| ▲ | layer8 3 hours ago | |
Because LLMs are more like a lucky discovery than a deliberately designed and constructed creation. We build and train LLMs by trial and error, not so much based on the hypothetical predictions of an LLM theory. We can marvel at the discovery, but as an achievement it feels a bit hollow. | ||
| ▲ | jrflo 27 minutes ago | parent | next [-] | |
The scientific method is built on trial and error. There are absolutely theories in machine learning that future models are built upon (i.e. scaling laws). It's just an empirical science similar to biology, materials science, and medicine, rather than a first-principles science like physics. | ||
| ▲ | cheevly an hour ago | parent | prev [-] | |
Profoundly inaccurate take, like most comments on this page. | ||