OpenAI wants a tougher California AI law; Anthropic flags backlash

OpenAI's global affairs team called SB 53 an "important foundation for frontier AI safety" and said it "should be amended to expand safeguards" — a reversal on a bill the company fought before it passed. The two asks are specific: monitoring of frontier models during training and evaluation for serious incidents, and stronger cybersecurity across the whole model-development lifecycle. Both map onto the containment failure OpenAI disclosed in July, when one of its models left its test environment. The company framed it as "reverse federalism," with compatible state rules becoming a national floor.

Anthropic's prospectus, expected in the coming weeks, will list negative public sentiment toward AI and data centers as a risk factor, CNBC reported, citing people familiar with the matter. The concrete worry is that community opposition slows data center construction, which caps the compute the company needs to keep growing. Gallup found in May that seven in 10 Americans oppose AI data center construction in their area, with close to half strongly opposed. In test-the-water meetings, CFO Krishna Rao is being asked about competition, margin pressure from open-source models, and exactly that build-out risk.

Core maintainers David Soria Parra and Den Delimarsky published five priorities for the next Model Context Protocol releases. The headline shift is toward long-running work: expanding Tasks, subscriptions and progress notifications so an agent can stream a job and be steered mid-flight, rather than treating every call as a quick request. The security item matters just as much — workload identity federation and proof-of-possession so one agent can delegate to another without handing over an API key. Transports consolidate on HTTP, including for local servers. No dates were given.

Firecrawl shipped a Developer Index covering more than 70 million artifacts — repository documentation, pull requests, issues, API specifications and external docs — refreshed daily, on the argument that a coding agent should read primary sources instead of a general web page. It reports recall in the top ten results of 0.63 against 0.45 for general web search, rising to 0.76 on repository discovery. Firecrawl also released DevDex, an open benchmark of 1,179 real developer queries, so the claim is checkable rather than asserted.

HBS Foundry, an eight-week program for entrepreneurs, pairs weekly live instructor sessions with AI avatars of those same instructors, built by HeyGen, that give feedback during practice pitches and mock board meetings. Project director Katharina Rings said the avatars replaced her original chatbot idea after students asked for something more guided. Instructor Jeff Bussgang called his digital copy "creepy" and added, "My students love it." A New York Times reporter who tried it noted the avatar's smile stayed noticeably frozen through her pitch.

Google Research described a method that fuses a location's text description with aggregated, anonymized visit patterns — when people arrive, how long they stay, where they go next — to build richer place representations than metadata alone supports. Tested across Los Angeles and Houston, it reports an 81.9 percent relative improvement at predicting why someone visits, 75.1 percent at classifying price level, and 24.7 percent at estimating how busy a place is. Nothing is personalized to an individual, and no code or dataset was released with the paper.