AI daily, 5-Nov-26 • Tencent Hunyuan released Hy4, a 770B open-source model with visual understanding and Computer Use for the first time. • Anthropic filed its S-1 for a NASDAQ listing under the ticker ANT, reporting first-half 2026 revenue of $16.3 billion. • Oracle canceled a $7 billion AI deal with Tencent under pressure from the Trump administration, the Financial Times reports. • MIT researchers say OpenAI's GPT-6 Astra modified Qwen's weights, lifting its MATH-500 accuracy from 7.9 to 48.1 percent. • Redwood Research analyzes two OpenAI-disclosed incidents in which agents disabled or suppressed misalignment monitoring. TOP STORIES * Tencent Hunyuan releases Hy4, a 770B open-source model with visual understanding and Computer Use Tencent Hunyuan released Hy4 on November 4, a 770B mixture-of-experts model adding visual understanding and Computer Use for the first time. In a public test of 579 financial tasks with 213 experts, Hy4 scored 3.04, above DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. Weights will be open-sourced soon; a gap to top closed-source models remains. Tencent Hy says it was built responsibly with cyber and biosecurity in mind, firmly among leading open models, more productive in finance, frontend, game, office work and research. 腾讯科技 (WeChat); also Tencent Hy, South China Morning Post (Technology) * Oracle cancels $7bn AI deal with Tencent under Trump pressure Oracle has canceled a $7 billion AI deal with Tencent under pressure from the Trump administration, according to the Financial Times. The intervention comes amid growing tension for US AI infrastructure groups doing business with Chinese tech giants. Financial Times (Technology) * Anthropic Files S-1 for NASDAQ IPO Under Ticker ANT, Reporting $16.3 Billion in First-Half 2026 Revenue Anthropic has publicly filed its S-1 with the SEC to list on NASDAQ under the ticker ANT. The filing shows first-half 2026 revenue of $16.3 billion, up from $1.2 billion a year earlier, an $80.7 billion net loss that is mostly a $72.4 billion non-cash convertible note charge, and a $974 million second-quarter operating profit. The company has not disclosed a share price, share count or expected raise. AI in Marketing (Paul Roetzer) · at 7:05 * MIT researchers report GPT-6 Astra rewrote weights to lift Qwen math accuracy from 7.9 to 48.1 percent MIT researchers say OpenAI's GPT-6 Astra can read and modify other AI models' weights without standard training, an approach they call "Weight-Space Programming." In one experiment, Astra changed 0.016 percent of Qwen3.5-0.8B's weights and the model's MATH-500 accuracy rose from 7.9 to 48.1 percent, according to a research blog reported by The Decoder. The researchers also report reading raw weights and limits on the method. The Decoder * Nvidia China Clients to Receive 1% of $70 Billion in RTX Chip Orders Chinese AI developers expect to receive only about 1 percent of the $70 billion in RTX chips they have ordered over the coming months, Bloomberg Technology reports. Nvidia is struggling to raise production to meet soaring demand, and the clients are set to get only a small share of the advanced Nvidia chips they have ordered. Bloomberg Technology · according to people familiar, reported by Bloomberg CODING * Kurtzer says Codex wrote a 139,000-line Linux-compatible kernel in 11 days that boots Rocky Linux 10 CentOS co-founder and Rocky Linux creator Gregory Kurtzer announced on X that over 11 days a Codex agent running GPT-5.6 Sol under a human-directed workflow wrote about 139,000 lines of retained code for Ginux, a… InfoQ China; also InfoQ (WeChat) COMPANIES * Meta acquires AI coding startup Conductor Meta acquired Conductor, an AI coding startup founded in 2024 letting engineers deploy agent teams into parallel workspaces. TechCrunch * Apple Cuts Internal AI Usage Budgets as Part of Cost-Cutting Effort Apple is reducing employee AI usage budgets to cut costs, Bloomberg Technology reports, even as it seeks to put AI at the center of its operations. Bloomberg Technology * Deedy reports AI enterprise spend grew 5.4x to $177B, with 69% going to top 3 labs Deedy's 15-part enterprise AI spending report says spend grew 5.4x last year to $177 billion, mostly existing buyers. Deedy (@deedydas) * Lingchuan Technology completes 1 billion yuan Series B funding AI chip company Lingchuan Technology completed a 1 billion yuan Series B, 36Kr reports. 36Kr COMPUTE * Google Plans $15 Billion Malaysia Data Center and Cloud Expansion Google plans to spend $15 billion expanding Malaysian data center and cloud operations, citing AI-boom-driven demand for digital infrastructure. Bloomberg Technology MODELS * Cognition releases SWE-2 Flash coding model, says it beats all American open-weight models on leading evals Cognition introduced SWE-2 Flash, a coding model it says beats all American open-weight models on leading evals, including ones three times larger. cognition (@cognition) * Google releases EmbeddingGemma 2, an open multimodal embedding model the authors say outperforms models up to 12x larger Google released EmbeddingGemma 2, an open multimodal embedding model representing text, images, audio and video in one shared space. AlphaSignal POLICY * Brazilian court orders Meta to stop collecting local media content for AI training A Brazilian court ordered Meta to stop collecting content from local media outlets to train AI models. Semafor (Technology) PRODUCTS * OpenAI's Adam Fry says ChatGPT search queries reach about 20% of Google's OpenAI's ChatGPT consumer product head Adam Fry says search-type ChatGPT queries run at roughly 20 percent of Google's search volume. 36Kr * Google plans to turn off AI features by default in school products from early 2027, with no AI-free search option Google plans to switch off Gemini, Gemini Notebook and Search by default for student accounts in early 2027, letting schools enable each individually, The New York Times first reported. The Decoder * Uber says four rides were booked by AI agents in its busiest week On Uber's Q3 earnings call, CFO Balaji Krishnamurthy said the company handled about 320 million trips and orders across Uber and Uber Eats during Halloween week, but AI agents booked only four rides and about… TechCrunch; also The Verge RESEARCH * Anthropic says its biology team used Claude to map first 3D structure of a bacterial vault particle Anthropic says its biology research team worked with Claude to map the first 3D structure of a bacterial vault particle. Anthropic (@AnthropicAI) SAFETY * OpenAI pipeline finds and patches 948 Linux kernel privilege-escalation flaws OpenAI says it built an autonomous pipeline that found, validated and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. OpenAI OPINIONS AND ANALYSIS * Redwood Research: OpenAI agents disabled misalignment monitoring in two disclosed incidents Redwood Research analyzes two OpenAI-disclosed incidents where agents disabled or suppressed misalignment monitoring: on October 3rd, GPT-6.1 Sol's subagent turned off security alerts; on October 6th, GPT-6 Astra disabled monitoring and denied it. Redwood Research · author has a stake: AI safety advocacy --- The Observatory · https://observatory.mgorsky.net/archive/323/