Good morning. The weight-editing story we covered yesterday has mostly turned into a correction cycle, with Simon Willison running his own experiments and a lengthy critique arguing the original post oversold its headline result. Elsewhere, Anthropic’s compute bill is eating more of its revenue just as it tries to court IPO investors, and Google has found a creative way to pressure schools into keeping AI turned on.
The Astra-edits-Qwen post, picked apart. A new critique argues that the viral weight-editing result isn’t what the headline suggests: the Qwen MATH-500 jump from 7.9% to 48.1% relied on Astra using Python tools and running the network, not reading raw weights. The genuinely tool-free capability is confined to tiny 17K–172K parameter networks. Mech interp researchers in the HN thread largely agreed, with one noting the Qwen improvement mostly reduces reasoning loops rather than teaching math, and others pointing to prior published work covering similar ground. Simon Willison separately showed Qwen3.8 27B can do the simpler task of guessing what a small network was trained on from its weights, which is interesting but a much weaker claim than the original framing implied.
Anthropic’s compute costs are getting worse, not better. The updated S-1 shows compute ate 58% of revenue in Q3 2026, up from 48% the prior quarter, per TechCrunch. Revenue grew from $11.5B to $17B, but Anthropic is paying premium short-term rates to “neoclouds” — $1.25B/month to SpaceX alone, with $5B of its $10B quarterly compute bill coming from these spot commitments. The company frames it as a demand problem rather than a structural one; IPO investors will have to decide if they believe that.
Anthropic wants to prove which model actually answered you. On the research side, Anthropic published FlashProve, an open-source cryptographic system for proving a given output came from a specific model without exposing the weights. The use case is straightforward: stop labs from quietly swapping cheaper models behind an API or inflating benchmark claims. No deployment plans yet, but the code and specs are out for external review.
Google ties Search to AI for schools. Starting next year, Google is making its classroom AI tools opt-in for K-12, which sounds reasonable until you read the fine print: schools that disable AI entirely also lose Google Search, because the features are now fused. TechCrunch has the details, and the NYT framing is less charitable. HN commenters pointed out Google has already done this with Gmail — you can’t turn off GenAI there without losing pre-LLM smart features either — and several suggested Kagi, which already offers free access to libraries, as the obvious off-ramp.
ChatGPT hits 20% of Google’s search volume. OpenAI told The Verge that ChatGPT now handles search queries at a fifth of Google’s rate — for context, Bing managed about 3% after a decade of trying. OpenAI’s definition excludes simple lookups like definitions and counts only queries that trigger actual web searches, so take the number with appropriate seasoning, but the trajectory is real.
Fal doubles to $16B as inference demand runs hot. AI inference host Fal raised $1.3B Series F led by Bessemer, with Nvidia participating, doubling its valuation from $8B just months ago. Annualized revenue crossed $800M on surging demand for video, image, and audio model hosting. Baseten, Fireworks, and Modal are all raising at similar multiples, though TechCrunch notes gross margins in the sector remain thin — compute is expensive whether you’re training or serving.
Nikkei: US tech giants’ off-balance-sheet AI obligations hit $3.2T. Nikkei reports that future commitments for data center leases and power contracts across US tech giants nearly doubled in six months to $3.2T — roughly 80% of their combined assets. HN commenters pushed back hard on the “debt” framing, correctly noting these are contractual future obligations, not borrowed money, and that aggregate customer backlogs (RPOs) sit at $2.33T as partial offset. Still, the growth rate is the story, and companies that used to generate free cash flow are now out raising capital.
Mechanism design as an alignment tool. MIT Technology Review covers a new MIT-led paper proposing “reverse game theory” for AI alignment — designing incentive structures where honest behavior dominates, rather than trying to interpret model internals. One concrete proposal: reward models based on how well their answers predict other models’ responses, which mathematically punishes sandbagging on safety evals. It’s an acknowledgment that interpretability and constitutional AI haven’t closed the deception problem on their own.
That’s the briefing. Anthropic’s S-1 continues to be more interesting than most companies’ entire investor decks — expect more pulled threads tomorrow.