Gregory Mostyn
speaker
336 appearances
1 recordings
1 series
first heard Jun 2026
last heard 29 Jun
Gregory Mostyn’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 Jun 2026 with 1.
Appearances
We don't want to be bogged down with, you know, like multiple levels, you know, tons of internal meetings.
I'm sure I think people come to an early stage startup because they have real agency and real ownership and they don't want to be sort of stuck in strategy sessions and meetings and those kind of things.
They want to be either building
or selling like i said um so yeah i think we're we're kind of at a point now where the next level up is going to have to require a reorganization but right now we're operating at an extremely high level and kind of bringing the best out of everyone so we're trying to keep it at that point before we sort of
US and we're hiring in the US and getting boots on the ground at the moment.
So I think that's going to be like the next level.
I think this article was inspired by a post by Bennett Evans, who's a brilliant technologist who has a really good perspective on AI.
I think it's a healthy amount of skepticism combined with how powerful this technology is.
I think his article was basically that anyone who's completely cracked the accuracy problem with AI is lying to you.
But that doesn't mean it's not incredibly useful.
So I think you need to have both of those things in mind when you're using AI.
So Wexler doesn't have like fabrications.
So what I would say is a fabric hallucination, we like to say internally is like if a human given the same information would not make this mistake, like a kind of intelligibility.
intelligent human then I would call that hallucination like it's an unexplainable situation where there's no sort of source etc and so that Wexler doesn't have that issue is it 100 accurate of course not and any any AI founder telling you it is is you know you need to sell them back the bridge they've just sold you it's really a question of like being able to very quickly independently verify the output checking it using it and then using that as you build up your output
In terms of practically how we managed to mitigate that process, I know I keep coming back to it, but it does come back to this fact extraction pipeline.
If you're not asking the models to do too much at one go, then you can massively mitigate context window issues leading to things being invented or fabricated.
we sort of say a to b to b to c to b to c sorry to c to d to e to f etc and stitch together rather than a to z in one go because in doing that process you're building up this kind of curated accurate database where it's not sort of having to grasp in the sort of fumble in the dark to find a piece of information which may not exist plus we have a huge amount of guard rails on the back end which is like only stick to that you know as you would expect and that's kind of what you can do without that kind of custom architecture plus the fact that everything
uh the sentence level and is you know easily verifiable within the platform so the workflow wexler would be to build a chronology and then we even have a verify button where you go through and you verify each of the facts that are in that chronology at the point of sort of finalizing it and that process of verification
doesn't just make sure it's accurate, but it also makes sure you've got your fingers on the documents.
Right.
Showing 261–280 of 336 · page 14 of 17
← Previous
Next →