That's impressive How much of the code did you write or fix yourself And has it been a pain to maintain or has AI been pretty good at handling that too?
I wrote under 5% of the code by hand. It was mainly telling the AI to build x and how to build it. With the new Fable 5.1, even the how part is getting replaced.
For this project, the maintaince is not too big of a pain, but we are working on another project where it's an absolute pain. Every PR is like 100k lines change, simple bug fix needs 600k tokens. We are probably doing a choppy job here.
At this point, the bottleneck isn't necessarily raw compute memory bandwidth, power, cooling, and getting workloads to efficiently use that much unified memory become just as important.
I think there's an important distinction between restricting who is allowed to work on mathematics and asking for responsible disclosure of AI-generated results.
Nobody can meaningfully control whether someone works on a mathematical problem. But if a model produces a potentially significant result and the authors are asking others not to immediately publish it, I can understand the motivation: independent verification, attribution, and avoiding a flood of unverified claims are legitimate concerns.
The harder question is what a reasonable disclosure process looks like. For example, requiring reproducible evidence and independent verification seems much more actionable than simply asking people not to test models on certain problems.
AI-generated posters only look horrible when lazy prompts force the tech to guess, but with the right human direction, strict styling limits, and clean layouts, they can be genuinely striking.
