openai: ai now reasons for days, weeks are next

openai researchers dropped some signals worth paying attention to

sebastien bubeck and ernest ryu joined the openai podcast with andrew mayne. here's the trajectory they laid out:

• 2 years ago – ai thinks like a high schooler on a problem for 2 mins
• today – ai thinks like a researcher grinding for hours or days
• next – weeks, then months of sustained reasoning

that's the automated researcher arc. it's already started

why math matters beyond math:
- broken reasoning in math = wrong answer, full stop
- models trained on math learn to catch their own mistakes
- that skill transfers everywhere: code, decisions, agent workflows

what this means if you're building with ai:
- agents can hold complex problems longer than ever
- models are getting good at asking questions, not just answering them

so reasoning is not going away. the rumors were wrong. openai and anthropic are partnering with compute providers specifically to handle more complex agent tasks. both labs are doubling down on reasoning, not replacing it. which also means the cpu bottleneck we've been tracking is very much still on the horizon