Aaron Levie
speaker
937 appearances
2 recordings
1 series
first heard Jul 2025
last heard 5 Sep
Aaron Levie’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Sep 2026 with 1.
Appearances
Mm.
Are
Yeah, so um I don't think we've we've uh like so we have internal benchmarks that deal with kind of general knowledge work and I would actually agree with you that that probably due to the cost difference, um you you'd probably argue that that um uh that that Opus five is now kind of an improvement over Fable because of the the the cost Delta.
Um on certain coding tasks um that that we had tested internally, Fable completely out outmatched uh four eight, like
Like in a in a in a way that you couldn't make up for by the by the you know token cost difference.
Um uh just like like being able to have superior solutions to hard technical problems.
Sorry, four A or Opus five?
Uh prior to having access to five, uh if we kind of compared like how big of a leap five was on on that, I don't have enough internal coding tests to to kind of compare against.
So so I had to see if like if how much is four eight meant to be, you know, kind of a superior coding um uh model versus a lot of the benchmarks that they just came out with were these kind of general knowledge work, um, kind of agentic tool use type benchmarks.
Um, so so I think it it kind of open question.
Um the one problem that Fable had and everybody's kind of tweeted about this already is uh is it will often, you know, again, kind of push you down to four eight uh for certain capabilities.
And then that becomes a problem because you're not, you know, you it's kind of hard in advance to know like did the did the thing kind of secu you know, trigger some security warning because it was looking through permissions, you know, access control code.
And so it kind of freaked out and and
And you know, isn't then giving you kind of fable level of intelligence.
Um, so I do think that Enthropic's gonna have to kind of work through that.
Um uh I've I've heard of uh, you know, folks in the biospace that that effectively can't use Fable um because it it it just again pushes their queries down um you know too frequently.
So you're not getting the the raw intelligence from the model.
Um and so, you know, I think I think we have to like I I I appreciated
Anthropic's ultimate proposal on on how we should kind of work through the these types of dynamics with the government, have a have a kind of a multi-point framework where we kind of agree on what what are these kind of uh capabilities that that are of different risk levels and test the models for them.
But there's there is kind of an interesting question though, which is like, well, what who gets to decide that risk framework and um and and how do we agree that the risks are actually, you know.
Showing 181–200 of 937 · page 10 of 47
← Previous
Next →