Moritz Sudhof
speaker
399 appearances
1 recordings
1 series
first heard Jul 2026
last heard 8 Jul
Moritz Sudhof’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 Jul 2026 with 1.
Appearances
Don't treat those just as um, you know, I'm I'm gonna send a little bit of feedback.
Um, actually from those I d you know, think about yourself as or the user as the human in the loop that's already doing evaluation, that's already telling you implicitly what do they want, how are they thinking about this thing.
Um, and any anybody in any organization that has access to their own transcripts or to a couple of test users transcripts, they can turn insights from that from just reading those into a clear set of behavioral requirements that they can sh um share with the engineering team.
Um, now in language that the engineering team will understand in terms of like, hey, this is these are decision behavioral decisions that the AI is making in these sessions that are leading to bad user experiences or bad user outcomes for XYZ reasons.
Um and if you can turn that into a list of requirements for a prompt or a list of eval cases um for the eval system, um you can actionably you can have a very actionable um uh input and uh impact onto how uh not just how you the engineering team thinks about the product, but actually the data assets that they have to improve it um and to uh
evaluate whether it's succeeding.
Because a big problem with AI and especially AI products where the the technical teams are um you know are very heavily involved and there's not a lot of UX designers or business stakeholders who are in the loop.
is that the definition of good for an AI product cannot just depend on technical metrics.
It cannot just depend on, you know, guardrails and latency and um and and response thoroughness or kind of generic metrics.
The definition of good has to come from business requirements, product requirements, real user, um, user needs that we have.
Uh, and those are most of the time not represented in the technical system.
And they they need people who are sit outside of the technical system to surface them.
And I I think the the single greatest exercise any team can do is take a single transcript, sit down in a room, look at that transcript together.
And I will guarantee you that if you have people with different job titles in that room.
They will all notice something different about what the AI could have done better or what they think the user actually needed in that moment.
No, no single person is right or wrong.
It's that we all have different expertise, we all have different perspectives, we all have a different um context on the user and on the business.
And it's only through taking all of those signals and all of that input together that we can actually understand how to build a great AI product.
Um, and so there's no substitute for just getting more and more business stakeholders, non-technical stakeholders to actually look at what's going on with the AI product, what what are users getting or not getting, and then to bring those learnings back to the technical team.
Showing 381–399 of 399 · page 20 of 20
← Previous