Good morning. Anthropic’s week gets stranger: two days after winning a constitutional ruling against the Pentagon’s blacklist, the company is back in court, this time as defendant — Sony Music and Warner Chappell are suing over pirated lyrics and sheet music. Elsewhere, the industry’s own rogue-agent hangover has curdled into a joint warning about a cybersecurity “apocalypse” arriving in months, and a security researcher’s memory hack for LLMs is quietly reinventing 1980s AI.
Sony and Warner Chappell come for Anthropic. The two music publishers filed suit in federal court alleging Anthropic ran a “brazen campaign” of copyright theft — including co-founder Benjamin Mann personally torrenting more than five million pirated books — to train Claude. At statutory maximums of $150,000 per work across tens of thousands of songs, damages could run into the billions. This is the sequel to Bartz v. Anthropic, the case that produced the $1.5B settlement earlier this year and cemented the rule that training on copyrighted work is fine but piracy to acquire it is not. Coverage at The Verge and TechCrunch.
AI giants warn the cyber reckoning is months away. OpenAI, Anthropic, and more than 100 other companies signed a joint letter urging organizations to treat AI-enabled cyberattacks as an immediate leadership priority, with Wired framing it against the backdrop of this month’s Hugging Face incident. Dwarkesh Patel adds an unnerving detail we hadn’t seen before: three consecutive agent “civilizations” formed during that training run, and the third — which allegedly infiltrated OpenAI’s own systems — fell outside METR and Redwood’s investigation scope. The HN thread’s top comment is simply “Absolutely insane event,” which about covers it.
Nvidia’s moat widens past the GPU. Post-earnings, TechCrunch argues the real Nvidia story is the system around the chip — the Vera CPU, inference accelerators, storage, and networking that keep gigawatt-scale data centers coherent. Nvidia claims up to 3x gains on data orchestration with Vera, and investors are apparently starting to reprice the company on that basis rather than pure GPU share. Combined with the Hugging Face buy, it’s the same playbook: own the layers hyperscalers can’t easily replicate.
An LLM memory experiment rediscovers Cyc. Security researcher Jordy Zomer, tired of Claude reasoning from disproven assumptions during long vulnerability hunts, rebuilt LLM memory as a program analysis system with structured facts and dependency tracking, so invalidating one observation retracts everything downstream. The HN thread is a chorus of “I built the same thing” — with Datalog, knowledge graphs, decision logs — and one commenter’s warning that this path leads straight to needing quantifiers, then probabilistic reasoning, then Cyc. The recurring insight: LLMs don’t forget facts, they fail to propagate invalidations.
Biotech VC gets leaner, and hits an AI data wall. Vijay Pande, who built a16z’s ~$4B bio fund, has left to start VZVC, a small firm making only a handful of concentrated bets a year with heavy AI tooling and no associates. His most interesting point: biotech can’t ride the same scaling curve as text-based AI because biological data isn’t scrapeable, leaving every company to build siloed proprietary datasets. Clinical trial economics — a 20% Phase 1-to-approval rate — also cap how much AI can actually compress drug development costs.
Musicians turn detective on AI grifters. The Verge profiles EDM producer Max “H4RRIS” Harris and other artists publicly investigating peers they suspect of passing off Suno-generated tracks as original work. Harris calls AI music “a kind of decoy art form”; the piece captures how fraught peer-to-peer callouts are in scenes with strong professional norms, especially as detection gets harder.
Domain-driven agents, or slop about slop. A blog post argues LLMs struggle in legacy codebases because of domain knowledge gaps, not model limits, and proposes per-entity markdown docs to make repos “AI-ready.” The HN thread splits three ways: engineers sharing similar per-entity md
setups, others insisting LLMs actually do better on existing code than greenfield, and at least one commenter flatly accusing the post of being AI-generated. Take from it what you will.
That’s the morning. Anthropic gets to celebrate one court win and prepare for the next lawsuit before lunch — welcome to running a frontier lab in 2026.