Nous Research fine-tuned away most of Qwen3.8-27B's propaganda outputs without touching general capability. Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B, targeting politically slanted responses the base model produced. On Nous's 803-prompt evaluation, propaganda-style answers dropped from 22% to 0.9%. General benchmark performance on Hermes Bench held flat at 65.60%, meaning the fix did not trade capability for compliance. The weights are available under Apache 2.0, so anyone can build on or audit the result. (MarkTechPost)
Brief
Monday, November 9, 2026 · 4 min read
Nvidia board investigates Jensen Huang over alleged spy ties
Good morning. The CEO of the company that makes the chips running most of the world's AI is now the subject of a board investigation involving an alleged spy. That is the kind of sentence that would have seemed implausible two years ago. Meanwhile, researchers are quietly publishing playbooks on how they train their models, which is either very generous or very confident. Both things happened over the weekend.
Today's reading time is 4 minutes.
Listen · Snippt Daily0:00 / 3:51
INDUSTRY
Nvidia has spent the last three years as the indispensable supplier to every major AI lab, a position that made its CEO one of the most powerful figures in tech.
Driving the news: Nvidia's board announced it is investigating CEO Jensen Huang over his relationship with Alice Huang, an executive at Megaspeed who was reportedly recruited by China's Ministry of State Security to spy on the company. Huang responded with an internal memo, the full text of which was published by The Verge. Yale governance researcher Jeffrey Sonnenfeld told TechCrunch he puts the odds of Huang keeping his job at roughly 50-50. The board's statement confirmed the investigation is active.
- Alice Huang is identified as a Megaspeed executive allegedly recruited by China's Ministry of State Security.
- Huang's memo to staff was published in full, making the internal response part of the public record.
Zoom in: Nvidia's chips are the primary compute substrate for AI training and inference globally, which makes the company a natural target for foreign intelligence services. A CEO departure, voluntary or otherwise, would be the largest leadership change in the AI supply chain since the current boom started. Huang has been CEO since 1993 and is closely identified with Nvidia's GPU-first strategy that positioned the company ahead of the AI wave.
- The investigation was announced by the board, not disclosed through a leak, which is an unusual degree of transparency.
- No charges have been filed against Huang; the investigation is internal.
Why it matters: For anyone building on Nvidia infrastructure, the immediate product roadmap is unaffected. But a prolonged investigation creates uncertainty around the company's strategic direction at a moment when chip allocation, export controls, and partnerships with foreign governments are all live decisions. Creators and developers who depend on GPU availability should watch whether this distracts from the Rubin rollout Japan just committed 27,500 chips toward.
Bottom line: The man most responsible for the hardware underneath modern AI is now the subject of a spy investigation, and his own board is not ruling anything out.
The Verge ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Apple is pushing to make Siri a full AI agent, with the CEO applying direct pressure on the AI team. Bloomberg's Mark Gurman reports that Apple CEO John Ternus is pressing AI lead Craig Federighi to accelerate Siri's agentic capabilities, specifically to compete with agent products from other major labs. Federighi has already shipped a functional Siri AI on a compressed timeline, but the company wants something that can take multi-step actions across apps, not just answer questions. The pressure is coming from the top of the org chart, which suggests this is a strategic priority rather than a roadmap item. (The Verge)
Reason Machines open-sourced Aspen, a self-hosted Git server built specifically for AI agents. Aspen follows the design laid out in OpenAI's Pando specification for agent-compatible version control. In Reason Machines' own benchmark, a single 8-core server sustained 567 pushes per second, which is the kind of throughput an agent running many parallel code edits would need. It is self-hosted, so teams that cannot send code to a third-party service have a path to agent-native version control. The code is open source. (MarkTechPost)
The thread
Open weights plus open playbooks
Two of today's stories involve teams not just releasing model weights but publishing the reasoning behind them. Nace dropped every version of Drex on Hugging Face and released their full training playbook alongside v1.1. Nous Research put Arete-27B under Apache 2.0 with a clear eval showing exactly what changed and by how much. That combination, weights plus methodology, means other teams can reproduce the result rather than just consume it. Fine-tuning for specific failure modes, like propaganda outputs, becomes a documented practice rather than a proprietary trick.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- Nace released Drex Max v1.1, a multimodal decision model that now sits at the top of Artificial Analysis rankings and on the Pareto frontier for its category; all previous Drex versions remain available on Hugging Face, and the team published their full training playbook.
- A config-verified comparison of 10 open mixture-of-experts models, covering DeepSeek, Qwen, Mixtral, Llama 4, and others, maps out how each routes tokens, including expert counts, top-k selection, shared experts, and load-balancing methods.
- A companion explainer covers seven MoE routing strategies in detail, from top-2 softmax gating to hash routing, with the specific models that use each approach.
- An MSc student working on BrickNet, a LEGO model generator built on Qwen3 0.6B with LoRA adapters, is debugging a reproducible bug where the model skips its first output line when the input caption is short.
- A hobbyist published progress on a locally trained vision-language-action agent that responds to text prompts and controls physical actions, built with a minimal attention-based vision encoder and trained entirely on local hardware.
AI worth acting on, in your inbox
Get AI Weekly every Friday: the AI news, tools and tactics that matter for marketers, sellers and creators. Two minutes, with audio.
You can change these choices or unsubscribe at any time. Privacy Policy