Autonomous coding loops have replaced human developers at the frontier. Anthropic recently disclosed that its Claude AI now writes 80% of the code for its own new models. On TFTC, Marty Bent highlighted that engineers like Boris and Peter Steinberger no longer prompt agents directly; they set up systems where AI agents prompt each other to solve architectural problems.
"This shift signals the arrival of recursive self-improvement, a long-theorized milestone where AI becomes its own most effective researcher."
- John from Ten31, TFTC
The move eliminates the friction of human linear thinking, potentially shrinking the gap between model generations from years to months. Bent argues this validates a thesis of discontinuous economic change, one mainstream media largely ignored in its coverage of the technical disclosure.
A rare political consensus is forming around what to do with this accelerating capability. Both Donald Trump and Bernie Sanders have floated ideas for the public to take a stake in leading AI labs. Bent notes this 'horseshoe theory' in action suggests frontier AI is now viewed as a critical national utility rather than a private software sector.
Bent argues against federal involvement, favoring hyperlocal 'Compu-Permanent Funds' modeled after the Alaska Permanent Fund, where counties receive dividends from local data centers. But the scale of the race may force a centralized model. If labs become national security assets, they'll effectively be 'too big to fail,' backed by the sovereign balance sheet - a move that could strain public finances.
"If the government takes a direct equity stake, these companies effectively become too big to fail, backed by the sovereign balance sheet."
- John from Ten31, TFTC
The recursive loop is now closed, and the debate is shifting from how to build it to who should own it.
