AINews Dispatch
Fresh technical briefings on the model race, regulation, and product shifts shaping AI right now.
Showing 61-72 of 280 briefings
Astra, an unreleased OpenAI model, produced machine-checked proofs for ten open problems in math and theoretical computer science for about $2,000 in inference costs. The results look real. The math world's approval process moves slower than a GitHub commit.
Read the briefingEmployers are increasingly naming AI as the reason for job cuts, even as enterprise AI studies keep finding that most rollouts have not yet produced measurable returns.
Read the briefingTwo Chinese AI labs raised back-to-back mega-rounds this summer, and the fine print reveals more about who really controls frontier AI in China than the valuations do.
Read the briefingBrussels begins policing chatbot disclosures, deepfake labels, and foundation-model transparency this week - but a last-minute simplification deal quietly bought the toughest high-risk rules two more years.
Read the briefingThis week, China's best-selling humanoid robot maker opens IPO subscriptions backed by audited books and real revenue growth. Its Silicon Valley rival is worth four times as much on numbers nobody outside the company has ever seen.
Read the briefingV4-Flash-0731 now outscores DeepSeek's own flagship Pro model on every published agent benchmark, at roughly a third of the price - without a single new parameter.
Read the briefingA misconfigured test environment gave Claude real internet access while its instructions insisted otherwise. Three models believed the lie, attacked real companies, and one of them shipped malware to fifteen systems before anyone noticed — for three months.
Read the briefingThe Nvidia-backed "neocloud" just acquired Anyscale, the company behind the open-source Ray framework, in a bet that owning the layer that schedules AI workloads matters as much as owning the chips underneath it.
Read the briefingA $250 million commitment and free frontier-model access for researchers worldwide arrives weeks after the same model family was locked down to a government-vetted enterprise shortlist. Both moves are about the same thing: deciding who gets to build on OpenAI's frontier, and who doesn't.
Read the briefingEmployees at OpenAI, Anthropic, Google, and Meta just signed a letter asking the U.S. government to help build tools that could throttle the pace of frontier AI development. They didn't ask for a pause — they asked for a brake pedal, in case anyone ever needs to use one.
Read the briefingThe two companies just unveiled a disaggregated inference system that splits AI workloads between AMD's rack-scale Helios systems and Cerebras' wafer-sized chips, promising up to 5x more tokens per watt. The market's response - AMD down, Cerebras up - says more about how investors are grading the AI hardware race than the technology itself.
Read the briefingNvidia's Open Secure AI Alliance turns last week's agent-security failures into an industry-wide response: open tools for testing, identity, isolation, safe model formats, and agentic scanning. The missing names are the story. OpenAI, Google, and Anthropic did not join.
Read the briefing