FIELD NOTES
What a dramatic week. Fresh off the midterms, tech stocks took a beating before bouncing back on Friday. And just as investors were catching their breath, another plot twist: Jensen Huang, Nvidia’s CEO, was reportedly targeted by a Chinese intelligence operative.
According to the New York Times, US intelligence suspects that an executive at Nvidia customer Megaspeed was recruited by Chinese intelligence to gather information about Jensen and potentially influence US chip-export policy.
The story first surfaced in late October, but fresh details emerged Friday night. Three congressional committee chairs have since requested a briefing with Jensen.
To be clear, there is no established evidence that Jensen knowingly cooperated with Chinese intelligence. Nvidia disputes the implications of the reporting.
The timing of the whole thing is just… odd.
Don’t worry, I’m not about to go all Emily Sundberg on you (the sensationalist telling of the story is that the executive is a woman, and there are allegations of romantic involvement). But there’s a bigger question here.
Nvidia is the lynchpin of the AI ecosystem. Its chips power frontier models, its roadmaps shape hundreds of billions in infrastructure spending, and Jensen has become an influential figure in US–China technology policy.
What happens when so much of the world’s AI infrastructure depends on one company, and so much of its commercial and geopolitical influence is concentrated in a few people? When one company sits at the center of the AI boom, does its CEO become a single point of failure?
And as for how the market reacts on Monday after the turbulent last week? If you’re reading this on Sunday, see you on the other side….
-Tara
THE DOWNLOAD
01 / Marvell to acquire Lumentum for $115 billion
Marvell announced a $115 billion acquisition of Lumentum, one of the biggest deals of the year and a significant bet on optical networking for AI data centers.
Marvell designs custom AI chips and networking silicon. Lumentum makes lasers and optical components that move data using light rather than electrical signals. Together, they would control more of the technology connecting thousands of chips inside increasingly large AI clusters.
As models scale, moving data between processors is becoming almost as important as the computation itself. Copper connections face bandwidth, distance and power constraints, making optics increasingly attractive.
02 / MIT researchers use an LLM to edit another model’s weights
Researchers at MIT CSAIL have demonstrated something unusual: an LLM that can inspect and directly rewrite another neural network’s weights, without conventional training.
Using GPT-6 Astra, the team identified a repetitive reasoning loop in Qwen3.5-0.8B and corrected it by modifying just 0.016% of the model’s weights. Its MATH-500 score jumped from 7.9% to 48.1%, while instruction-following performance nearly tripled. In other experiments, Astra identified what small neural networks were designed to do simply by inspecting their weights, and even wrote game-playing networks from scratch. There’s an important caveat: the Qwen improvement largely fixed a specific failure mode rather than teaching the model new mathematical abilities.
But the possibility is fascinating. Neural networks have traditionally been optimized through training algorithms. What if increasingly capable LLMs could instead inspect, debug and deliberately rewrite them?
03 / Researchers demonstrate an AI-operated molecular biology lab
Scientists at Shanghai Jiao Tong University have built AutoDNA, an autonomous laboratory that can coordinate and execute complex biology experiments, described in a new Nature paper.
The system orchestrates more than 20 types of instruments across over 11,000 experimental steps. It synthesized and sequenced DNA, completed a DNA-data-storage experiment, and detected a cancer-associated mutation. In one test, it ran concurrent experiments 3.2 times faster than conventional workflows.
Automated laboratories aren’t new. What makes AutoDNA interesting is how the researchers integrated AI planning with a diverse collection of instruments, allowing the system to coordinate multistep experiments with limited human intervention.
ALSO
Anthropic makes AI inference verifiable. FlashProve uses cryptographic proofs to verify that a particular model produced a response without exposing its weights. A potentially important trust layer for AI services, though proof generation remains computationally expensive.
OpenAI finds 948 exploitable Linux vulnerabilities. OpenAI disclosed an autonomous security-research system that identified 948 Linux kernel vulnerabilities, generated working exploits and proposed patches for human review, at a reported cost of $1.2 million.
Engineers recreate Flapping Airplanes’ GPU megakernel from a video. Two engineers at Magni open-sourced a reproduction of Flapping Airplanes’ unusual GPU demonstration, which trains a model, runs inference and plays Doom simultaneously inside a single kernel. They reportedly reconstructed it using the company’s demo video, without its source code. It doesn’t invalidate Flapping Airplanes’ broader technology, but raises an interesting question about how defensible clever GPU software optimizations really are.
NTT DATA signs 900MW power agreement. NTT and Bloom Energy plan to supply data centers with onsite fuel-cell power, beginning in 2027. Waiting for grid connections is becoming an increasingly expensive part of building AI infrastructure.
China blocks part of SoftBank’s ABB robotics acquisition. Chinese authorities prohibited the acquisition of ABB’s Chinese robotics business, complicating SoftBank’s planned $5.4 billion global deal. Industrial robotics is increasingly caught in the same geopolitical tensions as semiconductors.
Anthropic studies what AI agents should remember. New experiments suggest that selective, consolidated memories can outperform exhaustive activity logs. More memory isn’t necessarily better memory.
Together AI introduces continual inference optimization. Its new compiler system continuously tests and improves how models run on GPUs, reporting up to 2.6× speedups in selected workloads. AI inference performance is increasingly a software problem, not just a chip problem.
Micron’s Taiwan expansion hits a labor dispute. Construction at a planned $4.5 billion memory facility reportedly paused amid labor negotiations. A reminder that expanding chip capacity involves more than buying manufacturing equipment.