One Year of Bad Takes, LLMs That Write Neural Network Weights & Inside an AI Benchmark
transcript
show notes
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 Pod
02:17 News: Cathedral Says It Hacked Anthropic in 12 Hours
05:39 News: Inside the Artificial Analysis Intelligence Index v5
14:37 Retrospective: Our Five Worst Takes of Year One
21:27 Biggest Thing We Learned Doing the Show
33:43 Advice to Our Year-Ago Selves
37:01 Predictions for Year Two
43:22 Deep Dive: Sparks of Weight-Space Programming
1:00:02 Two Minutes to Midnight: A $5T Sell-Off
1:03:58 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