The White House rejected Anthropic's plan to let 70 more companies use its vulnerability-finding AI Mythos, citing security and compute shortages — partly because the government wants to run it itself.
OpenAI traced the GPT models' obsession with goblin metaphors to a reward signal meant for the 'Nerdy' personality. The signal contaminated training data, causing goblin mentions to spike 175%.
NVIDIA's new Blackwell GPUs cost nearly double per hour to rent, but a 65x boost in token throughput slashes per-million-token cost from $4.20 to $0.12 — at least with all software optimizations turned on.
Cambricon's Q1 revenue doubled and profit nearly tripled, driven by AI demand and US export controls that have cut off Nvidia and AMD from the Chinese market.
Amazon's Q1 EPS doubled to $2.78 on a $16.8B Anthropic gain, but adjusted earnings missed by a penny. AWS revenue surged 28%, while free cash flow plunged 95%.
Amazon CEO Andy Jassy says the company's in-house AI chip business could generate $50 billion annually if spun off, and all four generations of Trainium are essentially sold out, backed by orders from Anthropic, OpenAI, and Uber.
Emails and texts presented in court show Musk stopped OpenAI payments in 2017, fought for control, and used an ex-girlfriend on the board as an information pipeline.
Parag Agrawal's Parallel Web Systems raises $100M Series B, tripling valuation to $2B, as the startup builds specialized web search infrastructure for AI agents.
A new GEO paper tests three AI search engines — ChatGPT, Google AI Overview/Gemini, and Perplexity — and finds that citation count isn't everything: ChatGPT cites the fewest sources but absorbs their content most deeply.
Sam Altman argues per-token pricing will become obsolete, comparing OpenAI to an "intelligence factory" where customers care only about the final price and outcome, not the token count.
