Good morning. The week’s chip story keeps unfolding — Anthropic is now publicly accusing Alibaba of stealing Claude’s capabilities through distillation, while the Trump administration is asking OpenAI to delay GPT-5.6 and treating Anthropic considerably worse. Elsewhere, a former Databricks exec wants to throw out neural networks entirely and replace them with oscillators.
Anthropic accuses Alibaba of distilling Claude. Anthropic says Alibaba ran a distillation campaign between April and June, using roughly 24,000 fraudulent accounts to generate 16 million Claude exchanges and train a cheaper model on the outputs. The HN thread is unsympathetic: commenters note that model outputs aren’t IP-protected, that Anthropic itself trained on scraped web content, and that the actual mechanism here appears to be Chinese resellers offering pooled Claude Max access at 70-90% below list price through payment fraud. The Steve-Jobs-complaining-about-Xerox comparison came up more than once.
The White House wants GPT-5.6 slowed down, and Anthropic gets it worse. OpenAI will delay GPT-5.6 at the administration’s request, with the Office of the National Cyber Director approving enterprise access customer-by-customer during a preview window. TechCrunch reports the trigger is concern about frontier cyber capabilities — models that find and exploit vulnerabilities faster than defenders can patch. Anthropic got harsher treatment: an outright order to suspend Mythos 5 and Fable 5 access, with export controls extending to non-US Anthropic employees. The earlier “speed wins” posture is clearly over.
OpenAI’s Jalapeño, one day later. Following yesterday’s unveiling, more detail has trickled out — TSMC is the fab, Broadcom is claiming ~50% cost savings versus typical AI GPUs, and OpenAI is leaning hard on the “our own models helped design it” line that HN found mostly unverifiable. The skepticism we noted yesterday hasn’t gone anywhere: late-2026 deployment is a long horizon when quantization and architectural changes keep eating efficiency gains every quarter.
Unconventional AI bets on oscillators. Naveen Rao’s Unconventional AI released Un-0, an image generator built on coupled oscillator dynamics rather than standard neural nets, claiming a path to 1,000x lower inference power once they actually build the chip. The model hits FID 6.74 on class-conditional ImageNet 64×64, which is respectable. But the HN thread flags two problems: it’s currently running as a simulation on conventional hardware (so no energy benefit yet), and the n² scaling makes higher resolutions look prohibitive. One commenter summarized: “finally, a way to generate images that’s slower AND worse.”
Patronus raises $50M for agent stress-testing. Patronus AI closed a $50M Series B led by Greenfield Partners to build simulated digital environments — replicas of real websites and internal systems — where AI agents can be put through reinforcement learning loops to surface the shortcuts and silent failures they take on complex tasks. Revenue is up 15x year-over-year, with frontier labs as customers. The pitch is essentially crash-test dummies for agents before they touch production.
General Intuition raises $320M to train robots on video games. General Intuition, spun out of gaming clip platform Medal, raised $320M at a $2.3B valuation on the thesis that gameplay footage paired with exact button-press labels is a better training signal than video alone. The demo: a quadruped robot navigating their office using a model trained on gameplay, fine-tuned with just 8 minutes of real outdoor data. The CEO argues most competitors are inferring actions from video, which loses precision their action labels preserve.
Claude is gaining paid consumers, slowly. Indagari credit card data shows Claude’s paid consumer base has grown ~75% since January, and DataCamp says Claude course searches are up 18x in the last 30 days — outpacing ChatGPT course demand three-to-one. ChatGPT still dominates in absolute paid-consumer numbers by a wide margin, but the curve is interesting given Claude’s historical lean toward developers and enterprise.
AI2 looks at where hybrid models actually win. A new AI2 token-level analysis compares OLMo 3 against OLMo Hybrid under matched training conditions and finds the hybrid edges out on semantically meaningful tokens — nouns, verbs, pronoun resolution — while pure transformers stay ahead on verbatim recall when the answer appears earlier in the input. A useful explanation of why hybrid architectures keep tying transformers on benchmarks while feeling subtly different in use.
That’s the morning. The Anthropic-Alibaba fight and the GPT-5.6 delay are both going to keep generating heat — watch whether other labs get similar government calls this week.