Subscribe
Share
Share
Embed
The NeuralPod AI Podcast
The NeuralPod is all about deep diving into technical machine learning topics and showing its real-world impact, bridging the gap between AI-first companies and adoption.
We chat with subject matter experts across reinforcement learning, recommender systems, deep learning and generative AI. As well as chatting with business leaders, VCs, operations and strategy experts.
Who are NeuralRec.ai? NeuralRec is a recruitment staffing agency. We build niche ML teams and represent some of the globe's best ML talent.
Alberto: I think there are some things
that haven't changed and they will
never change, and some things that are
radically changing in the low level.
Everything is changing in
the, in the, in the specifics.
Everything is changing, in the skills.
Nothing has changed.
Chris: So, um, Alberto, welcome
to the, uh, the Neuro po.
It's, uh, it's a pleasure
to have you on today.
How are you doing?
I.
Alberto: Right.
Thank you.
Chris: You are welcome.
Um, so yeah, we're going to, um,
discuss, uh, your background today.
Um, some of the work you're doing
on agents and, um, but you've
also been a career coach, so
something we've got in common there.
Um, we can discuss that and, and the
market and as always, some future
predictions and, and where you see
the current markets specifically in.
Agen ai go in and know.
It's kind of like how long is a piece
of string at the moment, because,
um, the market's so, so fast moving.
But, um, it'll be great
to get your thoughts.
So, um, yeah, over to you Alberto.
You're a, you're an engineer by
training, but now AI researcher
and, and leading some products,
uh, projects over at Criteo.
Uh, do you wanna just give it a quick,
high level overview of your career and,
and how you got to where you are today?
Alberto: Yeah, sure.
So, so the thing is, uh, my case, I, I
have a kind of an, an orthodox career.
So I started my career like 20
years ago as telecom and engineer.
Telecom eng uh,
telecommunication engineer.
And I started working in
Telefonica, in, in Spain, like
as a, as a research engineer.
so the thing is that after five,
six years, I wanted to do something
else, something like beyond just
like, let's say plain engineering.
then actually I got back
to university to study pH.
Uh, but then I said, okay, it's, it's
kind of hard to mix philosophy with
a technical, my, my engineer life.
So then I started to find AI as
an answer to my, to my question.
So it's the closest that I could find
in the technical world the closest
that I could find to, uh, philosophy.
So actually that's why I got
interested in, in, in ai.
and then I enrolled,
uh, in a, in a master.
And to, to, to, to follow my, then my, my
PhD And what happened at that time is was
it was like 2020, uh, 20, 20 10, 20 11.
So, you know, the 20, the, the 2008
crisis, uh, hit then we are in Spain
in this post-crisis, uh, world.
Um, people were starting to get angry.
We had like super high levels of
unemployment rates in young people.
Like, um, the crime crisis
also hit very hot in Spain.
Like people were like evict,
evicted from their homes 'cause
they couldn't pay the mortgage.
And then also we have like many.
corruption cases.
Uh, and then people were like,
there was this feeling of this
generation that the democratic
system was not good enough, right?
So it started this the 50 may
movement, like 20, 20 11, and,
and was involved in this, right?
So I was like, study my master in ai.
And, and at the same time, uh, getting
involved in this, in this movement.
So what happened is that a group
of people, like more engineers and,
and the like, we started to create a
group to analyze the social movements.
Uh, online.
we started to do like social network
analysis methods to analyze, to
understand what what's going on, uh,
with this kind of social movements.
And so the, this got me
interested in sociology and
social network analysis and so on.
I had the opportunity to
start a PhD HC exactly.
On this, right on, on, on.
Um.
Analyzing, uh, uses behaviors, uh,
online with machine learning techniques.
So that's how I just jumped in, into
the, into this machine learning world.
Um, so I moved to France and I started
my PhD, and then I did a postdoc.
And then finally I just, uh, joined
joint c some later as a, as a
research engineer, as a researcher,
Chris: Nice.
Alberto: sorry.
Chris: Yes.
Um, makes sense.
And, uh, I guess that the recommended
systems connection to Criteo, it kind of,
um, made sense I guess with the types of,
uh, clients and projects they work on.
Um, all right, well, um, yeah,
let's talk in to, um, you
know, your career now at, um.
Criteo, and obviously you've got
a, a great scientific background
with you, your experience so far.
Um, your helps shape the scientific
agenda at Criteo, uh, specifically
around agents, which is a, a huge
buzz in the market right now.
A bit of a buzzword, um, but you are
very much chartering, uh, the unknown
there with some of the cutting edge
research that you're doing, um, and
developing, uh, real world products.
Um.
Do you want to just talk us through
kind of, um, some of the work around
shopping assistance that you're doing in
agent interfaces and specifically, you
know, everybody's kind of definitions of
agents may be different at the moment.
How is it affecting
e-commerce specifically?
And, um, you know, what type
of agents are you developing?
Alberto: the agent systems are
gonna change the way people shop,
people buy, and then therefore
it's gonna affect the Tels business
or the ad tech, uh, business.
And why?
Well, because and more
you will have people.
going directly to the website
to buy products, but they will
ask assistance, their agents to
do it by, uh, for them, right?
then it means that it doesn't make sense
for curtail to show any ad, uh, to the
agent because the agent is, is a bot.
So he, he doesn't care.
He's not impressed by your display.
you need to maximize now,
you need to convince.
The This middleman.
This middleman, who is the, the shopping
assistant you have the best product,
that you have the best recommendation.
So you have to find, you have
to be sure that you are giving
us super good recommendation.
And you have also to explain.
Why you think like all the, all
the, you have to give all the things
that you didn't give in traditional
recommended systems, such as why
you think this product is the best.
Comparing products, comparing
alternatives, like all, all doing all
these kind of product search that we do.
When you buy a computer, you are
comparing three alternatives.
You got super type of doing this.
So we need to, uh, give all
the information to the agent.
To, to, to make, to, to, to, yeah.
Put everything together for
the, for the user to make this,
to take this final decision.
Um, that's so in, in, in shopping, that's
one thing that's, that's will happen.
So in curtail, what aim for is
to be the best, recommend the
system for this kind of agents.
to, to call.
Right.
and these recommended systems,
these agents will, and we are seeing
already in open ai, we're seeing,
we are seeing already the first
movements towards these directions.
You will have this, like the main
open ai, Google and whatever.
So they will need these
catalogs and they will.
Not just contact retail, but they
will contact like many providers.
So that's why you really need to be
better than the, than than the others.
Right.
also we can also provide our own agents.
There is also why, like, why
not, like we're also exploring
these alternative on developing
some products in this direction.
Like, uh, doing this
end-to-end solution where we.
Give this agent to the
user in some context.
And we also give the the recommendations
and, and then also we are
developing agents for our clients.
So our clients now they have to deal
with, for instance, when you onboard
in any at the company, at you tell.
Well, have to create a campaign.
You have to upload your catalog.
You have to like do a lot of
bureaucracy to get things started.
And this is painful.
Like this is like, yeah, you
have to curate your catalog.
You have to decide your audience
to give all the, yeah, a lot
of information, metadata.
So we want to automate, we are
automatizing this so that we.
Let's say we shorten the cycles
from days to hours or minutes.
Chris: And, you know, just
take it back a second.
What, what do you think's
actually gonna add?
Happened to the ad tech industry now,
um, you know, AgTech, um, interfaces
and agents are being introduced.
Um, how's it gonna shape the future?
Will it just be a case of agents
will now be interacting, um, with
huge catalogs of data and, um, yeah.
What, what are your thoughts there?
Alberto: What do you I don't, I dunno
if I, can you, can you repeat the
Chris: In, in terms of the, the,
the ad tech industry, obviously
it's a, it is a huge industry.
Um, you know, if people aren't
looking at ads anymore, how do you
actually think it's, it's going to
be affected moving forward in terms
of the size of the industry and um,
Alberto: Yeah.
Chris: yeah.
Alberto: We have like different, different
markets, different realities, right?
So the traditional ads, they won't
disappear because you will still,
like with your own eyes, uh, read
a website or visit a retailer
website or things like that.
So it will exist.
and you will have this traditional
displays, um, even better because this
traditional now with lms, we're able
to understand more about the context.
For instance, like before it was
kind of hard to understand what
was the website talking about?
And, and now you can say, okay,
this website is talking about this
topic so I can put a more relevant
ad, um, in this, in this website.
Right?
Um, but then also you
have this space where.
The user is not interacting with the
website, but with the agent itself.
But yet still the, the, these
agents need some way to monetize
their, their, their platforms.
Right?
It's, it's expensive nowadays to
call an lm, even if it's, well,
the price is getting, uh, lower.
Uh, but, uh, but still, I mean, you, you
need a business model for this, right?
So, we think that many, many agents.
Chatbots in retailers, in WhatsApp, in
OpenAI, whatever, will need this the way,
this way to monetize, to make their space
available for, uh, for, for brands, right?
To put their, their, their
recommendations there.
So yeah, I think that the, the two walls
will, will, will, will quite co coexist.
Chris: Okay.
Interesting.
Um, and yeah, just, just to pick up on
where you left off, uh, a second ago,
uh, designing agents for commerce,
um, you've kind of touched on some
of the challenges there around it.
Expense, et cetera.
But what, what are the, um,
real challenges in developing
agents for e-commerce over,
uh, say traditional search and
recommendation engines and systems?
Alberto: You can, you can even
see there's an opportunity rather
than as a challenge if you want.
'cause the thing is that you can do more.
Yeah.
It's really, it's like you can do
like traditional recommendation if
you want, but you can also do more.
So you can use context that
you couldn't use before.
You can planning, you can do,
like now you can, I mean, you
can interact with the user.
That's super, uh, one,
one important thing.
Now the user now in many,
in many, in many, in many.
Use cases, the user just, uh,
for instance, in a retailer,
the user enters a keyword like,
I'm looking for milk for a bike.
And then you will have this
conversation, and now you will
have this conversational interface.
We'll say, okay, bike.
What kind of bike?
What it's for my daughter.
Okay.
What's.
How old is your daughter?
Like, give me all these clarifying
questions that will have the recommended
system to find a more relevant product.
And it will also be able to
remember your preferences.
It will remember if it's your personal
or you have a, you are locked in a, in
a, you have an accounting disservice.
It will remember your,
your, yeah, your preference.
What you said before.
If remember you have a family that you
live in London, that you don't like this
brand, that you like the other brand.
these things so that the user
doesn't, doesn't have to repeat
things to the agent, uh, every time.
Right.
So the experience will be much more
enjoyable for the, for the user.
And then on our side, on the
recommended side, we'll have to
be able to, yeah, to deal with all
these contents, with all the memory.
For instance, there is a lot of open
research in terms of how do you manage
the memory, what do you have to remember?
Memory recommended
systems was didn't exist.
You just have a classic machine
learning system where you have a
lot of examples, input, output, and
you train your system With this, now
you have an agent for every user.
So this agent.
We'll have to remember to have
a specific memory for that user.
And that means many
strategies of memory, right?
So you will have like episodic memory, uh,
contextual memory, like short-term memory.
Long-term memory, depending on
what, like the user didn't like
this product and he told me why.
So I should remember, it's somehow in
some kind of memory the user clicked here.
I should remember the user
said he has a dog, so I should
remember it somewhere, right?
And then.
Memory is not enough.
Then in every context, you need to be able
to retrieve the proper, uh, memory piece
that's gonna be relevant for this context.
So how do you retrieve this so that
the context, uh, the working context
of the, of recommendation is as
clear and as useful as possible?
Right.
And then also you have the adaptability.
That's another, another big,
big stream of research that's,
uh, that's starting these days.
So this agent needs to be adaptable.
Uh, you may change, you users
will change the preferences.
It's not the same.
For instance, uh, you shopping, uh,
patterns in Christmas, then in summer,
uh, maybe if it's your agent, you are,
you have been with your agent for years.
It'll have to adapt.
Your, your, I mean your, yeah,
your preference and so on.
They adapt.
You buy objects they have to remember.
So this kind of adaptability for
life, um, phases and even for,
for many different circumstances.
That's another thing that
agents need to, to deal with.
Chris: And, um, yeah, you've
already touched on how it's gonna
personalize and create a more enjoyable
shopping experience for people.
Um, and I guess nobody really knows,
uh, the answer to this question, but
how do you think it's gonna evolve over
the next couple of years in terms of
shaping people's experiences and adoption?
Um, you know, where do
you actually see it going?
Alberto: The agent shopping.
Um,
Chris: Yeah.
Alberto: I think what we are seeing now
that younger people, they're embracing it.
Uh, I mean, it's a very
typical pattern, right?
Like all the people, they are more
reluctant they, and, and the younger
people, they are already, like most
of them, they're already used to chat
with these systems every day and they
trust more in these systems and so on.
So I think what will happen is
that now we are seeing a phase.
We're seeing, we, we, we are, we, we have.
Been through a phase where people were
more, what if the system hallucinates?
Uh, what if like, yeah.
A lot of fears and a lot of reluctancy.
We will see that these systems were a
lot get better with time, much better.
Um, and then actually there was this
study that, You know, the, the capacity
of the systems, it's like being
multiplied by a factor of two every,
what was it, every, every seven months
I think it was the, the pattern, right?
So, um, then we will learn also to
better engineer the systems to increase
the, more the efficiency and so on.
So system, so people will, will
also trust more in the systems.
And I think, I mean, I, I think like, um.
Let's say in, in some years,
everyone will use an agent for
shopping their daily lives.
That would be my, my prediction.
Chris: Yeah.
And, um, I suppose touching on your
philosophy background here, you know,
if people are adopting agents more,
and this is completely unrelated to
machine learning, but do you think it
will affect the retail High Street and
people actually doing in-person shopping?
If you've got an, you know, an
agent doing all the legwork, you
know, you're just clicking buy at
the checkout, et cetera, um, yeah.
Do you think it'll affect the High Street?
Alberto: yeah, yeah.
I think it will affect, we already
seeing some movement from like, um,
plexity or Open ai, um, trying to like
opening their, and accessing to, to, to,
to, to catalogs, to product catalogs.
And, um, and then also doing the checkout
themselves without you leaving their
website, you are being able to buy a
product here in this catalog, another
product from other catalog, and then
you do everything, uh, with, with,
without leaving their, their website.
And that's challenge for retailers
because they, they want the opposite.
They want people go to
the website they also
Chris: Hmm.
Alberto: to control the user experience.
This is also interesting data for the
retailers, that they can learn from.
So there is gonna be a T of war,
and I don't think retailers are
gonna give up easily, uh, this,
this, uh, uh, this thing to, to
the, to the big, uh, tech companies.
Um, yeah, but they will.
But the thing is, I think, so it
will be an opportunity there are.
Is moving in this direction is the
position of s is helping retailers
actually to, to position themselves
very well in this new scenario.
So the challenge is how do you.
Monetize.
How and how do you prepare
your catalog to be accessible
for agents, but in a fair way.
So you don't want your catalog to
be stolen, consumed by agents, like
tracked by agents and doing the
checkout on any any other place.
And out of your, out of your
website, you, so how do you.
find a good, right.
How do you, so, and that means that
they need to create their catalogs to be
accessible for like, give some value for
agents to consume and then gain getting
paid for getting, uh, paid for it, right?
And that also implies the
idea of data marketplaces.
And that also some thing that we're
exploring cur retail, this idea of
data marketplace where, um, like our
brands, our partners, our clients will.
able to control the way they share
their products with third parties.
Chris: Interesting.
And, um, you know, there's some retailers
out there with, uh, amazing, uh, data
and AI teams, but do you think that in
general, imagine Criteo's got a lot of
different types of customers that, um,
you know, the average retailer is ready
for the agent wave of a user having its
own personal agent shopping assistant?
Um.
And people are ready for what's to come
Alberto: You mean if client, if Q tells
clients they're ready for what's gone?
Chris: it ju just in general, from
what you see in the marketplace, is
the average retailer ready for, um,
you know, the agent wave, do you think
Alberto: I
Chris: I.
Alberto: the market is ready yet,
but I think that we have no choice.
It's not a, it's, it's a,
it's like a force of nature.
It's gonna come, but you have
to adapt, and if possible, you
have to be first and you have to
be the leader to shape the way.
The world and the market is changing,
but, uh, but that's gonna change for sure.
And everyone, everyone is gonna be adapt.
I mean, if you and those
who don't adapt will.
As, as usual, but yeah, but well, that's a
Chris: Okay.
Alberto: companies as the role is to
help these companies, these retailers,
to, to, to adapt to this new scenario.
And I think there will be another,
like more opportunities for new
companies, for new, there will be
new business opportunities that
will appear in the, in the next
Chris: Hmm.
Alberto: idea of data curation, of how
do you, how do you create your data?
How do you create like good.
Understandable catalogs that
are easily searchable by agents,
something that needs still to be done.
Chris: Yeah.
Well you've, that segues nicely onto
the next section of the podcast.
You mentioned to me data is a competitive
moat and you know, a lot of people
want to implement the latest models.
And you said before models
are increasingly replicable.
Um, but unique data is a real edge.
Um, you know, what, what
do you mean by that?
Alberto: Yeah.
Yeah, exactly.
So what I mean is that a model in
ai, I mean, I don't know of any
company whose advantage is having the
best neural net, let's say, right?
So maybe there is, in some niche
market, you have a specialized
model for some specific use cases.
They will probably exist, it's
not the main strategy, let's say.
Right.
in companies like retail,
that's, that's gonna be the case.
So the model is not your mode, the model.
at some point we are seeing this
with agents with l lms, right?
So you are, you are having like, open
source models are today powerful.
And the closed source
models were one year ago.
So there is a, there
is a delay if you want.
But sooner or later we get there, right?
The, the open source
committee is gonna get there.
And that happens with, all AI models.
There are like, uh, common, common
knowledge there is no secret sauce.
There is some secret sauce at some point.
Now, open ai, for instance,
they train the models.
They post train, they create
the data, they create like super
Synthetic data, data sets, that
create very powerful models.
But you have behind like open
source companies that are like
catching up and saying, okay, you're
doing this, you're doing that.
Now we are going, we're
gonna replicate it.
That's something, for instance,
we saw with, with deep seek.
Right?
point, deep seek was okay.
Found the method to do the same thing
that you guys do with less resources,
and that's gonna happen every time.
So the, so the mode is the data.
So the thing is that the majority of
the, the most part of the data is, uh.
Has been already used by, I mean, most
part of the public data, all the public
data has been used by lms, has been
seen and used to train these, these
all lms, they all have seen already,
posts, all websites, all the newspapers.
That's already, that's,
that's done, that's used.
So what remains is two things.
Um, synthetic data.
one thing.
And then we are seeing like, um,
in research and like companies,
they are figuring out how to
create synthetic data is realistic
enough for your model to learn.
You bootstrap your data from some real
and then with some loops, with some,
uh, small techniques, you create some
realistic synthetic dataset, which
is way larger than the original one.
And then you can use your model,
you can train your model there.
So that's one main technique.
And the other one is private data.
So there is a lot of
private data that has.
A lot of value, that doesn't
have the public data.
For instance, commerce data, right?
The shopping patterns of users.
This is something that
you don't find on Reddit.
So this is something that Tel
has, Google has and and so on.
So having this data is gold because
then you have the same model or
more or less similar model to the
one that everyone has, but you have
fed this model with unique data,
and that's what creates your mode.
Chris: Hmm.
Interesting.
And, um, you know, what, what, what
are some of the most overlooked data
in, in ad tech, um, that you think.
Alberto: Sorry, can you
repeat the question, Chris?
Chris: Sure.
Yeah.
Yeah.
So what, what do you think some of
the most overlooked data a assets in
ad tech are, um, that are, you know,
the most important for the ENT era?
Alberto: The assets,
assets of data assets.
Data
Chris: assets.
Yeah.
Yeah.
I.
Alberto: I would say, well, one thing
is, but I dunno how it will play out,
but the with all those services, I think
that will be, that might be important.
So if your agent is isolated,
it can only do so much.
But if it can access to, uh, yeah, to
all services, like getting integrated,
sharing data all those services like,
you know, banks, your, if you are in
sports like your Strava, uh, data,
your, you know, uh, calendars, like
this kind of integration, I think that
will be, that will be key to give, uh,
your agent or any shopping assistant
or any personal agent, uh, value.
That's one thing.
The other thing is pro, pro descriptions.
That's, for sure.
And that's something we're working
very hard in, in cri retail.
Um, we, the catalogs today, they are not,
they were not ready for the Argent era.
So brands, they have to
explore the catalogs to many,
let's say a tech companies
The thing is that the traditional
display as a title, an image, a price,
it, but that's not enough for the agent.
The agent needs more information.
The product description, like clean,
clear, pro description today, the pro
descriptions that we have in our catalogs
are like robotics, like four words.
Um, like often they repeat the information
that is already in the title or in the
other metadata, so it's very useful.
Companies will have to provide more
data for, for agents to consume.
And Google also have to learn
to, to create this data.
Um, or for trouble, for instance.
For trouble is a specific, a
very challenging, uh, vertical
because the data is dynamic.
It's not like if you, when you go to buy
a lab, if you want to explore, catalog for
electronics, everything is nothing moves.
It's like you're looking for a computer,
just do some semantic, search, some
filtering, find your computers.
Then the, your agent is gonna think,
okay, is it relevant for the user or not?
But for travel, it's okay.
Go to go for, I go to
go on holidays to Italy.
Okay, let me take some flights.
No, there are no available
fights for flights for that date.
What about this, the next weekend,
what about a train instead of a flight?
What about, so this hotel is booked also.
So what about this alternative?
So you have to do this combination
of, of your, of, of multi-product
recommendation, right?
And, and you don't know
all the information.
You have to query APIs to see whether
the service is still available and so on.
So that's.
That's the other thing.
That's the other intent of data.
That's, that's the other challenge.
And, and then also the, the other data
asset will be, so before I talk about
this, the memory, uh, challenges.
So that's, that's the thing, right?
So the, the long-term use of
history, uh, past interactions,
um, personal information, family
data, um, all these kind of things.
And, and I think here, maybe
we'll start in a new era.
Where users that will be more willing
to give more information in exchange,
uh, of good quality services.
Chris: Yeah, that's an interesting point.
And um, I think historically people
have been quite hesitant about that.
But I guess if it was yours actually
to see a tangible increase in a better
shopping experience, you might be
more willing to, um, to pass over.
Interesting.
Cool.
Okay.
Thank you for sharing that.
Um, moving on to the next part, research,
engineering and, and the talent crossover.
Um.
Well, I think you've been fortunate
to work on, um, a, a couple
different sides of, of defense.
Um, but in particular your, your
research background is really strong
and, um, yeah, really keen for
you to share some insights there.
Um, you know, I've worked with.
Maybe hundreds of companies at this
point, and a lot of them define, um,
you know, the engineering capabilities
of a researcher in different ways.
You know, some of them almost
want, um, ML engineering, um,
levels of, of capabilities.
In your opinion, you know, what, what
should every good ML researcher be able
to do from a engineering standpoint?
Alberto: I think there are some things
that haven't changed and they will
never change, and some things that are
radically changing in the low level.
Everything is changing in
the, in the, in the specifics.
Everything is changing, uh, in the skills.
Nothing has changed.
So in the skill set, I think it's,
that's true for any engineer,
for any profession, probably.
So curiosity productiveness,
so the idea of.
I mean, I don't know how
many times I have changed.
My topic in terms of research product,
I started doing social network
analysis, then I did patient inference,
recommended systems, multitask
learning, then NLP, LLMs agents.
So every two years is kind of a new
cycle, hopefully building, uh, on
top of the, of the last one, right?
So there is some connection
between cycles, but, you have
to recycle a lot every time and,
and things are moving very fast.
So you have to be, you have
to have this curiosity to.
To be up to date with,
uh, with what's going on.
that will mean that that's the difference
for a, between a good engineer, a good
researcher, and a not as good one.
Right?
So the, the good ones
are the ones that are
updated with the last, um, trends,
the last techniques, and, and so on.
That's, and that's always true,
um, in the low level, in the more.
Technical level, everything is changing
because now, um, you are not priming.
I mean, in some, in some areas,
people are not perming anymore.
using, so this year especially has
been the boom of, uh, coding agents.
Right.
So now, before, like one year ago, well
two years ago, you were like coding
classically, then Chatt PT appeared and
you were like, okay, asking CT PT to.
Or any, or code or whatever to give you
some function to do whatever you wanted
and you copy paste it into your code.
you have this integrated into your, into
your coding interface, into your IDE.
now you are, you are in, and the, the,
the assistant is looking at your code
and it's creating the test, and it's
doing the interface and it's doing,
so it's, it's kind of augmenting your
capabilities, so being able to play with
it, to build with it, that's critical.
And still today we see some engineers.
I see some engineers that they are
not using these tools and they are way
slower than the ones that are using them.
So now we have a cap, a gap the ones
that are really using these tools and
the ones that are still not using them.
So that's super important.
Yeah.
Chris: Yeah, in.
Interesting.
Um, okay.
And, um, you know, in terms of
advising someone how to work with AI
right now, um, you know, in terms of
co-generation tools, for example, um,
you know.
Alberto: Hello,
Chris: often, how soon should DB implement
them into their, uh, daily workflows?
For example, um, would you advise
people to use 'em straight into
university or, um, you know, when,
when's the best time to, to start?
Alberto: To start using these systems?
Yes.
Chris: Yeah.
Alberto: I dunno, I don't
have the answer for this.
So the advantage of my generation,
let's say, is that we have
already built without agents.
So we have this practice and
now we know how to validate.
When the agent gives some code or some
solution, you can say, okay, this is
good, or this is not good, or change this.
You play the role as an architect.
Rather than the role
of a developer, right.
For people who start now?
I dunno.
I think yeah, for sure.
Like doing things themselves from
scratch, it's, it will always needed.
Um, but I mean, you only have a finite
amount of energy and enough time,
so you will have to sacrifice some
part of it and dedicate some time.
Some of this time you will have to move it
to learn and to use these, these AI tools.
Chris: Yeah, the, the rea the reason you
ask, I, I found a friend of mine, um, he
recently asked me to hire him, someone,
um, who'd done their PhD pre-chat, GPT.
Um, and I just wondered, um, is he quite
unique or is, is that potentially gonna be
a thing with, you know, someone's problem
solving ability or, or coding ability,
uh, where the people start asking for, uh,
candidates, you know, pre, pre, pre, um.
It's pre, um, open ai,
so, we'll, we'll see.
Alberto: yeah.
Yeah.
Chris: Um, cool.
And, you know, as a career coach,
what, what advice would you
give someone in general entering
into the age of AI right now?
Because it, the, like you say, some
things have changed, uh, some things
haven't, but you know, a lot has changed.
What, what advice would you give
someone coming into the market?
I guess I.
Alberto: Build things.
Get your hands 30.
Uh, so now you have the opportunity
of building things from end to end.
Like if you are a researcher or a, like an
engineer, you don't have any training on
front end, for instance, or on database.
now with the systems.
can use them to build all the other
things that you don't know how to
build or that would take you MOS learn.
So that's empowering, right?
So now you can create like apps, uh,
anything you want with the system.
So.
You can, you can focus on the
central part, that it's more
interesting to you of your profiles.
You are a researcher or ML engineer.
You can focus on your models,
your data, we have whatever.
And, and then you can have all
the frontend and everything
delegated to the, to the agents.
So, and the advantage of this, so it
changes everything because now, for
instance, in my case, it gives me.
I had never called it a
complex frontend before.
I had done like websites or CSS
played with some basic stuff, but
then now I, I am able to, yeah, like
to do complex databases, complex
frontends with react, uh, apps.
Um, a lot of API so now I can talk to,
to the engines in my company, uh, like
at the, at the, if at the same level,
but really it grounded things, right?
So now I understand much more what
they have in mind because I have
also built them on my, on my side.
And yeah, so it's gonna create,
I think it should create like
engineers that are more transverse
but know that are maybe specialized
in one thing, but it's still know.
How to do the other
things more than before.
Chris: Hmm, interesting.
Um, okay.
And do you, do you think potentially,
you know, with an engineer or researcher
able to do more with, um, you know,
these tools out there that will lead
to companies wanting bigger teams?
Because now people can do more or
actually, um, um, you know, companies will
want their researchers to do more, uh, do.
Do more with less people, basically
with the tools that are out there,
what, what are your thoughts on that?
Alberto: I think it would
depend the market, right?
me it's like, uh, it's
an economics problem.
Um, if I think if your
market is saturated.
Then you can do the same
thing with less people.
Because if you don't expect to grow your
market, then the only thing you can do
to grow your profit is to reduce cost.
So that means productivity
with less people.
If, there is a still market for you, then.
Will be room for you to, yeah.
To, to, to increase your team even.
Right.
um, yeah, so it will depend on the market.
Chris: Interesting.
Okay, cool.
Well, um, yeah, thanks, thanks
for sharing that as well.
Let's have, um, a little bit of
fun around future predictions.
Just a couple of
questions we ask everyone.
And you know, I think it'd be pretty cool
to revisit it in the next, uh, six to
12 months and, and see where we're at.
So, um, you know, what, what
trend in development, it doesn't
have to be in e-commerce.
Um, Orgen AI did mention Agen AI
to you before the show, but what,
um, yeah, what development in AI
are you looking forward to the most
over the next, um, coming months?
Alberto: Well, one thing that I
think would, would happen is glasses.
that's something I don't personally love.
don't see people like wearing
glasses rather than meta glasses.
But,
Chris: Yeah, even.
Yeah,
Alberto: but I think we will start seeing
this, uh, sooner or later, and that
will also affect the attack industry
also, because that's another space which
you might think of, ads or whatever.
So it's, it's another
Chris: I think I'd throw them in the bin.
If I got an ad in my glasses,
it'd definitely go in the bin.
Alberto: Either you have these or
you pay more expensive glasses.
So it's depending on, depending on
Chris: Yeah.
Fair?
Yeah.
Yeah.
Yeah, no, that's very good point.
Alberto: uh, but um, yeah, I
mean that, that's one thing.
And you know, I mean, we, we all know
that met times already trying to, to push
for these and other companies and Yeah.
the other thing that I would like to see.
It's about open source, open source
projects, and funded projects.
think it is very important now that
the open source community, uh, is a
strong as support, uh, from society
and governments and so on to avoid.
Increasing the gap between the
big tech companies with the
big private models, with big,
quantities of private data and so
on, and the rest of the world, right?
So AI should become a commodity,
like electricity, right?
So AI should be.
Available for everyone, for any
company to use the, to use it
to increase their productivity,
to build services and and so on.
So as much as possible you will have
then of course, like some, like the
very state of the art models, whatever.
So things that more
especially that will be.
Not community will be really the key,
the core business for some companies.
Sure.
But there are things that should be
like computing, like having, having,
uh, sort of a Chat GPT for everyone an
accessible price should be a comodity
And I think it'll become a commodity.
But I think and also like multilingual
that doesn't discriminate, that it's not
like super English focused and also deals
with other minority languages and so on.
So this is, we need also governments
to get involved here to guarantee that.
We, I mean, the data is not
biased, is not, has any, you
know, like it's ethical and so on.
And it's, uh, and that it's, uh,
and on top of this we train models
that are accessible for, for anyone.
Yeah.
No, that's one thing.
And the, the other thing is, so
the other thing is small models.
Uh, there is this other trend about, uh.
Yeah, like creating, now we
are like every, every time like
open AI or, or um, These, these
companies, they release a new model.
It's like even larger than the one before.
We know, we don't even know how large
these models are, the private ones.
so, but there is also the other
trend of trying to do the same with
smaller models that fit into your
mobile, into your computer and so on.
And that.
Use less energy.
That's super important.
'cause otherwise you are seeing
like some alman and these companies,
they're trying to, they are investing
in nuclear power because they say
we won't have enough energy to back,
uh, their, our, our AI systems.
So we need like a smarter
and smaller AI systems.
Um.
And the other one, the, the third
one I would say is autonomy.
Something similar to what I, talked before
about, uh, self adaptation and so on.
But the other trend will be like
having agents that will be, now you
are having agents that are able to do
some tasks that takes, you know, some
hours or, I dunno, let's like kind
of, this level of complexity, right?
Some kind of level of complexity.
This level of complexity and
composition of tasks will increase.
at some point you will see agents
that are like independent autonomous
for some days, and then they go back
to you say, okay, that is the answer.
That is my research the properly
formed for you and so on.
um, yeah.
Chris: Yeah.
Nice.
And out of interest, are you using
any agent tools in your personal life?
Um, not in work.
Alberto: In my personal life, I have
descriptions to every model to cloud.
I'm continuously
Chris: Okay.
Fair?
Alberto: I subscription to
Chris: Yeah.
Yeah,
Alberto: cloud.
I use
Chris: of course.
Alberto: code a lot for coding.
I, I also, also use for some, uh, cursor.
Um.
Well, basically these, basically
the programming ones, the
co-pilots for programming.
That's the main, these are
the main ones that I'm using.
And then I also, I try to,
Chris: Fair.
Alberto: to use cps uh, connect
to my Gmail, uh, through agents
to send emails, uh, automatically.
Um, share tickets in the company.
That's also useful.
Right?
We can now start dealing with your geox.
thanks to, thanks to agents.
And Yes, but,
Chris: Nice.
Cool.
Alberto: MCP actually, because the
thing is that we still don't have
enough tools for our agents to exploit.
once we will have tools to access to
your bank account, to your gm a to your
cleaning service, whatever, so it will
be your agents, your agents will be,
will become more and more interesting.
Chris: Yeah, it is gonna be
really interesting to see
how it evolves next year.
I think, um, there's some, some
predictions for a GI in 2027 and we'll,
we'll see if we actually get there
and what that actually looks like.
But it's certainly gonna be,
um, an exciting 18 months.
Um, okay.
And in terms of, you know, your research
background, um, and recent papers, um.
Is there anything that shaped your
thinking or any research direction
that you found especially interesting
or underrated in your career?
Alberto: I think there are, there
is, for instance, recent paper that
appeared last week from one single
author, uh, that is called Less is
More Recursive Reasoning Withstanding
Networks, and it's rethinking the way.
We are creating the, but the
basis to, to agents, right?
So the idea is as I, this, this
goes into the direction of doing
smaller models to the same thing.
So this is a paper that is, uh,
proposing like a new architecture.
Um, do the same thing with
much less, uh, um, resources.
Right.
So that's, that's one interesting thing.
And also the thing is, it's, it's
amazing that it's one single author
and I think that will also change.
So AI will the way we do research now.
Like you have to, research means
reading papers, uh, understanding
the state of the art, doing research,
uh, doing experiments and, and so on.
Especially in experiments.
These are heavy.
These are like painful.
Sometimes you have to spend days finding
a stupid bag that you, don't find, right?
So thanks to these kind of systems, I
hope will, will become, research will
become more enjoyable, faster, and
researchers will be able to focus on.
What matters, right?
Really the idea, how do you mix this
with this, this idea, with this idea.
Like more the innovation part.
and yeah.
What else?
That is the other trend about which,
which is called, which is called Regen,
generative information retrieval, and.
This is something that instead of,
because now when you are looking for
a product, what happens is that you,
you input your query in the system.
Then the system creates a
mathematical representation of this
query, what we call an embedding.
then you have done the same with any
candidate product, and it finds then
the closest products to your query.
So you embed your, your, your query.
And then you find the closest ones.
So generative information with
through systems, they're based
on LMS models and transformers.
What they do is they generate not
the embedding of the product to then
look for not the embedding of the
query to then look for the closest
products, but directly they generate
the ID product you're looking for.
And that's a paradigm,
shift if you want, because.
Then if you, you would be
able to generate the, so it's,
it's one step instead of two.
The system gives you the final answer.
Um, and then also you can say, okay,
now if I can generate the ID with
a generative system, like as NLM, I
could also generate the explanation
of why I recommended this product.
And I could also use reasoning,
uh, techniques, right?
To, to do, to increase my, the
quality of my recommendation given the
context or that I know about the user.
So, yeah, that's so generative.
Uh, AI is something,
interesting to follow.
Yeah, I would say,
Chris: Interesting.
I'll, um, yeah, if you can share the
link to that paper, I'll, I'll put
it in the show notes as well after.
That'd be great.
Um, and final question, just what, just
one book, any book, um, or products, um,
that's changed the way you think about ai.
Alberto: Mm-hmm.
Chris: Or, or, or, um, you know,
really challenged the way you think.
Alberto: there are, yeah,
maybe there are three books.
Let's say two categories.
of them is about, uh, climate crisis
and what's the connection between
AI and data and the internet with,
uh, with climate that thing that's
important that everyone in this, in,
in the industry is aware of, of this.
It's a French journalist,
so the book is in French.
don't think there is a translation
to other language for now,
or maybe to Czech, I dunno.
But the book, in any case called
the in French, it's the digital.
Um, so it's about, yeah.
How are um, computer centers,
data centers, um, like data, um,
all the internet, I mean, all the
data that we are processing with,
uh, in the, in the digital world.
And now even more with, with generative
ai, uh, what's the effect on the, on the.
On the climate, right?
What do we need to be the systems,
what's behind A GPO, right?
In terms of minerals, in terms
of the scarcity of, uh, of yeah.
Of some minerals and so on.
So what's, so that's very,
very interesting book.
And the other one also is about,
um, what they call it, the, it's
about the tive accumulation.
So I think what's happening now, so
the books are Caliban and the witch.
And the other one is the
ma, many headed hydra.
So they are kind of two related
books from, from historians, um,
that analyze how in the beginnings
of capitalism, the process of
primitive accumulation has started.
Right?
So the thing is that you had
some like common land that.
It was for everyone to like to use
to, to enjoy and so on to, to work on.
And then at some point what
happened is that you had this land
getting privatized and, uh, owned
by less and less people, right?
So you have this
accommodation process and.
In England actually, you have a very
famous movement of people trying to
resist to this, which is the digger.
And, and there are like movements
of people trying to out, getting
back the lands to the, to
the, to the common use, right?
I think the same is happening.
So one, once this happens
is very hard to undo it.
And um, now I think the
same is happening with data.
open AI and all these systems, they have
been built in public data, sometimes
in private data with no permission.
Um, and that's too late now.
So it's like once they are there,
everyone is using the systems
and there is no way to undo it.
Right.
Um, so So this kind of data
accumulation, it's happening now these
days and we have to be aware of that.
Chris: Yeah, I'm, I'm going
to check the last two out.
I lo I lo love a history
book to be honest.
So, um, yeah.
Thank you.
Um, alright, well, uh, Alberta,
that concludes today's episode.
Thank you so much for coming on
and, and sharing your knowledge
and um, yeah, being so open.
Uh, I really appreciate
and thanks for your time.
Alberto: Thank you, Chris for having me.
It's been a pleasure.
Chris: Thank you.
Alberto: Bye.