#167 - GPT-4o, Project Astra, Veo, OpenAI Departures, Interview with Andrey - Last Week in AI

Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Andrey (00:09):
Hello and welcome to the latest episode of Last Week in AI, where
you can hear us chat about what's going on with AI.
As usual, we will summarize and discuss some of last week's most
interesting AI news.
And as always, you can also check out our last week in AI newsletter
at last week in that AI for articles we did not cover in this episode,

(00:31):
I am one of your hosts, Andrey Kurenkov.
I finished my PhD at Stanford last year, and I now work at a
generative AI startup.
And once again, we do have a guest cohost who?
Jerry being on vacation, so I'll let her introduce herself.

Daliana (00:47):
Hello everyone. My name is Dalia Liu.
In my past life, two years ago, I was a senior data scientist
at Amazon. I built machine learning solutions for,
Amazon customers.
AWS customers, also were to, experimentation
a b testing, some areas in a general data

(01:10):
science domain. I left my full time job to host
my podcast called The Data Scientist Show, talking about data
science project in different industries and their career journey.
I also advised data and AI companies, community
building go to market.
And recently I focus career coaching, helping

(01:34):
data scientists find, career paths.
That works for them.

Andrey (01:38):
That's right. So another experience podcaster and interviewer of
many people out there in industry, similar to John Crone from last
week. And, as we've chatted, maybe a slightly different
background with not as much emphasis on AI and machine learning.
So your kind of takes and, additions to a discussion,
especially this week, which is very heavy, kind of consumer news that

(02:02):
affects all of us and not just we technical folk.
Will be fun to hear.

Daliana (02:09):
Yeah.

Andrey (02:10):
Before we dive in to a news, as usual, I do want to give a
call out. Shout out to some nice feedback
we got on Apple Podcasts.
As always, I do appreciate reviews.
Someone said great content presented in a fun, engaging way.
We do try to make it fun, although sometimes it can be hard.

(02:31):
A bit more technical stuff.
Someone so good overview of AI hype that is also
appreciated. And yeah, just as always,
appreciate reviews. And we also got a few comments on YouTube,
including someone saying that AI theme song was soothing as
well. And I am going to start including AI theme songs

(02:54):
at the end of every podcast, because that seems pretty fun.
We'll keep our usual FEMA at the beginning, so if you are a hardcore
listener who goes all the way to the end of a podcast, well,
you'll get that nice, fun tidbit to look forward to.
And just a quick FYI of this time, we are going to
try and record the video of this.

(03:18):
Partially because we will have a bit of a chat at the end
where I'll be going into my background so you
can go and look for that on YouTube.
Hopefully I do wind up editing the video and not just,
you know, cutting all of this out.

Daliana (03:36):
Yeah. My task straight is to ask people questions about their
careers. So I have to do this. Angie.

Andrey (03:42):
That's right. Well, let's get into news, starting
with our tools and Apps section.
And our first story to chat about is, of course, GBG
four. Oh, so on Monday, OpenAI had
a little video stream where they announced their
latest model, GPT four Omni, which is,

(04:06):
iteration on GPT four, natively trained to accept
audio as input and also output
images and text. And they had very impressive demos
of what it is capable of.
So they had this, like 20 minutes stream in which they showed
basically real time interaction via voice with this assistant.

(04:29):
Very much many people compared it to a movie her.
And I think that's pretty appropriate.
So you can hit a little button, the microphone goes on
and you can talk to AI and without almost any latency,
it processes your speech input and produces a speech
output that is very humanlike.

(04:52):
The intonations of voice, emotion, etc., are
very high quality, even maybe beyond what you've seen.
With typical speech synthesis.
And this is coming with all the intelligence of ChatGPT and GPT
four models before it, in fact, on benchmarks.
And I'll some of the technical numbers and so on, they,

(05:13):
say that it's even better across the board on all
these tasks.
And even beyond all that, they also
announced that this model would be, half
the cost and twice as fast as GPT four.
So that's exciting for me as someone and working with it and using the

(05:35):
API, like, that's a big deal to have GPT four quality
intelligence at twice the speed and half the cast.
So very, I think highly discussed.
And I thought very impressive.
Progress here from OpenAI.
I, I'm curious what you thought.
Diana.

Daliana (05:55):
Yeah. I watch some, demo
clips. I think the real time translation looks really cool.
And, one feature I'm very excited to try.
I'm not sure if it's available right now.
The, the capacity to view your
screen on your desktop.

(06:16):
I think that would be really cool to have, AI
companion to see what I'm working on.
Maybe call me out. And, I don't know
if you feel this way sometime.
It's very lonely to work on things by yourself, but
if you do something like pair programing or,

(06:36):
I try this. Just have a friend do some co-working with me.
Silent. I'm more productive because it just feels like, someone's
watching, so I think I might be more productive.
Maybe you can just chat with it.
Is it it her? Him?
I think that's a question. That's a feature I'm excited

(06:58):
about. On my phone, I also
played with the audio.
So it's, I think in the demo, it shows a female's voice.
I think they provide a few options.
They have males voice. Different type of a male's voice as well.

(07:19):
But I guess the default is female voice, because,
I think a female, virtual assistant sounds
less threatening psychologically.
Yeah. Would you take. Would you would you, if OpenAI developed
something like her, would you would you date an AI?

Andrey (07:39):
Well, many people already are, things like replica and many other
apps. And that's it's all whole own question.
And it's true that with the ability to do real time chat
and have is very, it's true that the female
voice in particular highlighted in the demo, was very friendly

(07:59):
and warm and kind of non-threatening.
And, some people even characterize it as sexy
or, you know, similar to, again, the movie her.

Daliana (08:11):
Yeah.

Andrey (08:12):
The the plot of a movie is that the main character falls in love with
an AI.
So yeah, one of the many implications of this announcement is
probably people, forming even deeper emotional bonds
with AI than they have been.
And there's been a lot of people forming strong bonds of AI

(08:32):
for a while now.

Daliana (08:33):
Yeah. I have ChatGPT open right
now. I want to see what was its
response if I ask.

Andrey (08:44):
Okay, let's.

Daliana (08:44):
Try it as it to.
Hello. I really like your voice.
Would you go on a date with me?

Andrey (08:52):
I'm glad you enjoy our conversations.
However, since I'm just an AI, I don't have the ability to go on dates
or form personal relationships.
I'm here to help with any questions or topics you'd like to discuss.
What's on your mind today?

Daliana (09:06):
So it doesn't.

Andrey (09:09):
Not right now, yeah.

Daliana (09:11):
Not right now.

Andrey (09:12):
That is how you make a lot of money.
So hopefully they.
I don't know. Keep thoughtful on that question because it is
I think it's possible to exploit people.
You've seen that before.
And I, you know, people, can maybe find a way to still
convince it, but we'll see.
And, there's some more details, worth noting.

(09:35):
So they did also announce a desktop app, that,
they demonstrated, for instance, coding assistance where it's,
you can, copy some of the code you're working
on and it will go straight to, which has to be the app.
So the workflow is a little bit streamlined.
There's no need to go to a website.

(09:56):
And they presumably will add that ability to look at your desktop
and check it out.
So yeah, this is a big news of a week I think.
And if you haven't looked at any of the demos, there are
quite a few showcasing live translation, showcasing two eyes
talking to each other and doing some duet singing.

(10:19):
It does have a, built in,
image processing capabilities. So they say it's natively multimodal.
So another thing they showed is that as you're talking to it, you can
show it images, have a video stream of what's in front
of you. And with your voice you can ask, okay, what am I looking at?

(10:39):
Or, even just show it equations and
say, can you help me work through this equation?
Things like that. And that also works seamlessly.
So.
Yeah, I think not.
Some people on Twitter in the community have expressed disappointment,
surprisingly saying, oh, this is not a GPT five.

(11:01):
This is, arguably OpenAI plateauing because the
intelligence and the benchmarks aren't that far off from what we've
seen. But personally, I'll say I was very impressed.

Daliana (11:13):
Yeah. What, what kind of feature do you think
you want to use immediately in your personal life or
in your workflow?

Andrey (11:24):
That's an interesting question.
I think personally, I do
use chat bots a lot already via, the text
input, and I don't think I will in the near term,
kind of want to use the conversational aspect of it in general,
but perhaps over time, especially for just random

(11:47):
questions and thoughts that are not related to coding.
For instance, I might try it out more and see if it's
fun to just interact with AI in this way, which wasn't
really possible prior to this sort of thing.

Daliana (12:05):
Yeah.

Andrey (12:06):
And moving on to the next story, which is from Google and
very related to read GPT four oh announcement.
So just a day after the open AI event,
Google held its own big event with a slew of announcements
related to AI, which we'll cover, over the next

(12:27):
few stories. But the first one we'll start with, and that is
most related to GPT four oh, is Project Astra,
a real time multimodal AI assistant that
can be in almost real time without almost any lag.
Listen to your voice.
Look at video inputs, and

(12:51):
answer questions about what you're seeing.
I am basically describing what you just went over GPT four.
Oh, and that's because it is very similar in some ways
to GPT four. Oh, so in some of the demo clips
Google shared, it wasn't quite as emotive in
its, voice output, and it isn't quite as real time was

(13:13):
a bit more of a lag going on, but
still. It seems like OpenAI and
DeepMind have been working on something pretty similar,
and OpenAI beat Google to official announcement.
But Google then had the unveiling of this asteroid.
And in fact, there were some fun examples on Twitter from people

(13:37):
working on it of, for instance, someone chatting with Project
Astra as they were watching the live stream of OpenAI,
and, people walking around the office
of Google with Project Astra and showcasing what it can do.
So, you know, now we'll have,

(13:57):
two of these types of chat bots
coming soon, at least.
And Google has announced Gemini Live, which is a
voice only assistant for easy back and forth conversation
and a new feature in Google Lens for video based searches.
So lots of updates.

(14:20):
Astra is still in early prototype phase,
so they aren't rolling out as soon as GPT for oh, which is
pretty much out to many people already.
But again, I was pretty impressed by what Google DeepMind have
done here, and very few companies are capable of

(14:43):
something like GPT four. Oh, it seems like DeepMind is, even if they
are a little bit behind.

Daliana (14:48):
Yeah. How do you think those works?
Do they literally have spies in each other's company and pick the
same day to do the announcement?

Andrey (14:58):
Well, the Google IO announcement, I guess we knew that
the event would be happening, that there would be a bunch of an
announcement. I do wonder if OpenAI planned to have VR unveiling
the day before to steal the thunder from under
Google to some extent.
But in terms of product development, I'm guessing they just both

(15:20):
realized that real time voice interaction seemed like
a killer. Next, advance that they should both target, and
both of them went for it.
Because from a technical perspective, without getting too
much into it, some of the things regarding like
real time processing are very impressive.

(15:40):
But on the other hand, we've had years and years of research on
training multimodal, models that accept multiple
modalities beyond just text with text and images and now
with audio. So it makes sense that the trend
is very much been to have more and more modalities as input, more and

(16:00):
more modalities as output, and make sense at both.
OpenAI DeepMind kept pushing in that direction.

Daliana (16:08):
Yeah. And I also saw the Gemini 1.5 Pro
is going to offer, 2 million token
context window, the largest of any chat bot in the world.
And the GPT four is at one 28,000
context window.
And Google is working towards and and limited context

(16:32):
window size.
What do you think? Do you think it's useful for consumer use
case or it's more for storing all the knowledge
in the, in the world. What does it mean for a limited context window.

Andrey (16:46):
Right. Well, I guess in practice it probably won't be truly
unlimited, right? You can't have an infinitely long input and expect
to allow them to.
Or now I guess not all of them.
Now it's a multimodal model.
Expect it to process that correctly, but.
Even at 2 million tokens, right?

(17:06):
That's a whole bunch of books.
And I do think for less
sort of consumer applications, but for many industrial
applications and for many jobs, at that point, you can
input many documents, like maybe even your entire
code base.

Daliana (17:27):
Yeah.

Andrey (17:28):
And have the AI process that and apply.
And that's been one of the limitations I found when working with AI
for coding, for instance, is that it doesn't have the context of
your current code base and your current company and all this stuff.
So if it can process a whole bunch of documents, when
processing your input and question, I think it'll be much more

(17:52):
effective. And in that sense, having a longer context
window and also in addition to that, maybe retrieval
that a lot of people have also worked on is to some extent
a game changer.
Moving on to the lightning round where we'll try to keep it a bit
shorter because there is a lot to still get through.

(18:13):
And the first story is another announcement from the
Google event and that is regarding their search experience.
Google is now rolling out AI overviews which
previously was known as Search Generative experience.
It's a slightly less nerdy moniker there, and it is

(18:35):
pretty much similar to what I have been experimenting for a while now
where when you do a Google search of these for some queries, there
will be, Gemini model to process your input and
produce the outputs kind of add to top that is not
just links, but an actual AI response
that has processed the contents of some websites, produced a response,

(18:59):
and then there are some attached links for you to follow.
And they are have been experimenting
with a search, generative experience for a while
now. They are saying it will start to really roll out to more people
and, be in more Google searches.

(19:20):
Again, something we I guess knew was coming and something
that, you know, will be interesting to see, to what extent
people will just stay on Google rather than going into chat bots
or things like perplexity, because Google already
will have AI will tend.
And then the next announcement, because again, there were many

(19:43):
from Google, is that they have unveiled their
own Sora type model with video.
So they have produced some clips
that show it generating pretty high resolution
HD videos, for various things.
A lot of sort of similar things we've seen from saw, with

(20:07):
clips of tracking shots of various kinds, some
surreal imagery of like llamas wearing glasses, this
sort of stuff. And the video is pretty smooth,
pretty different from most previous AI
outputs where, you know, even just a few months ago,

(20:28):
or last year, video from my I was clearly AI generated.
There are lots of ways to see where it was AI
with this it's much more natural and convincing.
Although I will say looking at the clips, it's nowhere near the
quality saw.
In terms of it's still pretty clearly AI from this

(20:50):
VR model, even if it is a lot better.
And alongside video, Google has also announced imagine
three, the latest iteration of their image text
to image model, which, similar to reality free and other
things, is producing higher quality outputs
and more complex inputs that

(21:14):
previously would not have worked as well.
Have you taking a look at this or in general when sort of was
released, how did you react?
Where was your mind blown by how AI
video generation is progressing?

Daliana (21:31):
Yeah. I guess because there are previously a lot of other AI,
you know, video generation products like, PCA
and, I don't know, I just
I saw people have some comparison, but I got some already
and are used to, the quality and,

(21:54):
it's it's very interesting, I think when it just,
AI generated the video when it was just launched, you definitely
feel. Oh, wow, this is a game.
And then you saw some YouTubers
would add sometimes, kind of similar to like a B-roll.
I generated the video. You just feel.
Oh, yeah. So obviously it's, AI generated.

(22:18):
I think for our humans perception,
if you are able to generate something like cartoon, that's really
cool, like 2D anime style and,
that's fine. But if you want to make us believe
it's real kind of footage, I think there's

(22:40):
still a long way to go. It has to be, like, near
100% kind of real, because I think a human
eye does the same. If it's just slightly off, it feels
just not real. That that's how I feel.

Andrey (22:55):
Yeah. And definitely with this video
announcement, if you look at the videos, there is just some
less consistency between the frames, some sort of weird eye
blurring. What happens? Yeah, that is pretty obvious.
So saw the some videos were convincing.

(23:16):
But again, I agree that in most cases you could still tell that it was
AI. So and that's why, you know, we highlight
kind of showcases focused more on the surreal.
And that's kind of magical.

Daliana (23:29):
Yeah.

Andrey (23:30):
Very saturated imagery rather than the sort of
things we often see in TV and movies.
And now one more announcement from Google
before we move on were even more we have an ongoing, cover
in this, but the last one we will cover is
their new music, AI sandbox.

(23:51):
So in addition to everything else, they are unveiling a new
music making tool.
It will accept text inputs and generate short audio clips
or stems based on the prompt.
And so this is a little bit more cater towards
music making as opposed to videos we've seen that generate

(24:15):
entire songs that are more, I would say consumer
related. And this again, as we have a lot of announcements
from this event isn't being available.
It's just a sort of demonstration video, but we're
working on it, and it may be a long time until any of us
are able to try it, but.

(24:38):
Google? Sure, if nothing else.
Announced a lot of stuff yesterday.

Daliana (24:42):
Yeah. And I.
Did you watch the ACM man's interview on the
podcast last Friday?

Andrey (24:50):
I have it now.

Daliana (24:51):
So, one of the hosts asked him about,
copyright. I think they he mentioned
they're not sure about how to credit artist, and then
they're kind of afraid to get into the area.
So I think they're not doing any music generation.
Generation. I think it's, in this perspective,

(25:15):
might be a bold move for Google because, for example,
one of the, topic they discussed was if someone want to generate
a song in Taylor Swift's style,
even if it's not using any Taylor Swift's
music directly, tweeting those music,

(25:36):
but learning her style from the news articles, her
her maybe her lyrics.
Is this what is have any copyright issue
related to, you know, Taylor Swift?
I think that's, interesting.
And, yeah, I think I might play with this because I have some idea
generated some fun songs, though, talking about data science

(25:59):
struggles.

Andrey (26:01):
Yeah. Yeah. A lot of podcasters I guess.
Yeah. Do this and I think it's a good point.
And maybe that's why they focused more on these short
audio clips or stems and more instrumental things like viola
or clapping rather than generating songs
of lyrics, because copyright is, especially in music,

(26:26):
kind of thorny. And one story of a section
that is not about Google and it's about the other
big player we often talk about on throwback.
So on topic, did not have any of these sorts of massive announcements
going on, but they did have a couple.
So first, they have announced a prompt engineering tool

(26:48):
that helps you craft the best input to
and from Vic Cloud to be able to get the most
out of it. They also are launching Anthropic in Europe
now, so expanding to a wider user base and we'll get
to a new, different announcement related to their company

(27:08):
and business in just a short while.
But Dropbox still in the race.
But where these announcements of the live voice assistance
and almost real time audio input, it does seem like OpenAI and Google
are maybe leading the pack right now with
the most cutting edge AI.

Daliana (27:29):
Yeah. What do you think about the, prom generation?

Andrey (27:34):
I think it'll be interesting, because a lot of people.
Probably still aren't using chat bots in some sense, and I think
this is one of these things that people will learn how
to craft the inputs.
So this might help with that.
At the same time, what I found is with

(27:57):
regards to this topic of prompt engineering and crafting prompts
in general, what I found is very in most cases,
what you want is just to be very clear and
you know, almost the way you would communicate to another person, just
like lay out your task in a clear terms and it'll do what you

(28:19):
want. And so in some sense it's intuitive
and I'm not sure this is required unless you have some very tricky
things where you might need to know some of more advanced strategies.

Daliana (28:31):
Yeah. And, I feel as the model
is getting better, maybe eventually.
Do think we still need, a lot of prompt engineering.
And, I remember a couple months ago,
the prompt engineer. It's a very hot title.

(28:52):
You saw something like 500 K salary.
I'm. I'm not sure whether this is just,
kind of a research role for the short period of time, I feel as the
model is getting better and better, maybe the engineering on the
prompt is going to the effort.
There is going to be reduce.
Also, I feel, this feature is

(29:14):
important if you are doing something very repetitive,
kind of tasks. But I also feel it takes away
of our ability to think the specific thing you want to ask.
So I think, for example, if you want to, doing something
like create an architecture for like a workflow

(29:35):
or a pipeline, maybe I would say it's a good idea to
at least think about it a little bit, because you might get
influenced by the prompt generator, which is only trained on previous
data. And maybe it can serve
as a tool. If you feel stuck, it can give you some inspiration or
you already have some idea about like the type of question you want

(29:59):
to ask and then use it to double check.
Oh, if I'm missing something.

Andrey (30:05):
Right? Yeah. I do think, personally that the idea of prompt
engineers a role is was a bit of a fad and
not something that will be here long term.
Already I still feel that just being a clear communicator
is.
Enough to be a good crafter of prompts, and that'll

(30:27):
be more and more viscous, as these models are trained to be aligned
to what humans want them to do.
So. That's my take is no,
you don't need prompt engineering as a unique skill set.
You need to be a good communicator, which is true in many jobs,
as is, and something that isn't easy necessarily

(30:50):
for people. So this is just going to make it more important
to be good at communication, than it already is.

Daliana (30:58):
Yeah.

Andrey (30:59):
And now moving on to applications and business.
We have some more exciting news, starting with OpenAI AI.
And this came out just after a couple, 1 or
2, I think maybe one day after the announcement of GPT four, oh,
with the news that Ilya Sutskever,
the chief scientist and one of the co-founders of OpenAI, is

(31:23):
officially leaving the company.
This is, of course, following up on the drama from last year, where
Ilya and several other members of the board,
for a brief, window of time
made it so Sam Altman was not the CEO of OpenAI.
And then Ilya at the time, of course, said that he

(31:46):
regretted that action that arguably hurt OpenAI.
So now he is departing to do some other ventures.
He never did go back to work after that incident, although he was
still an employee, and interactions on Twitter at least were
very friendly. There was no big drama, Yulia posted

(32:07):
saying.

Daliana (32:07):
Yeah, very corporate speaking.

Andrey (32:09):
Yeah, yeah, yeah, positive.
And it was great to be up to. And Sam Altman responded saying that he
is great and that, there's now a new chief scientist
replacing, Ilya in that role.
And the one bit of drama that did happen that's worth noting is
after Julia's announcement, Jen Leakey, who

(32:32):
co-leads the Super Alignment team and is one of the people who
published the initial alignment paper, at least one of the
early ones from OpenAI on RL.
A chef also said that he's resigning and
less corporate speak. I will say, yeah, I resign.
Yeah, and that's it.
So that does indicate maybe some tensions going on.

(32:56):
These two more. So then Ilya leaving, which I think is
not as surprising.

Daliana (33:03):
Yeah. Also on the idling
podcast last week, they asked Sam again about,
what happened there. I think he didn't provide a lot of more
information, but he did say, yes, there was a
conflict in terms of, culture.
And previously a lot of board members

(33:27):
have experience in nonprofit, come from
that world. And maybe since OpenAI
is not like a nonprofit anymore.
So there is he the way he frame it is, culture
clash, I think is probably around
AI safety. And although they didn't promote a new

(33:50):
chief scientist, I did see they promoted
a, now, like, the promotion I did mentioned
this guy. He is the director of research.
Jacob had Choki.
And, are you familiar with this person?
I think he's been. It looks like he's been director of research since

(34:12):
last October, and, he might take over
kind of the chief scientist role or has more influence
in the research.

Andrey (34:23):
That's right. Yeah. He is announced as the new chief
scientist, in fact. And he similar to Ilia,
has a pretty strong tracker track record with research.
And he's been with OpenAI since 2017.
So kind of a long time employee.
OpenAI started, I believe, late 2015, 2016.

(34:46):
So he's been there from early on before GPT even
happened.

Daliana (34:50):
Okay. I'm just looking at his LinkedIn.
Okay. He joined as a research lead in 2017.
Previously was a postdoc fellow at Harvard.
He was a software engineer intern at Facebook.
And, yeah, I think he did his, college

(35:10):
graduate degree in computer science from University of Warsaw.

Andrey (35:15):
And the next story is also about employees
and who is leading what companies this time with regards to on topic.
And the announcement of someone being hired rather than leaving the
company. And the person is Mike Krieger,
who is now joining on as chief product officer.

(35:35):
And he is a pretty notable figure.
He was a co-founder and CTO of Instagram and later
also artifact, which, was a personalized used app that
was acquired by Yahoo!
And he has an announcement was said to oversee
product engineering and product management and product design efforts

(35:56):
as we work to expand.
This is the language from announcement as we work to expand our
suite of enterprise applications and bring cloud to a wider audience.
So once again, showcasing that now, these
are not just research labs, they are very commercial and they are
very much seeking growth and seeking income.

(36:20):
And with this announcement, I think, Tropic does position themselves
more strongly in that aim.
Next up we are moving out to Lightning Round.
And finally, a story not about anthropic
or Google or OpenAI.
This one is a story about unit three, a Chinese robotics company,

(36:42):
and they have released details about their second humanoid
model with G1 humanoid agents that
notably will be priced at $16,000,
which is very, cheap.
It's, cheaper than their first iteration, which was 90,000
and presumably cheaper than pretty much any other humanoid you can

(37:04):
hope to get.
And if you want to look at it, it's it is human in shape
and in some ways similar to what we saw from Boston Dynamics, with
very wide range of motions being able to rotate its torso all the
way around and just doing all sorts of flexibility.
Part of why it will be costing less is it is smaller

(37:28):
of any human. It's like almost child sized.
And it compared to some of our other humanoids we've seen, the
numbers aren't necessarily as impressive where it can't,
carry necessarily as much weight
in its arms.
The battery life will have two hours per charge and etc., but,

(37:51):
yeah, definitely notable. We've been talking a lot about humanoids.
I don't know if you've seen this brand, Dalian, of a lot of robotics
companies being funded and announcing humanoid robots
in development. But this one is adding
to that trend. And, I think with the low cost
definitely making it seem more possible, we'll be seeing more humanoid

(38:13):
robots out in the wild.

Daliana (38:15):
Yeah. I think, after I
Model Ridge, kind of converge to a level, I think maybe
we might talk about it later.
The next, I would probably put it in the real,
real world. And, I think robotics probably is the next
frontier. And, robotics is really hard.

(38:37):
I remember reading something, saying, you know, it's easy to do
some live some like, something heavy, do those things.
But for those tasks, for example, putting things in a dishwasher
is extremely hard for, robots to have the,
precision. And, yeah, I think there are still a lot
of interesting challenges for researchers to solve.

(39:02):
And, I also look at this robot, I think intentionally they
design it in a way that not looking like a human
so doesn't have a face, doesn't have any skin.
I think it's kind of branded as a task focused
robot. I'm curious whether

(39:26):
I think there are, I saw this robot in Japan.
Those, but I haven't seen it from bigger companies to try
to develop.
Oh, I think Tesla would launch the robot.
But I don't think that looks like a human.
I'm kind of curious. Are people.
I don't know, that's a weird thing. I don't I think it's.

Andrey (39:49):
True that in general, in commercial,
like, big businesses. And what we've been seeing, the trend has been
when you make a humanoid robot, it looks like a robot.
Yeah. It has bare metal and usually
only at most, like an abstract face with a screen.
And while there have been demonstrations of things

(40:13):
that look a little bit more humanlike, especially from Japanese
researchers, that isn't something that companies are
seeking to do, I think, and partly because the
aim is to, get these things doing tasks
and work and not, you know, socially interacting with us.
So that's, definitely a bit more far off than having

(40:36):
just these robots moving stuff and solving chores for us.

Daliana (40:41):
Yeah. And, I think if I hug a human,
the human interaction and the release oxytocin, I wonder
if there's research on if I should handshake with a robot.
Would I release oxytocin, would boost my mental health as a
create some sort of connection.

Andrey (41:01):
Maybe.
And next story. We are moving on to robot Taxis
with a few stories on that, starting with cruise.
And the news is that they will start testing in the Phenix
area with human safety drivers on board.
For a little bit of context, cruise had,

(41:21):
major incident last year in a crash that essentially
halted their efforts and rollout of their self-driving cars
for now, quite a while.
So this is pretty notable to show them getting back on the roads
and slowly trying to roll things out again.
It seems very carefully since they are adding these human safety

(41:45):
drivers. So hopefully
I'm a fan of robotaxis, so I hope cruise can get back
in the game, so to speak.
Yeah.

Daliana (41:57):
I only tried Waymo like three
weeks ago.

Andrey (42:02):
For the.

Daliana (42:03):
First time.
I think I'm yeah, a little bit conservative.
And I feel, oh, I want a company to collect more data before
I actually try it.
And I also, I think, when we evaluate those accidents,
we tend to have a higher standards for

(42:24):
robotaxi.
We don't necessarily compare that to the accident
ratio for human drivers.
Sometimes you think think about it is like a little unfair for the
researchers, but I do think it's necessary to,
evaluate those accidents, even if it's statistical speaking, it's

(42:47):
it's safe. But I think understanding the the real costs is important.

Andrey (42:53):
Speaking of which, the next story is that the
National Highway Traffic Safety Administration is now probing
Amazon owned Zoox, which also is developing self-driving
vehicles after two crashes.
So Zoox, since March, has been expanding its vehicle testing
in California and Nevada to include a wider area and

(43:16):
be able to drive and higher speeds at nighttime.
And it seems that now its vehicles have been
involved in two crashes.
And while they were driving, that resulted in minor injuries to
motorcyclists.
Zoox, unlike Waymo or Cruise, did not try to roll out a commercial

(43:38):
offering yet. So they just are testing.
And, not too many details on this yet as to
whether vehicles cost it or not.
But we have been talking with more and more examples
of crashes, and this is adding to that trend of, as
you say, the standard for self-driving vehicles is high.

(43:58):
And while our story for a section again on robotaxis,
lots of stories on that this week.
And this time it's about Waymo.
And it is also under investigation from
that administration after some crashes and mishaps,
apparently there have been 22 reports of crashes or potential traffic

(44:19):
safety law valuations.
And this intersection will aim to evaluate the software's ability
to avoid collisions with stationary objects and its response
to traffic safety control devices.
This, yeah, is following up two days after
the Zoox announcement, so this administration sure has a lot of work

(44:41):
to do with regards to self-driving vehicles, it seems.
And on to our next section projects and open source.
And our first story is going back to Google.
They had some announcements on this front as well, in addition to all
the product announcements.
And what they announced was first a preview of Gemma

(45:02):
two. So Gemma is their main open source,
language model that we've covered previously.
Now they're saying that in June I'll be rolling out, Gemma two
that will have some larger variants and
will, presumably be much better.

(45:22):
And besides that, the kind of bigger news on this
front is that they announced Polly Gemma, which is an
open vision language model.
So unlike Gemma, which is just a language model, this can accept image
inputs as well as text, and they are now
releasing it through all the usual platforms of GitHub and hugging

(45:46):
face. People are able to now build on top of this,
and this is pretty notable because there are
much fewer open source, high quality vision language models
compared to open language models, of which
we are now many.
So yeah. Google.

(46:09):
Continuing to push in this direction.
And this is almost now a new competitive front with meta, of
course, also releasing a lot of models.
Seems like to be, seen.
Seriously, this is one of the things they're investing in.
And another big story for open source
models. The next one is Falcon two, which is

(46:33):
the UAE's new AI model release.
So I think last year, or even
maybe two years ago, Falcon was one of the first
large language models to be open sourced.
It at the time was pretty big news to have, language
model where that many billions of parameters out there in the

(46:57):
open prior to that becoming much more of a thing with
things like clamor.
And so now they have launched a second iteration of Falcon with
Falcon to 11 B and Falcon to 11 B via
LM. Another visual language model like poly Jamma.
And the numbers are as usual

(47:19):
with model releases, pretty good.
They say that Falcon 211 B outperforms It Matters demo three,
and performs on par with Google's GEMA for those,
sizes of models.
And both of these models are open source and provide unrestricted
access to developers world wide.

(47:40):
So yeah, open source continuing to.
Push forward with more and more models that people can build upon and
continue to improve.
Pretty big news these two combined this week.

Daliana (47:58):
Yeah.

Andrey (48:00):
And just one more story for this section.
And this one is coming from a hugging face.
And it is about, software library rather than
an actual, model.
So as we covered last week, they released, model
for robotics, the robot. This week they are announcing

(48:21):
Transformers Agents 2.0, which is
a framework to make it so you can have agents
that can iterate based on past observations to complex,
to complete, solve complex tasks.
And they show this agent framework that, for

(48:43):
instance, using Lama 370 B instruct agent, it can
outperform GPT four based agents in the GUI leaderboard.
As with other software announcements, this one is notable just
because Huggingface libraries are used very often
by people after building software and agents is

(49:05):
one of the challenges we haven't quite solved
with language models and with AI in general.
It's still very much an ongoing effort.
So having this, library for people to build
upon could accelerate that significantly.
Moving on to research and advancements, which we'll have just a couple

(49:26):
stories, not too many, this week.
The first one is the platonic representation hypothesis,
a pretty interesting conceptual paper, not a new breakthrough and
performance, but very interesting ideas.
So the key idea being presented in
this paper is that platonic representation hypothesis that

(49:49):
says that neural networks trained for different objectives and on
different data and modalities, are converging to a shared statistical
model of reality in their representation space.
In neural networks, when you give it an image or give it some text
that gets mapped to a big set of numbers, the representation

(50:09):
and what this paper shows is that across various
models trained with different data sets and so on,
as you get bigger, as you get more performant, as you get more,
able to do multiple tasks, the representations
converge and become more and more similar.

(50:30):
And so there is this hypothesis that there is
a one true kind of ideal representation of
underlying reality, where images and text and so on are just
kind of, projections from reality.
And they have a lot of details in this paper as to why

(50:51):
this initial hypothesis may be true with things like
the capacity, hypothesis that if a non optimal
representation exists, then larger models that can explore
more possible solutions in terms of representations will.
Find that optimal or get closer to that optimal representation.

(51:12):
And then in addition to that, if are,
where's the multitask scaling hypothesis that an increasing number of
tasks, well, be subjected to learn representations
that can solve all of those tasks, and the simplicity bias
hypothesis, saying that larger models can fit the

(51:33):
data in different ways and will generally
tend towards the simplest possible solution.
It's a very interesting kind of insight and hypothesis, not
quite proven in this, but there are some numbers showing that
as you get bigger and as you, get more multitask,
the nearness and similarity of these different

(51:57):
models gets bigger.

Daliana (52:00):
Yeah, it is interesting like there is only the
one source of truth is also although it's from different
dimension text vision, but it feels like
the law of large numbers.
If you have, you know, now everything kind of covered
in the middle. So.

Andrey (52:20):
Right. And it also relates to some extent,
to research in neuroscience, where it's been
known for a while that representations of
images, for instance, do in neural nets do
correlate with how representations of images happen in

(52:40):
the human brain. And you can actually do a sort of mapping.
They're not the same, but there are shared characteristics.
And in this paper, they do go slightly into how
seems like neural nets are in, for instance, how they represent
color are getting more similar to

(53:01):
humans. So that is another kind of point and favor
of this hypothesis. Yeah.
And next up the other research paper will cover is Sutra
scalable and multilingual language model architecture.
And in this paper they show how you can train
model on over 50 languages while having

(53:24):
good performance across most languages, or even being generally much
better at English than other languages.
And the key technical bit is that we essentially separate language
from actual intelligence.
We begin by having a language encoder
and then later have a language decoder basically saying in the

(53:45):
input phase, we'll take the language and first process
the language itself, and then the language model
will just learn to think in terms of concepts, so to speak,
in a more abstract space where all languages kind of map onto the same
thing and the language model reasons on that.

(54:07):
And when they have an evaluation, they show that compared
to GPT four, for instance, it's not as good
at English or Hindi at all, but it is more consistent across
many more languages. So things that GPT four doesn't
do quite as well on for instance, Tamil or Telugu.

(54:29):
This model does do very well on.
And this. I do think it's pretty notable because there are many, many
languages in the world. And if there's an approach that makes it so,
model is effective or almost equally effective in all
of them, that would be very useful.
Moving on to the policy and safety section.
And the first story is about a bipartisan Senate bill

(54:53):
on AI security.
This is the secure AI act of 2024, and it
would require the National Institute of Standards or Technology to
update the National Vulnerability Database and the Cybersecurity
and Infrastructure Security Agents to update their
common vulnerabilities and exposure program.

(55:15):
In addition, the national security agents would be agency
would be tasked with establishing an AI security center to
provide an AI testbed for research for private sector and academic
researchers, and develop guidance to prevent counter
AI techniques.
It seems like, some of this would go into effect pretty soon.

(55:38):
This would have 30 days after the enactment of legislation
to evaluate how to do this.
And it kind of.
has some emphasis on public private communications to stay
updated on threats and safeguard against threats facing

(56:00):
infrastructure.
We've been talking about a lot of AI bills being introduced in the
Senate bill, this one just being the latest one.
So it seems like, with AI being as big as it
is, there's a lot more efforts going into that on the policy
front in the US.
And the next story is about the UK AI

(56:22):
Safety Institute, and it has released an open source tool set
called inspect that is designed to strengthen AI safety and facilitate
the development of AI evaluation.
They say that, this is the first AI safety
testing platform spearheaded by a state backed body

(56:43):
that is released for wider use.
And as part of this, we do release data sets for evaluation,
test solvers to carry out with test and scores
to evaluate, to work, these,
things going through the tests and aggregate the scores into metrics.

(57:04):
And this would be open source and possible to expand of
more Python packages.
So yeah, now, we'll see if
the big companies will actually use this and, release metrics
on safety and, the kind of metrics
that are part of this.

(57:25):
And on to the Lightning round. One more story in the section,
and it is about protesters fighting to stop I
and how we're split on how to do it.
A group of activists called pause AI, what we have covered in the past
has protested recently and called for

(57:45):
the halt in development of large AI models
because they believe it could pose a risk to humanity's
future. Family protests have been taking place globally,
including San Francisco in New York, Berlin, Rome and Ottawa.
And it seems that.

(58:08):
Yeah. I don't know what you
know. And this story goes into.
Yeah. Not being sure or some members of the
movement disagreeing, on the necessary weight.
Some people are even considering sit ins at
I developers headquarters, with OpenAI being example,

(58:31):
where these protesters would just sit outside their offices as
part of a protest. So, still definitely a small effort
in general, but I wouldn't be surprised if there's going
to be more calls like this by more people to just
say, AI is moving too fast.
Just stop and let us catch up.

(58:53):
Yeah. And to the last section, Synthetic Media
and Art. And the first story is once again going
back to Google. And they did have one announcement on this front
alongside the rest.
The story is that Google's invisible AI watermark will help identify
all, identify generative text and video, and this

(59:15):
is an expansion on their AI watermarking
technology, since I'd may say that,
this was first announced in August, but
now, Google has enabled Sinfield
to inject an audible watermarks into AI generated

(59:37):
music. And this will generally expand to any
modality, including, potentially text in the near future.
So. Similar to other news you've covered in terms
of for a meta open AI, all deciding to include
watermarks to be able to detect synthetic imagery.

(59:57):
And at least if the metadata is fair,
kind of classify it as such.

Daliana (01:00:04):
Yeah. Do you think they need to, unify
in one watermark, maybe kind of become like a standard?
So every company having the same thing?

Andrey (01:00:16):
I think so, personally, and a lot of companies are.
Already doing this with Sea Tupa, where they
have a standard that they've collaborated on.
Google and the and the meta have not
adopted that standard. Exactly.
We've, taking a different route and,

(01:00:39):
and possibly it's may not be easy,
but it probably would be ideal for there to be a standard
that everyone adopts.
Next story how one offer pushed the limits of AI copyright.
This is about Alisa Shupe, who has successfully registered.
I registered a copyright for a novel she wrote

(01:01:03):
using OpenAI's chat. GPT this
novel is AI machinations down called Webs and Typed Words,
and it is among the first creative works to receive a copyright
with AI, generate text, and being at least a part
of the creation, and the US

(01:01:24):
Copyright Office has granted the copyright registration, but
only recognized her as we offer a V selection,
coordination, and arrangement of text generated by artificial
intelligent intelligence, which means that no one can copy
the book about permission, but the actual sentences and paragraphs
themselves are not copyrighted and could theoretically be rearranged

(01:01:47):
and republished as a different book.
So yeah, this is still an open question on copywriting
stuff. You write with AI, and it seems like we have a bit more clarity
of this happening.
I'm curious, Diana, for your interviews.
Have you started doing any prep for using AI models to

(01:02:08):
generate questions or at least like research background?

Daliana (01:02:12):
Yeah. So because the people I interview, a lot of them don't
have any public information online.
So it does require me to sometimes do a 30 minute pre-show
chat. But, I have a play with a oh, for example, generated
a question in the Tim Ferriss style.

(01:02:33):
And, but sometimes the AI
generated was a very verbose.
I have to, I think I got inspired by the idea.
And then I will still probably use
a lot of my own, content.
But I do use anthropic sometimes to help me summarize

(01:02:56):
a what would be a good chapter for me to put it on YouTube,
highlights because, I don't know, I haven't tested
with GPT two for. Oh, but previously I find anthropic
has given me better results.
When it process like longer context window compared
to, to defeat ChatGPT often

(01:03:19):
ignore something in the middle session, only focusing on the
beginning and end.

Andrey (01:03:25):
Yeah, it makes sense.
A couple more stories before we move on to a bit of an interview
section. So the first one is stellaris
gets a DLC about AI which features I created
voices. Stellaris is a video game and the very latest
downloadable content for Machine age uses generative AI technologies

(01:03:49):
to create some assets, including generating voices for an AI and
talking Unnest and AI player advisor, which of course,
is. This is a controversial topic, especially voiceover and video
games. So this drew some pushback.
And the game's director actually had to address it and reassure the
players that the AI voice generation,

(01:04:13):
doesn't mean that voice actors, want at least receive royalties
for every line created.
So pretty significant.
And that this is a big game.
This is, a big developer, not huge,
not like, you know, millions of players, but still,
not an indie release necessarily.

(01:04:35):
And, this is among the first examples I'm
aware of, of, commercial professional game
studio adopting some AI tools.
And the last story we'll be covering is about an
AI film festival, the second annual film festival
organized by generative AI startup runway that,

(01:04:59):
after it ran, showcased the top ten finalists films,
all of which incorporate AI in some form.
So, for instance, doing AI generated backdrops, animations,
synthetic voiceovers and special effects.
And the title of article says that humanity triumphed
over tech because there is some editorial evaluation

(01:05:22):
of the finalists films, and the
claim is that the limitations of current AI tools were evident in the
films, with some scenes clearly being a product of an
AI model, and that some films were constrained
by limitations of AI with disjointed scenes and a lack of control

(01:05:43):
over generative models.
But, despite these limitations, some
of the films were able to still be good
due to strong scripts and performances.
So the argument is that human contributions
still make the difference, right?

(01:06:03):
We still don't have AI that can generate films that are compelling,
and I would argue that that's probably going to remain true for a
while. But, we are
still kind of pretty early on. And, I did look at a couple
of videos and if not for analysis, it's interesting to see
people try to start, and to

(01:06:26):
come up with how to utilize limitations with its current,
how do you realize technology with its current limitations to produce
something compelling?
Alrighty. Well that's it for venues for this episode.
Slightly fewer stories. Unusual you do just because you wanted to
focus on the exciting new announcements.

(01:06:47):
So with the remaining time, since television,
does host the.

Daliana (01:06:54):
It up the show.

Andrey (01:06:56):
Data science, a show in which she interviews many.
People about their paths, throughout
their careers and, kind of learnings.
Seems like would be fun if Diana did, like, a short
little interview of me.
So listeners can get a sense for

(01:07:18):
my background and my thinking and I.
So let's see how it goes. Diana.
You can go ahead and take over.

Daliana (01:07:25):
Yeah. So since we're talking about art, when I was looking at your,
personal website, I notice you like, films.
You like music.
So I'm curious if you didn't become, AI researcher, what would
be another career you wanted to get into?

Andrey (01:07:42):
That's a very interesting question.
Yeah, I do feel there's a good chance I would have wanted to pursue
something in the creative arts.
Perhaps with film. I do enjoy, video editing
in particular. So I could see myself doing that.
And another possible route that I may still pursue at some point

(01:08:04):
is writing fiction.
Okay. Because I do enjoy reading a lot, and I've written some short
stories and a lot of essays over the years.
So I do like writing, even though it can be very hard.
Yeah. I think there's a lot of possible paths for me that are
not technical.

Daliana (01:08:22):
Yeah. Have you publish any short stories?

Andrey (01:08:27):
Yeah. Just a couple, a few, in fact, last year
and even before the release of ChatGPT, me and a friend
started a little newsletter project called stories by
AI, where we published one short story per week
with the idea to experiment with how we could use AI to

(01:08:48):
aid in the creation of short stories and, still have
our kind of control over the short stories and be the authors, but
use the tools to, see how
they could help. And as part of that, I did publish, I
think about maybe 3 or 4 things I wrote

(01:09:09):
up and we project we stopped around
May of last year because it was kind of coming to a point
where ChatGPT and these other things were
so good, and so many people were creating AI content.
It didn't seem like an interesting thing to explore

(01:09:31):
necessarily anymore, but I did have a lot of fun
publishing and writing those few things.

Daliana (01:09:38):
Nice. And, what kind of
experience do you want people to have when they read your stories
or in the future, watch some videos
you're created.

Andrey (01:09:53):
I think personally, what I like a lot is when
things are interesting, when things sort of like give you a moment
of, being taken aback by some new idea
or some interesting notion or something you didn't expect.
So my if I do write something or make more videos for

(01:10:14):
YouTube, which I did for a while and released a couple of things,
I think would be a heavy.
Maybe even intellectual component or conceptual component to it
and less sort of an action or.

Daliana (01:10:33):
Do you have example?

Andrey (01:10:35):
Well, the short story is I, I have put out,
I guess, examples of that where I think the last
one I wrote was conceptually,
a chronicle of in the
near future, a few decades from now.

(01:10:55):
It's kind of a bleak one where it's just dried up, of how
humanity descended into decades long wars over
resources due to a combination of climate change
and, militaries being,
continuously more automated with more and more robot soldiers.

(01:11:16):
So essentially, you got into a dynamic where you have
endless war. We have all the sides just manufacturing robots
to send out in a battlefield to fight over,
resources that are increasingly necessary as climate
change and technology racing happens.

(01:11:37):
So that's an example. And there was a little twist of, like, the
narrator, of that one was actually an AI that wrote
propaganda and was aware of all these secret details and so
on. Yeah. So stuff like that.

Daliana (01:11:53):
And it sounds like you do think about like ai a
law even in your art creation.
So you talk a lot about AI safety.
Where where do you stand. Are you more of, you know,
Duma or.

Andrey (01:12:11):
Yeah. We've gotten into this with Jeremy, in the past,
sometimes. And we are very much it's
we have, opposite views of my regular co-host, where my
ergo co-host is very concerned about AI, even
potentially exterminating humanity.

(01:12:31):
And personally, I'm very skeptical.
We should be very worried.
For numerous reasons. We had a whole episode, if listeners want
to go into that, I think, like later last year
on AI safety and alignment.
So personally, I think there are a lot of concerns that are legitimate
to have misinformation with scamming, which is already

(01:12:55):
happening, with people,
you know, getting addicted to, romance
with AI and moving away from human attraction.
Yeah, a lot of these very significant things.
But I'm very skeptical that there is a real.

(01:13:15):
Pathway. Even if it got to human level, I for that
lead to, significant number of people being, as
Zoomers say, exterminated.

Daliana (01:13:27):
Yeah, but you're less worried compared
to the people on the East Room side.
It seems like you're also very excited about the, development.

Andrey (01:13:40):
So in some ways, I think even
as someone who's been in the eye for a while.
It's come to a point over pace of AI.
Progress is a little jarring.
And, it's there's so many impacts it will have on so many
people like voice actors, like illustrators,

(01:14:02):
like copywriters, and offers in general.
And.
It's it's hard not to both be excited by things.
And I'm someone who uses ChatGPT a lot, and I really like how it helps
me do some of my work and some of my writing, but

(01:14:23):
that comes with some costs.
As with previous, technological revolutions.
So it's, it's, a sense of excitement at.
What people will do with these very powerful tools, combined
with some worry about the people who will be hurt

(01:14:44):
by your tools.

Daliana (01:14:46):
Yeah. And you said you have been in AI research for a while.
Can you tell us how did you get into AI and what was your journey
like?

Andrey (01:14:55):
Sure. It's interesting.
It goes back a long while.
I guess the starting point, you might argue, was in high school
where we had a robotics club where we competed
in this, program called first that has various
high schools building robots to compete in these kind of sports, like

(01:15:19):
games. And I was part of a software team and a reward
some of the very simple AI not not at all like neural nets
and so on. So that was my starting point.
And then at that point, I didn't think
of myself as like heading towards working in AI and

(01:15:40):
the.
Reason I kind of got more into it was later on in
college when I took the intro to AI class.
I really got excited and curious and
impressed by a lot of what I was learning, and that led me
to, doing a research internship the summer

(01:16:02):
after that, and then actually joining the lab of a professor
as an undergrad researcher and then doing
another internship doing research.
So I think it's I sort of naturally gravitate towards
it. Then I try to be a software engineer for a while and,
and let's say I wasn't quite as excited by it.

(01:16:24):
So I went back to Stanford for a masters and PhD that,
also wound up working on AI and robotics.
So it's it was a sense of kind of gravitation towards it,
you might say.

Daliana (01:16:38):
So you got an internship?
After high school, like in college?
In a lab. How did you get your first internship?

Andrey (01:16:46):
Right, so they are.
I think they're called rescues.
Many universities offer research.
Summer, internship, including Stanford, for instance.
So you can apply as a student to work in,
research lab and be guided by graduate students, on

(01:17:07):
a project. And so, one of the big ones,
that does this is from CMU and their robotics lab.
It's something called, risks.
And that's been going on for a while.
So I saw I think I saw a flier for it, like,
piece of paper hanging on some professor's door when I

(01:17:29):
walked by to ask him questions and just.
Yeah, it really I would not have been aware of it as a possibility.
I don't think, had I just not seen that, as I was walking around
campus. And that led to a lot
of other stuff.

Daliana (01:17:47):
Yeah.
I love those stories.
And you you from California?
You grew up.

Andrey (01:17:55):
I grew up in a mix of places.
I was actually born in Ukraine, in Crimea.
And when I was five, we moved to Israel.
And then when I was 12, my family moved to the US, to Georgia.
So my undergrad was at Georgia Tech, and it wasn't until after
my undergrad that I moved out to the Bay area, where,

(01:18:18):
you know, a lot of the software jobs and exciting tech stuff
is happening. So now I've been in the Bay area for almost nine
years, and given that a lot of AI is here, it
seems like I might stick around for a while longer.

Daliana (01:18:33):
Yeah. And, can you tell us a little bit what you're currently
working on in your company?
Is this the first job you took after you finish your PhD?

Andrey (01:18:44):
Right? Yeah. So this is it.
It is a first thing I did after my PhD.
In fact, I started there a couple months before my final
defense, as I was finishing up.
And we are working on a platform where people
can create games of AI and publish them for other people

(01:19:05):
to enjoy. So as someone who has done
YouTube videos, podcasts, photography,
all these sorts of things, some writing,
I one of the things I wanted or considered doing
after a PhD was to going

(01:19:26):
to some company that is
using very, very impressive advancements in generative AI to
enable creativity, for more people
to build cool stuff.
And I also like video games.
And it just so happened that there was this early

(01:19:47):
stage startup with like eight people that,
was founded by a former lab mate who did my,
his PhD with the same advisor.
So in some way, like my PhD.
LED to me being connected to the startup and and led to me
having that as an option because they were still tiny, right?

(01:20:10):
I would not have known of our existence had that not led to
it. And, I wonder is a little like I've been there
a year now and.
We don't have for the listeners of podcasts wondering if they can try
it out. Let's just say it's not ready yet.
Yeah, it turns out to be a hard problem.

(01:20:31):
But.
You know, I think it's good to.
Take on hard problems and hopefully
in not too long a time frame, I'll be able to share
something cool for people to, let them do stuff
I would not be able to do without I.

Daliana (01:20:51):
Yeah, that would be awesome.
So when you were taking the offer, have you considered,
say, joining an AI research lab in
Google or Microsoft?
OpenAI.

Andrey (01:21:07):
I did think about it.
But.
I think ultimately I decided to.
You could argue, pivot or move away from research towards joining
a startup because especially towards the latter years
of my time as a PhD student, I enjoyed research, I liked it,

(01:21:29):
and I would probably enjoy working at Nvidia or
Google. But I did miss like building something
over a longer time horizon in research.
Typically you do a project, to write a paper and
you work on it maybe eight months, maybe a year,

(01:21:50):
and that's kind of it.
You move on to the next paper, to the next project, and you keep doing
that many times over.
And so there's a sense where.
You don't kind of build something over a long term, or you might,
with a succession of papers building on the same topic or questions.
But who you're impacting is other researchers and other people

(01:22:15):
trying to ask questions. So I felt like.
I would like to use my skill
set of knowledge and and the recent advancements in AI to
build something that impacts more people more
directly.

Daliana (01:22:36):
Yeah. And.
When you started content creation, building a podcast
and the news that a while you were at, Stanford.
So what made you want to start this
newsletter and the podcast?

Andrey (01:22:57):
Yeah, there's a bit of a history behind it where
the kind of seed forest that led to the podcast
was, I started working in late 2017,
actually. And, you know, the motivation at that point
was that we were starting to get into a lot of AI

(01:23:19):
hype, and at the time, a decent portion of the coverage
in the media was sensationalist or inaccurate.
In some ways. This was, you know, just a year after AlphaGo and
at that time seemed like a lot of coverage,
overemphasized the kind of impressiveness.

(01:23:40):
Yeah, or in some other ways wasn't quite right and didn't point
out that would explain things quite right.
So at that point, I started
something called Kind of today where we published,
overviews of recent, AI headlines where we
tried to point out things that were inaccurate and then just provide

(01:24:04):
better explanations that were informed by understanding AI.
That, in turn, led to starting the, last week in AI
newsletter, because I already needed to be up to date of
news to know what to write about.
And at that time, there wasn't really there
were some long running AI newsletters already, but

(01:24:27):
nothing had quite aggregated as much.
So we started that like mid 2018
actually, and that's still running.
And then the podcast happened in.
March of 2020, just before Covid hit.

(01:24:48):
Yeah. And the motivation for that was.
I guess it seemed like a natural fit.
It was at a point where I was starting to roll out
and impact more regular people, and not just researchers.
And I had, by that point done YouTube for
kind of a while with maybe ten videos

(01:25:12):
and felt that I would enjoy doing a podcast and
producing it and and so on.
So we started, now over four years ago.
Which is kind of crazy, and it's been a lot of fun.
So we kept at it just because it is fun and
very useful actually, to just for myself to keep track of and use.

Daliana (01:25:35):
Yeah. That's awesome. I remember last time I saw you, I asked you,
hey, you work on this newsletter and podcast.
I think you are also involved in some other AI related publication,
like are you tired?
And you're like, oh yeah, I'm so tired, but
I don't want to start doing any of those.

Andrey (01:25:57):
Yeah. And it's worth pointing out that,
these are like team efforts to some extent.
So. And the I newsletter was one of a person who helps
out and, on the podcast, there's also
co-hosts that help, share their efforts and,

(01:26:19):
yeah. And then my other side project, the gradient,
I have sort of taken a bit less of a role to try
and make time for other things.
So, yeah, I think, it's a challenge sometimes.
I did like as part of focusing on this, I
stopped making YouTube videos. The last one I made was late 2020,

(01:26:43):
and that was fun. And, you know, I actually somehow
got to like 70,000 views on my last YouTube video
so I could see myself doing more of that.
But you have to pick your battles and you only have so much time.
And I tried to make time for enjoying life as well
and not just producing content.

(01:27:04):
So.

Daliana (01:27:06):
Yeah. So what are the things you do to enjoy life?

Andrey (01:27:11):
A lot of it is, appreciating art or entertainment.
So lately I've been playing some video
games. I've been watching some cool TV shows,
for example.
Currently I'm watching through Rome on HBO, and that's fun
to see. Depiction of a lot of the history and major events

(01:27:35):
of that time. And I'm playing this game called Sky, which
is a very wholesome, kind of beautiful game with a lot of nice scenery
and social mechanics.
And then, I'm a huge fan of films.
So I watch a lot of the classics, and
I know a lot of directors names and stuff like that, and,

(01:27:59):
try to make time to go to arthouse theaters in San Francisco, for
instance, to see new releases.
So that's a lot of it. And, aside from that,
I do like.
Going to concerts.
And. Doing things with friends, you know, going on

(01:28:22):
hikes and so on.

Daliana (01:28:24):
Yeah. So it seems like you do spend a lot of time thinking about the
Roman Empire.

Andrey (01:28:30):
Lately? Yes. Lately, yes.

Daliana (01:28:35):
And. So what are something you are struggling with
right now?

Andrey (01:28:42):
Yeah. Well.
The startup life has its challenges, right?
Where I think.
Early on. You know, at a point where we are right now
where you would say
probably pre product market fit, pre growth,

(01:29:06):
pre like knowing if still work you got to be
able to sort of live with that certainty of maybe
all this work we doing will lead to failure as is the case
with most startups. And we just won't be able to solve it.
And that does lead to some stress. And, and, you know, in a startup

(01:29:26):
environment, there's a lot of.
Uncertainty and even like.
Hi. How you should do things. What things you should do.
There's a lot of questions you need to proactively think about.
And, take charge of
trying to steer the ship.
Yeah, when there are very few people.

(01:29:49):
So that's the case. And then.
Yeah, it's it's.
I'm sad that I can't do more in life, but I can't make YouTube videos.
At least.
Currently, if I really tried, I might be able to, but,
maybe I have less energy when I was younger than when I was younger.

(01:30:10):
So my output has to some extent decreased where
I have not writing much recently, I am not making YouTube videos.
I do enjoy doing this podcast a lot, but I do wish
I could, or I could manage to carve out more time for creative
outlets. Yeah. But you know, yeah, you gotta pick your battles.

Daliana (01:30:32):
Have you missed a week since you started for
podcasting?

Andrey (01:30:37):
Yes. We have had some weeks where, just didn't make
sense. In fact, we had a hiatus, in late
2021 I believe, or 2022,
for a few months, before the release of ChatGPT, where
am I? The original co-host of a podcast?

(01:30:59):
Sharon. Joe had to move on.
And there was no co-host, so we just paused for a while,
but then Jeremy came on as his venue co-host.

Daliana (01:31:09):
Oh, nice. Yeah.
Yeah, I can resonate with.
I recently also slowed down the production of my own podcast
because I'm trying to figure out what are the things I want to do.
For example, spend more time with, career coaching course I'm
building and learning new,

(01:31:31):
coaching skills and doing personal development for myself.
And, also realize I actually want to create
more, YouTube videos.
I don't have my own channel. Maybe I'll create one for myself.
And it does feel very scary when you have a great momentum
and then you all of a sudden stop, you feel like, oh,

(01:31:53):
am I going to disappoint people when I stop a podcast?
Am I have quote unquote quitter?
But on the side, it also feels oddly freeing in a way that,
oh, I don't have to be on this podcast treadmill
anymore. Who make the rules?
I have to publish every, every week.
I'm for for my own podcast.

(01:32:15):
So, yeah, for own case.
Do you feel, oh, maybe you can have other
host, so take over for certain weeks
and, or do like bi
weekly. I mean, right now there are so many I knows weekly they're
even daily I podcast. Have you thought of ways to scale

(01:32:38):
the podcast and give yourself more freedom?

Andrey (01:32:43):
Yeah, a little bit.
I'm still currently the editor of a podcast as well,
so I spend a couple hours each week post recording,
doing that. So that's one thing I thought about.
I think maybe one of my flaws, or
arguably things I could do better at is now that we have some

(01:33:06):
listenership, there's probably
a potential for more monetization and sponsors, and I'm taking this as
slightly more professional direction, and I really resist that in some
way. Like, I like the idea of just doing this for fun and for the pure
motive of, helping people keep up with I.

(01:33:27):
But, we'll see.
Maybe I do still enjoy editing it as well and having the
full kind of creative control over it, but perhaps in the future
I'll find a way to do this a bit more faster and potentially
do more like start. Go back to doing interviews in
addition to covering news or do YouTube videos or whatever.

Daliana (01:33:51):
Yeah. As I'm preparing this episode with you, I can see how
much effort you, you put in.
Do you remember particular apple common
or I don't know, do people email you, message you, main thing
or something that you feel are is so worth it.
Do you remember something like that?
What do you say?

Andrey (01:34:13):
Yeah, it's.
I mean.

Daliana (01:34:17):
You help me get through a pandemic without an audience.

Andrey (01:34:21):
Yeah. I don't know about that, but, there's it's hard to
pick one out because, in general.
We get a so much like, we covered a
few of them in the beginning of the episode, and we do get.
I, you know, maybe I'm self-critical.

(01:34:42):
Sometimes when I release something, I'm not sure that
this is really that good or it's, you know, we
go through like, 35 articles in rapid succession and,
there's really not many podcasts out to ever do this kind of thing.
Like, usually you cover maybe a few stories.
Right? We are pretty different.

(01:35:03):
And, some times you're not sure that maybe we should do
things like, totally differently and we better, but
it's nice to see at least a decent number of people like
that. We cover a lot, and then they say things like, this is great.
This is your top source for I news.
This is helping me keep up with I, as a journalist

(01:35:26):
or as, a PhD student, things like that.
So.
Yeah. That's that's been very.
Nice to see. And, you know, in the first couple
years, we were tiny and didn't have much of a listenership, and
we just did it for fun.
So it's over the past year, pretty much, a year and a half

(01:35:50):
that started happening and been nice to see.

Daliana (01:35:53):
Yeah. I really, you know, respect
that you want to keep this simple and pure not to have,
sponsorships.
I don't know, maybe someone audience, video editor or something.
They want to volunteer or help out or,
or. I don't know if you're open.

(01:36:13):
I'm sure if people want to do something like buy you a coffee or
donation, you can buy better equipment.
You can, you know, put it back to the podcast or have a haircut
or something.

Andrey (01:36:24):
Yeah. And I will say, we I
suspect we've had a good number of paid subscribers on our Substack.

Daliana (01:36:32):
Oh, nice.

Andrey (01:36:33):
Because of the podcast.
Partially. And so if you do really want to support and are,
you know, a big fan, then one way would be to go over to the last
week in that I cite and
we do have paid, subscriptions enabled on the Substack.
And we don't do

(01:36:56):
like that. That's more pretty much for peer support.
We have, an archive of editorials you can get access to if
you do that. But aside from that, there's not many perks.
We release everything to everyone.
Yeah.

Daliana (01:37:12):
Yeah. And if you like, leave a comment.
Subscribe, subscribe on YouTube.
Five star review. I think those things definitely helps.
And speaking of haircut, and before we wrap up, is your hair
naturally blond?

Andrey (01:37:25):
No, no. Your hair? No, it's partially bleached.
Oh, you can see.

Daliana (01:37:30):
Yeah, it's our signature.

Andrey (01:37:33):
I don't know about that, but. Yeah, I have dyed my hair a few times
over the years, and.

Daliana (01:37:38):
They're going for some crazy color.

Andrey (01:37:41):
Not. It's kind of boring.
I did dark blue last time.
Okay? Like a grayish thing.
So I don't think I'll ever go for pink or purple or something,
but who knows?
I think it's fun. I do try to do a little bit self self-expression.

Daliana (01:37:58):
I can do about gradient.
Oh, sorry. Gradient.
No pun intended.

Andrey (01:38:03):
That's fun. I think that's probably pretty hard to pull off.
Yeah, in terms of process.
But aside from bleaching my hair, a fun fact about me is I have
how many?
Seven tattoos.
Oh, wow. So I'm not quite as nerdy as I might seem.
Or maybe I am, but.

Daliana (01:38:24):
Do you have anything you can show us? Is it like a, you know, any
parts available? Oh. What's that?
It's like a tree with a.

Andrey (01:38:31):
Cube of a tree.
And then I have this.

Daliana (01:38:35):
Wow. What is this? Also a tree, but like.

Andrey (01:38:38):
Like a neuron tree.

Daliana (01:38:41):
Is this, like, the kind of the brain stem from Westworld?

Andrey (01:38:44):
This is where I just found a photo of a neuron, and.

Daliana (01:38:48):
Yeah.

Andrey (01:38:49):
I did that, so.

Daliana (01:38:50):
Oh, wow. I don't know if that make you more nerdy or less morally
nerdy. Cool.

Andrey (01:38:56):
Both. Yeah.
Yeah.

Daliana (01:38:59):
Yeah. Awesome.
So. Yeah. What is something you're excited in your life
or career in the next couple of months?

Andrey (01:39:09):
I'm excited, I think.
Hopefully to.
Really figure out how to
in the startup. We're working on we've had some very productive
discussions on, you know, the stuff we've built so far.
Maybe in some ways.

(01:39:31):
We underestimated the difficulty of what we were taking on.
And so I'm excited to
hopefully.
Have a lot more insight and manage to build the right thing,
so to speak. And then aside from that.

(01:39:53):
I have.
Been thinking about maybe getting back to writing more.
Doing some essays and maybe even some short fiction.
Just to spread out sort of my focus.
So, we'll see.
I'll try to see if I can do that.

Daliana (01:40:13):
Yeah. That's. That's awesome.
Anything else you want to tell your audience, or do you want them to
know?

Andrey (01:40:24):
Not. Not really. I think.
I will say, yeah.
It's always humbling to think that so
many people listen. Yeah.
And that a lot of people seem to appreciate this thing
a lot. That to me is,

(01:40:45):
just a fun thing to do.
Our thing I, enjoy doing.
So, I do want to thank people who
do reach out or leave reviews and so on because,
Charles, nice to have done something that a lot of people
are if this in some small way touched by.

Daliana (01:41:05):
Yeah.
Cool. Yeah. This is fun. Maybe sometime I invite you to my podcast,
or you can interview me or talk more about your journey
someday.

Andrey (01:41:16):
Maybe. Yeah. That was fun.
But we are getting a bit far into this recording, so I think
we'll go ahead. And that is where.
And this will be the end of this episode
of last week, and I so thank you, Diana, for co-hosting and for
doing this fun interview segment that hopefully our

(01:41:38):
listeners will enjoy.

Daliana (01:41:39):
Yeah, thanks for inviting me.

Andrey (01:41:42):
And as we mentioned in our conversation, and as I say at the end,
we would like you to share a podcast or leave a review
to do those nice things.
But more than anything, we do like knowing that people listen
and benefit from us doing this.
So please do keep tuning in.

Unidentified (01:42:03):
I was.
Without me and.
I'm bringing all the news to you.
Me?
Too.

(01:42:25):
In. Next time on that island breeze
you'll be.
Take.

Andrey (01:42:34):
I want you to march on.

Unidentified (01:42:37):
I know the angst.
And and and and and and.

All Episodes

#167 - GPT-4o, Project Astra, Veo, OpenAI Departures, Interview with Andrey

Episode Transcript

Popular Podcasts

United States of Kennedy

Dateline NBC

Bookmarked by Reese's Book Club

.css-15opob5{left:0;position:absolute;top:0.8rem;} All Episodes

.css-14f5ked{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:2;overflow:hidden;}#167 - GPT-4o, Project Astra, Veo, OpenAI Departures, Interview with Andrey