Post ATeJwra3Yzh9IQjPeq by ianchanning@mastodon.social
 (DIR) More posts by ianchanning@mastodon.social
 (DIR) Post #ATeFip3NlKTBaWUvDc by simon@fedi.simonwillison.net
       2023-03-15T18:36:23Z
       
       0 likes, 0 repeats
       
       The thing that's surprised me most about GPT4 is how much more finely grained knowledge it has baked into it than 3/3.5It knows the (real) addresses of restaurants in my area!You can ask for people's Twitter accounts and it often knows those - even roughly how active they are
       
 (DIR) Post #ATeGL1M2TaPwk7RG3k by simon@fedi.simonwillison.net
       2023-03-15T18:43:27Z
       
       0 likes, 0 repeats
       
       Here are two examples - one asking for the Twitter handles of three people (two real, one made up) and a second example that asks for popular restaurants in Half Moon Bay with their name, address and a pithy review
       
 (DIR) Post #ATeGWErGbtLzoEOViy by ben@bluetoot.hardill.me.uk
       2023-03-15T18:44:44Z
       
       0 likes, 0 repeats
       
       @simon The cut off date comment answers the question i was going to ask...
       
 (DIR) Post #ATeGireaCBS7bdTwA4 by simon@fedi.simonwillison.net
       2023-03-15T18:47:16Z
       
       0 likes, 0 repeats
       
       ... well that's embarrassing: I just tried the same thing against GPT3.5 and it did much better than I expected it to - though it did hallucinate a Twitter profile for Michael TurianskiSo this isn't actually much of a new GPT4 capability after all
       
 (DIR) Post #ATeGwugTizVB3k4FXc by SnoopJ@hachyderm.io
       2023-03-15T18:50:13Z
       
       0 likes, 0 repeats
       
       @simon I really do wish we knew what the number of parameters is but I suspect it hasn't been reported for pretty much this reason
       
 (DIR) Post #ATeHLsjA8XRYELEFdo by simon@fedi.simonwillison.net
       2023-03-15T18:52:45Z
       
       0 likes, 0 repeats
       
       @SnoopJ It turns out 3.5 can do the same trick, and we know that was 175 billion
       
 (DIR) Post #ATeHn9OkGOsFkhgK5Q by SnoopJ@hachyderm.io
       2023-03-15T18:58:44Z
       
       0 likes, 0 repeats
       
       @simon right, my suspicion is that 3→4 may not be as much of a big step in terms of #params as 2→3 was, but there's nothing to base that off of other than the noticeable omission of it and the similar performance on some tasks.Seems possible that the lion's share of the "more bigger-er" effect at work here is in the corpus/training rather than the architecture
       
 (DIR) Post #ATeJwra3Yzh9IQjPeq by ianchanning@mastodon.social
       2023-03-15T19:23:46Z
       
       0 likes, 0 repeats
       
       @simon the power of having an edit button 😂 Still good find in GPT 3.5 anyway.