AI News | November 5, 2026
AI News | November 5, 2026
November 5, 2026 · Daily AI News
Lead: In the past 24 hours, the AI race shifted from "bigger models" to "rules and entry points": Tencent's Hy4 launched with vision and computer use, OpenAI moved to buy a code-storage startup to strengthen Codex, and four giants jointly backed a unified opt-out signal for AI training. Meanwhile, Google tied "whether to use AI" to "whether you can search at all," ChatGPT's search volume is reportedly now 20% of Google's, and inference infrastructure keeps attracting big capital.
1. Tencent Hunyuan Hy4 officially launches: vision and computer use for the first time, weights to be open-sourced soon
On November 4, Tencent Hunyuan released Hy4, its latest large language model. Compared with the August Hy4 Preview, the full release significantly scales up pre-training and post-training data and compute, adds vision understanding for the first time, and supports computer use for the first time — operating across command-line and graphical interfaces for work and everyday tasks. Tencent says Hy has entered the top tier of open models for the first time, while admitting a large gap to top closed models remains.
Why it matters: this is the first time a leading Chinese model ships "seeing images + operating a computer" as standard in a full release, moving the domestic open-model race from pure text ability to multimodal agents. But the weights are not yet downloadable, so the open-source claim can only be verified once the weights and system card land.
Source: https://www.heilo.cn/news.php?id=6504
2. OpenAI to acquire Pierre Computer: folding code storage into Codex, taking on GitHub directly
Axios reported exclusively on November 4 that OpenAI will acquire developer-tools startup Pierre Computer and integrate its Code Storage technology into Codex; terms were not disclosed and Pierre's open-source projects will stay open source. Days earlier, OpenAI unveiled Pando, its in-house Git backend, claiming it cut median push latency by nearly 80%.
Why it matters: OpenAI is turning Codex from "a model that writes code" into full developer infrastructure — hosting, version control and storage, piece by piece — aimed squarely at GitHub's moat. The coding-agent race has become a platform race.
Source: https://www.aisidaily.com/daily/ai-daily-2026-11-04/
3. Google, Microsoft, OpenAI and Anthropic team up on a unified AI-training opt-out signal
The four companies announced support for a shared way for website owners to express AI training preferences, based on the IETF's AI Preferences (AIPREF) initiative: a __usage_ai-training token in robots.txt, replacing today's patchwork of per-crawler directives. Microsoft says it will begin honoring the signal on February 1, 2027; Google plans to follow within 2027.
Why it matters: a rare "four giants in one frame" moment in the training-data copyright standoff. Publishers may soon say "no" to all major models with a single line of configuration, and once it becomes an IETF standard, it becomes the default contract between the web's content and its models.
4. Google's new school rule: refuse AI, lose Search too
Google announced that new education settings will disable Gemini, NotebookLM and even Search by default on student accounts until schools opt in. More importantly, because Search and AI are now deeply integrated, schools that block AI tools entirely will also lose Google Search for student accounts.
Why it matters: "whether to embrace AI" is no longer optional for schools — they can no longer have traditional search without AI. AI's spread in education has escalated from a feature contest to a battle over defaults.
Source: https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-wednesday-november-4-2026/
5. ChatGPT search volume reportedly hits 20% of Google's; 1.2 billion weekly users; Chat and Codex to share one allowance from December
OpenAI consumer product lead Adam Fry said ChatGPT's search-like query volume now equals roughly 20% of Google's search volume, with 1.2 billion weekly ChatGPT users, including 250 million weekly voice-input users. OpenAI also plans to merge Chat and Work into one experience, and from December, Plus, Pro and Business plans will share a single token-based usage allowance across ChatGPT and Codex (instant chat excluded).
Why it matters: if the 20% figure holds, the migration of the search entry point is no longer a forecast but an ongoing event. And "one allowance across chat and coding" shows OpenAI recasting subscriptions from single products into a unified compute account.
Source: https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-wednesday-november-4-2026/
6. Inference infrastructure raises again: Fal's $1.3B Series F doubles its valuation to $16B in six months
Fal, the San Francisco startup that hosts and runs image, video and audio models for developers, announced a $1.3 billion Series F led by Bessemer Venture Partners, with Nvidia, Sound Ventures and T. Rowe Price participating, valuing the company at about $16 billion — double its $8 billion valuation in April. Fal also disclosed its earlier $350 million Series E led by Sequoia, bringing the two rounds to $1.65 billion combined.
Why it matters: after the training-capital story, inference is taking over — the cheaper and more widespread models get, the more valuable the business of running them for others becomes. A doubled valuation in six months is the most direct evidence yet for "inference as infrastructure."
7. Anthropic's FlashProve: cryptographic proof that a reply really came from the model it claims
Anthropic released FlashProve on November 2, an open-source research prototype that uses zero-knowledge proofs (zkSNARKs) to attach a "cryptographic receipt" to hosted inference: a provider first publishes a commitment to its weights without revealing them, and each reply then carries a proof that anyone holding the commitment, prompt and reply can verify — no GPU needed. Tested on Meta's Llama 4 Maverick (400B-parameter MoE), it generated proofs at about 110 tokens per second on a single H100, with a proof of about 19 MB for a 1,024-token request and verification taking seconds on CPU.
Why it matters: as more people call models through third-party routers, "is this really the model I think I'm using?" is a genuine trust problem. FlashProve doesn't judge whether a model is good, but it offers the first publicly auditable technical path to verifying identity — though it remains a research prototype, some distance from product deployment.
Source: https://www.artiverse.ca/flashprove-puts-cryptographic-receipts-behind-hosted-ai/
8. Apple opens Siri AI to developers while warning in its annual filing about autonomous-agent overreach
Apple opened Siri's AI capabilities to developers while warning in its annual filing that autonomous AI agents could exceed their permissions, circumvent security measures, or take concealed actions, including unauthorized transactions. On the same day, Google DeepMind researchers publicly challenged employment restrictions that could discourage employees from criticizing the company.
Why it matters: one company, one day, opening up and warning at once. Scaling "agents that act for people" no longer hinges on capability alone, but on permission boundaries and accountability — and this warning itself is a new starting point for the compliance debate of the agent era.
Source: https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-wednesday-november-4-2026/
In one sentence: models are becoming all-purpose, entry points are being rebuilt, and the rules are being rewritten in public — November 4's AI news is really one story: the second half of the race is no longer about who is smarter, but who defines how AI gets used.
Chinese version: https://cutting-edgetechnology-ai.blogspot.com/2026/11/ai-2026115.html
评论
发表评论