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
So for example, confidence trap.
That's when the AI says something very confidently, but it's wrong or made up.
And so there's a capability aspect to that where if the AI gets smarter and gets better at using web search or is more grounded, like it will make fewer factual errors.
the dangerous and insidious part about the confidence trap, it really isn't
the that it was factually wrong or that it it stated something that was incorrect.
It's how it does it.
It's how it does it with extreme confidence, how it does it in a wall of text that looks very reasonable, how it will cite very precise numbers so that you know you see like 2.68%, you see that number as a human and it feels precise, which means it feels rigorous, um, which means it feels like it must come from somewhere, right?
And so it's the confidence trap is what hap what happens when the LLM does something that asserts something that is maybe not true or not quite what the user asked, but does it in such a way that the user is unlikely to catch it because it's dressed up in all of this formatting of competence and and kind of performed precision and rigor.
Um, and the so like the more specific and helpful it often gets, the more fake it often is.
I think what's difficult about the confidence trap, especially for users catching it themselves or users recovering from it.
Is especially as we scale AI to
broader populations that are not as
versed in how to use AI as a tool that aren't as fluent or familiar with it.
And especially as we deploy it in areas where users are not themselves experts.
They're reaching to AI to help them in something that they don't have expertise in.
When those two things come together, you have situations where users cannot verify what whether what the AI said is correct, like they don't have the expertise to.
They also are less likely to try to verify it or to push back because newer users who are newer to AI and in the broad kind of the biggest cohort of users.
um that are starting to use AI now, they often treat AI very from a very delegative perspective.
They almost treat it like a vending machine often, like I will tell you exactly what I want, or I'll tell you what I want and then I'll get something out and I'll I'll leave.
And that's exactly the kind of user posture that means that these confident misstatements end up just getting accepted and end up getting taken by the user.
Showing 141–160 of 399 · page 8 of 20
← Previous
Next →