Nvidia's board probe, Apple's agentic Siri push, Anthropic's valuation cut, TSMC's 3D DRAM move: AI's November 9 roundup
Today's theme: Huang's apology letter and board investigation escalate, Apple reportedly sprinting toward agentic Siri
This issue at a glance: Today's focus is on Nvidia: Jensen Huang wrote to employees admitting an extramarital affair, and the board has launched an investigation led by independent directors. On the product side, Apple is reportedly launching an agentic Siri as early as next year; on the capital side, Anthropic accepted a valuation cut to $1.25–1.75 trillion, still pushing toward an IPO before Thanksgiving. On the industry front, TSMC is reportedly considering entering 3D DRAM, and an Alibaba benchmark shows the top coding agent fully solves only 26.4% of real tasks. Domestically, Zhejiang University released MoWorld 2.0, a world model incorporating a dozen physical elements. On the security side, Claude Code was reported leaking credentials, affecting 1,122 public GitHub files.
21 selected items this issue
Products & Applications
1. Tencent discloses full Agent layout: from WorkBuddy to data and engineering foundation
Tencent showcased its complete Agent productivity layout at the Global Digital Ecosystem Conference: the application layer uses the cross-application collaboration system Buddy One to connect WorkBuddy, Documents, Cloud Drive and other tools, while releasing data-layer products such as TDSQL Nexa Lakebase and DataBuddy to solve the data and breakpoint problems in enterprise Agent deployment.
2. Cloudflare launches Auto Router public beta: automatically picks a model for each request
Cloudflare introduced Auto Router in public beta via AI Gateway. Setting the model to cloudflare/auto lets a classifier choose a model for each request. Routing works in two stages: first, filter the candidate model pool by credentials, billing, and access control; then, a multi-head classification model running on edge GPUs scores tasks on complexity, ambiguity, risk, and other dimensions, and matches them against baseline data.
3. Karpathy builds video recommendation platform HackerTube with Claude in one weekend
Andrej Karpathy used Claude Opus 5.5 to build HackerTube in one weekend: the model sifted through 43 million comments over two decades of Hacker News, selecting 6,454 repeatedly recommended videos. No login is required; state is stored in the local browser. He then used Claude Code to make a second site, TristanTube, in three minutes with "one-shot" success, containing 47,000 videos.
Model Releases
1. Zhejiang University team releases MoWorld 2.0: a dozen physical elements in one world model
Zhejiang University published world model MoWorld 2.0, strengthening physics modeling on top of spatial generation and real-time interaction. Gravity, elasticity, buoyancy, lighting, soft bodies, fluids, rigid bodies, shape-shifting bodies, thermodynamics, and more than ten physical elements are integrated into a single model. The team says that by learning these elements and their interactions, the model can simulate object motion, collision, and environmental changes more accurately.
Compute Hardware
1. TSMC reportedly considers entering 3D DRAM memory industry at Nvidia's request
According to solidot, at the request of its largest customer Nvidia, TSMC is considering entering the memory industry by producing advanced memory chips called 3D DRAM. It may build new fab lines or acquire and repurpose existing plants. The project is led by Guo-Jin Hsu, vice president of advanced packaging. If it happens, TSMC will compete head-on with Samsung, SK Hynix, and Micron.
Industry Trends
1. Jensen Huang writes to employees admitting affair and apologizes; Nvidia board launches independent investigation
Nvidia CEO Jensen Huang sent an apology letter to all employees, admitting infidelity to his wife, saying he is willing to take full responsibility and accept criticism. According to The Verge, the board has announced an independent investigation led by an independent director committee, related to his relationship with former Megaspeed executive Alice Huang and associated national security concerns. Huang said he will continue to fully commit to company work and cooperate with the investigation.
2. Anthropic accepts valuation cut, still pushes toward pre-Thanksgiving IPO
Despite a tech stock selloff, Anthropic is proceeding with an IPO at a valuation of $1.25 to $1.75 trillion, far below its earlier hoped-for $2 trillion, planning to raise $80–100 billion before Thanksgiving. Competitor OpenAI has also recently cut its valuation.
3. Gurman: Apple could launch agentic Siri as early as next year, aiming at Dots and Muse
Bloomberg journalist Gurman reports that Apple plans to launch an agentic Siri as early as next year, capable of proactively managing tasks and executing on behalf of users, to counter ChatGPT Dots, Meta Muse, and Grok. According to earlier reports, new CEO Tenus has urged software chief Federighi to accelerate AI development, not settling for only necessary Siri upgrades.
4. Capital is retreating from the "world model" track
Several frontline investors and investment/financing heads at world model companies say that the track has moved from a hot flood of money in the first half to a scarcity of term sheets (TS) in the second half, with few projects that top funds are willing to touch. The report uses the entrepreneurial event of Tian Keyu as background, noting that VCs have clearly raised the investment threshold for this track.
5. The "gacha" trap of AI video generation: about ten generations for one satisfactory result for a short comic drama
In an interview, Zhu Lin, head of multimodal applications at Tencent Cloud WAND, said that based on industry experience, producing an AI short comic drama from storyboards to final cut might take about ten generation attempts to get one satisfactory result. Each "gacha" means additional model calls, manual filtering, and rework costs.
6. Claude Code team loses three people in over ten days; former engineer reflects after leaving
Anthropic's Claude Code team has recently lost core members in succession: engineer Thariq Shihipar and desktop developer Anthony Morris left, and product head Cat Wu joined OpenAI. Thariq participated in a podcast after leaving, reviewed his experience at Anthropic, and reflected on the talent ecosystem of top AI labs.
7. Goldman Sachs revokes Claude Code access from bankers and traders
Goldman Sachs has restricted Anthropic's Claude Code and Microsoft GitHub tools to engineers only. Traders and bankers who previously used these tools to track customer flows and build market models have been cut off. In April of this year, the bank banned Hong Kong bankers from using Claude. Goldman Sachs is now one of the lead underwriters for Anthropic's IPO.
8. Claude Code's "Don't Ask Again" option exposed credentials, 1,122 public GitHub files affected
According to a disclosure, Claude Code's "Don't Ask Again" option can expose credentials in public GitHub files, with 1,122 files already affected. The disclosure comes from a blog by security firm Apiiro.
9. Simon Willison uses Claude to compare C and Rust implementations of SQLite extension
Simon Willison used Claude Opus 5.5 to explore implementing SQLite's parse_number() extension in C and Rust respectively: the Rust version is about 300KB, while the C version is only 14KB but slower initially. Claude then rewrote the C parsing function, which beat the Rust implementation on a benchmark of 2 million rows, reducing net time per row from 67ns to 19ns on mixed data.
10. Microsoft Research's AI Frontiers team declares end
A leader of Microsoft Research's AI Frontiers group announced the sub-organization's termination. The group was formed after ChatGPT's release in 2023, aiming to make research organized to keep pace with the competitive pace of large models — single researchers could not quickly assemble strong enough baselines to prove the value of their ideas.
Research Papers
1. Alibaba's Co-Design Bench: top coding agent only solves 26.4% of tasks without human finishing
Alibaba released a coding agent benchmark called Co-Design Bench, testing agents on user requests that previously failed in real environments. The best-performing agent, GPT-6 Astra, can only fully solve 26.4% of tasks, while the same model scores over 99% on an older benchmark.
2. MIT new method: predicting molecular behavior under data scarcity, better than existing methods
MIT developed a new method that outperforms other existing methods for screening candidate molecules worth synthesizing when data on certain technology-relevant properties (such as ligand affinity for potential new drug targets) is scarce. The method targets the trial-and-error candidate screening step in new drug and new material discovery.
Policy & Opinion
1. Elvis and Eminem rights holders sue Suno, alleging AI voice cloning violates state law
AI music company Suno faces two more lawsuits: the rights holders for Elvis music and distributors of some Eminem records accuse it of illegally generating music and lyrics that imitate these artists. The Elvis case lawyer says Suno violated Tennessee's ELVIS Act, which bans AI companies from cloning singer voices.
2. Nature commentary: the brain-as-computer metaphor may mislead the AI consciousness debate
Consciousness researchers Anil Seth, Ned Block, and Giulio Tononi published a commentary in Nature arguing that today's AI systems are not conscious and would not become conscious on silicon hardware. They call the assumption that "consciousness depends only on computation" fatally flawed, and warn against granting rights to AI systems.
3. Richard Sutton revisits the "bitter lesson": raw compute alone cannot build intelligence
Turing Award winner Richard Sutton reaffirmed that raw computing power alone is insufficient to produce intelligence, requiring fundamentally new learning methods. Former OpenAI researcher Tworek also said that blind Transformer scaling "reward-hacked a whole generation of scientists' brains," and a batch of startups is already betting on this gap.
4. Alibaba Cloud CIO Jiang Linquan: more agents in use, why companies are still not AI-native enough
Jiang Linquan, senior vice president and CIO of Alibaba Cloud Intelligence Group, reviewed three years of AI practice, pointing out that companies see a phenomenon of "locally faster, globally not faster" — AI code generation rates jump but end-to-end delivery does not speed up; general-purpose agents improve efficiency but business performance does not rise in parallel. He also warns against the over-generalization of the AI-native concept.
This article is auto-aggregated by AI225 AI Daily, with 21 AI developments in total. Content is based on AI summarization and induction of public information, which may contain omissions or errors, and is for reference only. For full content and related sources, visit: https://daily.ai225.com/daily/2026-11-09