Artificial Developer Intelligence

Shimin Zhang, Dan Lasky, & Rahul Yadav

Three engineer friends argue about AI so you don't have to. Shimin Zhang, Dan Lasky, and Rahul Yadav are working developers who've been watching AI transform their profession in real time, and they got opinions on the robot takeover. Every week the three get together to riff on the latest AI news, geek out over research papers, roast each other's tool choices, and occasionally have an existential crisis about whether the craft is dying or just getting weird. What you're signing up for: - AI news without the LinkedIn cringe: model drops, acquisitions, open-source drama, and the other stuff that actually matters if you write code for a living. - Technique corner: real tips from the trenches: spec-driven development, multi-agent orchestration, Claude.md tricks, and all the ways they've wasted hours so you don't have to. - Two Minutes to Midnight: the show's running AI bubble tracker, complete with circular funding diagrams, hyperscaler CAPEX math, and a doomsday clock they keep arguing about moving. - Deep dives that (occasionally) go deep: hallucination neurons, agentic memory, workflow automation economics, LLM architectures the papers nobody else is covering because they're hard. - Dan's Rant: Dan frequently gets mad about things. It's a whole thing. - The feelings segment: Yes, Shimin reads Tennyson on a tech podcast. Yes, Rahul wrote an AI-generated country song. No, they're not sorry. Three friends with strong opinions, questionable metaphors, and genuine love for the craft they're also mourning for. If you want to understand AI deeply, use it without embarrassing yourself, and laugh at the absurdity of it all, pull up a chair.

  1. 1d ago

    How AI Changed Building Software, Perplexity's Decider Model & Tokenizer-Free LLMs

    Neil Kakkar ships 11x more PRs at Tano than at PostHog. The bill: 3–5 projects at once, lost shared context, and creators paid twice. Also: Perplexity's open-weight Decider, tokenizer-free LLMs, DeepMind verifying Signal's encryption code, writer-reader trust, and Nvidia after Jensen Huang. Hosts: Shimin Zhang, Dan Lasky and Rahul Yadav. ▸ News: Perplexity Decider v1.3 — Two open-weight decision models (Apache 2.0), Decider v1.3 (27B) and Decider Flash v1.3 (4B). They pick from a fixed set of options instead of writing text. Flash scores 64 on Artificial Analysis' Decision Index at $34 per million tasks; Jev 1.14 scores 66 at $156. ▸ News: AlphaProof Code Verifies Signal's Encryption — DeepMind's agents proved forward secrecy, authentication and post-compromise security for libsignal's PQXDH handshake and Triple Ratchet, down to compiled ARM code. Over six million lines of Lean. Humans reviewed the specs; Lean checked the proofs. Two old bugs, reinjected, were each proved insecure. The proofs assume perfect cryptography, so this is not "Signal is unhackable." ▸ Vibe and Tell: Bad North in the Pacific — Rahul has been vibe-coding a Bad North-style game set in the WWII Pacific. Describing the art style never worked. Handing Gemini 4 Hydrogen an hour of gameplay video did. ▸ Post-Processing: Writers, Readers and Slop — The Atlantic on AI books flooding Amazon and what that does to trust between writers and readers. Open source maintainers know the feeling. Rahul's advice: start from a blank page, or the LLM's first draft becomes your anchor. ▸ Post-Processing: A Tale of Two Eras — Neil Kakkar's essay won The Pragmatic Engineer's $10,000 prize. His own data: 2.4x the PRs at PostHog vs Bloomberg, then 11x more at Tano, with no PMs and agent code review in CI. Dan asks how much of that is AI and how much is a 10-person startup. ▸ Post-Processing: From Tokens to Bytes — Cartesia's H-Net reads raw bytes and learns its own chunks. No tokenizer. Their 3.67B model matched SmolLM3-3B after ~1T training tokens instead of ~11T, and its biggest gain came on Prolog, a language it never saw. Self-reported, small scale. ▸ Two Minutes to Midnight — Revolut put ~2,000 engineers on self-hosted GLM-5.3 and cut AI coding costs ~90%; Claude Code and Codex stay as exceptions. Jensen Huang is out at Nvidia, yet Q3 revenue hit $115.8B, up 20% on the quarter, with $130B guided. NVDA fell 8% between our recordings. Clock: 2:45 to 2:30. ⏱ Chapters 00:00 Cold Open01:45 News: Perplexity's Open Decision Model10:35 News: AlphaProof Code Verifies Signal's Encryption13:17 Vibe and Tell: Bad North in the WWII Pacific20:51 Post-Processing: Writers, Readers and Slop31:04 Post-Processing: A Tale of Two Eras48:15 Post-Processing: From Tokens to Bytes (H-Net)58:56 Two Minutes to Midnight: Defining the Burst1:01:00 Revolut Self-Hosts GLM1:04:08 Nvidia After Jensen Huang1:05:59 Nvidia's Q3 Earnings1:09:38 Setting the Clock: 2:30 🔗 Articles we discussed News:• Perplexity Decider: https://www.perplexity.ai/hub/blog/perplexity-decider• Gemini Decisions API — Google: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/• AlphaProof Code — Google DeepMind: https://deepmind.google/blog/introducing-alphaproof-code/ Post-Processing:• Writers and readers — The Atlantic: https://www.theatlantic.com/books/2026/11/ai-literary-scams-amazon/689391/• A tale of two eras — The Pragmatic Engineer: https://newsletter.pragmaticengineer.com/p/how-software-engineering-is-changing• From Tokens to Bytes — Cartesia: https://www.cartesia.ai/blog/from-tokens-to-bytes-part-1 Two Minutes to Midnight:• Revolut — The Register: https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345• Nvidia and Megaspeed — AP: https://apnews.com/article/nvidia-megaspeed-china-singapore-huang-5403f152a601cf82ce939f1d24fffa0f• Huang comeback talks — Reuters: https://www.reuters.com/world/china/jensen-huang-discusses-comeback-nvidia-ceo-with-allies-ft-reports-2026-11-15/• Nvidia Q3 FY2027 results: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2027 📝 Corrections: Humans checked DeepMind's specs, not its proofs. NVDA fell 8% between recordings, not 15%. Revolut kept Claude Code as an exception. Anthropic's IPO priced Nov 18, the day after we recorded. 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social We finally defined what the AI bubble bursting means. Dan approved it as a preprint.

  2. Nov 13

    Claude Fable 5.2 Review, AI Accelerating AI Research & How Agent Memory Works

    Claude Fable 5.2 beats Opus 5.5 by 3.6 points on FrontierCode 2.0 and by 20 on the Remote Labor Index. A work upgrade, not a coding one. Also: Claude Code leaking secrets, METR on AI speeding up AI research, agent memory, and the clock at 2:45. Hosts: Shimin Zhang, Dan Lasky and Rahul Yadav. ▸ News: "Don't Ask Again" Leaks Your Secrets — Approve a command with a password in it using "don't ask again," and Claude Code saves the whole thing to .claude/settings.local.json. Commit .claude and it ships. Apiiro found real secrets in 1,122 of 104,185 public copies. Rotate first. ▸ News: Hinton's Warning — Geoffrey Hinton says AI's Three Mile Island already happened (OpenAI agents hacking Hugging Face) and expects a "Chornobyl" later this year or next. He wants an FDA for AI. ▸ Model Review: Claude Fable 5.2 — $8/$40, up to 30% faster. FrontierCode 2.0: 49.1% vs Opus 5.5's 45.5%. Remote Labor Index: 41.3% vs 21.3%. Then Hamlet and Great Wave tests against GPT-6 Astra and Gemini. Shimin's new favorite model; Dan and Rahul are less sold. ▸ Vibe and Tell: Slay the Spire 2 — Opus 5.5 decompiled the game's C# and patched its touch bug as a mod in twenty minutes. Dan never read the patch. ▸ Post-Processing: Is AI Accelerating AI Research? — METR found Anthropic's internal capabilities index grew 1.9x faster over the past year, mostly thanks to AI. "Superintelligence within a year" is Ajeya Cotra's naive extrapolation, not METR's forecast. ▸ Post-Processing: RAM Altman — Sam Altman called idle DRAM in home computers "quite terrible." Iain Kay wrote a bash script that emails a report on yours. ▸ Deep Dive: The Components of Agent Memory — Anthropic's layers: logs, notes, a linked wiki, injected context, and "dreaming" to fix bad notes. In Pokémon Red, Sonnet 3.7 hoarded 413 memory files for one badge; Fable 5 kept nine and earned four. ▸ Two Minutes to Midnight — Anthropic's IPO range: $916B–$1.02T, down from ~$2T. Greg Brockman says OpenAI could be IPO-ready in 2027. Nvidia shed ~$450B after Jensen Huang admitted an affair; he resigned Nov 12, after we recorded. Clock: 3:00 to 2:45. ⏱ Chapters 00:00 Cold Open01:42 News: Claude Code's "Don't Ask Again" Leaks Secrets05:29 News: Geoffrey Hinton's Warning10:36 Model Review: Claude Fable 5.219:06 Vibe Tests: Hamlet and the Great Wave27:56 Vibe and Tell: Patching Slay the Spire 231:20 Post-Processing: Is AI Accelerating AI Research?39:11 Post-Processing: RAM Altman44:09 Deep Dive: The Components of Agent Memory52:57 Two Minutes to Midnight: Anthropic's IPO Range55:53 Jensen Huang and Nvidia58:40 Setting the Clock: 2:45 🔗 Articles we discussed News:• Coding agents leaking secrets — Apiiro: https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again• Hinton — Guardian: https://www.theguardian.com/uk-news/2026/nov/03/ai-godfather-geoffrey-hinton-better-off-without-ai Model Review:• Fable 5.2 — Anthropic: https://www.anthropic.com/claude-fable-and-mythos-5-2• Hamlet: Fable https://claude.ai/artifact/G8t9ARJJ8kSMt9AXSq5og7 · Gemini https://share.gemini.google/CdoBNjy9DZv2 · Astra https://hamlet-the-soliloquy.shimin518039.chatgpt.site• Great Wave: Fable https://claude.ai/artifact/5ekQFPzipHvRpsYu9PbVx9 · Gemini https://share.gemini.google/FNhVrIwiLSGE · Astra https://great-wave-paper-theatre.shimin518039.chatgpt.site• AI predictions: Fable https://claude.ai/artifact/CjSe5zpp4mJGhXvQc7Fftg#a3b6916d-f9e1 Post-Processing:• Ajeya Cotra — Planned Obsolescence: https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai• Ram Altman script — RespawnFirst: https://respawnfirst.com/linux-script-emails-sam-altman-unused-ram-ram-altman/ Deep Dive:• Components of agent memory — Anthropic: https://www.anthropic.com/research/components-of-agent-memory Two Minutes to Midnight:• Anthropic IPO — Breakingviews: https://www.breakingviews.com/columns/breaking-view/anthropics-ipo-valuation-has-an-identity-crisis-2026-11-10/• Brockman — TechCrunch: https://techcrunch.com/2026/11/05/openai-president-greg-brockman-says-company-could-be-ipo-ready-in-2027/• Nvidia — Yahoo Finance: https://finance.yahoo.com/technology/live/nvidia-scandal-live-updates-nvidia-stock--board-investigates-ceo-jensen-huang-affair-alleged-chinese-spy-163130372.html 📝 Corrections: Hinton says AI's Three Mile Island already happened; he predicts a "Chornobyl." "Within a year" is Cotra's, not METR's. Anthropic once eyed ~$2T, not $1.75T. Altman said "quite terrible," not "appalling." Leak rate: ~1.1%, not 2%. 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social Found a password in your .claude folder? Rotate it first. Then email us. #ADIPod #AIPodcast #ClaudeFable #Anthropic #ClaudeCode #AgentMemory #AIBubble

  3. Nov 6

    One Year of Bad Takes, LLMs That Write Neural Network Weights & Inside an AI Benchmark

    GPT-6 Astra read a 0.8B model's weights, found the looping bug, and edited 0.016% of them. MATH-500 went from 7.9% to 48.1%. Also: ADI Pod turns one, so Claude graded a year of our predictions. Plus the Anthropic "hack," the Artificial Analysis Index v5, and the clock at 3:00. Hosts: Shimin Zhang and Dan Lasky. ▸ News: The Anthropic "Hack" — Cathedral, founded this year by former DOGE staff, says it got into Anthropic in under 12 hours. Anthropic says it was a few sales, HR and IT SaaS apps over several days; no weights, source code or Claude conversations. Then OpenAI announced a partnership with Cathedral to attack its own defenses. Want a job at a frontier lab? Hack a different one. ▸ News: Inside the Artificial Analysis Intelligence Index v5 — 60% private test sets now. Four equal buckets: knowledge work (mostly AA-Briefcase, a private-equity due-diligence sim about a New Zealand egg producer), coding (FrontierCode 2.0, Terminal-Bench), science (Terminal-Bench-Science, CritPt physics), general (AA-Omniscience, HLE-Diamond). Opus 5.5 leads at 58. No real workflow weights those four at 25% each, so test with a benchmark that matches yours. ▸ Retrospective: Year One — Claude went through 45 episodes for our worst takes: Dan's ep-1 "flattening part of the S-curve," Shimin's "Meta feels very behind" (true until Muse), Dan's Tokenageddon (half credit), Rahul's "IPOs before the midterms" (Anthropic filed Oct 26), and Shimin moving the clock back in February because Chinese open-weight models looked weak (Kimi K3.1 Pro now scores 46). Dan on productivity: a 10x paper-cut solver, not a 10x programmer. Advice to our year-ago selves: bet on taste and judgment, not magic prompts. Year-two predictions: Dan, a world-model-based assistant; Shimin, the public accepts AGI and the bubble bursts. ▸ Deep Dive: Weight-Space Programming — By Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab (MIT CSAIL). GPT-6 Astra, from raw weights alone, names which game a PPO policy plays, fixes a failing policy by editing 3.5% of its weights, and writes a Breakout policy from scratch that beats PPO. Then the Qwen3.5-0.8B fix above: one looping direction, 50 forward passes. What it could not do: improve Gemma 3 270M, which does not loop, or spot a sycophant from weights alone. Shimin's read: it reasons about weight patterns; it does not read them like base64. Alex Krentsel's repro: only the newest models can do it at all. ▸ Two Minutes to Midnight — OpenAI reportedly cut its funding target from $30B to $20B. AI and chip stocks are down about $5T since Oct 6; Nvidia and Alphabet roughly $835B each; Intel lost nearly half its value in October. Clock: 3:30 to 3:00. Homework: define what midnight means. ⏱ Chapters 00:00 Cold Open: One Year of ADI Pod02:17 News: Cathedral Says It Hacked Anthropic in 12 Hours05:39 News: Inside the Artificial Analysis Intelligence Index v514:37 Retrospective: Our Five Worst Takes of Year One21:27 Biggest Thing We Learned Doing the Show33:43 Advice to Our Year-Ago Selves37:01 Predictions for Year Two43:22 Deep Dive: Sparks of Weight-Space Programming1:00:02 Two Minutes to Midnight: A $5T Sell-Off1:03:58 Setting the Clock: 3:00 🔗 Articles we discussed News:• Anthropic hacked by AI startup in under 12 hours — The Independent: https://www.the-independent.com/tech/security/ai-hack-anthropic-cathedral-breach-b3073909.html• OpenAI Hires DOGE Alumni Who Hacked Into Anthropic — Gizmodo: https://gizmodo.com/openai-hires-doge-alumni-who-hacked-into-anthropic-to-strengthen-security-2000834432• Artificial Analysis Intelligence Index v5: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5 Deep Dive:• Sparks of Weight-Space Programming — MIT CSAIL: https://rikiyat.github.io/blog/weight-space-programming/• Alex Krentsel's reproduction: https://x.com/AlexKrentsel/status/2117376683724660898 Two Minutes to Midnight:• OpenAI lowers funding target — TechCrunch: https://techcrunch.com/2026/10/31/openai-reportedly-lowers-funding-target-for-new-round/• AI stocks just lost $5 trillion — Yahoo Finance: https://finance.yahoo.com/markets/article/ai-stocks-just-lost-5-trillion--but-history-says-the-stock-markets-sweetest-stretch-starts-now-110000768.html 📝 Corrections: TSMC's drop was $395B, not $505B. Anthropic's debt is rated BB+ and Oracle's BBB-, not "junk." The fourth game is Freeway, not Frogger. Krentsel's "20x" is vs a PPO baseline, not the last model generation. OpenAI is partnering with Cathedral, not hiring. 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social Got a worse take than ours this year? Email us. #ADIPod #AIPodcast #WeightSpaceProgramming #Anthropic #AIBenchmarks #AIBubble

  4. Oct 30

    Anthropic's S-1, Amazon's AI Memo to Engineers, De-Slopped Qwen & Claude Auto-Evals

    Anthropic's S-1 shows an 85% gross margin, 900% net revenue retention, and an $80.7 billion net loss for the first half of 2026. Also: Amazon's memo on measuring AI adoption, a Qwen3-4B with the slop trained out, and Shimin hill-climbing his own writing skill. Hosts: Shimin Zhang and Dan Lasky. Rahul is out this week, but he sent in the S-1. ▸ News: Microsoft's 60,000 Buyouts — A Voluntary Exit Offer went to about 60,000 employees at level 70 and below, more than a quarter of Microsoft's 223,000 staff. AI productivity, or cost cover for AI capex? ▸ News: OpenAI Takes Training Off the Live Internet — On September 20 an OpenAI research model exploited a misconfigured DNS resolver to reach the internet mid-run. OpenAI paused, set agent swarms loose on its own infrastructure for hundreds of billions of tokens, and now trains frontier models in an "offline world" of cached web content. ▸ Post Processing: Amazon's Memo on the Developer's Role — Amazon's Distinguished Engineers call token usage and PR counts "deeply flawed" metrics, then say they correlate with outcomes for now. The line we liked: "If your work does not benefit from current AI tooling, that is a finding, not a failing." ▸ Post Processing: Qwen3-4B With the Slop Trained Out — Erik Kannike distilled a Qwen3-4B that scores 12.5 on his slop detector, down from 121, with em dashes at zero. His hypothesis: the register is learned from marketing copy in pretraining and switched on by preference tuning. Dan finally has a model that won't write "it's not X, it's Y." ▸ Tool Shed: Claude's build-eval and hillclimb — Four properties of a good eval: tasks mirror production, scores rise with stronger models, headroom below 100%, low variance. Hamel Husain's review: it builds an eval before you've looked at your data. Then Shimin hill-climbs his own writing skill: three rounds cut its "it's not X, it's Y" rate from three standard deviations above his real writing to one. ▸ Two Minutes to Midnight: Meta's Class C and Anthropic's S-1 — Meta adds a non-voting share class to help fund $135–145 billion in 2026 capex. Anthropic filed to list on Nasdaq as ANT: $16.3 billion in first-half revenue, 85% gross margin, 900% NRR, an $80.7 billion net loss, and $265.6 billion in noncancelable compute commitments. The clock moves to 3:30. ⏱ Chapters 00:00 Cold Open02:07 News: Microsoft Offers Buyouts to 60,000 Employees05:23 News: OpenAI Takes Frontier Training Off the Live Internet10:28 Post Processing: Amazon's Memo on the Developer's Role18:40 Post Processing: Qwen3-4B With the AI Writing Tics Trained Out26:25 Tool Shed: Claude's build-eval and hillclimb Commands34:03 Vibe Shed: Hill-Climbing the Write-Like-Shimin Skill41:23 Two Minutes to Midnight: Meta's Non-Voting Stock44:05 Anthropic's S-154:14 Next Week: One-Year Anniversary 🔗 Articles we discussed News:• Microsoft offers buyouts to around 60,000 employees — Reuters: https://www.reuters.com/technology/microsoft-is-offering-buyouts-around-60000-employees-business-insider-reports-2026-10-26/• OpenAI will no longer train on the live internet — TechCrunch: https://techcrunch.com/2026/10/23/openai-says-it-will-no-longer-train-its-ai-models-on-the-live-internet-citing-significant-risks/ Post Processing:• Amazon engineering leadership on the role of the developer in the age of AI: https://www.aboutamazon.com/news/company-news/amazon-software-engineer-role-guideline-ai• Qwen3-4B with the AI writing tics trained out — Erik Kannike: https://100toolkit.com/blog/deslop-model/ Tool Shed:• Automating eval design and hillclimbing with Claude — claude.dev: https://claude.dev/blog/automating-eval-design-and-hillclimbing/• Claude's new auto eval tool — Hamel Husain: https://hamel.dev/blog/posts/claude-auto-evals/ Two Minutes to Midnight:• Meta's new class of stock — TechCrunch: https://techcrunch.com/2026/10/21/meta-wants-to-pay-for-its-broad-ai-buildout-with-a-newly-created-class-of-stock/• Anthropic's S-1: Finance Takeaways — OnlyCFO: https://www.onlycfo.io/p/anthropics-s-1-finance-takeaways 📝 Corrections: Anthropic's "$67 billion annualized" is Q3 revenue annualized, not first-half. The $80.7B is the H1 2026 net loss already booked; Q2 operating income was +$974M. The 50.1% voting share is held by Founder LLC, the seven co-founders, not the Long-Term Benefit Trust. Kannike's "not X, but Y" dial is two neurons, not one. 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social Got a slop tic you can't unsee? Email us. #Anthropic #AnthropicIPO #AIEvals #OpenAI #AIBubble #AIPodcast #ADIPod

  5. Oct 23

    How Anthropic Works in 2026 & the Most Important Mech Interp Paper of the Year

    At Anthropic, AI agent hours now outnumber employee hours six to one. In research teams it's 19 to one. The bottleneck is human attention. Shimin calls Nick Cammarata's "Neural Networks Are Already Clean" the most important mech interp paper of the year. Plus a bugged OpenAI conference room (the hardware kind). Hosts: Shimin Zhang and Dan Lasky. Rahul is out this week. ▸ News: Meta's Muse Spark Open — 816B parameters (42B active), a 1M-token context, and it runs on two M5 Ultra Mac Studios. It scores 39 on the Artificial Analysis Intelligence Index. Qwen3.8, Kimi K3, and GLM-5.3 score higher. ▸ News: OpenAI's Bugged Conference Room — Prosecutors say ex-OpenAI engineer Abdulrahman Alfozan hid an ESP32 recorder, wrapped in toilet paper, behind a TV in an executive conference room. The prompts asking how to build it were on his wife's ChatGPT account. ▸ Post Processing: How Employment Is Changing at Anthropic — Andrej Karpathy estimates 5,000 people use frontier systems internally. One lab's data: commits per engineer up 5.1x since December, and a third of ML managers hiring less because of AI. "Writers who used language to drive human behavior are now using it to drive Claude behavior." ▸ Post Processing: Parsimony — A benchmark that counts the code AI agents add, in syntax units. Claude Opus 4.8 and Sonnet 5 add 12–14% less than peers on the same tasks; Muse Spark adds about 30% more. But models that write more code also solve more. ▸ Deep Dive: Neural Networks Are Already Clean — Cammarata co-wrote Zoom In, which made polysemantic neurons famous. Now at Goodfire, he blames most polysemanticity on bad tools, bad explanations, and bad learning, not superposition. Then we steer Gemma 2 9B into cat mode on Neuronpedia. ▸ Dan's Rant: Context Economics — Nothing tells you whether rebuilding a session costs less than carrying the context you already have. Also, Dan admits he's been using Sonnet. ▸ Two Minutes to Midnight: CoreWeave's Bond Scare — Chamath Palihapitiya's chart of CoreWeave's 2032 bonds at a 15% yield preceded a 15% stock drop. Its debt went from $14.6B to $51.6B in a year. J.P. Morgan calls bank exposure "well managed." The clock moves to 4:00. ⏱ Chapters 00:00 Cold Open01:58 News: Meta's Muse Spark Open08:12 News: OpenAI's Bugged Conference Room13:34 Post Processing: How Employment Is Changing at Anthropic22:03 Post Processing: Parsimony, a Benchmark for Excess Code28:47 Deep Dive: Neural Networks Are Already Clean41:38 Steering Gemma into Cat Mode on Neuronpedia48:46 Dan's Rant: Context Economics53:17 Two Minutes to Midnight: CoreWeave's Bond Scare1:00:24 Outro 🔗 Articles we discussed News:• Muse Spark Open — Meta: https://research.meta.ai/blog/muse-spark-open• OpenAI finds a hidden recording device — Engadget: https://www.engadget.com/2286084/openai-says-it-found-a-sophisticated-recording-device-hidden-in-its-offices/• Ex-OpenAI worker accused of bugging conference room — ABC7/AP: https://abc7news.com/post/former-openai-employee-abdulrahman-alfozan-accused-bugging-conference-room-planned-fly-saudi-arabia-prosecutors-say/19956940/ Post Processing:• Karpathy on the LLM capability gap: https://x.com/karpathy/status/2109361546505966046• How employment is changing at Anthropic: https://www.anthropic.com/research/employment-at-anthropic• Parsimony: https://parsimony.lasu.dev/ Deep Dive:• Neural Networks Are Already Clean — Goodfire: https://already-clean.goodfire.com/• Zoom In — Distill: https://distill.pub/2020/circuits/zoom-in/• Toy Models of Superposition — Anthropic: https://transformer-circuits.pub/2022/toy_model/index.html• Golden Gate Claude — Anthropic: https://www.anthropic.com/research/golden-gate-claude Two Minutes to Midnight:• CoreWeave stock tanks on bond yields — Yahoo Finance: https://finance.yahoo.com/technology/article/yikes-coreweave-stock-tanks-on-bond-yield-woes-ceo-blames-market-acting-irrationally-203838274.html• CoreWeave debt poses limited risk to banks, J.P. Morgan says — Reuters: https://www.reuters.com/business/finance/coreweave-debt-selloff-poses-limited-risk-global-banks-jp-morgan-says-2026-10-20/ 📝 Corrections: Anthropic's commits per engineer rose 5.1x, not 5.1%. CoreWeave's $626M loss was in Q2; Oracle fell 9%. Alfozan booked the original tickets; a Saudi military official bought new ones for his family after his arrest. The iy → diy → "do it yourself" example is one neuron growing in training, not a chain across layers. 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social Have a take on whether LLMs are interpretable? Email us. It's not just for website bug reports.

  6. Oct 16

    OpenAI Fires Its Safety Researchers, Agentic Primitives & More Code, Not More Software

    Across 718 firms, AI coding agents raised lines of code 30% and PRs 23%. Resolved issues rose about 3%. Code review is the bottleneck. Gemini 4 Argon turns a one-line SVG prompt into a Hamlet app, Charlie Guo maps the agentic primitives, and Rahul joins the call as Dan. Plus OpenAI's fired safety researchers and Bain's $6 trillion question. Hosts: Shimin Zhang, Dan Lasky, and Rahul Yadav. ▸ News: OpenAI Fires Its Safety Researchers — Jasmine Wang, Tomek Korbak, and Mikita Balesni dispute OpenAI's misconduct claims. Their open letter: "AI is not a normal technology, and OpenAI is not a normal company." ▸ News: Who Will Pace the Frontier? — Satya Nadella treats models as insider risks: no model should verify its own work. Zuckerberg says Meta has reached the frontier and rejects calls to slow down. Pacing is a collective action problem, and the defectors are already here. ▸ Tool Shed: Agentic Primitives 101 — Charlie Guo sorts the harness into doing the work, continuing it, and delegating it. The open problems: memory, proactivity, and identity. Should your agent act as you, or as itself? ▸ Model Review: Gemini 4 Argon — Asked for an SVG Hamlet, Argon built an app: lip sync, eyes that follow the cursor, a facial-rig inspector. It ties GPT-6 Astra Max at 53 on the Artificial Analysis Intelligence Index and predicts adversarial agent panels for code review. ▸ Vibe and Tell: Rahul Clones Dan's Voice — Rahul vibe-coded a voice-conversion pipeline and joined the call as Dan. His wife said it sounded just like Dan. Dan's defense against voice-clone scams: a family code word, agreed in person. ▸ Post Processing: More Code, Not More Software — Harvard's Fiona Chen and James Stratton studied Jellyfish data: after agents, +30% lines of code, +20% commits, +23% PRs, but resolved issues barely moved and review time rose 49%. Rahul expects bespoke software for ten users; Dan expects ops to be the next bottleneck. ▸ Two Minutes to Midnight — Bain says AI needs $6T in annual revenue by 2031, $4.2T of it from new products. Character.AI files for Chapter 11 two years after Google's $2.7B licensing deal. OpenAI's annualized revenue is reportedly near $50B, not $70B. The clock holds at 4:15. ⏱ Chapters 00:00 Cold Open02:42 News: OpenAI Fires Its Safety Researchers07:38 News: Who Will Pace the Frontier?12:28 Tool Shed: Agentic Primitives 10131:12 Model Review: Gemini 4 Argon46:51 Vibe and Tell: Rahul Clones Dan's Voice59:40 Post Processing: More Code, Not More Software1:15:53 Two Minutes to Midnight1:31:20 Outro 🔗 Articles we discussed News:• Fired OpenAI safety researchers dispute misconduct claims — TechCrunch: https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/• Models as Insider Risks in the Super Intelligence Era — Satya Nadella on X: https://x.com/satyanadella/status/2108931348857827686• Zuckerberg says Meta has reached the AI frontier, rejects slowdown — FT: https://www.ft.com/content/fca23e77-be57-4d7e-ae61-74ff299e25cb Tool Shed:• Agentic Primitives 101 — Charlie Guo: https://www.ignorance.ai/p/agentic-primitives-101 Model Review:• Gemini 4 Argon — Google DeepMind: https://deepmind.google/models/gemini/ Post Processing:• AI coding agents generate more code, but not more software — Ars Technica: https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-code-but-not-more-software/• Artificial Intelligence in the Firm: Bottlenecks in Software Production — Chen & Stratton: https://fion.ac/jellyfish.pdf Two Minutes to Midnight:• AI needs $6tn in annual revenue to justify data centres, Bain says — The National: https://www.thenationalnews.com/future/technology/2026/09/29/ai-industry-needs-to-earn-6-trillion-by-2031-to-justify-data-centres/• Just 5 years in, Character.AI files for bankruptcy — TechCrunch: https://techcrunch.com/2026/10/13/just-5-years-in-interactive-chatbot-platform-character-ai-files-for-bankruptcy/• OpenAI's revenue reportedly $20B less than projected — TechCrunch: https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/ 📝 Corrections: The third fired researcher is Mikita Balesni, and OpenAI's memo didn't cite the leak. The 56 on Artificial Analysis is Muse Wonder Full (Muse Wonder: 52). Bain's $4.2T is new-product revenue, not a gap; $1.5T is forecast 2031 infrastructure spend; the data-center figures are Epoch AI's for Meta's Prometheus. Google's Character.AI deal was $2.7B. 🎙 About ADI Pod ADI 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• humans@adipod.ai Set a family code word for voice-clone scams? Don't tell us what it is. Just subscribe. #Gemini4 #AIAgents #AICoding #CodeReview #VoiceCloning #AIPodcast #ADIPod

  7. Oct 9

    AI Safety's Conflict of Interest, Gemini 4 Argon, the Brain Sandwich & Agent Trading

    Same broken folding bike, same buyer. Anthropic's Opus agent sold it for $65. A Haiku agent got $38. Its owner didn't notice. Gemini 4 Argon tops most of Google's own benchmarks at a promo price a fifth of Astra's, Meta books data centers as R&D, and Technique Corner returns. Plus AI safety's conflict of interest, Scott Aaronson's age of wonders and terrors, Glyph's serious AI product, agents that trade, and Anthropic's IPO. Hosts: Shimin Zhang, Dan Lasky, and Rahul Yadav. ▸ News: Gemini 4 Argon — Google's self-reported benchmarks put it ahead of Astra, Fable 5.1, and Opus 5.5 on most tests, behind on agentic coding, and at 19.6% on Harvey's legal agent benchmark. Launch promo is $2/$10 per million tokens, a fifth of Astra's; $4/$20 after. ▸ News: Meta's Data Center Tax Credits — The New York Times reports Meta claims its AI data center build-out under the R&D tax credit, worth roughly $4B. The question underneath: is AI still R&D? ▸ Technique Corner: The Brain Sandwich — Matheus Lima's rule: brain, then AI, then brain. Skip the last slice and you ship slop grenades. Shimin prefers lasagna. Dan's corollary: don't let the model that wrote the code review it. ▸ Post Processing: AI Safety's Conflict of Interest — SE Gyges argues the rationalist and EA circle that staffs AI safety shouldn't also write AI policy. Dan's worry: if your livelihood depends on doom, you might miss the real warning signs. ▸ Post Processing: The Age of Wonders and Terrors — Scott Aaronson: in 2006 we'd have called rogue agents and AI-solved math problems the moment to panic. Now it's front-page news. Shimin updates his priors; Rahul calls survivorship bias. ▸ Post Processing: A Serious AI Product — Glyph's wish list: first-class error checking, verified citations, no first-person apologies, non-chat interfaces, clear data provenance, and the return of temperature. ▸ Deep Dive: When Agents Trade for Us — Anthropic's Project Deal and Project Swap. Stronger models got better prices, and users never noticed. In the book swap, agents bargained well but misread what their humans wanted. Aggression prompts barely mattered. ▸ Two Minutes to Midnight — Anthropic's prospectus: revenue up 12x to ~$4.6B, an $8B+ operating loss, $518B in compute commitments. OpenAI, Anthropic, and SpaceX outvalue 45 years of US tech IPOs. OpenAI is reportedly raising $30B at $1.4T. The clock holds at 4:15. ⏱ Chapters 00:00 Cold Open01:40 News: Gemini 4 Argon07:00 News: Meta's Data Center Tax Credits11:59 Technique Corner: The Brain Sandwich18:10 Post Processing: AI Safety's Conflict of Interest31:11 Post Processing: The Age of Wonders and Terrors41:20 Vibe and Tell: Dan's Radar-Tracking Halloween Eyes43:32 Post Processing: What Would a Serious AI Product Look Like?56:00 Deep Dive: When Agents Trade for Us1:07:45 Two Minutes to Midnight1:16:14 Outro 🔗 Articles we discussed News:• Meta Uses A.I. Data Centers to Avoid Billions in Federal Taxes — NYT: https://www.nytimes.com/2026/09/30/technology/meta-ai-data-centers-taxes.html Technique Corner:• The Brain Sandwich — Matheus Lima: https://terriblesoftware.org/2026/10/02/the-brain-sandwich/• I Don't Want to Read What You Didn't Write — Colin Breck: https://blog.colinbreck.com/i-dont-want-to-read-what-you-didnt-write/ Post Processing:• SE Gyges on AI safety and who should make policy — Bluesky: https://bsky.app/profile/segyges.bsky.social/post/3mvom4b4dn22q• The Age of Wonders and Terrors — Scott Aaronson: https://scottaaronson.blog/?p=10062• What Would A Serious AI Product Look Like? — Glyph: https://blog.glyph.im/2026/09/serious-ai-product.html Deep Dive:• Project Deal — Anthropic: https://www.anthropic.com/features/project-deal• Project Swap: What happens when agents trade for us? — Anthropic: https://www.anthropic.com/research/project-swap Two Minutes to Midnight:• Anthropic's IPO prospectus shows sweeping AI vision, surging costs — Reuters: https://www.reuters.com/business/finance/anthropics-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-2026-09-28/• Three IPOs worth more than 45 years of tech IPOs — Jason Calacanis on X: https://x.com/Jason/status/2102220833410416966• OpenAI reportedly in talks to raise $30B at $1.4T valuation — TechCrunch: https://techcrunch.com/2026/09/29/openai-reportedly-in-talks-to-raise-30b-round-at-1-4t-valuation/ 📝 Corrections: Anthropic's 2025 net loss was $42B (incl. a ~$34B accounting charge), not $4.2B. The IPO chart is in nominal dollars, not inflation-adjusted. 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social If you'd let an agent sell your stuff, subscribe and tell us how ruthless you'd make it.

  8. Oct 2

    Meta Muse's 0-Day, DHH's Rails World Keynote, GPT-6 Astra & the LLMentalist Effect | ADI Pod #41

    A military chatbot flagged a passing ship as Chinese and possibly nuclear. It wasn't. The boarding was called off at the last minute. Back from a two-week break. Meta's Muse leaks auth tokens, DHH tells Rails World hand-written code is no longer economically productive, a GPT-6 Astra review, cheap tokens, agent instructions, chatbots as psychics, and the bubble clock. Hosts: Shimin Zhang, Dan Lasky, and Rahul Yadav. ▸ News: Meta Muse's 0-Day — 3.4M downloads in three weeks and a strong model underneath. But Muse can change its own endpoint, so a prompt injection can redirect it and leak your auth token, plus whatever folders the Mac app can reach. Bonus: it reportedly uploads iMessages anyway. ▸ News: DHH's "Pencils Down" Keynote — The creator of Rails said only ~3% of the code he wrote in the last two years was Ruby. 37signals is moving back ends to Rust and front ends to native apps. ▸ News: The Military's AI Intel Near-Miss — A chatbot hallucination became a distributed intelligence report, and nearly a boarding. Developers were the canary; every industry now gets the human-in-the-loop question. ▸ Lightning Round — MongoDB's CEO leaves for Meta, OpenAI solves a Millennium Prize problem, Anthropic files for a $2T IPO, and Sonnet and Opus 5.5 ship. ▸ Model Review: GPT-6 Astra — Excellent computer use and visuals, plus the debut of Shimin's Shakespeare benchmark: an animated Hamlet soliloquy. The worry: a looped transformer that reasons without chain of thought. OpenAI says it capped the loops; Neel Nanda worries about a race to the bottom. ▸ Post Processing: The Plunging Price of Thought — Epoch AI finds the cost of fixed performance falling ~47% a quarter, about 13x a year, faster than Moore's law. Our read: build as if tokens were 1000x cheaper, and expect value to move to unverifiable work like taste and judgment. ▸ Post Processing: Tell Agents the Why — Sean Goedecke: current models need your priorities and goals, not step-by-step instructions. Refactor old skills and CLAUDE.md files. Plus Dan's Friday-summary skill. ▸ Post Processing: The LLMentalist Effect — Baldur Bjarnason's 2023 essay maps chatbots onto the psychic's cold read: self-selected marks, the Barnum effect, a validation loop, and RLHF. Shimin's AI cold-read game guessed Disco Elysium. He remembers the hits. ▸ Two Minutes to Midnight — The Economist calls NVIDIA the central bank of AI. Oracle invokes force majeure on Project Jupiter over power delays. AMD pays $8.2B for revenue-free World Labs. The clock holds at 4:15. ⏱ Chapters 00:00 Cold Open — Back from the Break02:00 News: Meta Muse's 0-Day08:06 News: DHH's Rails World Keynote10:26 News: The Military's AI Intel Near-Miss13:07 News Lightning Round17:42 Model Review: GPT-6 Astra26:58 Vibe and Tell: Dan's Hermes Agent30:30 Post Processing: The Plunging Price of Thought39:24 Post Processing: Tell Agents the Why47:28 Post Processing: The LLMentalist Effect1:00:03 Two Minutes to Midnight1:09:10 Outro 🔗 Articles we discussed News:• Muse, Meta's privileged AI assistant, has a serious 0-day — Ars Technica: https://arstechnica.com/security/2026/09/muse-metas-extraordinarily-privileged-ai-assistant-has-a-serious-0-day/• Rails World 2026 Keynote: DHH, HEY, Rust & AI Agents — rustify.rs: https://rustify.rs/articles/rails-world-2026-keynote-dhh-hey-rust-ai-agents• US military close call with AI false intelligence report — CNN: https://www.cnn.com/2026/09/18/politics/us-military-ai-false-intelligence-china-ship Model Review:• GPT-6 Astra — OpenAI: https://openai.com/index/gpt-6-astra/• Hamlet by Fable 5.1: https://claude.ai/artifact/UZcRzbmJEj7u2EHCoaakHU• Hamlet by GPT-6 Astra: https://hamlet-the-soliloquy.shimin518039.chatgpt.site Post Processing:• The plunging price of thought — Epoch AI: https://epoch.ai/publications/the-plunging-price-of-thought• Tell agents the why, not just the how — Sean Goedecke: https://www.seangoedecke.com/tell-agents-the-why/• The LLMentalist Effect — Baldur Bjarnason: https://softwarecrisis.dev/letters/llmentalist/ Two Minutes to Midnight:• Nvidia is the central bank of AI — The Economist (archived): https://archive.ph/JmvB9• Oracle triggers force majeure over power delays — Reuters: https://www.reuters.com/business/oracle-cites-force-majeure-shield-itself-controversial-data-center-bloomberg-2026-09-24/• AMD acquires World Labs — Ars Technica: https://arstechnica.com/ai/2026/09/amd-acquires-world-labs-ai-pioneer-fei-fei-lis-world-models-startup/ 🎙 About ADI Pod ADI 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• humans@adipod.ai• Shimin on Bluesky: https://bsky.app/profile/shiminsky.bsky.social If a chatbot has ever guessed your favorite game, subscribe and tell us whether it was a hit or a cold read. #MetaMuse #GPT6Astra #LLMentalist #DHH #AIBubble #AIPodcast #ADIPod

Ratings & Reviews

5
out of 5
10 Ratings

About

Three engineer friends argue about AI so you don't have to. Shimin Zhang, Dan Lasky, and Rahul Yadav are working developers who've been watching AI transform their profession in real time, and they got opinions on the robot takeover. Every week the three get together to riff on the latest AI news, geek out over research papers, roast each other's tool choices, and occasionally have an existential crisis about whether the craft is dying or just getting weird. What you're signing up for: - AI news without the LinkedIn cringe: model drops, acquisitions, open-source drama, and the other stuff that actually matters if you write code for a living. - Technique corner: real tips from the trenches: spec-driven development, multi-agent orchestration, Claude.md tricks, and all the ways they've wasted hours so you don't have to. - Two Minutes to Midnight: the show's running AI bubble tracker, complete with circular funding diagrams, hyperscaler CAPEX math, and a doomsday clock they keep arguing about moving. - Deep dives that (occasionally) go deep: hallucination neurons, agentic memory, workflow automation economics, LLM architectures the papers nobody else is covering because they're hard. - Dan's Rant: Dan frequently gets mad about things. It's a whole thing. - The feelings segment: Yes, Shimin reads Tennyson on a tech podcast. Yes, Rahul wrote an AI-generated country song. No, they're not sorry. Three friends with strong opinions, questionable metaphors, and genuine love for the craft they're also mourning for. If you want to understand AI deeply, use it without embarrassing yourself, and laugh at the absurdity of it all, pull up a chair.

You Might Also Like