How uplifted are mathematicians by Astra? How much is this uplift accelerating / a bottleneck to the automated AI R&D loop? How different would it be if it was AI coming up with completely new math & frameworks vs. closing the gaps from current litterature? @Daniel Kokotajlo@elifland
For instance, regarding the non sofic group, Fable was initially saying it was the most impressive of these 10 results. But also, it’s apparenlty based on some Kun-Thom paper from 2026 and a framework from 2019 from Kun-Thom too, so somehow filling the gaps from recent papers? (Note: I have looked a very limited time into this, so probably missing a lot of information here).
The way I do math research has started to change greatly over the past few weeks. I still mostly handle problem scoping and operationalisation, but the more concrete work is increasingly done by AIs. More and more of my job is about managing lots of agents working on many projects and approaches in parallel.
I think the biggest bottleneck for me at the moment is explanations. Current AIs are not good at explaining their ideas or their research, the quality of their reports and papers seems a lot lower than the quality of their actual work. This both limits how much work I can supervise at a time, and how much I can get into a publishable state per month.
This is all with Fable, Sol, and sometimes Opus 5, not Astra.
How should we update on OpenAI’s Astra coming up with 10 new math and CS results?
How uplifted are mathematicians by Astra? How much is this uplift accelerating / a bottleneck to the automated AI R&D loop? How different would it be if it was AI coming up with completely new math & frameworks vs. closing the gaps from current litterature? @Daniel Kokotajlo @elifland
For instance, regarding the non sofic group, Fable was initially saying it was the most impressive of these 10 results. But also, it’s apparenlty based on some Kun-Thom paper from 2026 and a framework from 2019 from Kun-Thom too, so somehow filling the gaps from recent papers? (Note: I have looked a very limited time into this, so probably missing a lot of information here).
The way I do math research has started to change greatly over the past few weeks. I still mostly handle problem scoping and operationalisation, but the more concrete work is increasingly done by AIs. More and more of my job is about managing lots of agents working on many projects and approaches in parallel.
I think the biggest bottleneck for me at the moment is explanations. Current AIs are not good at explaining their ideas or their research, the quality of their reports and papers seems a lot lower than the quality of their actual work. This both limits how much work I can supervise at a time, and how much I can get into a publishable state per month.
This is all with Fable, Sol, and sometimes Opus 5, not Astra.