#235 - Sonnet 4.6, Deep-thinking tokens, Anthropic vs Pentagon

Our 235th episode with a summary and discussion of last week's big AI news! Recorded on 02/27/2026 Hosted by Andrey Kurenkov and Jeremie Harris Feel free to email us your questions and feedback at andreyvkurenkov@gmail.com and/or hello@gladstone.ai Read out our text newsletter and comment on the podcast at https://lastweekin.ai/ In this episode: Model and tool updates highlight Anthropic’s Sonnet 4.6 (1M context; strong ARC-AGI-2 results), Google’s Gemini 3.1 Pro (major ARC-AGI-2 jump and multimodal demos), xAI’s Grok 4.2 beta (multi-agent debate), plus Anthropic’s Claude Code “Remote Control” and Perplexity’s multi-agent “Computer” coordinator. Compute and business moves include Meta’s reported up-to-$100B AMD chip deal with warrant/equity incentives, MatX raising $500M to build specialized transformer chips shipping in 2027, World Labs raising $1B for world-model/3D environment tech, and a new startup raising $100M to simulate/predict human behavior. Infrastructure and geopolitics cover Stargate data-center delays amid OpenAI/Oracle/SoftBank control disputes and cash concerns, and China’s plan to scale 7nm/5nm wafer output despite yield and tooling constraints. Research and safety/policy discuss optimizer gains from masked updates, “deep thinking tokens” as a reasoning-effort signal, LLM attractor-state behaviors in bot-to-bot chats, mechanistic interpretability of counting/line-wrapping, methods to map task difficulty to human time horizons, plus Anthropic–Pentagon contract tensions, Anthropic’s report on distillation attacks (DeepSeek/Moonshot/Minimax), and OpenAI’s report on disrupting malicious use. A thank you to our current sponsors: Box - visit Box.com/AI to learn more ODSC AI - go to odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026. Factor - head to factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a year Timestamps: (00:00:10) Intro / Banter (00:01:52) News Preview Tools & Apps (00:03:20) Anthropic releases Sonnet 4.6 | TechCrunch (00:11:24) Google Rolls Out Latest AI Model, Gemini 3.1 Pro - CNET (00:14:54) Elon Musk says Grok 4.20 public beta is now available: Capabilities of AI chatbot offered by xAI - The Times of India (00:18:06) Anthropic just released a mobile version of Claude Code called Remote Control | VentureBeat (00:21:01) Perplexity announces "Computer," an AI agent that assigns work to other AI agents - Ars Technica Applications & Business (00:23:40) Meta strikes up to $100B AMD chip deal as it chases 'personal superintelligence' | TechCrunch (00:27:05) Nvidia challenger AI chip startup MatX raised $500M | TechCrunch (00:31:00) World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows | TechCrunch (00:33:07) Simile Raises $100 Million for AI Aiming to Predict Human Behavior (00:33:52) Stargate AI data centers for OpenAI reportedly delayed by squabbles between partners — sources say OpenAI, Oracle, and SoftBank disagreed on who would have ultimate control of the planned data centers (00:36:43) China to increase leading-edge chip output by 5x in two years, report claims — aims to lift 7nm and 5nm production to 100,000 wafers per month, targeting half a million monthly by 2030 Research & Advancements (00:40:33) On Surprising Effectiveness of Masking Updates in Adaptive Optimizers (00:48:03) Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (00:54:52) models have some pretty funny attractor states (01:01:41) When Models Manipulate Manifolds: The Geometry of a Counting Task (01:05:16) BRIDGE: Predicting Human Task Completion Time From Model Performance (01:12:00) NESSiE: The Necessary Safety Benchmark -- Identifying Errors that should not Exist (01:13:15) The least understood driver of AI progress (01:21:45) The Persona Selection Model: Why AI Assistants might Behave like Humans Policy & Safety (01:25:04) Anthropic CEO Amodei says Pentagon's threats 'do no