People of AI

Contains genuine technical substance - WER history, the cascade vs. native audio architecture distinction, full-duplex/open-mic research goals, and latency targets - but the useful signal is heavily diluted by host restatements, filler affirmations, and conversational padding throughout.

“what we've been calling a cascade system. And now it's multimodal. You clobber the entire thing into one system and it's just audio in and it's audio out”

“it was something like 20% which is 1 out of 5 words at the time...it got to like 10%, uh, 8% and then for Google Home, at some point that was like, okay, this is like completely usable, 5%”

The Astra memory bug anecdote and the user demanding Astra inside Google Maps navigation are genuinely fresh practitioner observations, but the overarching thesis - LLMs enabling human-level fluid dialogue - is widely circulated and no contrarian or first-principles argument is advanced anywhere in the episode.

“it remembered from a past conversation that it couldn't see”

“she's like, this is great. I want this in Google Maps for navigation. And I'm like, what? Like, that doesn't make any sense”

Bibo Xu is a genuine deep practitioner - original Google Home PM, now lead PM on Gemini Multimodal and Project Astra at Google DeepMind - who has demonstrably done the work at scale across a decade; the conversation does not fully unlock her technical depth but she speaks credibly from the inside throughout.

“I was the original pm, uh, for Google Home, the first version of a lot of these assistant y kind of technologies”

“I joined Ah, DeepMind as the Gemini audio modeling PM”

The WER trajectory from 20% down to 5% and the 500ms latency target are concrete and useful, and named projects (Project Mariner, Android Computer Control, Gemini Live API) ground the discussion, but the vast majority of evidence is anecdotal with no user-research data, adoption metrics, or rigorous outcome numbers.

“word error rate...it was something like 20% which is 1 out of 5 words at the time...it got to like 10%, uh, 8% and then for Google Home, at some point that was like, okay, this is like completely usable, 5%”

“the goal is to get under 500 milliseconds for human dialogue”

The hosts ask reasonable scene-setting questions and occasionally identify a useful thread (interruptability, politeness norms), but they habitually restate or amplify the guest's answers rather than probe or challenge them, and the rapid-fire closing section produces no new insight.

“No, I mean, that's so cool. And that opens up so many different avenues for so many different people. Right.”

“That's amazing. You're getting your MBA, you happen to kind of go into this linguistics kind of class and then it ends up kind of changing, is it fair to say kind of change the trajectory”