GPT-6 Astra hands-on: compute is the bottleneck and CoT observability is eroding

thedealdirector · x · 2026-09-06

After extensive testing of GPT-6 Astra, the author draws four takeaways:

Model performance and outcome quality are accelerating.
Compute remains the primary constraint: GPT-6 trained on 100k GPUs, and balancing multi-billion-dollar training runs against inference serving is critical — especially since user interaction feedback loops feed training data.
GPT-6 is a very close step toward AGI, but it will incentivize other labs to obscure CoT observability; training next-gen models on current ones without understanding their logic brings us closer to uncontrollable behavior.
(Fourth point truncated in source.)

More from AGI Musings