From Business Insider · Katherine Li · · 1m 30s
Claude now leads a quarter of Anthropic's AI research
Anthropic's chatbot is doing more of the work to build its own successor, raising questions about AI safety and control.
In brief
Anthropic's chatbot is doing more of the work to build its own successor, raising questions about AI safety and control. Claude has jumped from doing almost no AI research work in March to leading 26% by September, marking rapid progress toward systems that could improve themselves without human intervention…
Claude now leads a quarter of research
Anthropic says Claude completes 26% of its AI research and development work with human supervision, up from below 1% in March.
“Claude now "leads" 26% of its AI research and development work”
What 'leading' research actually means
Claude can complete most of a task from a high-level prompt while humans supervise. The company says 90% of work is at or above 'AI collaborates' level.
“Claude can complete most of a task from a high-level prompt while a human supervises”
The path to self-improving AI
Anthropic describes recursive self-improvement as a model fully autonomously building its successor. The company is tracking how close it is to this milestone.
“a model fully autonomously building its successor”
Safety concerns drive transparency
Anthropic CEO Dario Amodei called for frontier AI companies to coordinate on slowing development to give safety measures time to catch up.
Anthropic's safety monitoring in action
Automated monitors reviewed over 1 billion decisions in August, blocking about one in 47,000. Roughly 50 high-priority cases are escalated for human review each week.
“0.002% of more than 1 billion decisions analyzed in August, which amounts to about one in 47,000, were blocked”
Calling for industry-wide transparency
Anthropic says other frontier labs should publish similar metrics and allow third parties to verify them, giving society visibility into AI development.
What matters most
Claude has jumped from doing almost no AI research work in March to leading 26% by September, marking rapid progress toward systems that could improve themselves without human intervention.
