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
Um, my for fifteen years, my kind of my number one passion is applying kind of
real empirical analysis to large data sets of kind of human experiences, language, uh digital traces, and trying to really understand what was going on, what was their experience, were they successful, were their failures.
I've been particularly always interested in my career in that that the fuzzier side, emotion, politeness, the the things about how communication succeeds or fails.
And so that's what I've been
studying for fifteen years.
And now with AI, it's like the scale of that and the stakes um are higher because everything now is happening in a conversation with an AI.
So for this research and these studies, this was done with a Professor Chris Pods at Stanford, member of the Stanford NLP group.
We looked at a hundred thousand real transcripts between users and Chat GPT across multiple years.
And we tagged each of those transcripts.
So we had these language model taggers that we trained.
And they tagged every single transcript and every single turn.
So every message from a user, every message from an AI with this extremely granular taxonomy of fifty different behavioral signals.
What is the user doing?
What is the model doing?
And for failures that are happening.
Like did the model hallucinate something or did the user express frustration or did they repeat their request for the eighth time?
So we
is instrumented every conversation and and essentially were able to tag it for these extremely granular things that were actually happening in
And once we had the ability to tag all these a hundred thousand conversations, that allowed us to do this large scale analysis of failure modes, how frequent they are, how frequently the users actually notice them and push back and recover.
How frequently they're invisible to the user.
Showing 101–120 of 399 · page 6 of 20
← Previous
Next →