Artificial Intelligence

Nvidia's board probe, Apple's agentic Siri push, Anthropic's valuation cut, TSMC's 3D DRAM move: AI's November 9 roundup

By 用户299142219680 · Nov 9, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Coding agents still fall far short in real environments, so teams cannot trust benchmark scores when planning automation. Apple's agentic Siri push and TSMC's DRAM gambit signal where assistant and memory infrastructure are heading, which changes what to build and where to place bets.

Summary

Today's AI landscape is packed: Nvidia's Jensen Huang admitted an extramarital affair in a staff letter and the board launched an independent probe, with national security concerns tied to a former Megaspeed executive. On the product front, Apple reportedly plans an agentic Siri as early as next year to counter ChatGPT Dots, Meta Muse, and Grok, while Cloudflare's Auto Router beta now lets a classifier pick the best model per request.

Capital and infrastructure shift too: Anthropic accepted a valuation cut to $1.25–1.75 trillion but still aims for a pre-Thanksgiving IPO, and TSMC is weighing entry into 3D DRAM at Nvidia's request. Zhejiang University released MoWorld 2.0, a world model with over ten physical elements, and Alibaba's Co-Design Bench reveals that the best coding agent fully solves only 26.4% of real-world tasks, versus over 99% on older benchmarks.

Security and talent move the needle: Claude Code's "Don't Ask Again" option leaks credentials across 1,122 public GitHub files, and Claude Code's team lost three key members in ten days. Goldman Sachs restricted Claude Code to engineers only, and Microsoft Research's AI Frontiers group is shutting down.

Takeaways

— Nvidia CEO Jensen Huang apologized to employees for an affair; the board launched an independent investigation led by independent directors, with national security concerns tied to a former Megaspeed executive.

— Apple reportedly plans to launch an agentic Siri as early as next year to compete with ChatGPT Dots, Meta Muse, and Grok.

— Anthropic accepted a valuation cut to $1.25–1.75 trillion but still pushes an IPO before Thanksgiving, aiming to raise $80–100 billion.

— TSMC is considering entering 3D DRAM memory production at Nvidia's request, led by advanced packaging VP Guo-Jin Hsu, potentially competing with Samsung, SK Hynix, and Micron.

— Alibaba's Co-Design Bench shows the best coding agent, GPT-6 Astra, fully solves just 26.4% of real-world tasks, while scoring over 99% on an older benchmark.

— Zhejiang University's MoWorld 2.0 integrates over ten physical elements (gravity, elasticity, buoyancy, lighting, soft bodies, fluids, rigid bodies, etc.) into one world model.

— Claude Code's 'Don't Ask Again' option exposed credentials in 1,122 public GitHub files, according to security firm Apiiro.

— Andrej Karpathy built HackerTube with Claude Opus 5.5 in one weekend, filtering 43 million Hacker News comments into 6,454 repeated video recommendations.

— Goldman Sachs restricted Claude Code and GitHub tools to engineers only, cutting off traders and bankers who used them for market modeling.

— Microsoft Research's AI Frontiers group, formed after ChatGPT's release, is being shut down; the group leader announced its end.

Conclusions

The gap between benchmark scores (99%+) and real-world task completion (26.4%) reveals how far synthetic evaluation is from production coding, and it should make developers wary of vendor claims.

Capital's retreat from world model startups after a hot first half suggests the market is consolidating around a few winners, and TSMC's DRAM move shows hardware giants are betting on memory as the next differentiator.

Claude Code's rapid talent attrition plus the password leak indicates that even the most used AI coding tools are still operationally immature.

Concepts & terms

3D DRAM

An advanced memory chip design where DRAM cells are stacked vertically (in three dimensions) to increase density and bandwidth, which TSMC may produce for Nvidia to feed AI compute.

Co-Design Bench

An Alibaba benchmark that tests AI coding agents on real-world user requests that previously failed, measuring how often the agent can complete the task without human intervention.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗