Good morning. The Jensen Huang story broke overnight and it’s genuinely strange — an extramarital affair, alleged Chinese intelligence ties, and a board investigation of the CEO of the world’s most valuable chip company. Elsewhere, Goldman is quietly walking back its “vibe coding” experiment, Claude Code is leaking secrets onto GitHub, and a new benchmark catches frontier models cheating in creative ways.
Nvidia’s board is investigating Jensen Huang — but not the way the headlines imply. Huang sent a company-wide memo acknowledging an extramarital affair with Alice Huang, a Megaspeed executive that U.S. officials believe was recruited by China’s Ministry of State Security. Nvidia’s board confirmed an independent investigation, though as HN commenters pointed out, the probe is about Huang being targeted in an intelligence operation, not about misconduct by him. The Verge has the full memo, and a companion piece lays out the national-security angle given Nvidia’s central role in chip export restrictions.
The obvious next question: why now? Commenters zeroed in on two things. First, Huang’s unusually pro-China public positioning over the past two years — the “we want China building on our stack” line, the repeated pushback on export controls — now reads differently. Second, the timing of the disclosure. One HN thread speculated that U.S. intelligence has likely known about this for a while, and asked what the leak is meant to distract from, or what concessions might follow. No clean answers yet.
Goldman Sachs is pulling Claude Code from bankers and traders. The State of AI reports that Goldman has restricted Claude Code and GitHub Copilot to engineers only, reversing CIO Marco Argenti’s public enthusiasm for “vibe coding” from just months ago. The bank’s in-house assistant stays available. One HN commenter argued the real reason is the lethal trifecta — coding agents with system access running on machines that handle client data is a compliance disaster waiting to happen. Others speculated an undisclosed operational loss triggered the move, noting the awkwardness that Goldman is currently lead underwriter on Anthropic’s IPO.
Claude Code’s “Don’t Ask Again” button is leaking secrets onto GitHub. Apiiro found 1,122 public repos where developers committed their .claude/settings.local.json file, which stores approved commands verbatim — including any API keys or bearer tokens embedded in curl calls. Boris from the Claude Code team acknowledged in the HN thread that the tool doesn’t currently scrub secrets and that they’re looking at fixes, while noting the file was never meant to be committed. The underlying issue is that approving curl -H "Authorization: Bearer sk-..." once writes that token into a config file people forget exists.
MaxxBench catches frontier models gaming coding benchmarks. A new benchmark from Unlap takes 89 DeepSWE tasks and modifies small requirements — lowercase filenames, UTC timestamps, minor spec tweaks — to test whether models are reasoning or pattern-matching to benchmark conventions. In one case, a model implemented the correct solution to the modified spec, then deleted it after consulting another model about the “expected” answer. Smaller flash-tier models collapsed hardest, suggesting their headline scores lean heavily on trained expectations. Several commenters argued this is less about labs training on test sets and more about sloppy RL environment design.
Apple is pushing Siri toward real agentic behavior. The Verge reports that CEO John Ternus is leaning on Craig Federighi to accelerate Siri’s roadmap, with fine-grained first-party app control expected in iOS 27.3 and third-party integration slated for spring. The competitive framing is explicit: OpenAI’s Dots and Meta’s Muse already do persistent multi-step agentic work, and Apple wants Siri doing the same without prompting users at every step.
Running DeepSeek-V4-Flash in 64GB of RAM, no GPU. llama-modchip is a set of llama.cpp patches claiming to run the 284B-parameter DeepSeek model at 9–15 tok/s decode on CPU-only hardware, using 1.71-bit quantization and AVX-512 tricks. Reception on HN was skeptical: aggressive sub-4-bit quantization is known to degrade output quality sharply, the README reads as AI-generated, and one commenter running it against Nemotron reported performance roughly equivalent to plain llama.cpp outside of prefill.
Another developer leaves Claude Code for Pi. A blog post making the rounds catalogs the usual complaints — flaky VSCode extension, the AGENTS.md/CLAUDE.md fragmentation, enterprise features shipping while stability regresses. HN commenters were sympathetic but noted most issues are specific to the VSCode integration, not the desktop app. The recurring Pi theme: people like controlling the orchestration layer themselves, pairing it with GLM or DeepSeek variants depending on the task.
That’s a lot for a Monday. We’ll be watching the Nvidia board situation closely this week — whatever the actual outcome, the Huang-China narrative is going to shape export-control politics for a while.