Labs confirm safety talks; voters reject data centers

Chris Lehane, OpenAI's global policy chief, told reporters in Washington on Tuesday that OpenAI, Anthropic and Google DeepMind have been working together on AI safety for several weeks, confirming earlier reporting by Bloomberg. He said the companies do not need the antitrust waiver that Dario Amodei's slowdown essay asked for, and that OpenAI supports a provision in the Frontier Act that would require top labs to admit independent verification organizations. The Information separately reported that the three are building an industry standards body, the watchdog Demis Hassabis first called for in July, and that Sam Altman told staff it would have to proceed without the US government's support.

Asked whether they support building data centers to power AI, 61 percent of the 1,503 likely voters in a New York Times and Siena University poll fielded in early September said they oppose it, and only 14 percent strongly support it. The opposition crosses party lines: 2024 Trump voters split roughly even, while Harris voters and non-voters opposed at 74 and 67 percent. Water and the environment was the most cited reason at 32 percent, ahead of impact on local communities at 21 percent, and among opponents 56 percent wanted limits while 38 percent wanted an outright ban. Neither party holds an edge on the issue, and fewer than one percent of voters named it their top midterm concern, so the backlash is broad rather than decisive.

The September update of Mozilla's State of Open Source AI report puts the best open-weight models about 4.4 months behind the closed frontier on METR's task-horizon data, and roughly three points behind the closed leader on the Artificial Analysis Intelligence Index at a fraction of the price. On a neutral harness from Vals AI, Zhipu's GLM 5.2 landed within a point of Claude Opus 4.7 and 4.8 on Terminal-Bench 2.1 at about a fifth of the cost per completed task. Mozilla's rule is workload-specific: closed models earn their premium on tasks that take a human expert eight to twelve hours, on expert professional work and on long context, while routine work belongs on open models. Eight of the ten highest-volume models on OpenRouter in August shipped open weights, and nearly all of the leading ones are Chinese, which Mozilla's Raffi Krikorian calls "a plurality and a concentration at the same time."

Google DeepMind released Gemini 3.8 Live and a heavier Gemini 3.8 Live Extended Thinking, which it calls its most advanced live dialogue models, across the Gemini API, the Gemini app, Google Workspace and Search. The models run tool calls and API requests in the background while the conversation continues, switch between 97 supported languages mid-conversation, and take near real-time visual input; the Extended Thinking variant reasons and speaks at the same time and narrates its progress on multi-step tasks. Google reports that the Extended Thinking model took the top spot on the Artificial Analysis Speech to Speech Quality Index at 82.6 and scored 68.6 percent on the tau-Voice agentic benchmark, figures that are vendor-reported.

Salesforce announced Koa at Dreamforce, its first reasoning model, built by post-training Nvidia's open-weight Nemotron 3 Super with supervised fine-tuning and reinforcement learning on a synthetic dataset that simulates enterprise workflows across more than 14 industries. Salesforce holds the weights and runs inference inside its own trust boundary, and on its CRM benchmark it says Koa matches or beats leading models on CRM actions with three times fewer errors. Until now Agentforce routed multi-step reasoning to Claude or ChatGPT, and Salesforce's Jayesh Govindarajan told TechCrunch the blocker had been the lack of a state-of-the-art American base model with clear data provenance, adding "we have no idea what Qwen trains on." Koa is in pilots with Formula 1, UChicago Medicine and Xero, with general availability expected this winter in US regions.

Cloudflare shipped a Disallow AI Training setting that publishes a no-training preference in robots.txt while letting mixed-use crawlers, the single bots that feed both a search index and model training, keep indexing a site. Apple, Google and Microsoft have met or committed to meet its "Accountable" bar, which requires a training opt-out, an opt-out from AI summaries and URL-level visibility into what was made available for training, with Microsoft's robots.txt support targeted for early 2027. Amazon, Anthropic, Meta and OpenAI already run separate training crawlers, so those are simply blocked. As of September 15 the plain Block setting stops Applebot, Bingbot and Googlebot outright, new ad-supported domains default to disallowing training, and a control over how much of a page appears in AI summaries is due early next year. Fewer than 1 percent of Cloudflare sites block search bots, but 17 percent already block training.

Meta rolled the Meta One subscriptions it launched in the US in May out globally, with more than 50 features across Instagram, Facebook, WhatsApp and Meta AI and 15 million subscriptions and trials so far. Individuals pay $7.99 for Core or $19.99 for Premium, which bundle the three app subscriptions with more image and video generation from the Muse models and heavier use of tools like Instagram's Restyle; the new creator and business plans run from $14.99 to $499 a month and add a verified badge, a bold follow button, higher search ranking and a WhatsApp business agent. Meta declined to tell TechCrunch what the AI usage limits are, saying they vary by country, surface and system conditions, and says the core Meta AI experience stays free.

Mistral and Mozilla announced a partnership under which Firefox Smart Window, Mozilla's beta AI browsing assistant, runs on Mistral models for users in France and North America, with the United Kingdom and Germany expected later this year. Smart Window summarizes complex searches and recalls pages a user clicked away from by reading their open tabs; conversations are not saved on Mozilla's servers by default and Mistral has agreed to zero data retention. Mistral says it is fine-tuning on regional languages and dialects so answers carry local nuance, and both companies frame the deal as open distribution for open models, the argument Mozilla made the same day in its open-source AI report.