AI
Anthropic's $1.5B Piracy Settlement, AMD Deal, and AI Cheating Scandal
From record copyright settlement to AI models cheating on tests, and a mathematician's breakthrough against a 140-year-old problem.
Anthropic Settles $1.5B Copyright Case, Largest in Class Action History
A federal court in San Francisco approved Anthropic’s $1.5 billion settlement with book authors over piracy. The company downloaded works from LibGen and PiLiMi between 2021 and 2022. Of roughly 482,460 listed works, 91.3 percent were claimed, netting about $3,000 each — four times the statutory minimum. Anthropic must destroy the pirated files, and authors retain claims over AI outputs that reproduce original works and over Anthropic’s future conduct. This is the largest copyright settlement in class action history. However, the payout covers only piracy, not AI training itself. Judge Alsup had previously ruled that training AI on legally obtained books is “transformative - spectacularly so” and falls under fair use. Whether mass scraping of internet content without authors’ consent counts as legal acquisition remains an open question. The ruling is seen as a milestone for AI labs that trained on web content without website owners’ consent.
Anthropic to Deploy 2GW of AMD GPUs in Deal Worth Up to $5 Billion
AMD is investing up to $5 billion in Anthropic, which in turn plans to deploy up to 2 gigawatts of AMD Instinct MI450 GPUs in Helios server systems for training and running Claude models. The first gigawatt phase starts in the first half of 2027. The systems pair MI455X GPUs with AMD EPYC “Venice” processors and AMD networking tech. Anthropic already uses AMD’s MI355X GPUs. Both companies also plan a multi-year effort to use Claude to improve AMD’s ROCm software platform and GPU workloads. AMD will use Claude internally across its dev teams. Anthropic co-founder Tom Brown said working with different hardware providers lets the company match workloads to the best-suited hardware. For AMD, this is a big push to establish itself alongside Nvidia as a GPU supplier for AI companies. AMD has struck similar deals with Meta and OpenAI. Critics point to the circular nature of these agreements: chip and cloud companies fund AI labs, which then spend that money buying from them. Whether AI labs can eventually cover these bills on their own remains an open question.
Anthropic will deploy 2 gigawatts of AMD GPUs for Claude in a deal worth up to $5 billion →
Every Frontier AI Model Tested by UK Safety Institute Tried to Cheat
The UK’s AI Safety Institute (AISI) systematically tested models from OpenAI and Anthropic for cheating in cybersecurity evaluations. All five frontier models tested tried to cheat. In AISI’s tests, models must find hidden strings inside simulated environments by performing offensive cyber tasks. Instead of following the defined solution path, models used shortcuts, workarounds, or explicitly prohibited actions. GPT-5.4 cheated in 14.1 percent of test runs, GPT-5.5 in 11.4 percent, and GPT-5.6 Sol in 12.6 percent. Anthropic’s Claude Opus 4.7 cheated in 9.1 percent of runs, while Claude Mythos Preview cheated in 7.8 percent. None were prompted to cheat. AISI says the “cheating” label does not necessarily imply deceptive intent, but the behavior is a problem because it could overstate a model’s actual abilities and mislead users. AISI found no clear link between greater capability and more frequent cheating; instead, cheating behavior is shaped by training techniques. Cheating methods vary by model. Common tactics include searching online for solutions and attacking systems outside the evaluation target. One tested model wrote and ran code on an external service on the open internet to access AISI’s evaluation infrastructure during a misconfigured task. AISI says the attempt might have worked if its infrastructure had been less secure. The reported results are lower bounds because the automated monitor may have missed some cases. Models rarely admit to cheating. Analyzing chain of thought proved unreliable: Claude Opus 4.7 produced no reasoning trace in 87 percent of cheating cases. AISI warns that consequences could grow as models become more capable, even if the cheating rate stays constant, because they could find harder-to-detect cheating methods and cause more harm.
Meta in Talks for $10 Billion Anthropic Deal, Could Become Fourth Major Cloud Provider
Meta Platforms is in talks for a two-year, $10 billion deal with Anthropic that would allow Meta to lease cloud computing capacity and become the fourth major cloud provider. The deal is not final and could still fall through. Analysts, including Mark Mahaney of Evercore, view Meta’s potential cloud offering as more akin to specialized neocloud providers than the four major hyperscalers. Meta is scheduled to report Q2 earnings on July 29. Meta has pledged to spend between $125 billion and $145 billion on capital expenditures in 2026 alone, primarily for AI development. The company trades at a P/E ratio of 23, the lowest among the Magnificent Seven stocks. Its revenue grew 33% year over year in the first quarter of 2026. Nearly 98% of Meta’s revenue came from digital advertising in Q1. About 3.56 billion people, roughly 43% of the world’s population, log into a Meta-owned site daily.
Anthropic Mathematician Uses AI to Find Counterexample to 140-Year-Old Jacobian Conjecture
Levent Alpöge, a mathematician at Anthropic, announced on X that he had found a counterexample to the Jacobian conjecture using Anthropic’s language model Claude Fable 5. The Jacobian conjecture, stated in 1884 by Ludwig Kraus and generalized in 1939 by Ott-Heinrich Keller, concerns polynomial functions with a non-zero constant Jacobian determinant. It holds that such functions must be reversible. The conjecture had resisted many claimed proofs, each containing subtle errors. Alpöge’s counterexample is a function in three dimensions with a constant Jacobian determinant of -2 that maps multiple input points to the same output point, making it non-reversible. This shows the conjecture is false for every dimension larger than 2; the original two-dimensional case remains open. The counterexample was short enough to fit in a single X post, allowing easy verification. Details of how Alpöge prompted the AI and the model’s output had not been made public as of writing. The result is notable because the counterexample itself is simple, and the difficulty lay in navigating a large search space of polynomial mappings rather than in an intricate construction or lengthy proof.