OpenAI wants its Georgia data center to act as a grid backup during emergencies. OpenAI announced a pilot called the Grid Protectors Program, tied to an upcoming data center in Effingham County, Georgia. The company says it will free up to 1,000 megawatts of power during extreme weather or grid stress events, enough to run air conditioning in roughly 700,000 homes. The pitch is that large compute campuses, which are usually criticized for straining local grids, can be designed to release power back when demand spikes. Whether utilities will structure contracts around that promise is the open question. (The Verge)
Brief
Tuesday, November 3, 2026 · 4 min read
Anthropic blocks Claude Code from cloning existing games
Good morning. Copyright law and AI tools have been circling each other for years, mostly in courtrooms. Today the fight moved one step earlier in the chain: into the system prompt itself. Anthropic decided it was easier to block the behavior than to litigate the consequences. That is a notable place to draw the line.
Today's reading time is 4 minutes.
Listen · Snippt Daily0:00 / 4:06
INDUSTRY
AI coding tools have been quietly fueling a wave of retro game reconstructions, and rights holders have been watching.
Driving the news: Anthropic has updated Claude Code's system prompt to prohibit the tool from recreating existing video games. The restriction is narrow: Claude will still help users build original games or work on their own titles. What it will no longer do is take a classic game and reproduce it, whether through decompilation, recompilation, or porting. The change was first spotted by Tom's Hardware and confirmed by The Verge.
- The block lives in the system prompt, meaning it applies to Claude Code specifically, not Claude broadly.
- Retro game reconstruction has become common enough that Anthropic apparently felt a policy response was warranted before a legal one arrived.
Zoom in: Claude Code launched as a terminal-based agentic coding tool, and developers quickly found it capable enough to handle substantial reverse-engineering tasks. Reconstructing a game from memory or from a ROM is exactly the kind of multi-step coding job that agentic tools handle well. Nintendo and other publishers have spent years pursuing fan-made ports and decompilation projects in court, so the legal exposure for a company whose tool assists that work is real.
- Decompilation projects for titles like Super Mario 64 and Zelda: Ocarina of Time have already drawn Nintendo cease-and-desist letters.
- Anthropic's move is a policy decision, not a technical one - a determined user with a different tool faces no new barrier.
Why it matters: For developers using Claude Code for legitimate game work, nothing changes day to day. For anyone who was using it to reconstruct a commercial title, the path just got harder. The broader implication is that AI tool makers are now making editorial calls about what their products will help build, not just what they will say.
Bottom line: Anthropic did not wait for a lawsuit to decide that helping clone Nintendo games was not worth the risk.
The Verge ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Deep Principle released a 753-billion-parameter open-weight model aimed at scientific research. Hydrogen-1.0 is built on top of Z.ai's GLM-5.3 base, post-trained on chemistry, materials science, and life sciences data including laboratory records and agent discovery trajectories. The weights are open, which matters in scientific computing where reproducibility and local deployment are often requirements. It reports a Frontier Math score competitive with closed models in its class, though benchmark comparisons in specialized science domains are notoriously hard to verify independently. (MarkTechPost)
OpenAI merged ChatGPT's Chat and Work tabs into a single usage pool. ChatGPT previously split its interface into separate Chat and Work tabs, each drawing from different allocation buckets. As of this week, both Codex and standard chat pull from the same pool. For users who were carefully rationing Codex credits separately, the math on their daily limits just changed. OpenAI framed it as a simplification, which it is, though it also means heavy Codex use now competes directly with regular chat. (TechCrunch)
Trajectory raised $45 million to let non-engineers train their own AI models. The one-year-old startup is now valued at $300 million on the back of a pitch that custom model training should not require a machine learning team. The market it is targeting is the large middle of enterprise software buyers who want domain-specific models but cannot staff the work. At $300 million on a year of existence, the valuation reflects how much appetite there is for that promise rather than proven scale. (TechCrunch)
The thread
AI tools are acquiring editorial guardrails
Two of today's stories show AI tool makers making deliberate choices about what their products will and will not do, independent of what is technically possible. Anthropic added a system-prompt block on game cloning before any lawsuit forced the issue. OpenAI merged usage pools in a way that quietly changes how much of each capability a user can access. Neither change is a technical limitation - both are policy decisions baked into the product layer. For people building with these tools, it means the ceiling on what you can do is increasingly set by a product manager's risk calculus, not the model's capability.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- OpenAI built Pando, an internal Git-compatible backend for its monorepo that cut median push latency by nearly 80% and now moves close to a petabyte of Git data per day - a scaling problem that emerged directly from Codex's volume of automated commits.
- A research paper on arXiv tested how well a widely used consumer LLM could assist surveillance tasks, finding it capable of generating functional code and detailed guidance for monitoring individuals at scale.
- MIT Technology Review covered 'reverse game theory', a mechanism-design approach to AI alignment that tries to make rule-following the mathematically rational choice for capable models rather than relying on training alone.
- Zyphra released a sparse attention method called ZHA that retrofits frozen large language models with roughly 4 billion tokens of distillation, matching dense attention on reasoning and long-context retrieval benchmarks.
- OpenAI raised its security bug bounty payouts and added bonuses for report quality, researcher loyalty, and priority campaigns, as part of a broader internal security sprint with Cathedral Systems.
- Omen AI acquired Aquatic Labs for $16 million to add liquid cooling expertise as data center water usage draws more scrutiny.
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