AI Is Not Inventing Itself Yet. It Is Making Researchers Faster.
Meter estimates roughly double the pace at Anthropic over the past year, most likely from humans using AI.

The a16z Show · Can AI Learn to Improve Itself? | Recursive Self-ImprovementThe phrase self-improving AI sounds like machines inventing their own successors. Right now it means something plainer, and more useful: AI helping human researchers run their experiments faster.
John Schulman of Thinking Machines, an AI startup, defines it broadly as using AI to assist and improve research. Mark Chen of OpenAI calls it an accelerant, work that shifts the curve for training and testing models.
In practice, the help is narrowly scoped. A researcher tells a coding system called Codex to run a particular experiment in a particular way, or to watch a set of experiments for specific results. The system then writes the code, runs the test, and reports back, without being asked to invent its own research agenda.
Schulman described what that looks like day to day. He writes a long, detailed prompt for an AI helper, lets it work for days, and gets back reasonable experiments and interesting ideas that still fall short of the goal.
Progress comes when he steps back in. By coding himself, reading the results, and asking questions the system would not have thought to ask, he said, he gets much further.
That gap explains a split estimate from Chen. He put AI at about 80 percent of the way along the path to general intelligence, but only 25 to 50 percent of the way toward self-improvement defined as models feeding back new ideas on their own.
One outside estimate, from Meter on November 5, put the pace of progress at Anthropic at roughly twice the rate of a year earlier, most likely because human researchers were using AI rather than because research was fully automated.
Questions people ask
What would full recursive self-improvement look like?
It is the idea that an AI system could improve the research process that builds better AI systems, creating a loop. In the full version, models would come up with new ideas, test them, and use the results to make stronger models without much human help. Researchers say we are not there yet.
Why is human judgment still the bottleneck?
Schulman and Chen say models are still weak at what they call taste, picking ideas that will matter over the long run. Schulman says in the discussion that intuition like that comes from personal experience doing research and reading past work in the field. Automating experiment code is further along than automating that choice of what to try.
Does a twofold speedup really matter that much?
As an illustration only, rather than a finding in the source, halving the time each step takes would let a team try more ideas in a year. Whether that kind of gain compounds depends on having good ideas to test and good ways to judge them. That is why researchers are focused on whether AI can learn to propose ideas, not only carry out instructions.
