Daily AI Briefing
9 stories, ranked by importance. Generated at 07:01.
At a press event on Monday, OpenAI consumer product lead Adam Fry said queries that trigger web search in ChatGPT now equal roughly one-fifth of Google's search volume, using OpenAI's definition of a search query. Fry also said ChatGPT reached 1.2 billion weekly active users after adding about 200 million in two months, underscoring how quickly AI-native search is scaling versus traditional engines.
The VergeChina's National Development and Reform Commission said it is accelerating legislation to govern AI across the full model lifecycle, from data and compute to deployment. Senior official Wang Ruomeng said the rules aim to guard against loss of control over increasingly powerful systems, covering major domestic developers.
Bloomberg LawGoogle said Gemini, Gemini Notebook, Gemini in Classroom, and Google Search will be off by default on K-12 student accounts until school administrators explicitly enable them. The company also said Search and AI are deeply integrated, so schools that refuse AI tools will not offer student Search access, while rolling out Guided Learning in Search and Classroom to coach students instead of giving direct answers.
Google BlogAt its inaugural Forge conference, Fireworks unveiled Ultrafast inference in public preview on NVIDIA Blackwell GPUs, citing speeds above 1,500 tokens per second on DeepSeek V4.1 Flash in demos, plus multi-day prompt caching and background mode for agents. It also announced expanded European capacity on NVIDIA Vera Rubin and Grace Blackwell platforms beginning in 2027, with CEO Jensen Huang joining the keynote.
Fireworks AI BlogTraversal, which builds AI agents for enterprise site reliability engineering, announced $120 million in new funding at a $1 billion valuation. Kleiner Perkins, Redpoint, and Khosla co-led the round, with existing backers Sequoia and others increasing stakes, bringing total funding to $173 million.
Business WireHungary is seeking EU member support for the EU Frontier AI Safety and Security Team (EU FAST), which would monitor AI models during development before they reach the market. MLex reports the initiative is framed as filling a gap after recent security incidents involving major U.S. labs.
MLexResearchers including Microsoft affiliates posted WebHarbor on arXiv as an environment that embeds live websites to train and evaluate GUI agents on realistic, changing interfaces. The work targets a gap between static simulators and production web UIs that shift over time.
arXivFlower Labs introduced Endeavor 1.1, positioning it near leading closed models with stronger reasoning and coding scores versus Endeavor 1.0 on benchmarks such as GPQA and HumanEval. The release adds EU-hosted inference, image inputs, readable thinking traces, and private deployment options for enterprises.
Flower Labs BlogAnthropic detailed FlashProve, a research system that uses zkSNARK-style proofs so users can verify a reply came from a committed model without exposing weights. The company reported proof generation around 110 tokens per second on Llama 4 Maverick in its prototype and released code under Apache 2.0, with no production deployment planned yet.
TechCrunchThe clearest competitive signal of the day was OpenAI's claim that ChatGPT search traffic has reached roughly 20% of Google's, alongside 1.2 billion weekly users, reinforcing that AI chat is becoming a parallel discovery layer. Policy moved in parallel: China outlined lifecycle-wide AI rules, the EU debated pre-market frontier-model testing via Hungary's EU FAST proposal, and Google reshaped U.S. K-12 access by linking student Search to administrator choices on AI tools. On infrastructure and tooling, Fireworks' Forge event highlighted NVIDIA-backed ultrafast inference and regional expansion, while Traversal's billion-dollar valuation showed continued investor appetite for production-grade AI agents. Trust and evaluation also advanced through Anthropic's FlashProve research and new agent benchmarks like WebHarbor, even as regulators and educators tightened controls over where AI is allowed to operate.