Anthropic is building cryptographic proof of which AI model wrote what

Anthropic has published an open-source system called FlashProve that generates cryptographic signatures for AI outputs, mathematical proof that a response came from a specific model rather than a cheaper substitute quietly swapped in behind the scenes. The target is impersonation and hidden model-swapping as AI outputs circulate beyond their origin. Takeaway: content provenance is shifting from policy to math. For anyone publishing or auditing AI-generated work, machine-verifiable signing like FlashProve is worth tracking as an emerging default expectation. TechCrunch

Anthropic's amended filing prints the compute bill: $895 billion committed

Following yesterday's responsible-scaling headline, Anthropic's amended S-1 (filed November 2 with financials through September 30) puts hard numbers on the arc: $16.3B in half-year revenue against an $80.7B half-year loss, compute commitments now at $895B, and cloud deals with Google, Amazon and Microsoft running up to 20 years. Growth fell to 47%. Takeaway: the ratio to watch for AI company health is now compute-committed-to-revenue. It gets reprinted with every amendment. TechCrunch

Palantir's skyrocketing quarter hides a foreboding signal

Palantir's Q3 crushed consensus (revenue up roughly 96%) and the press release said so in as many words. Yet the shares dipped after earnings. The foreboding signal: the market is pricing saturation inside the skyrocket, and perfection is no longer enough. Takeaway: a beat can be a sell signal when expectations outrun any possible print, growth stocks live and die on the second derivative. TechCrunch

Inference startup Fal doubles its valuation to $16B as demand surges

Fal, which hosts image, video and audio models for creators and developers, raised $1.3B across Series E and F announced together, doubling its April valuation to $16B as inference demand keeps climbing. Takeaway: the picks-and-shovels trade keeps paying. The money quietly settles in the capacity layer. TechCrunch

Data centres are buying old jet engines to keep the lights on

Stuck in years-long queues for grid connections and new turbines, some AI data centres are converting scavenged aircraft jet engines into generators. The US now has 378 GW of gas-fired power capacity in development, up 50% in six months. In Australia, the Climate Council warns the data-centre boom could undercut the switch to renewables. Takeaway: for Australian readers this is now a domestic planning fight, not a foreign curiosity, the Climate Council piece is the local brief. TechCrunch · New York Times · Climate Council

🏠 Local models (Hugging Face)

  • ccore/Qwen3.5-4B-3D-Coder (605 likes), a 4B-parameter coding model aimed at 3D work; runs comfortably on a modest GPU.
  • baidu/Unlimited-OCR-2 (591 likes, ~190K downloads), local OCR for scanning documents without sending them to a cloud API.
  • LiquidAI/d1-2B (201 likes), a 2B-parameter experiment for tiny-hardware tinkerers. Context worth noting: Simon Willison reproduced the new "weight-space programming" paper on his laptop using Qwen 3.8 27B at 72-88% accuracy, open models reading other models' weights on consumer hardware is now a weekend project. simonwillison.net