|
|
|
|
|
|
by impossiblefork
597 days ago
|
|
|
No, I don't think it is really expensive in terms of hardware. It's probably only a hundred million dollars for renting computing resources. The problem rather, is that there are actually some difficulties with training very large LLMs still, and then there's data and data curation; and no one has reproduced something like O-1 (I think I know how O-1 works, but I haven't confirmed it), so I actually think it's knowledge about LLM training that is the difficulty. |
|
|