IMO the coolest thing about ACT-1 is that we got it to look up stuff it doesn't know. You *could* try googling this but it doesn't work for me, and there are lots of other reasons to be excited about what this means for future models: ⬇️

Sep 15, 2022 · 7:08 PM UTC

4
19
5
141
0
First, it's just pretty sophisticated behavior. The model has to search for star wars, disambiguate to the movie, see that Mark Hamill plays Luke, and then look for his age before giving a useful response.
1
9
Second, it stops the model from having to make stuff up. If it can query an up-to-date tool, you don't have to worry it won't know about new things that have happened since you trained it.
1
1
7
Third, we can get the model to do things like this as an inner-loop in order to improve behavior on some higher level task, e.g. by looking up documentation when it's confused about what to do!
1
6
Fourth and finally, this *already* seems like a thing that Google can't do reliably. I'm sure there are permutations of the query that will work, but models keep getting better...
2
7
Replying to @gstsdn
Thinking similar thoughts.
NirSD
 @nirsd
13 Sep 2022
Who's building the Google Search-killer with LLMs as the search interface? Google Search returns overly SEO-ed 2 page answers for the simple questions, users hate it but have to tolerate it because the friction to improve your search is very very high.
1
2
The friction in finding information has been steadily increasing, ACT could easily be better than the naive user at understanding intent and finding it.
Replying to @gstsdn
I read that as ACT-R at first and got *super* confused about how it got that much more capable in the last five or so years
Replying to @gstsdn
Good luck competing with plug-in Gpt bud