The Loop Closed on Cost First

Scroll to the bottom of OpenAI’s GPT-5.6 release notes and there is a benchmark table headed “Self-improvement.” One row is labeled RSI Index. The new flagship scores 57.9. The model it replaced scored 41.7.

Recursive self-improvement spent sixty years as a thought experiment. It now has a scoreboard in a product launch, filed between the cybersecurity results and the multimodal ones. The coverage went to the price cuts.

Companies benchmark what they intend to optimize.

I wrote up why I think the loop is already closed, and why most people are watching the wrong variable. Everyone is looking for self-improvement in capability. It arrived first in cost, and cost is the highest-leverage place for it to arrive, because power is the binding constraint now. Roughly 2,300 gigawatts sit in US interconnection queues with waits running four to seven years. Tokens per watt is the only term anyone can move, tokens per watt is software, and software is what these systems are best at.

The receipts, briefly. GPT-5.6 rewrote OpenAI’s production GPU kernels and designed, trained, and supervised the draft model that makes it generate faster. Google’s AlphaEvolve has been recovering 0.7 percent of Google’s worldwide compute for over a year and put a circuit change into an upcoming TPU. Anthropic says Claude authored more than 80 percent of the code merged into its codebase in May.

The strongest argument against all of this is in the piece too. METR published a study ten days ago finding that autonomous agents are still not economically competitive with humans on a frontier-style optimization problem. I take it seriously, and I explain why I think it bounds a narrower claim than the one I am making.

Also in there: what would change my mind, and which of it is measurable inside twelve months.

Full write up: https://www.teamignite.vc/blog/the-loop-closed-on-cost-first