AI Is Starting to Build the Next Generation of AI
Google's Gemini 4 Neon and OpenAI's GPT-6.2 Astra and Sol are competing on more than model intelligence. Both are pushing toward better task completion with fewer tokens and lower costs.
But the more consequential development is in OpenAI's new system card.
GPT-6.2 Astra is the first externally released model OpenAI rates at the High threshold for AI self-improvement. Its internal evaluations show it can save 42.9% of baseline training time on a small-model training task and close 74.8% of a performance gap in a post-training task.
This is not autonomous recursive self-improvement. Human researchers still define objectives, select ideas and judge results.
Mark Chen and John Schulman identify the remaining bottleneck as research taste: knowing which ideas matter, not simply executing experiments faster.
My thesis: AI research itself is becoming a new source of compute demand—and a new productivity multiplier.
The economic test is not tokens generated or experiments run. It's the cost of producing a verified improvement in model capability.
As research becomes more automated, independent evaluation, reproducibility and oversight become more valuable, not less.
The next AI cycle may be driven by systems that help build their successors. But the ability to verify those improvements will determine how fast that cycle can safely compound.
#stocks
$GOOGL.US