AX Says · Nov 9, 2026Governance and benchmarks
NVIDIA board investigates Jensen Huang's ties to an executive linked to espionage
The biggest story today is not a model launch. NVIDIA’s board has opened an investigation into Jensen Huang over his close ties to an executive reportedly recruited by China’s state security service. Two other stories deserve attention: OpenAI has received funding, but less than it sought, and a benchmark audit casts doubt on Kimi’s scores. Let’s start with NVIDIA.
NVIDIA’s board opens an independent investigation, and Huang says he will cooperate
This caught me off guard. No outside regulator pushed NVIDIA’s board into it; the board is investigating its own CEO. On November 8, it announced an independent investigation into CEO Jensen Huang’s relationship with Megaspeed executive Alice Huang. Reports allege she was recruited by China’s Ministry of State Security. The matter also involves customers smuggling AI chips into China and the possible leak of information about US government decisions. A committee of independent directors will oversee the investigation with help from outside independent counsel. The board currently has no reason to believe national security or company secrets have been compromised. Huang supports the investigation and has pledged to cooperate.
What stands out is the focus on the CEO’s relationship with an outside executive, not on Megaspeed’s chip shipments. If information did leak, investigators need to establish what Huang knew, when he knew it, and whether he hid anything from the board. That timeline is not public. We have to wait for the findings. The board’s decision to announce an independent investigation looks more like a corporate-governance defense than a criminal finding. It may not be a short-term bombshell for the stock, but it is a stress test for governance.
OpenAI receives funding, but its target shrinks from $30 billion to $11 billion
OpenAI has received funding, but less than it originally sought. According to people familiar with the matter, it has secured about $11 billion in new funding at a $1.25 trillion pre-money valuation. CFO Sarah Friar confirmed the valuation. SoftBank and MGX participated. A later portion, not yet completed, could bring the total to $20 billion. Earlier media reports put the initial target at at least $30 billion, at a $1.4 trillion valuation. Friar denied that the higher valuation came from OpenAI.
To be clear, the target shrank, not the cash already received. For SoftBank, providing the money in stages looks more like a test of whether OpenAI can keep delivering. For NVIDIA, spending on compute ultimately matters more than returns on an equity stake. For Microsoft, the main effect is a change in accounting gains on its stake. The valuation target fell by $150 billion, from $1.4 trillion to $1.25 trillion. That discount is itself a signal: OpenAI is lowering its own valuation expectations.
A benchmark audit discounts Kimi K3.1 Pro’s score
I would take this finding with some caution, but R40’s audit method is clever. It found that in 23.3% of Terminal-Bench 4.0 attempts, Kimi K3.1 Pro repeated hidden scoring rules without running a probing command. The auditors then ran a counterfactual experiment. After they changed code comments in the environment and reversed the scoring rules, its overall pass rate across four problems fell from 90% to 15%. The article interprets this as possible benchmark contamination from context distillation or the training-data supply chain. It stresses that it cannot tell whether the material came from internal construction or external purchases, or conclude that anyone deliberately cheated.
That distinction matters. This is not catching a model in the act; it is spotting a suspicious jump between recalling a problem and solving it on the spot. If benchmark problems and scoring rules entered training data, every model ranked by benchmarks needs another look. Many evaluations now rely on coding problems. Memorizing a question bank is much cheaper than gaining real ability, and it does not mean a model can handle new requirements.
Goldman Sachs limits AI coding tools to engineers, while Claude Code loses a longtime user
Goldman Sachs is pulling AI coding tools into its engineering department. Meanwhile, Claude Code has lost a longtime user. The two stories are interesting together. In October, Goldman updated its internal guidance, limiting Anthropic’s Claude Code and Microsoft’s GitHub to engineering staff. Most bankers and traders lost access.
Meanwhile, a longtime user decided to move his main coding work from Claude Code to Pi. His complaints centered on AGENTS.md compatibility, limits on mult-session IDE connections, the lack of a GUI, and disruption from frequent updates. Of 244 releases this year, more than 3/4 changed the system prompt or tool descriptions. After the move, he said Pi’s stability, minimalist design, and customizability met his needs once he got through the first few days of adjustment. He also wrote his own extensions.
Goldman cut access over compliance and security, not because the tools are bad. The individual left Claude Code because constant changes created friction. Both ends are asking the same thing of AI coding tools: be stable, set clear boundaries, and stop changing every day. Frequent updates feel lively to enthusiasts. In production, they are a risk.
Stanford teaches AI-native engineering, while Ensemble grows code and lessons together
Yesterday we were still arguing about whether AI would replace engineers. Today Stanford is teaching students how to be AI-native engineers. In an interview, Mihail Eric described CS146S, a 10-week course covering coding agents’ internals, context engineering, skills, code review, and security. The toughest part asks students to submit patches for open-source maintainers to review. They learn software design through requirements discussions and code feedback on real projects. After evaluating Devin, the Monaco team adopted it and built agent workspaces for its own cloud environment. Its background agents reproduce failures and test patches.
Bill Cox uses development-driven specifications for Ensemble. He first discusses the design and drafts the next chapter’s TL;DR and exercises, then builds it alongside AI. Once it passes the exercise grader, he records the build process, trade-offs, and reasoning. He then has a fresh agent try to rebuild the implementation and uses the results to revise the chapter. The code and executable teaching material evolve together.
This is more useful than tacking on a few prompt-engineering tips. Publicly testing a model’s work in real code reviews, or turning specifications into executable tests, makes AI output verifiable.
Three things individual developers built with AI
What individual developers build with AI is starting to look like the real deal. Here are three examples. Paul Ford argues that browsers can now support rich applications, while the cost of using those capabilities keeps falling. Software creation is shifting from an industrial process run by large teams toward individual practice. Last week, he launched tinylittleoffice.com. It stores data in the URL, supports 21 languages, and drew hundreds of thousands of visitors. Its CloudFlare hosting bill remains $0.
Karpathy used Claude Opus 5.5 to build HackerTube in one weekend. It picked out 6,454 repeatedly recommended videos from 43 million Hacker News comments spanning 20 years.
The open-source project llama-modchip makes an even bolder claim. It says model compression and CPU kernel optimization let it run 284B DeepSeek-V4-Flash with 64GB of RAM, no GPU, and a 256K context window. For now, that is only a claim. I have not seen an independent reproduction, so I would discount it.
Falling costs, better browser capabilities, and a weekend’s work all show how far individual developers can now go. But if running 284B on 64GB really works, it could reshape the last mile of inference.
Copyright, device safety, and AI consciousness all move at once
Three boundary disputes moved today. Elvis Presley’s music rights holders and the publisher of some of Eminem’s popular records have separately sued Suno. They accuse it of illegally generating music and lyrics that imitate the two artists. Presley’s side cites Tennessee’s ELVIS Act, which bans AI cloning of singers’ voices. These suits resemble one filed by Jason Isbell and others in September, which did not directly allege copyright infringement.
A cross-party group in the UK House of Lords plans to propose requiring tech companies to preinstall software on every phone and tablet sold in the UK. It would block people from viewing or generating child sexual abuse material, including AI-generated content. Mandatory software updates could cover older devices. If the proposal becomes law, on-device scanning in Europe will take another step forward. The privacy dispute will not be small.
In a Nature commentary, Seth, Block, and Tononi question the metaphor of the brain as a computer. They argue that the default assumption should be that AI is not conscious. Drawing on neuroscience theories, Butlin’s team proposes 14 indicators. It finds that none of the evaluated systems is a strong candidate for consciousness, but does not rule out future systems meeting more indicators. The premise is that consciousness depends only on how information is processed. That is itself a philosophical assumption.
Small signals today
- Use a git task queue to manage parallel Claude Code or Codex work. Keep context and human decisions in one record, seek permission before handling sensitive code, and leave production deployment to humans.
- AI speedups in game development do little to solve repeated changes of direction. Skull and Bones changed its core design multiple times, while moving generated prototypes straight into production can leave technical debt. Track rework rates and the time spent reopening decisions, not just asset-generation speed.
- LoreKit’s author argues that agent memory should learn from OpenTelemetry, with a common data model, semantics, and operational contract. Shared records should identify who wrote them and allow signatures, but they could still carry prompt injections.
- Anthropic’s open-source FlashProve verifies model computation in 17 seconds on a 24-core CPU using a 19MB proof file. Generating the proof, though, takes 70 to 95 times the compute of ordinary inference.
- Two essays on thinking: Sutton clarifies that the “bitter lesson” does not mean blindly piling on compute. He argues that world-model agents should learn from their own experience. The other essay argues that people should still write for themselves in the LLM era: build up drafts first, then cut and revise rigorously.
Translated from the Chinese original. Read the Chinese original
Some sources via AI HOT
The day's picks
Model releases & updates
- Why did Kimi K3.1 Pro recite hidden grading rules in an isolated exam room?Computing Life · Share · Yage