AI Brief, November 9: Nvidia investigates Huang, Goldman pulls Claude Code, agentic Siri

Nvidia's board investigates Jensen Huang, Goldman Sachs restricts Claude Code to engineers, Apple speeds up an agentic Siri, and Suno faces new suits.

Nvidia's board has opened an independent investigation into CEO Jensen Huang over a relationship with an executive US officials believe was spying for China, putting leadership uncertainty at the center of the AI compute supply chain. Elsewhere, Goldman Sachs reportedly pulled Claude Code from traders and bankers, Apple is pushing Siri toward agent behavior, and Suno faces two more lawsuits.

Nvidia board opens independent investigation into CEO Jensen Huang

Nvidia's board said Sunday it is starting an independent investigation, overseen by a committee of independent directors with outside counsel, after reports that Jensen Huang was the target of a foreign intelligence effort. The Verge reports the inquiry concerns Huang's relationship with Megaspeed executive Alice Huang, who was reportedly recruited by China's Ministry of State Security; Huang sent a memo to the company apologizing for being unfaithful to his wife. The board said it has no reason to believe national security or Nvidia secrets were compromised and expressed confidence in Huang. TechCrunch reports Yale's Jeffrey Sonnenfeld called Huang's career salvageable but put the odds at "about a 50-50," and that the Chinese Embassy called the allegations "groundless smears and slanders."

Why it matters: Nvidia sets the roadmap and supply for most AI compute, so uncertainty at the top is a planning risk for anyone whose capacity or hardware timeline depends on it.

Sources: The Verge AI / TechCrunch AI / TechCrunch AI

Goldman Sachs restricts Claude Code and GitHub to its engineers

Internal guidance updated in October restricts Anthropic's Claude Code and Microsoft's GitHub to Goldman Sachs engineers, cutting off traders and bankers, Bloomberg News reported according to The State of AI. The report says the change followed a review of productivity and risk concerns; strats retain access, and Goldman's in-house AI assistant remains available firmwide. Traders had used the tools to monitor client flows, model market moves and positions, and build materials. In July, CIO Marco Argenti had talked up "vibe coding" by non-technical staff, and the report notes Goldman and Morgan Stanley are lead underwriters on Anthropic's IPO per the S-1 filed October 26.

Why it matters: It is a concrete signal that regulated firms will gate agents that write and run code by role, so builders selling into those environments need review ownership and information barriers built in.

Sources: thestateofai.com / Hacker News discussion

Gurman: Apple pushes Siri toward agentic tasks and third-party app control

The Verge, citing Bloomberg's Mark Gurman, reports CEO John Ternus is pressuring AI czar Craig Federighi to accelerate the Siri roadmap and deliver an agent that can compete with OpenAI's Dots or Meta's Muse. Fine-grained control over Apple's first-party apps, such as rotating and cropping images in Photos and organizing Safari tabs, is expected in iOS 27.3, with third-party app integration likely in the spring. Gurman says Apple wants Siri to pursue ongoing goals rather than one-off commands and believes it could become more agentic as early as next year.

Why it matters: If Siri gains third-party app control on device, agent builders targeting iOS will be competing with, or routing through, a system-level assistant.

Sources: The Verge AI

Elvis and Eminem rights holders file new lawsuits against Suno

The Verge reports that rights holders of Elvis Presley's music and the publisher of some of Eminem's biggest records have sued Suno, claiming it illegally generates music and lyrics that mimic those artists. The Presley suit alleges violations of Tennessee's ELVIS Act, which makes it illegal for an AI company to clone a singer's voice. Suno spokesperson Cat McCormack said the company exists to help people create original music, does not allow prompts for specific artists' names and songs, and screens uploaded audio and lyrics with third-party providers. Suno already faces suits from Sony, UMG, Jason Isbell, Reservoir Media and others.

Why it matters: State likeness laws, not just copyright, are becoming the main exposure for generative audio products, which affects how voice and style features are designed and guarded.

Sources: The Verge AI

SLA adds spatial alignment to VLA pre-training without changing inference

A new paper introduces SLA, a spatial-language-action model initialized from a pretrained pi05 policy that jointly optimizes action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. The authors curate and unify heterogeneous robot datasets, and use feature-dependent gates to aggregate visual embeddings from multiple VLA layers for spatial supervision plus probabilistic state dropout to reduce overfitting to demonstration trajectories. The teacher model uses partially available depth and camera calibration alongside RGB. All auxiliary components are training-only, so pretrained VLA weights can be reused and the original inference architecture is preserved.

Why it matters: Robotics teams can try spatial supervision during continual pre-training without rewriting their deployment stack or paying extra inference cost.

Sources: Hugging Face Daily Papers

DPDM uses a frozen shared predictor to lift agent decision accuracy

Decoupled Predictive Decision Making (DPDM) pairs frozen language-model representations of current and historical records with a shared pretrained conditional predictor that maps evidence and queries to task-defined targets, with no parameter updates at deployment. The authors use a single frozen LimiX-2 predictor across five task categories. With Qwen3.5-27B as the semantic encoder, they report mean accuracy across 50 benchmarks rising from 78.80% with few-shot prompting to 84.92% at the methods' respective evaluated history budgets, and 82.33% with identical histories and queries. On ToolPRMBench's BFCL and ToolSandbox tasks, they report absolute gains of 34.26% and 28.87%.

Why it matters: It points to a way to improve tool selection and reasoning assessment from accumulated deployment logs without fine-tuning or retraining a model.

Sources: Hugging Face Daily Papers

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.