| ▲ | hansvm a day ago | |
My best guess is it's sincere. That's been a core observation of LLMs for awhile. Filtered random search is basically an inefficient gradient descent. If you apply even a poor-quality "gradient" to a lot of problems with infinite compute (with large LLM budgets), it doesn't take that good of an LLM to accomplish great things from time to time. | ||
| ▲ | foltik 19 hours ago | parent | next [-] | |
“Shine light on a random clump of atoms for long enough and it shouldn't be surprising that you get a plant.” | ||
| ▲ | TheOtherHobbes a day ago | parent | prev [-] | |
We've graduated to motivated stochastic parrots. | ||