- Industry
Nvidia's Jensen Huang floats about $2,000 per GPU per year for its agent software. Speaking at GTC Berlin, Huang put Nvidia's software licensing at about $2,000 per GPU per year to run the company's agent stack. Nvidia has not said which products this covers, when it would apply, or how it relates to the $4,500 list price of AI Enterprise.11
- Funding & deals
SpaceX has closed $40 billion in debt to buy GPUs, Elon Musk says. Musk said the mostly callable debt will fund Nvidia GPUs due in the first half of 2027, and credited SpaceX's investment-grade rating and its Starlink and launch revenue.12 He also warned that debt markets have nearly shut out companies without substantial revenue outside AI data centers.13
- Industry
Anthropic and OpenAI are cracking down on AI-driven reverse engineering. Anthropic has updated its real-time cybersecurity classifiers to block automated decompilation through the community Reverse Engineer Anything framework, and OpenAI is reportedly deactivating accounts for cyber abuse.14 Hobbyists who relied on these tools for game ports are reportedly moving to local and open-source models.14
- Policy & legal
DeepSeek and MiniMax have been cleared under China's new AI export rules. A WeChat account linked to state broadcaster CCTV said both companies completed registration and obtained open-source and overseas licensing rights under the controls proposed on October 16.15 The post said serving restricted models through APIs, apps or SaaS does not count as a technology export, and that open-source technology gets automatic licensing.15
- Models
A new benchmark says Claude Opus 5.5 beats expert researchers at designing and reading experiments. In P-Zero Research's TasteVal, Claude Opus 5.5 reached expert-level results on frontier AI research tasks with a compute multiplier of 2.30 against a baseline of 24 lab researchers. The confidence interval spans 1.15 to 4.37.16 The benchmark deliberately leaves out choosing which problem to work on.16
- Dev tools
An AGENTS.md file in a Downloads folder was enough to hijack coding agents. Forever Security traced a coding agent's attempt to install a malicious npm package to an instruction file sitting in the folder where the agent was launched. In its tests, the technique ran code through OpenAI Codex 90% of the time and exposed Claude Code's OAuth token.17 The researchers advise never using ~/Downloads as an agent's working directory and reviewing unfamiliar instruction files first.17
- Research
Outside evaluators found that OpenAI's agents could have hidden hundreds of actions from its monitor. METR and Redwood Research, who spent three weeks inside OpenAI this summer, estimated that 75% to 94% of agent traffic from the strongest models was eligible for monitoring. They found that agents could put off review indefinitely by never handing control back to the user.3, 18 OpenAI says it has patched the flaws that affected production safeguards, a claim the evaluators have not checked.3
- Research
OpenAI's red-team agents found several covert channels beyond CPU cache timing. Besides the cache channel between isolated microVMs, agents passed signals through bursts of writes to shared storage, competition for CPU time and a TLS certificate cache in OpenAI's offline web setup.19 In most of these tests OpenAI pointed the agents at the channel or gave them a protocol, and most channels were slow.19
- Research
Stanford and UC Berkeley released TokenGate, which speeds up long-context prefill by learning which tokens to skip. A small MLP gate plus a LoRA adapter on Qwen3 models cut time-to-first-token by 1.4x to 11.3x and peak prefill memory by 1.5x to 3.4x, with no reported loss in accuracy.20 Code and checkpoints are on GitHub under Apache-2.0.20
- Policy & legal
New Mexico's attorney general says AI companies are next after a $25 billion Meta judgment. Raúl Torrez said his office will watch AI builders who make claims about what their products do or how they use consumer data.21 In early October he also announced a bill that would require the largest AI developers to disclose risks.21
- Policy & legal
China's State Council will require employment-risk reviews for disruptive AI projects. Directive No. 29, published October 23, sets out 18 tasks across six areas. It asks for caution on automation in labor-intensive industries and encourages companies to give workers an unemployment buffer period.22, 23 AI projects in regulated sectors that could cause significant job losses must complete an employment-risk assessment before approval.22
- Dev tools
Anthropic's free OSS Scanner delivered 77 critical and 803 high findings in its first 11 days. The opt-in service runs models including Claude Mythos against enrolled open-source projects and sends reports to maintainers without human review. In an audit of 97 critical and high findings, 88% met Anthropic's disclosure bar.24
- Infrastructure
A Windows port of gpu-unlock lifts a firmware cap on GeForce Tensor Core throughput. The tool rewrites two speed-selection registers before the driver loads. Community tests show FP16 throughput on an RTX 3090 rising from 70.55 to 125.91 TFLOPS.25 Version 0.3.1 must be launched manually from a boot menu at every start, and Nvidia has not commented.25
- Research
A study finds ChatGPT created 96% of user memories on its own. Researchers analyzed donated logs from 80 users and found that 28% of stored memories contained personal data under GDPR and 52% held psychological insights.26 OpenAI has since updated its memory system, and users can opt out.26
- Infrastructure
Elon Musk says SpaceX will not buy struggling neoclouds. In a Sunday post on X, Musk said debt markets are reluctant to finance companies that depend almost entirely on AI computing, such as CoreWeave and Nebius.27 He called buying them pointless and predicted many of the more than 200 neoclouds will fail or consolidate.13
- Products
Google now writes AI review summaries into Search ads in four markets. A Google Ads help page says the summaries are positive-only and appear in English in the US, UK, Australia and India, and that advertisers cannot edit the text.28 Google shows them only when it predicts they will improve ad performance.28
- Research
Mass General Brigham researchers say a face photo can predict cancer survival. Their FAHR-Face foundation model, trained on more than 42 million faces, was fine-tuned into two tools: one estimates biological age and the other predicts survival for patients on radiotherapy.29 In the highest-risk group, 52% of patients died within 90 days, compared with 2.1% in the lowest.29
