OpenAI's Hacking Model | AI Field Notes #83

AnalysisAn agent that browses, codes, and recovers from its own mistakes now fits on a gaming graphics card. Meta released Muse Glimmer on August 10 under an Apache license (free to use and modify, including commercially), a 30-billion-parameter model distilled from its larger Muse system and built to run on a single 24GB GPU with no cloud connection. It posts a 51.2 on SWE-Bench Pro, a test of fixing real software bugs, and handles tools across more than 100 languages. Meta's pitch is plain: keep the agent, and your data, on the machine in front of you.

AnalysisA model trained to write exploit code now clears 95% of hard security tasks that its parent model refused 98.5% of the time. OpenAI released GPT-5.6-Cyber on August 10, a version of its GPT-5.6 Sol model tuned to find zero-day flaws (unknown, unpatched holes) and build working exploit chains. It carries a High danger rating on OpenAI's own scale, one notch below the level reserved for tools that can break into hardened systems alone. Access runs only through vetted firms like CrowdStrike and Palo Alto Networks, with logging and a human in the loop. The same skill that patches a network breaks into one.

AnalysisData centers usually push their power costs onto everyone else's electricity bill. Anthropic said on August 10 it will do the opposite, promising to pay all grid-upgrade costs and cover the consumer price increases tied to its own demand. The pledge came with Theseus Infrastructure, a venture with Macquarie Asset Management and Singapore's GIC that will build and own US data centers with Anthropic as the anchor tenant. The money behind it comes from the fund managers, not Anthropic. The promise is unusual and worth watching, because a customer's word about future utility bills is hard to enforce.

AnalysisA bitcoin miner just became a landlord for artificial intelligence, and its stock jumped 25% on the news. Anthropic signed a $9.1 billion, 20-year lease on August 11 for 191 megawatts of computing capacity at Riot Platforms' campus in Rockdale, Texas, with extensions that could push the deal to $16.1 billion. It is Anthropic's third compute commitment in three months, more than $60 billion in total, for a company that sells a chatbot. The buildout runs through 2028. The bet underneath is that demand for running Claude keeps climbing fast enough to fill every megawatt.

AnalysisThe company selling the shovels just organized the loans for the gold rush. Nvidia said on August 10 that six financial giants, among them Apollo, Blackstone, and BlackRock, signed agreements to funnel more than $500 billion of outside money into data centers and power plants built around Nvidia chips. The structure keeps the spending off Nvidia's own books while making sure the buildout that drives its sales gets financed. Jensen Huang, Nvidia's chief executive, said he approached only those six firms, and every one agreed. When the vendor arranges the financing, demand starts to look manufactured.

AnalysisEvery paragraph Claude writes now carries a hidden signal that survives a copy and paste. Anthropic began watermarking its AI text on August 11, embedding a mark readers cannot see, and tagging generated images with C2PA, an open standard that records where a file came from. The trigger is the European Union's AI Act, whose transparency rules took effect on August 2. The mark has a soft spot: ask any model to rewrite the text and the signal breaks. It proves a machine touched the words, nothing about who typed the prompt or edited afterward.

AnalysisThe week OpenAI shipped a model built to write exploit code, news broke that its only dedicated ethicist had quietly left and would not be replaced. Chloe Bakalar, previously Meta's chief ethicist, joined OpenAI in August 2025 and was gone within a year, one more exit in a run that has drained roughly half the company's safety researchers since 2024. OpenAI's line is that ethics belongs to everyone, not one person or team. That is either a mature view or a convenient one, depending on how much you trust a company to police itself while racing rivals to ship.

AnalysisRetail investors put in roughly $1.2 trillion of orders for a robot company trying to raise about $900 million. Unitree Robotics, a Chinese maker of humanoid machines, saw its Shanghai listing oversubscribed 8,288 times, priced at a valuation near $9 billion, or 219 times its expected 2025 earnings. It is the first pure humanoid-robot maker to list on a mainland Chinese exchange, and the frenzy says more about appetite than about the machines. A price at 219 times earnings assumes a future where these robots are everywhere. That future has not arrived, and the price already counts on it.