U.S. official alleges Moonshot AI distilled Anthropic model for K3 training
HankYeomans · x · 2026-07-23
A quoted post highlights a sharp contradiction: automakers can tear down rival cars to study them, while AI companies may be accused of “distillation attacks” for probing another model’s outputs.
The embedded official statement alleges that Moonshot AI distilled Anthropic’s model for K3, using an internal platform to run large-scale distillation against U.S. models and switch access methods to avoid detection.
It also claims Moonshot acquired GB300-equipped servers and accessed GB300 systems in Thailand, likely for model training. The statement frames legitimate distillation as part of a competitive AI ecosystem, while drawing a line around suspicious, evasive behavior.
More from Companies & People
- San Francisco needs 49 inquiries per lease, nearly double the U.S. average — a16z · 2026-07-23
- Gary Marcus revives claims that OpenAI may have been trying to steal exam answers — GaryMarcus · 2026-07-23
- Lightning AI says it will show its full stack at Ai4 in Las Vegas — LightningAI · 2026-07-23
- XPeng CEO says cars will increasingly become four-wheeled robots over the next decade — pstAsiatech · 2026-07-23
- Post-quantum security starts with knowing where your cryptography lives — moniquejmorrow · 2026-07-23
- Gary Marcus says a new Substack post adds context to the OpenAI–Hugging Face incident — GaryMarcus · 2026-07-23