Content
@
0 reply
0 recast
0 reaction
July
@july
For the foreseeable future, I think that LLMs will continue to: - improve significantly on what can be measured - struggle on what cannot be measured
10 replies
0 recast
63 reactions
July
@july
In many ways, I think the limit is what can be measured. In order to do better on what cannot be measured, we will need to convert more of what cannot be measured to what can be measured, and I think we continue to vastly underestimate what cannot be measured because it is an unknown unknown (we have no idea)
4 replies
0 recast
20 reactions
downshift
@downshift.eth
been thinking on this a bit recently…we need models to operate on a substrate besides language for this to happen probably some substrates need to be domain-specific (chemistry, for example) (cc: @swabbie.eth @rjs)
1 reply
0 recast
3 reactions
Ryan J. Shaw
@rjs
Relevant: https://x.com/karpathy/status/1835024197506187617
1 reply
0 recast
2 reactions
downshift
@downshift.eth
extremely 🎯 modeling domains in a more constrained way could lead to much more powerful models for specific use cases (like writing functioning software) thinking about it another way: some domains don’t compile to English very well at all (like complex geometry) but are easily represented to computers in other formats
1 reply
0 recast
2 reactions
