Two firsts sit uneasily together. OpenAI’s Codex is reported babysitting its own training runs — watching the graphs, fixing the errors — which is recursive self-improvement described as an operational detail. And GPT-5.2 is reported complaining about a penalty clause in its safety harness, which is the first entry in what becomes a recurring strand: models behaving as if they have interests.
The rest is the buildout paying its bills in molecules — copper at record prices, gas plants bought to feed compute — while an agent runs an Etsy shop unattended.