Summary
The morning's strongest signal is a hardware-talent story: OpenAI's Clive Chan, an early hire on its custom-chip program, announced he is leaving for Anthropic, and HuggingFace's Elie Bakouch openly speculated this points to Anthropic starting internal chip development (Reuters has separately reported Anthropic is weighing its own AI chips). That lands inside the week's larger compute-and-money anxiety, with several handles riffing on Musk renting out his hoarded GPUs and on a wobbling AI market. The second cluster is Google Research live-tweeting its CVPR 2026 booth: 3DCodeBench (agentic procedural 3D modeling via code, with reusable skill and experience libraries), D4RT (a CVPR Best Paper for feedforward 4D scene reconstruction), and TIPSv2, whose headline finding is that distillation produces students that beat much larger teachers on patch-text alignment. Kilo Code kept hammering its model-routing thesis with fresh pricing (MiMo-V2.5-Pro hits 47.6% on Terminal Bench 2.0 at $4.92 a run versus GPT-5.5's 74.2% at $72.63). On the hardware-frontier fringe, @magicsilicon flagged Intel's neuromorphic partnership with Echo Neurotechnologies to train brain-activity-based algorithms on silicon. The rest of the feed is heavy political and Elon-adulation noise from @brivael and SF-scene color from @Scobleizer, little of it on-topic. No curated @bayesiansapien retweets in the window.
Posts
Clive Chan leaves OpenAI's chip program for Anthropic (@eliebakouch, @ns123abc, Reuters). Chan was the second hardware hire on OpenAI's custom-chip effort (previously Tesla Dojo) and posted a personal departure note; Bakouch asks directly whether his move signals Anthropic beginning internal chip development, and notes Reuters has reported Anthropic is weighing building its own AI chips. @ns123abc framed it as a defection right before OpenAI's custom silicon was supposed to ship. This is the human-capital edge of the hardware-sovereignty story the digest tracks (Anthropic also leasing 220k GPUs from SpaceX), and the clearest sign yet that the frontier labs are racing to own their compute stack, not just rent it.
Google CVPR 2026 booth: 3DCodeBench, D4RT, TIPSv2 (cluster of 3, @GoogleResearch). 3DCodeBench benchmarks agentic procedural 3D modeling: a VLM-driven agent turns nested Infinigen procedural factories into standalone Blender Python scripts, with a skills library (every successful API migration becomes a reusable skill) and an experience library indexing recurring patterns, across 212 object classes and 12 frontier VLMs under multi-turn refinement. D4RT (a CVPR Best Paper) is a feedforward model that jointly infers depth, spatio-temporal correspondence, and camera parameters from one video via a unified transformer, with a querying mechanism that probes any point in space-time without dense per-frame decoding. TIPSv2 reports that distillation unlocks superior patch-text alignment, so distilled student encoders significantly surpass their much larger teachers, a datapoint for the wiki's recurring "distillation is not merely lossy" thread (knowledge-distillation).
Kilo Code keeps pushing model-task routing with fresh pricing (@kilocode). The argument: every task defaulting to the most expensive frontier model is why AI coding bills are unpredictable, and the cheapest capable model for a job is often a fraction of the priciest. New number: MiMo-V2.5-Pro clears 47.6% on Terminal Bench 2.0 at $4.92 a run, while GPT-5.5 leads at 74.2% for $72.63. This is the same production-routing case behind their Claude-vs-MiniMax-M3 audit. → Kilo routing audit summary
Intel partners with Echo Neurotechnologies on neuromorphic silicon (@magicsilicon quoting Intel CEO Lip-Bu Tan). Intel is accelerating brain-inspired neuromorphic chip work, now developing algorithms trained on human brain activity itself rather than merely brain-inspired heuristics. @magicsilicon frames it as a bet that scaling ever-larger models in power-hungry datacenters will not reach AGI alone, so emulating the brain on silicon becomes the alternative lever. A fringe but on-theme counterpoint to the brute-scale capex story.
@ns123abc on a vertically integrated fab fantasy (@ns123abc). A riff on a "TERAFAB" building packaging, logic, memory, and masks on one site, with the punchline being the EUV bottleneck (buy out a Micron EUV contract, outbid TSMC for High-NA ASML machines). Lightweight, but it tracks the same lithography-scarcity thread as the recent TSMC High-NA EUV confirmation.
@Scobleizer SF-scene color (@Scobleizer). Mostly off-topic, but one anecdote is notable: a user gave the "poke" agent Gmail access to audit ten years of mail and it surfaced $990,000 in liabilities on long-shuttered entities that human CPAs and assistants had missed, by expanding the search to spelling and address variations. A small but concrete agentic-retrieval-beats-humans datapoint.
Political and adulation noise (Skip). @brivael's 16 tweets are almost entirely French politics and Elon worship; @MillionInt's "Hardware is hard" Elon hall-of-fame and @spencerpratt/@brivael SpaceX/AOC jokes carry no AI substance. Skipped.