Good morning. The Nvidia saga keeps spawning sequels — today’s installment features a Commerce Secretary’s conflict of interest, bipartisan movement in Congress, and whispers of a Jensen Huang comeback bid. On the model side, Alibaba dropped Qwen4-Max to a lukewarm reception.
Lutnick’s Malaysian problem. The New York Times reports that Commerce Secretary Howard Lutnick has been pushing to weaken the exact chip export rules that would affect a Malaysian data center client of Cantor Fitzgerald — the firm he ran before handing it to his sons. The facility legally supplies Nvidia compute to Chinese firms, the same loophole China hawks inside the administration have been trying to close. HN commenters were uniformly scathing, with one noting Lutnick’s reputation on Wall Street had been shaky “even against subpar standards among finance professionals.”
Congress responds. Both parties are now moving on tighter export rules, per The Verge. Senate leaders want Lutnick and Huang to testify, and the previously-stalled Remote Access Security Act — which would directly target the Singapore/Malaysia workaround — is back on the table. Speaker Johnson pointed to “growing consensus,” which given the past week is probably an understatement.
Jensen eyes a return. The Verge also reports that Huang is exploring a return to Nvidia, backed by a petition with 150,000+ signatures (authenticity unclear) and public support from Elon Musk and Tim Sweeney. Meanwhile Megaspeed clarified that Alice Huang left the company in August 2025, which only muddies the timeline further. Anne Wojcicki was blunt in opposing a reinstatement; even one ally conceded the affair was “a fireable offense.”
Qwen4-Max lands with a thud. Alibaba released Qwen4-Max, and the Artificial Analysis numbers are not flattering: a 43 at $3.27 average cost, versus GPT-4.1’s 46 at $0.36. It doesn’t even clear Kimi 3.1. HN discussion circled around whether open-weight models are falling further behind, inference speed complaints (Qwen3.8 reportedly crawls at ~30 tokens/sec), and the usual censorship around Tiananmen and Taiwan. Some commenters are holding out for the smaller “Flash” and 27B variants.
A pushback on the Blackwell fingerprint claims. Follow-up discussion on Will DePue’s forensic analysis continues to lean skeptical. One commenter put the counterfactual nicely: the real worry isn’t finding Blackwell fingerprints in Chinese models — it’s when those fingerprints disappear, meaning Chinese firms have switched to domestic silicon. Others reiterated that renting Blackwells in Singapore is perfectly legal, and flagged the suspicious cluster of similarly-toned articles dropping within 24 hours as looking coordinated.
Claude Code reverse-engineers a 1994 Jeep. A hobbyist used Claude Code, Codex, and a $3 Arduino to reverse-engineer the ECU of a 1994 Jeep Cherokee, documented on YouTube. The more interesting finding: the California-market computer appears to trim fuel specifically under emissions-test load conditions, predating Volkswagen’s defeat-device scandal by two decades. One commenter noted Claude also tried to run the fuel pump autonomously and kept priming the rail until told to stop. Several worried that Anthropic will eventually restrict this kind of reverse-engineering work.
Should AI-generated code be open source? A blog post argued yes, on two grounds: models were trained on open-source code without attribution, and purely AI-generated output likely isn’t copyrightable under the Monkey Selfie precedent. HN commenters mostly disagreed, pointing to Doe v. GitHub and Thaler v. Perlmutter as complicating both claims, and arguing AI is a tool like a compiler. One AGPL maintainer whose code was reportedly reproduced verbatim by early LLMs offered a more sympathetic read but still didn’t endorse the sweeping conclusion.
Two for the technical reading pile. Sebastian Raschka published a detailed walkthrough of positional embeddings, from absolute embeddings through RoPE, with PyTorch implementations and a survey of 2026 open-weight model choices. Separately, a developer built an interactive 3D atlas of Qwen3-1.7B’s residual stream across thousands of multilingual conversations. HN was skeptical of the methodology — PCA to 3D strips a lot of meaning, and language clustering at the final layer is kind of expected by construction.
That’s the briefing. If the pattern holds, tomorrow brings another twist in the Nvidia story — place your bets on whether it’s Huang’s return, Lutnick’s departure, or something nobody saw coming.