Four proposed conditions for AI to repeatedly improve itself

The four-part test

Ghodsi’s framework requires four conditions to hold together:

1. The next model uses fewer resources, including fewer GPUs, to train.
2. Training the next model takes less time.
3. The next model is more intelligent or accurate.
4. Those three gains can be repeated “again and again and again.”

“All at the same time, not just any of them,” Ghodsi said in the excerpt. As an illustration, he described a future model that might take half the time and half the resources to train while also becoming more intelligent. That combination, repeated over successive generations, is the speedup he is interested in.

Why code generation is different

Ghodsi drew a line between that recursive cycle and AI writing software. He said AI already writes “90-some percent” of the software at Databricks, then argued that automating the remaining share would not by itself create the same kind of accelerating model-improvement loop.

In his framing, the two ideas answer different questions. AI-generated code changes how people produce software. Recursive improvement would require the next model to use fewer resources and take less time to train while becoming more intelligent, and for those gains to repeat.

Not necessarily superintelligence

Ghodsi also stopped short of claiming that satisfying the four conditions would guarantee superintelligence. “I don't even know if that necessarily leads you to superintelligence,” he said.

That caveat keeps the proposal narrow: it is a test for a repeatable improvement process, not a claim that any current system has achieved it or that such a loop must end in superintelligence.