Critics question Anthropic's training traceability: incompetent or unwilling to check?

davidmanheim · x · 2026-09-09

In a thread about Anthropic's training timelines, a researcher argues that if they continued training through a later date and claimed to use a model trained on the data, then either they lack training safety competence and traceability, or they are unwilling to verify. The thread also notes it is now "generally known" that Anthropic runs many internal fine-tuned model variants at once—including non-Astra variants discussed in connection with the Hugging Face hacking incident.

Original post →

More from Models

Models channel →