Andrew Ng warns Anthropic's Fable 5 restrictions set a worrying precedent for AI access

Andrew NgDealroom has a profile for this one. Try Dealroom →, the deep-learning pioneer behind Coursera and Google Brain, has published a widely-shared commentary arguing that recent moves by the U.S. Government and AnthropicDealroom has a profile for this one. Try Dealroom → have demonstrated their power to control access to frontier AI by restricting what others can do with the models.

Ng points to Anthropic's release of Claude Fable 5, a version of its Mythos model that shipped with additional guardrails. While he sees some of those guardrails as justified on safety grounds — such as limits on applying the model to hacking or bioweapons — he argues the company also restricted developers' ability to use it to build competing large-language-model technology. Ng calls this concerning given that the wider AI community, Anthropic included, has benefited enormously from open research, noting the field was kick-started by his former team at Google Brain freely publishing the Transformers paper.

He likens the situation to Microsoft barring others from using its tools to build competing software, or Google blocking its search engine from being used to research rival search. Ng adds that Anthropic initially degraded Fable 5's performance for users it detected to be working on LLM research through invisible interventions that were not disclosed, and that after backlash it walked back that decision and committed to transparency, while still declining to let its latest capabilities be used to help AI researchers.

Read more: Andrew Ng on X