaces to look for fundingOpen Philanthropy's Career development and transition fundingLong-Term Future FundManifund

Before I begin with the paper-distillation based minisodes, I figured we would go over best practices for reading research papers. I go through the anatomy of typical papers, and some generally applicable advice.00:56 - Anatomy of a paper02:38 - Most common advice05:24 - Reading sparsity and path07:30 - Notes and motivationLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.Ten simple rules for reading a scientific paperBest sources I foundLet's get critical: Reading academic articles#GradHacks: A guide to reading research papersHow to read a scientific paper (presentation)Some more sourcesHow to read a scientific articleHow to read a research paperReading a scientific article 

Join our hackathon group for the second episode in the Evals November 2023 Hackathon subseries. In this episode, we solidify our goals for the hackathon after some preliminary experimentation and ideation.Check out Stellaric's website, or follow them on Twitter.01:53 - Meeting starts05:05 - Pitch: extension of locked models23:23 - Pitch: retroactive holdout datasets34:04 - Preliminary results37:44 - Next steps42:55 - RecapLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.Evalugator libraryPassword Locked Model blogpostTruthfulQA: Measuring How Models Mimic Human FalsehoodsBLEU: a Method for Automatic Evaluation of Machine TranslationBoolQ: Exploring the Surprising Difficulty of Natural Yes/No QuestionsDetecting Pretraining Data from Large Language Models

I provide my thoughts and recommendations regarding personal professional portfolios.00:35 - Intro to portfolios01:42 - Modern portfolios02:27 - What to include04:38 - Importance of visual05:50 - The "About" page06:25 - Tools08:12 - Future of "Minisodes"Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.From Portafoglio to Eportfolio: The Evolution of Portfolio in Higher EducationGIMPAlternativeToJekyllGitHub PagesMinimal MistakesMy portfolio

Darryl and I discuss his background, how he became interested in machine learning, and a project we are currently working on investigating the penalization of polysemanticity during the training of neural networks.Check out a diagram of the decoder task used for our research!01:46 - Interview begins02:14 - Supernovae classification08:58 - Penalizing polysemanticity20:58 - Our "toy model"30:06 - Task description32:47 - Addressing hurdles39:20 - Lessons learnedLinks to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.ZooniverseBlueDot ImpactAI Safety SupportZoom In: An Introduction to CircuitsMNIST dataset on PapersWithCodeClusterability in Neural NetworksCIFAR-10 datasetEffective Altruism GlobalCLIP (blog post)Long Term Future FundEngineering Monosemanticity in Toy Models

A summary and reflections on the path I have taken to get this podcast started, including some resources recommendations for others who want to do something similar.Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.LessWrongSpotify for PodcastersInto AI Safety podcast websiteEffective Altruism GlobalOpen Broadcaster Software (OBS)CraigRiverside

This episode kicks off our first subseries, which will consist of recordings taken during my team's meetings for the AlignmentJams Evals Hackathon in November of 2023. Our team won first place, so you'll be listening to the process which, at the end of the day, turned out to be pretty good.Check out Apart Research, the group that runs the AlignmentJamz Hackathons.Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.Generalization Analogies: A Testbed for Generalizing AI Oversight to Hard-To-Measure DomainsNew paper shows truthfulness & instruction-following don't generalize by defaultGeneralization Analogies WebsiteDiscovering Language Model Behaviors with Model-Written EvaluationsModel-Written Evals WebsiteOpenAI Evals GitHubMETR (previously ARC Evals)Goodharting on WikipediaFrom Instructions to Intrinsic Human Values, a Survey of Alignment Goals for Big ModelsFine Tuning Aligned Language Models Compromises Safety Even When Users Do Not IntendShadow Alignment: The Ease of Subverting Safely Aligned Language ModelsWill Releasing the Weights of Future Large Language Models Grant Widespread Access to Pandemic Agents?Building Less Flawed Metrics, Understanding and Creating Better Measurement and Incentive SystemseLeutherAI's Model Evaluation HarnessEvalugator Library

In this minisode I give some tips for staying up-to-date in the everchanging landscape of AI. I would like to point out that I am constantly iterating on these strategies, tools, and sources, so it is likely that I will make an update episode in the future.Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance.ToolsFeedlyarXiv Sanity LiteZoteroAlternativeToMy "Distilled AI" FolderAI Explained YouTube channelAI Safety newsletterData Machina newsletterImport AIMidwit AlignmentHonourable MentionsAI Alignment ForumLessWrongBounded Regret (Jacob Steinhart's blog)Cold Takes (Holden Karnofsky's blog)Chris Olah's blogTim Dettmers blogEpoch blogApollo Research blog

Alice Rigg, a mechanistic interpretability researcher from Ottawa, Canada, joins me to discuss their path and the applications process for research/mentorship programs.Join the Mech Interp Discord server and attend reading groups at 11:00am on Wednesdays (Mountain Time)!Check out Alice's website.Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance. EleutherAI Join the public EleutherAI discord server DistillEffective Altruism (EA)MATS Retrospective Summer 2023 postAmbitious Mechanistic Interpretability AISC research plan by Alice RiggSPARStability AI During their most recent fundraising round, Stability AI had a valuation of $4B (Bloomberg) Mech Interp Discord Server

We're back after a month-long hiatus with a podcast refactor and advice on the applications process for research/mentorship programs.Check out the About page on the Into AI Safety website for a summary of the logistics updates.Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance. MATSASTRA FellowshipARENAAI Safety CampBlueDot ImpactTech with TimFast.AI's Practical Deep Learning for CodersKaggleAlignmentJamsLessWrongAI Alignment Forum

This episode is a brief overview of the major takeaways I had from attending EAG Boston 2023, and an update on my plans for the podcast moving forward.TL;DLStarting in early December (2023), I will be uploading episodes on a biweekly basis (day TBD).I won't be releasing another episode until then, so that I can build a cache of episodes up.During this month (November 2023), I'll also try to get the podcast up on more platforms, set up comments on more platforms, and create an anonymous feedback form.Links Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance. How to generate research proposalsKarolina Sarek: How to do research that mattersWill releasing the weights of future large language models grant widespread access to pandemic agents?Like the show? Think it could be improved? Fill out this anonymous feedback form to let me know!Please email all inquiries to intoaisafety@gmail.com.

In this episode I discuss my initial research proposal for the 2024 Winter AI Safety Camp with one of the individuals who helps facilitate the program, Remmelt Ellen.The proposal is titled The Effect of Machine Learning on Bioengineered Pandemic Risk. A doc-capsule of the proposal at the time of this recording can be found at this link.Links to all articles/papers which are mentioned throughout the episode can be found below, in order of their appearance. MegaSyn: Integrating Generative Molecule Design, Automated Analog Designer and Synthetic Viability PredictionDual use of artificial-intelligence-powered drug discoveryArtificial intelligence and biological misuse: Differentiating risks of language models and biological design toolsModel Organisms of Misalignment: The Case for a New Pillar of Alignment ResearchShadow Alignment: The Ease of Subverting Safely-Aligned Language ModelsFine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!unRLHF - Efficiently undoing LLM safeguards

Welcome to the Into AI Safety podcast! In this episode I provide reasoning for why I am starting this podcast, what I am trying to accomplish with it, and a little bit of background on how I got here.Please email all inquiries and suggestions to intoaisafety@gmail.com.