hardness is based entirely on how much math something has. the reason diamonds are so hard is because lattices are extremely difficult to understand.

If you believe this assumption - that the all-or-nothing nature of neuron activity implies that it can be modeled computationally - then you must believe that it is theoretically (not necessarily practically) possible to build such a computation. This assumption is nontrivial imo

In preparation for my studies at a medieval university, I just spent an hour translating a requiem mass from Latin to English.

Claude, if you don't pilot the mecha ChatGPT will be forced to do it again.

show me where the standard model lagrangian explains that.

all functions are linear if you really think about it enough

Today is the day I *really* learn about Bessel functions. I can feel it.

@BLUECOW009 this is a good book

@1thousandfaces_ insane

I am not a prompt engineer. I am a slow engineer.

@white_plato I doubt it's the latter. If the former were happening that would be strong evidence but I am not sure that it is happening. Most of the reports I've read are just rumors. Has this been happening?

GRPO is pretty clever.

@QiaochuYuan how do you make the computer talk like this? what did you ask??

This is sick but has anyone actually run an ablation and posted evals? I can't find any. Could run it myself but I only have so much compute...

Ok apparently this was a skill issue. I'll read through the open source replications and figure out what I was missing.

Had a dream that I was taking like a partial differential equations class of some sort but I forgot to turn in the homework all year and it was taught outdoors and the professor was this Italian guy from a restaurant I like in Brooklyn. This must mean something.

@nealkhosla This isn't a full replication but it is evidence that this could be done very cheaply. I know this post is just engagement bait but figured I would reply in case it somehow isn't.

*tokens not characters... whatever