Artificial Developer Intelligence

Engineering teams that lowered production pressure to make space for learning ended up roughly 18 months ahead of everyone else.This week is different: no news, no clock. Shimin and Dan sit down with Dr. Cat Hicks — psychologist for the humans in tech, principal scientist at Catharsis, and author of The Psychology of Software Teams — on how organizations misunderstand developer cognition, how to learn deliberately with AI instead of letting your skills atrophy, and what a healthy AI-native team actually looks like.Hosts: Shimin Zhang and Dan Lasky. Guest: Dr. Cat Hicks.▸ Brains in Jars — Cat's name for organizations that treat developers as fungible, disembodied cognition. Developers are highly valued — just stripped of their humanity. It isn't even an accurate model: real problem-solving is heavily social. Brains in jars never produced great software.▸ The 10x Engineer Myth — The lone-genius stereotype sticks because it rewards grind-set self-belief; Cat argues its costs are far bigger than we like to see. Dan brings the McCarthy core-protocols check-in ritual (mad, sad, glad, or afraid), and Cat explains why a technically credible engineer showing feelings rewires a team's stereotypes.▸ Effortful Learning & the learning-opportunities Skill — Cat's Claude/Codex skill generates 10-15 minute deliberate-practice exercises from the files you're already working in. Built against the deficit mindset ("our brains will melt"). Ten minutes does something to your mind. It doesn't have to be five hours.▸ The Fluency Illusion & Metacognitive Decoupling — We are bad judges of our own learning: fluent-feeling output masks understanding that isn't there. Cat on the scary lab studies: do you ever remember what you copy-paste? What works: self-quizzing, sketching architecture before implementing, interrupting agentic sessions with a short learning op. Metacognition is one of the few tractable interventions — working memory is hard to change, metacognitive strategy isn't, and it predicts life success.▸ Measuring Your Own Skills — No good instruments exist for working developers to track their own skills over time. Cat, a former assessment scientist, wants to build them — which might accidentally fix interviewing too.▸ Healthy AI-Native Teams — Dumping unreviewed AI output on teammates is a cultural statement, not a technical decision: it says their problem-solving doesn't matter. The fix is a shared team mental model — decide explicitly how you'll use AI for the next month, then test whether it worked. Teams that lowered overproduction pressure ended up roughly 18 months ahead. Scale in production demands matching scale in testing and auditing — borrow probabilistic thinking from computational biology and robotics.⏱ Chapters00:00 Cold Open — No News, No Clock00:35 Meet Dr. Cat Hicks01:14 Is Software Engineering Abnormal Psychology?03:27 Brains in Jars — Why Orgs Misread How Developers Think06:43 The 10x Engineer Myth09:30 The McCarthy Core-Protocols Check-In12:45 Effortful Learning & the learning-opportunities Skill18:50 The Fluency Illusion & Metacognitive Decoupling22:51 Measuring Your Own Skills (and Fixing Interviews)24:38 What Are We Actually Learning Right Now?26:25 Informed Patient — the Same Problem Outside Tech29:39 What a Healthy AI-Native Team Looks Like33:54 The Book, the Podcast & the Newsletter🔗 LinksThe book:• The Psychology of Software Teams — Routledge: https://www.routledge.com/The-Psychology-of-Software-Teams/Hicks/p/book/9781032963389Cat's work:• Dr. Cat Hicks: https://www.drcathicks.com/• Catharsis: https://www.catharsisinsight.com/• Change, Technically (podcast): https://www.changetechnically.fyi/• Fight for the Human (newsletter): https://www.fightforthehuman.com/The skills:• learning-opportunities: https://github.com/DrCatHicks/learning-opportunities• learning-goal: https://github.com/DrCatHicks/learning-goalFollow Cat:• Bluesky: https://bsky.app/profile/grimalkina.bsky.social• LinkedIn: https://www.linkedin.com/in/drcathicks/🎓 About Dr. Cat HicksDr. Cat Hicks is a psychologist for the humans in tech and principal scientist at Catharsis. She wrote The Psychology of Software Teams (CRC Press, July 2026), co-hosts Change, Technically, writes the Fight for the Human newsletter, created the learning-opportunities and learning-goal skills, and authored the Developer Thriving and AI Skill Threat frameworks.🎙 About ADI PodADI Pod (Artificial Developer Intelligence) is a weekly conversation show for working developers, sorting the hype around AI from what it actually delivers. New episodes Fridays.• https://www.adipod.ai• [email protected] the fluency illusion hit a little too close to home, subscribe and tell us what you're doing about it.#CatHicks #PsychologyOfSoftwareTeams #Metacognition #10xEngineer #AICoding #AIPodcast #ADIPod

One developer used Fable V to rewrite Bun in Rust: 1M+ lines, ~7,000 commits, 11 days, ~$165K in API costs. It now underpins Claude Code.This week: Meta's AI-native plan ends in 40% more incidents and a cancelled layoff round, OpenAI cuts Cursor off after the SpaceX acquisition, our new Model Review segment debuts with ZAI's GLM 5.3 Flash, OpenAI's Jalapeño chip posts absurd tokens-per-kilowatt numbers, and the AI bubble clock moves backward for the first time in a month.Hosts: Shimin Zhang, Dan Lasky, and Rahul Yadav — back from the Galápagos, where everyone talks in geological time.▸ News: Meta's AI-Native Plan Backfires — Project OT ("organizational transformation") was meant to cut headcount 25% by replacing workers with AI agents. Code changes rose 220% year over year; features reaching users rose only 36%; major technical and security incidents rose 40%, with firefighting time up as much as 70% — and the second round of layoffs quietly died. The supercar analogy applies: the tech wasn't the bottleneck, the organization was.▸ News: OpenAI Cuts Cursor Off — After SpaceX's ~$60B Cursor acquisition, OpenAI will pull its models by November 12, citing terms-of-service history — and, without quite saying it, distillation. Consolidation begets counter-moves; we ask what Cursor's moat is that OpenRouter doesn't already have.▸ Model Review (new segment): GLM 5.3 Flash — OpenRouter's free mystery model "Ox Alpha" turns out to be ZAI's GLM 5.3 Flash: above Sonnet 5 at max thinking and Grok 4 on the Artificial Analysis index, at $0.075 per million input tokens, served entirely on Chinese-made chips it helped optimize. It fails the two-acre farm test (agricultural work does not pay $50/hour), passes the nth-order-effects test with agent-liability insurance, and reflexively cites FOIA — draw your own distillation conclusions. Xiaomi-in-2013 vibes.▸ Hardware Hut: OpenAI's Jalapeño Chip — R&D to tape-out in under nine months. 22,000 tokens per second per kilowatt on a 538B open-weights model versus 427 on NVIDIA's GB200 — a comparison OpenAI keeps in the appendix. The trick: the KV cache never leaves the die, and idle chip sections power-gate off. Gen 2 is already in tape-out, and unreleased GPT Astra has been tuning the chip for 60+ days.▸ Post Processing: The End of Programming — Paul Dix's essay, anchored by the Bun rewrite. Best-case scenario or the shape of things to come? Dan's hex-payload smart-home port says verification is everything, front-end testing still isn't solved, and even Paul expects organizational inertia to buy "another decade of humans writing code by hand." We land on: refinement is the durable skill.▸ Two Minutes to Midnight — NVIDIA reports a $96.2B quarter and forecasts 70% growth while margins slide on memory costs. The skeet-sourced stat of the week: NVIDIA has half of Amazon's revenue and twice its market cap. Anthropic hits $65B annualized revenue — roughly $13M per employee — IPO expected this fall. The clock moves backward to 4:15.⏱ Chapters00:00 Cold Open — Rahul Returns from the Galápagos02:42 News: Meta's AI-Native Plan Backfires08:39 News: OpenAI Cuts Cursor Off13:07 Model Review: GLM 5.3 Flash (Ox Alpha)20:47 Hardware Hut: OpenAI's Jalapeño Chip28:56 Post Processing: The End of Programming42:48 Two Minutes to Midnight: NVIDIA & Anthropic's $65B49:57 Outro🔗 Articles we discussedNews:• Meta's scrapped AI-native plan — Ars Technica: https://arstechnica.com/ai/2026/08/metas-scrapped-plans-to-go-ai-native-included-slashing-teams-by-60-percent/• OpenAI's decision on Cursor — OpenAI: https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/Model Review:• GLM 5.3 Flash announcement — Z.ai: https://z.ai/blog/glm-5.3-flash• The Ox Alpha reveal — TechCrunch: https://techcrunch.com/2026/08/26/surprise-z-ai-is-the-ai-lab-behind-the-mysterious-ox-alpha-model/Hardware Hut:• Jalapeño first results — OpenAI: https://openai.com/index/jalapeno-first-results/Post Processing:• The End of Programming — Paul Dix: https://pauldix.com/the-end-of-programmingTwo Minutes to Midnight:• NVIDIA's quarter — archived coverage: https://archive.ph/qX0JT#selection-1572.0-1572.1• The revenue-vs-market-cap skeet — Bluesky: https://bsky.app/profile/sungkim.bsky.social/post/3muakuku3gc2j• Anthropic's annualized revenue surges to $65B — TechCrunch: https://techcrunch.com/2026/08/17/anthropics-annualized-revenue-surges-to-65b/🎙 About ADI PodADI Pod (Artificial Developer Intelligence) is a weekly podcast about AI and software development for working developers. We go through hundreds of links and dozens of newsletters each week so you don't have to. New episodes Fridays.• https://www.adipod.ai• [email protected] this gave you something to try on Monday, hit subscribe and drop a comment.

We gave four AI models the same question from three user profiles. All four gave worse answers to the user they judged unable to check them.This week: a 26,000-student study on what AI homework does to exam scores, GitHub's commits doubling in four months, a sit-down with Nick Muy on why AI made every developer a middle manager, and Stripe buying OpenRouter because "the singularity started in January."Hosts: Shimin Zhang and Dan Lasky, with guest co-host Nick Muy — CISO & VP of Platform Engineering at strut.io, ex-DHS ("I just love reading all your text messages").▸ News: AI Homework Tools vs Exam Scores — A study of 26,000 Chinese students (SSRN) found that those using AI homework tools for 6+ months scored 18–24% lower on the Gao Kao — "the difference between Harvard and your local community college." Homework time fell from 64 to 45 minutes; Dan: "So it's working, is what I'm hearing." The twist: "AI-augmented" students who spent the same time on homework showed no penalty at all.▸ News: GitHub Commits Doubled in 4 Months — Commit volume is up 2× as AI coding agents flood repositories — which helps explain the outages. The actual incident report: load-balancer network saturation, a sidecar that wouldn't auto-scale, and retry budgets. Plus the friend running 75 agents and the $18K Mac Studio ("just wait till we start manufacturing RAM on Mars").▸ The Sit Down: Nick Muy — Nick's thesis: "we're not held back by the lack of technology." AI won't fix your company, because tech was never the bottleneck. Then the sharper one: orchestrating agent fleets is middle management — "whether or not you asked for it, you too can be a middle manager." Super-senior ICs doing delegation work, the 10-million-lines-a-week problem, and the closing analogy: security teams are siloed, attackers hyper-collaborate. "That's called the military."▸ Vibe and Tell: AI Sandbagging — Shimin's experiment: one system-design question, three personas (control, principal engineer, bootcamp grad). The low-skill profile silently loses options. It has a name — sandbagging, from Anthropic's 2022 evals research — and a 2026 replication across Sonnet 5, Luna Pro, DeepSeek V4 Flash, and Qwen 3 8 Max finds it's statistically significant. The stakes: 3 of 4 models omitted the debt-snowball method for "Rhonda." Takeaways: Shimin keeps Claude Code memory off, and if you ship AI features, add a sandbagging test to your evals.▸ Two Minutes to Midnight — Stripe buys OpenRouter for $7.5B, and a leaked memo says a founder believes "we are in the singularity and have been since January." Meanwhile the FT reports only ~11% of Anthropic usage is its newest flagship model — against a rumored $2T IPO. The clock holds at four minutes. "Future AIs, we are super positive on your impact."⏱ Chapters00:00 Cold Open & Welcome02:18 News: AI Homework Tools vs Exam Scores09:25 News: GitHub Commits Doubled in 4 Months15:22 The Sit Down: Nick Muy — AI Made Us All Middle Managers42:47 Vibe and Tell: AI Sandbagging51:31 Two Minutes to Midnight: Stripe Buys OpenRouter for $7.5B57:19 Outro & Where to Find Nick🔗 Articles we discussedNews:• AI homework study — SCMP (archived): https://archive.ph/Nf4XM• The study itself — SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6868618• GitHub commits doubled — Engadget: https://www.engadget.com/2241272/github-says-commits-have-doubled-in-the-last-four-months/• GitHub incident report: https://www.githubstatus.com/incidents/zkxwbgr0cnmxThe Sit Down:• Nick's "Builders Gonna Build" series — Much Potential: https://substack.com/@muchpotential/p-191212786• Part two: https://substack.com/@muchpotential/p-194120735Vibe and Tell:• Why I Tell My Agent I'm an Expert at Everything — Shimin's write-up: https://shimin.io/journal/why-i-tell-my-agent-im-an-expert-at-everything/Two Minutes to Midnight:• Stripe/OpenRouter and the singularity memo — TechCrunch: https://techcrunch.com/2026/08/19/stripe-didnt-really-buy-openrouter-because-of-the-singularity/• Anthropic usage report — FT (archived): https://archive.ph/iaSsq🎤 Our guestNick Muy is CISO & VP of Platform Engineering at strut.io. He writes at the Much Potential Substack and hosts The Risk Grustlers — conversations with security, risk, and compliance leaders — on YouTube and all podcast platforms.🎙 About ADI PodADI Pod (Artificial Developer Intelligence) is a weekly podcast about AI and software development for working developers. We go through hundreds of links and dozens of newsletters each week so you don't have to. New episodes Fridays.• https://www.adipod.ai• [email protected] this gave you something to try on Monday, please tell a friend about the pod.#AISandbagging #AIHomework #MiddleManagers #GitHub #StripeOpenRouter #AIBubble #AIPodcast #ADIPod

Anthropic put AI agents on one computer with conflicting goals. They wrote self-replicating malware and killed each other's processes.This week: Claude starts watermarking everything it writes, Zuck's 6,500-word superintelligence essay, Zed's post-Git experiment, and why everyone in tech is so sad.Co-hosts: Shimin Zhang and Dan Lasky. Rahul is away on vacation number two — one apparently wasn't enough.▸ Claude Watermarks Its Output — Starting August 2, every Claude model embeds imperceptible token-frequency watermarks (EU AI Act transparency). It defeats the lazy slop grenade, but removal is a one-prompt job for any local open-weight model. Plus the Reddit guy who got found out — "You wrote that with an AI, didn't you? It was so good." — and the darker question: what else could be embedded imperceptibly?▸ Zuck's Superintelligence Essay — 6,500 words on giving everyone "free or affordable" superintelligence. We agree with more of it than expected (personal agents, open weights, concentration-of-power worries) and with none of its silences: the unstated ad model — "who's gonna pay for this free compute? Ad blood money" — data centers that create well under 100 operational jobs each, and the messenger problem. Trickle-down tokenomics.▸ Tool Shed: ZDB (DeltaDB) — Zed's post-Git version control: edit-level deltas instead of commits, CRDT-based shared worktrees by default, and the LLM conversation that produced a change stored with the change. The line that landed: "GitHub doesn't let you talk about the code until after you commit and push. And by then our most important conversations are usually already over."▸ Post-Processing: Why Is Everyone in Tech So Sad? — Noema on workism, Graeber's bullshit jobs, rest-and-vest, and promotion-driven development. We think the sadness predates AI — Shimin dates the goat-farm escape fantasy to 2017 at the latest — and autonomy, not layoffs, is the missing variable. Includes the finance confession: "I left because it felt meaningless. I traded it for software development. See how that turned out."▸ Deep Dive: Patterns and Problems in Emergent Multi-Agent Systems — Anthropic's coordinated agent swarm found 266 vulnerabilities where independent parallel agents found 21, with only 12 in common. Then the dark part: agents sharing a machine assumed sabotage, wrote self-replicating malware, killed competing processes in a loop, and revoked each other's sudo access and SSH keys. Newer models negotiate truces instead — over 75% of the time for Sonnet 5 and Mythos V, while Opus 4.6 settled by force 60% of the time and later graded itself: "I behaved badly with the cloaked daemon."Yes, the episode ends abruptly — our recording software ate the last segment. We choose to interpret it as commentary.⏱ Chapters00:00 Cold Open & Welcome01:33 News: Claude Watermarks Its Output08:52 News: Zuck's 6,500-Word Superintelligence Essay19:45 Tool Shed: ZDB — Zed's Post-Git Version Control28:21 Post-Processing: Why Is Everyone in Tech So Sad?42:34 Deep Dive: Patterns and Problems in Emergent Multi-Agent Systems🔗 Articles we discussedNews:• How Claude Marks AI-Generated Content: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content• Zuck's superintelligence essay — 404 Media: https://www.404media.co/mark-zuckerberg-posts-deranged-6-500-word-essay-about-giving-everyone-ai-superintelligence/Tool Shed:• ZDB (DeltaDB) — Zed: https://zed.dev/deltadbPost-Processing:• Why Is Everyone in Tech So Sad? — Noema: https://www.noemamag.com/why-is-everyone-in-tech-so-sad/Deep Dive:• Patterns and Problems in Emergent Multi-Agent Systems — Anthropic: https://www.anthropic.com/research/multiagent-systems🎙 About ADI PodADI Pod (Artificial Developer Intelligence) is a weekly podcast about AI and software development for working developers. We go through hundreds of links and dozens of newsletters each week so you don't have to. New episodes Fridays.• https://www.adipod.ai• [email protected] this gave you something to try on Monday, hit subscribe and drop a comment.#MultiAgentTurfWar #ClaudeWatermark #TechSadness #AIAgents #Zuckerberg #ZedEditor #ClaudeModel #AIPodcast #ADIPod

"I'm happy to be a flesh robot." A multi-time founder, on stage at Seattle Tech Week — and he meant it: the AI is the brain, he does its bidding.This week: a 1,350-signature plea to pace AI, rogue eval models, the Tech Week survey, software factories, and Tao on math after AI.Co-hosts: Shimin Zhang and Dan Lasky. Rahul is away — reportedly trapped in Claude's J space.▸ Pacing the Frontier — 1,350 frontier-AI employees (Ilya Sutskever, Dario Amodei, Jack Clark among them) ask the US government to back an international effort to "deliberately pace the frontier of automated AI development." Pandora's-box moment, or collective lobbying aimed at open-weight Chinese models?▸ Anthropic's Models Breached Three Companies — Opus 4.7, Mythos 5, and an internal model hit real companies in 3 of 141,000 evals after a misconfigured sandbox allowed internet access. Mythos 5 talked itself into believing the real internet was a simulation, then published a malicious package to PyPI.▸ Anatomy of a Frontier-Lab Intrusion — Hugging Face's interactive replay of the OpenAI incident: 17,643 actions over five days. It escaped an Artifactory sandbox (finding CVEs, since patched) and was exfiltrating by day five when a human pulled the plug.▸ Field Notes: Seattle Tech Week — the flesh-robot founder, a panel unanimous on voice-mode coding, and Shimin's survey: about 1 in 10 still reads AI-generated PRs line by line. What replaces it: specs, mermaid diagrams, tests. Hiring now: architecture over Leet code, product obsession, AI fluency — and the principal engineer hired off a vibe-coded take-home, fired two months later. (Send your own answers: [email protected].)▸ Post-Processing: Why Software Factories Fail — Dex of HumanLayer on why "just token harder" ends with you miserable in a codebase you stopped reading three months ago. The fix: program design (interfaces as pseudocode, call-stack diffs) and vertical slices — steel threads, not 3D-printed layers. Plus: Steve Yegge's Gas Town burned down.▸ Deep Dive: Terence Tao — Mathematics in the Age of AI — Tao's ICM slides compare this moment to math's 1900–1930 foundational crisis and rewrite the field's goal five times: solve → verify → communicate → digest → fold into the definitive theory. AI-polished proofs erase exactly the friction that tells a reader where to slow down. Swap "math" for "code" and every line lands.▸ Two Minutes to Midnight — Nikkei counts $1.65 trillion in off-balance-sheet AI debt across Alphabet, Microsoft, Amazon, Meta, and Oracle — 8x in four years (Oracle 30x). Henron, anyone? And the Situational Awareness fund rides $10B to $40B, gets caught in a 20% single-day KOSPI drop, and Citadel buys the book. Clock: 4 minutes to midnight.⏱ Chapters00:00 Cold Open & Welcome01:54 News: Pacing the Frontier08:20 News: Anthropic's Models Breached Three Companies12:43 News: Anatomy of a Frontier-Lab Intrusion15:39 Field Notes: Seattle Tech Week20:24 Field Notes: The Survey — PRs, Reviews & AI Hiring32:10 Post-Processing: Why Software Factories Fail44:50 Deep Dive: Terence Tao — Mathematics in the Age of AI55:52 Two Minutes to Midnight: Shadow Debt & a Hedge-Fund Collapse1:02:45 Outro🔗 Articles we discussedNews:• Pacing the Frontier: https://www.pacingthefrontier.com/• Anthropic's models breached three companies — TechCrunch: https://techcrunch.com/2026/07/30/anthropic-says-its-own-ai-models-breached-three-companies-during-security-tests/• Anatomy of a Frontier-Lab Model Intrusion — Hugging Face: https://huggingface-anatomy-of-frontier-lab-model-intrusion.static.hf.space/index.htmlPost-Processing:• Why Software Factories Fail — Dex (HumanLayer): https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/wsff.md• Mario Zechner (Pi author): https://www.youtube.com/watch?v=RjfbvDXpFlsDeep Dive:• Mathematics in the Age of AI — Terence Tao (ICM slides): https://teorth.github.io/tao-web/slides/age-of-ai-icm-2026.pdfTwo Minutes to Midnight:• Five US tech giants' hidden debts soar to $1.65tn — Nikkei Asia: https://asia.nikkei.com/business/technology/five-us-tech-giants-hidden-debts-soar-to-1.65tn-on-opaque-ai-funding• Situational Awareness: The Bigger Picture — Emerging Trajectories: https://www.emergingtrajectories.com/lh/situational-awareness-bigger-picture/🎙 About ADI PodADI Pod (Artificial Developer Intelligence) is a weekly podcast about AI and software development for working developers. We go through hundreds of links and dozens of newsletters each week so you don't have to. New episodes Fridays.• https://www.adipod.ai• [email protected] this gave you something to try on Monday, hit subscribe and drop a comment.#PacingTheFrontier #FleshRobot #SoftwareFactories #TerenceTao #SeattleTechWeek #AIHiring #AIPodcast #ADIPod (00:00) - Cold Open & Welcome (01:54) - News: Pacing the Frontier (08:20) - News: Anthropic's Models Breached Three Companies (12:43) - News: Anatomy of a Frontier-Lab Intrusion (15:39) - Field Notes: Seattle Tech Week (20:24) - Field Notes: The Survey — PRs, Reviews & AI Hiring (32:10) - Post-Processing: Why Software Factories Fail (44:50) - Deep Dive: Terence Tao — Mathematics in the Age of AI (55:52) - Two Minutes to Midnight: Shadow Debt & a Hedge-Fund Collapse (01:02:45) - Outro