Language Log

« previous post | next post »

Marc Andreessen's recent appearance on Joe Rogan's podcast presented a striking example of AI promotion (or AI hype, as you please). We can discuss his extraordinary claims and predictions another time. My topic this morning is something Andreessen does that AI still can't do, namely talk like a human being. I'm referring to the way that humans talk in spontaneous conversation, not in fluent reading or in well-rehearsed presentations, which AI text-to-speech can imitate increasingly well.

I brought up a similar point a couple of years ago, in "Spontaneous SCOTUS"  (3/2/2024). The background there was an effort to recreate U.S. Supreme Court oral arguments, from transcripts made before the court starting making tape recordings in 1955. Actors read the transcripts, voice morphing was applied, and the result was a success.

But a problem remained, as the cited post notes:

There are many things besides individual vocal identity that transcripts leave out. And even a skilled actor reading a standard transcript will not put most of these things back in, neither in a generic way nor in a way that's faithful to the original.

These omissions include things like

• Filled pauses (uh, um)
• Filler words (you know, I mean, so, like… )
• Silent pauses (and pseudo-pauses)
  not in a reading-style relation to message structure
• Rapid initial repetitions (in- in- in the- the, …)
• False starts (“that was my= uh the last time”…)
• Non-speech vocalizations (laughs, sighs, grunts, … )

Such things are usually called "disfluencies", but I've argued that this is a misleading term, partly because they're typical of fluent spontaneous speech, and partly because  they often have communicative content.

And whether or not they have communicative content, these phenomena are common in effective spontaneous rhetoric from persuasive speakers like Paul Krugman and Elon Musk, as well as U.S. Supreme Court justices.

I've therefore suggested that the term "spontaneities" should be substituted for "disfluencies". And from the very beginning of the Rogan interview, Marc Andreessen show us that he's a fluent spontaneitizer:

yeah so these guys are driving around in cars
and yeah they're switching cars, whatever
they- yeah and they're- and they- they went to like a dozen locations
and like fight- and tri- tried shooting-
shooting at buildings and people and houses and- and all kinds of stuff and so
okay so these guys running around so they-
there's this system called Flock
which is one of our companies and- and what they do is-
kind of like in the movies you- you take all the municipal cameras
and traffic cameras and everything
and you feed them into an A- AI and the AI is able to
first find a license plate
in- in real time so you can- you can find that
but- but second you can actually find a car even if you don't have the license plate
you can find like distinct markings of the car
it'll- you- on the car it'll track the car
and so this- this thing is deployed it's- this- it's sold to city governments
it's used all over the country

Here's another example, from a 6/27/2024 interview, illustrating that Rogan exhibits the same sorts of spontaneities that Andreessen does:

A: well the i-  the- the internet idea was all the computers are peers
right so there's no- there's no single node
right and so there's just four computers that talk to each other
which was the basis of what the internet is today
four computers talk to each other now it's four billion computers talk to each other
but it was that same idea
R: and h- did they store things individually
like did you have access to each individual computer's data?
or did they have a collective data base?
A: i- i- it- they- you know they had a combination- I mean I- this is very- original- the- these- these were very simple systems as compared to what we have today
so these were very basic implementations of these ideas
that they- they would've- they had very simple what's called store and forward email
uh they had very simple what's called file retrieval
so if there's a file on your computer and you wanted to let me download it I could download it
they had what was called telnet where you could log into somebody else's computer and use it
R: So you are messing around with this stuff and you guys create- was it the very first web browser or the first u- like used by many people web browser?
A: yeah it was the first- it is a productized s- uh it was the first browser used by a large number of people
um it was the first browser that was really usable by a large number of people
um it was also one of the- one of the first browsers that had integrated graphics
the- the actual first browser was a text browser
uh the very first one which basica- was a- was- which was the prototype
that Tim Berners-Lee had created
so- it- but it was- it was just very clear at that point like
we- we now have the g- we have- we have Windows we have the Mac we have the GUI
right we have graphics like
and- and then we have the internet
and we need to basically pull all these things together
which is what Mosaic did

I have no doubt that AI systems — even the simplest "stochastic parrot" forms — could be trained to talk that way. But the underlying mechanisms would obviously be very different.

June 4, 2026 @ 8:48 am · Filed by Mark Liberman under Artificial intelligence, Prosody, Psychology of language

Permalink