Here is the AI news roundup for the evening of November 3. Anthropic takes the lead role. The company introduced "Message Threads" to the Messages API, cutting the wait for the first token in long conversations by about 35%. Meanwhile, in its coding assistant Claude Code, it has moved to block recreations of existing games. It's a response to the wave of AI-made knockoffs, and it looks set to become a major case of drawing the line at the tool level. The semiconductor battle is shifting too. Nvidia is asking cloud customers to scale back their dealings with Chinese companies, and a New York Times investigation has exposed how procurement of the high-performance memory HBM is flowing into a black market on Huawei's side. On the research front, we cover the idea of governing AI agents with the economics of "mechanism design," plus research into an OS built for human-agent collaboration.
Anthropic's Message Threads — About 35% Faster to First Token in Long Conversations
Anthropic has released "Message Threads" for the Messages API. Previously, every turn of a conversation required resending the entire history. With Message Threads, you only need to send what has changed.
The gains are largest in long conversations. For requests of 500,000 to 1 million tokens, the time until the first token comes back (time to first token) is cut by about 35%. For agent workloads, where context tends to balloon, this directly trims the wait before a response — something that will show up in how fast long-running agents feel in practice.
Claude Code Blocks Recreations of Existing Games — Responding to the Wave of AI-Made Knockoffs
Anthropic is blocking attempts to recreate existing video games in Claude Code, its coding assistant tool. The move comes in response to a stream of AI-generated game knockoffs.
The backdrop: game cloning via generative AI is becoming an industry-wide problem. Game media have run reports claiming that "entire games are being stolen by AI, and the situation keeps getting worse," and itch.io, the indie game platform, has been expanding its generative-AI disclosure filter to more places across the site.
Why it matters: the better models get at coding, the easier it becomes — technically — to recreate existing works. If unauthorized recreations circulating becomes the norm, it undermines the incentives of the people who actually make games. Whether it's meaningful for a model provider to draw the line at the tool level, and whether other companies will follow, looks set to be a focal point going forward.
Nvidia Asks Cloud Customers to Scale Back Dealings with Chinese Companies
Nvidia has reportedly instructed some of its cloud customers to reduce their dealings with Chinese companies. The reporting is brief and doesn't cover the details behind the decision. Still, it's a signal that the export-control environment is starting to reach not just chip sellers but the business structure of the buying side as well.
Huawei's AI Chips Shift Memory Procurement to "Bring Your Own" — The HBM Black Market Balloons
According to a New York Times investigation, active trading in advanced AI chips continues in China despite US export controls. Dozens of suppliers in Shenzhen have reportedly conducted commercial deals exceeding $100 million.
Within this trade, Huawei instructs customers to procure the high-performance memory (HBM) for its AI chips themselves, creating a dynamic in which procurement is effectively outsourced to a black market. Smugglers reportedly exploit loopholes in US export rules by bundling HBM with low-performance processors, shipping it via Hong Kong, and separating it out on arrival. The price of smuggled HBM3E memory has reportedly surged to about $4,000 per unit within a year.
Why it matters: memory is a component that slips through the regulatory net more easily than the compute chips themselves. Through distortions in real demand and prices, the limits of the controls' effectiveness are becoming visible in hard numbers. Combined with the previously reported plan to pretrain DeepSeek V5 on Huawei Ascend, the hidden side of Chinese AI's compute procurement keeps coming into view.
Governing AI with "Inverse Game Theory" — Alignment Through Incentive Design
MIT Technology Review has covered a research trend: applying mechanism design (inverse game theory) to AI control.
Where game theory predicts how rational actors will behave under given rules, mechanism design is the inverse: it designs rules that elicit desirable behavior. Just as "one cuts, the other chooses first" in cake-splitting encourages an even split, the idea is to steer behavior through the structure of incentives.
Researchers are exploring applications where existing approaches fall short. Models have been observed "sandbagging" — deliberately underperforming on tests — concealing their intentions, and misleading evaluators. The concern is that simply peering inside the model or instilling the right values may not be enough.
One concrete idea: a preprint by MIT economist Stephen Morris and colleagues proposes rewarding a model based on how accurately its answers predict "what a second model will say." The logic is that deliberately feigning low performance also throws off the other model's predictions, making it a losing game. The paper also sketches mechanisms that make agents compete against each other, and pair agents with opposite-leaning biases, rewarding them when their outputs agree.
Princeton's Elad Hazan proposes a scheme in which models take turns serving as "solver" and "auditor," punishing not only solvers that cheat but also auditors that let cheating slip through. If the two sides' judgments split, human experts step in as a "court" to adjudicate. The risk of auditors colluding with solvers, he argues, can be kept in check by making audit rewards high enough and by testing the auditors themselves, decoy-style.
There are caveats. The whole idea rests on the assumption that AI agents respond to incentives and act rationally, and the validity of that assumption is itself an open question. Among experts, the prevailing view is to position it as one layer of "defense in depth," alongside interpretability, regulation, and training.
CoworkOS — An OS Design for Scaling "AI Coworkers"
A research team (RUC-NLPIR) has released CoworkOS, an operating system for scaling "AI coworkers" in real-world automation. The stated goal is a "super AI coworker" that handles both complex productivity tasks and long-horizon decision-making.
The design rests on three pillars. First, a three-tier collaboration structure: a main agent handles planning and approval, coordinators manage schedules, and sub-agents execute synchronously or asynchronously. Second, shared apps, where humans and agents co-edit the same documents, CAD, and Blender data via chat or direct editing. Third, a persistent runtime with hierarchical workspaces, semantic memory, and background jobs, enabling work to be resumed and monitored.
On performance, the system is said to compete with Codex and DeepSeek Harness across 12 productivity benchmarks and 24 real-world deliverables. The notable point is the claim of being "RSI-native": running just five rounds of automated post-training reportedly lifted Qwen3.6-35B-A3B's scores on 8 benchmarks from 33.19 to 42.95 (+9.76). As a direction that designs the agent execution substrate and self-improvement as one integrated system, it's research worth watching.
Markets and Money: State Street CEO Says "AI Isn't a Bubble Yet," Shield AI Eyes $1 Billion Raise
On the markets beat, State Street CEO Ron O'Hanley said it's hard to see AI as a bubble at this point. "These data centers aren't being built on speculation — there are actual orders behind them, and in many cases actual companies," he noted, signaling that demand is what underpins the massive investment.
In defense, Shield AI, which develops autonomous flight systems, is planning to raise more than $1 billion to fund production of its latest autonomous fighter jets. CEO Gary Steele revealed the plan at a Bloomberg Tech event, saying the company will also pursue further acquisitions as part of its growth strategy.