Builders shipping voice interfaces get a mechanistic account of why fluent audio models still fail at trivial deterministic tasks, and where the pipeline (not the model) is at fault.
“You think it's more likely that I'm a superhuman speed traveler than you got the time wrong.”
Julia Turc
“However, they're LMs at the end of the day, they just don't have this concept of time.”
Juman Mal
“Look, we put a man on the moon, so I'm pretty sure we can build a voice agent that sets a timer.”
Julia Turc
“the fact that you can't really control these LLMs very well, it just exposes a lot of potential dangers”
Juman Mal
“we have loads of engineers who will solve this problem. And that sounds like a solution. It's not an actual solution because they haven't solved the thing”
Juman Mal
Checking sign-in…
Loading comments…