Anthropic’s Sept. 17 report on its internal AI research and development rated Claude in the “lead” category for 26% of measured work in August and found no task fully autonomous.
The report also recorded AI involvement at Level 3 or higher in more than 90% of tasks. Those figures describe AI’s role in Anthropic’s own work; they do not compare AI performance with that of human researchers.
Andreessen Horowitz co-founder Ben Horowitz said a cycle of models creating improved successors would depend on four conditions: lower computing needs, faster training, stronger capabilities and continued gains in later generations. “Zero of those four have been met,” he said in a podcast episode released Oct. 16.
He also cited heavier training demands and constraints on GPU supplies and electricity. For a process to count as full recursive self-improvement, a model would have to create its successor independently.
The Anthropic assessment measures models’ roles in research tasks. Horowitz’s four conditions describe a higher threshold: models driving a continuing cycle of independently created successors.
