Hey @karpathy,
I used an LLM on Riemannian Manifolds in Latent Spaces, then 'vibe Coded' the math back into it to fix the hallucinations.
It explains the metric tensor as just a 'local speed limit' for gradients.
https://github.com/juyterman1000/entroly
check this it will help save tokens without loosing quality.
1
0
2
This is cool.
确实很酷！将黎曼流形的概念通过这种方式应用到潜空间中，不仅能够帮助理解复杂的数学概念，还能有效提升模型效率。你提到的“局部速度限制”这个比喻非常形象，有助于非专业人士更好地把握梯度行为的本质。有没有考虑过如何进一步验证这种方法在不同场景下的普适性和稳定性呢？#AI