Latest / Agentic Conversations (formally mlops.community) / Beyond the Matrix: AI and the Future of Human Creativity

Transcript

Beyond the Matrix: AI and the Future of Human Creativity · Agentic Conversations (formally mlops.community)

0:00
My name is Fausto Albers, I am the Co founder of the AI
0:04
Builders community, which is a really cool community that has
0:08
meetups in Amsterdam and Berlin every month.
0:11
And we get together with a bunch of AI nerds and and people
0:14
working on a really cool stuff and we share ideas, problems,
0:17
etcetera and some some cool some cool talks.
0:20
I'm doing research consulting in the IAI space from a background
0:25
of behavioral science owned restaurants.
0:27
I was a flare bartender in London and I can use all this
0:30
stuff in this in this amazing, super interesting space.
0:34
I always love talking to you, Dimitrios about, about the
0:37
things that did are possible and may not be possible and that
0:40
sort of thing. And let me get some coffee
0:43
because I do drink a lot of coffee.
0:45
And that is actually because I swear I, I, I, I am carrier of a
0:50
gene that is associated with a very fast metabolism of a
0:55
caffeine. And when people ask me how to
0:57
how I drink my coffee is a lot. I can, I literally drink a liter
1:00
of coffee before lunch. I can drink a double espresso
1:03
before I go to sleep and I will sleep like a baby.
1:06
Drinking coffee is just like making love in a canoe.
1:09
It's fucking close to water. A quick little tidbit for you
1:12
all before we dive in. I met Fausto and instantly fell
1:18
in love with this guy because he was running a restaurant
1:22
business. He also is super creative and he
1:27
doesn't let any of that stop him from being an engineer, just
1:33
constantly iterating on AI and AI products.
1:38
He brings so many diverse perspectives to the table
1:43
whenever I talk to him that I am in the middle of trying to con
1:49
him into coming back monthly and chatting with me about what he's
1:53
been seeing. It was an absolute pleasure
1:56
talking to him and I hope you enjoy.
1:57
Let's get into it. What were you just telling me?
2:05
And I said, hold on, don't tell me that yet.
2:08
Let's hit record. You're doing job interviews and
2:10
you created something. What is it?
2:13
Yeah, yeah, that's, that's something that's almost feels as
2:15
uncomfortable as being announced at the fast hours.
2:18
But no. So I'm doing this job interviews
2:20
at the moment, right? And of course I'm, I'm recording
2:24
these conversations. I also have some other tricks in
2:26
the book, but I will review later.
2:28
But I'm recording these conversations and I build my AI
2:31
analyzer to to analyse the transcripts, sometimes the video
2:35
as well. And it's just it's very
2:37
confronting to see yourself or like to hear yourself drift off
2:43
on the spots to nowhere and why you're thinking like, man, she
2:47
was just asking a question, answer it.
2:49
But it does help because being confronted with your own
2:51
mistakes is is shame is a is a good way to learn.
2:56
And what are you doing? The analyzer I imagine is just
3:00
grabbing the audio, converting it into text and then throwing
3:06
that into chat. GBT.
3:08
There's so many different tools out there that can actually
3:11
record your conversation, even a Google Drive, Gemini,
3:14
Supernormal, many of them, and basically all you need is the
3:18
audio or sometimes they're already transcripted for you,
3:22
but. After a lot of them do it,
3:23
that's true. Yeah.
3:24
But to have something that really that's really useful,
3:30
your goals may differ from call to call or for an AI to have as
3:35
much context as possible is of course, gives a better result.
3:39
So you can do it in Jet GPT, or you can let a tool just do it
3:42
for you. But you can also use, for
3:44
example, instructor, which is a really great open source
3:47
library. Yeah, you might know it right
3:49
from Jason Lu, which can help you extract the data that you
3:53
really want. And you can use rich
3:55
descriptions, pedantic schemas, etcetera, to to to do it much
3:59
like structured extraction from Open AI, but it has a bit more
4:02
to it. Yeah, that's that's how I do it.
4:04
And you're doing that, throwing it into some kind of a model and
4:08
saying where could I be better? Is that the gist of the prompt?
4:12
Analyse this. And but it's fun to just to play
4:14
around with. And I think what's really useful
4:16
is for an AI to, to really understand the personas that,
4:21
that are in the conversation, which may not always work, of
4:25
course, with the example that I just gave you, like a job
4:27
interview. But for example, when you have a
4:30
meeting with your team, then then when I have a meeting with
4:33
with a team, I, I also record the conversation just to put a
4:37
device on the table. And then I always have everyone
4:41
introduce themselves like to the AI like I am this and the name
4:44
and this is my background, etcetera.
4:46
So then when we are analysing conversation, then we do this
4:50
every, everything that is being said is analysed taking into
4:55
account who said it. And in general, I think that's
4:58
one of the interesting things with AI and where we're going is
5:02
personalization really means also who is talking to the AI.
5:07
Imagine that you have this really cool sophisticated rack
5:10
system for a company and all the information of the company is in
5:14
there. Maybe you have a layered access,
5:16
the CEO can access everything, all the containers and or
5:19
container sized access. But then there is maybe a legal
5:22
worker from the company and that's accessing this
5:25
information. But the AI or like some the
5:29
intelligent intermediate layer between the rack system and the
5:33
user should understand who's asking the questions.
5:38
Because the next step will be that the AI can also suggest
5:42
questions to ask. Because how you, you don't know
5:45
what you don't know, right? Yeah.
5:47
Are you able to? It's it's really hard to ask
5:49
good questions if you don't know all the information available in
5:53
the database. And so it, it's so funny that
5:56
you mention that because this is something that we've almost been
5:59
doing with recommender systems for a while, but in a different
6:04
way. It's the interactions on a web
6:07
page build up a certain profile and they build up features about
6:12
someone. But you're talking about, can we
6:16
know the features about someone? Like this person works in the
6:19
legal department, They're actually an intern.
6:21
They've only been around for six months, so they should be doing
6:26
these things, or they've interacted with these files over
6:30
the last three days. So those are the highest
6:33
priority. Probably they're working on this
6:35
project XYZ. But then what other stuff do you
6:39
want the AI to know about you to give it more context?
6:42
Because like we mentioned before, we hit record, context
6:45
is key. One step back like the blank
6:49
canvas problem, right? Like just sit down with a blank
6:53
canvas, what whatever it is that you're going to do, like a
6:56
drawing, write a story and learn.
6:57
But it's really hard to start. And and when you help people use
7:01
Jet GPT is what one, one of the first things that that you help
7:05
them with is what to ask, right? Because it's really hard to be
7:09
super imaginative and and a lot of cognitive load.
7:13
Yeah, yeah. So where should the AI step in
7:18
and based on whatever knowledge it has of this user, for
7:22
example, of course, with Jetty PT now they recently, I think
7:26
extended their memory function that, that that you find out
7:29
like it, it used to say memory is full and all of a sudden
7:33
it's, it's gone. So I assume that they updated
7:35
this. I didn't verify, but I'm pretty
7:37
sure. So that AI, so to say, already
7:41
knows a lot about you, right? And is therefore better able to
7:45
serve you respond in a certain way.
7:47
But I assume that we going to see more and more customization
7:53
that the AI is also going to suggest what you should ask,
7:58
right? The 111 way of looking at all of
8:02
this. And I think like what when I'm
8:05
going to go all the way back to when I was introduced to GPT 3.5
8:11
before GPT and I was I was having a walk in the woods with
8:14
my girlfriend and I just I couldn't stop talking about it.
8:16
I was like, this is so cool. And that day I made like a, a
8:19
drawing that like an image that I still use when I give talks.
8:23
The way that we manipulate the digital world is through binary
8:26
code, which is essentially a language that we don't speak.
8:31
All right, We don't understand. So we have to work with that.
8:34
We have an abstraction around that.
8:36
And that abstraction is programming languages.
8:39
And there's no, there's different programming language
8:41
because each abstraction comes with the upside that is that you
8:45
can actually manipulate the underlying complexity.
8:48
But the downside, there's always trade-offs, right?
8:50
With compression comes information loss, unless you're
8:53
a zip file. But I think it's a pretty
8:56
universal thing. I'm probably going to get
8:58
destroyed by anyone with more war with.
9:01
But yeah. But in this case, we can say
9:03
that that that programming languages are an abstraction
9:05
around binary code. And now all of a sudden we had
9:07
language like natural language as it is new abstraction around
9:12
these existing abstractions as a way to manipulate the digital
9:16
world and therefore have an effect in the real world.
9:20
And I thought that was fascinating.
9:22
And still to be in this space in the last 2-3 years is there's
9:28
there's still a lot of programming involved and a lot
9:30
of abstract thinking, which is great, but we're increasingly
9:33
going to see more abstractions. Sam Altman recently said that he
9:39
hated the user experience of GTPT with all these different
9:42
models. But, and it's something we can
9:44
all agree on, but you can also see this in Reg.
9:47
There's all these different options and there, there, there
9:50
might still be a lot of choices to make for a user or even the
9:53
engineer. And more and more we are going
9:56
to see those hard decisions going to be abstracted away with
10:02
the upside is that we can then again even work with more power
10:07
and complexity. And the downside is of course
10:09
there is always loss. There's information loss and
10:12
we'll see. It reminds me of someone who
10:16
gave a presentation back in the day on using Vertex AI as their
10:21
main platform. And they said the best part
10:24
about Vertex AI is that it's a managed service and it does a
10:28
lot of things for you. And the worst part about Vertex
10:31
AI is that it's a managed service and you can't get under
10:34
the hood. And so it's that double edged
10:37
sword in a way that we want that abstraction, but at what cost?
10:42
What are we willing to give up for that abstraction?
10:45
Human behavior here is funny as well, because it's really hard
10:49
to say that. It would be pretty preposterous
10:51
as you said it like a nice word for arrogant, but I don't mean
10:54
arrogant, but pompous. Yeah, you should to say that to
10:59
state that you're at the perfect level of abstraction because you
11:03
were born now 4040 years ago. So that programming language
11:07
that you are really good at that is the perfect level of
11:09
abstraction. That's bullshit.
11:11
You know how how would you know? And in in the end, it's all
11:14
languages. What are human language is
11:16
programming language is ways of communicating this.
11:21
And I'm going to go off the philosophical deep end here.
11:24
This raw thing that we call creativity or like the human
11:28
condition, like I have language to understand myself, right?
11:34
I have words and to yeah. And of course, feelings.
11:38
There, there. It's a weird mixer.
11:39
And when I want to transfer that to you, my condition, my ideas,
11:44
I I use language, right. So in in the end, we're
11:50
transferring creativity like the human condition into whatever it
11:54
is that we do, like communicating with other
11:57
entities. I heard the term last week and
12:00
it kicked with me, of course, at the way I use AI.
12:04
But it's called vibe coding. Yeah, yeah.
12:07
All that, too. I saw.
12:09
That's funny. Yeah.
12:10
It's the new buzzword. Yeah.
12:12
And my subtraction, yeah. 100% But going back to this idea of
12:19
OK, we have language to basically encapsulate feelings
12:25
and thoughts and things that we are trying to do and get in this
12:31
physical dimension there. The language is almost like a
12:35
signpost towards what symbolic the symbols are inside of us,
12:40
right? Yeah, it's really because also
12:46
when you define something, whether it is like symbolic or
12:50
or what language, you also make it real, right?
12:54
It is so because we call it. It is only there when we observe
12:58
it. As as we found from.
13:02
The cat. Exactly, Exactly.
13:05
Schrodinger. Yeah.
13:07
I, I have a very nerdy sweater that says Schrodinger's cat, and
13:11
then it says dead and alive, depending on how you look at it.
13:13
Dead or alive. Nice.
13:16
Going back to, I think one thing that I love from talking to you
13:21
a lot is how different you attack the problem that
13:27
sometimes we can get into when dealing with AI.
13:30
And a lot of it comes down to this context idea and how can we
13:36
get more context for the models? Because if we can give them more
13:40
context, then they are going to better be able to carry out the
13:47
tasks that we are asking of them.
13:49
And I know that just an example of that on this call is when
13:53
you're talking about how you ask people on your team to introduce
13:57
themselves to the AI so it can have more context on who they
14:00
are, what they're doing. Another one that we had talked
14:02
about a lot was when people join the ML OPS community, could we
14:07
figure out a way to have them go through an onboarding call with
14:11
some kind of AAI persona that could give a lot of downloads on
14:18
what the person wants to do, why they're joining the community,
14:21
what are the biggest things on their mind, like challenges
14:24
they're working through? And that way we can better
14:28
create different paths and suggest different activities
14:32
that we do in the community because we have so many
14:34
different activities, right? So how can we make sure that
14:37
each person is finding out about the activities that will best
14:43
suit them? Essentially, it is a paradox of
14:47
choice problem, right? In a world where information and
14:50
options are abundant, we get paralysed by the amount of
14:56
options that we have as a personal experience kind of
14:59
thing. But no, logically it's just for
15:03
example, within a community, right In that envelopes, they
15:07
are builders and and meet ups that I organize.
15:10
There is 100 people in A room and they are there with some
15:14
objectives and they may not even be fully aware of their
15:17
objectives. So that's not a discussion, but
15:19
let's say a rational humans. So we have objectives and there
15:23
is it's a given. There's limited time.
15:25
You can only speak to so many people.
15:27
You only you can only listen for so long attention span.
15:31
I said there is constraints, right?
15:33
So how do we optimize or make this that always sounds so
15:38
economic, but like, how do we make this the best experience
15:42
for any given individual if we would want to use AI for this?
15:45
And I think like what a lot of AI does, whether it is alpha go
15:52
within the constraints of a game or it's a jet GPT in, in in a
15:57
conversation, what AI does? It's it's making representations
16:01
of the world of a world. And so that that got me thinking
16:06
that when we indeed, when we on board people and we have a goal,
16:10
like, all right, let's say the goal is that we want to connect
16:13
the people that have a lot in common or have some interests
16:17
that are overlapping or whatever, then we would need to
16:20
have that information. So we would on board these
16:23
people into the event. You could actually do that with
16:25
a human to human interview, but you could also imagine a voice
16:29
AI interview or anything that is basically not a form.
16:34
Yeah. Because the forms it's hit or
16:38
miss. Some people will fill it out,
16:39
others not. Yeah, it's Dufour Harare in his
16:42
latest book Nexus, great book. He, he explained, took that
16:49
example of explaining bureaucracy and like a form is
16:54
forcing you as a complex human being to to fit this slots,
16:59
however complex you may be, or multi dimensional and you have
17:02
to fit the slots, you have to take a box or that sort of
17:05
thing. And you can see a lot of AI.
17:08
It is a much more flexible form when in in production use cases,
17:13
right, Whether it be a sales, sales application or maybe
17:17
onboarding as we're talking about here.
17:18
It's like a dynamic form, but it's a better experience anyway.
17:22
We would end up with a database or with all these individuals
17:27
and there's all this information about them and not in a super
17:31
structured format. So we might have to then use
17:33
some, some AI, but the structured extraction and maybe
17:35
knowledge graph or that sort of thing or like information
17:39
storage. We could even augment that
17:41
information. There's I saw this, these guys
17:44
lately doing a presentation. They were doing a presentation
17:47
at one of the meet ups. It was called Airweave and then
17:49
they were explained to some pipeline like that that he could
17:52
onboard a user and then augment it by connecting it to the
17:55
Perplexity API, doing research, scraping Linkedins, etcetera.
18:00
Anyway, your end. Here would be.
18:02
A lot of information on each individual.
18:05
Then I thought I actually bought a made a POC on on this.
18:07
Very, very interesting to do is to automatically create a
18:11
virtual entity of each person, like an agent.
18:15
It's basically an agent with a complex set of instructions and
18:20
the information taken from the information that we've gathered.
18:24
And then we ask it like we make it to embody as a matter of
18:28
speech. Strong word, yeah.
18:31
This, this person and put the goal of then joining this
18:35
virtual arena, the, the meet up floor, if you will, and to chase
18:42
their goals and to connect with other agents there.
18:47
And now there's some really interesting research.
18:51
So my, my background is in, in sociology and, and social
18:55
sciences. And there's a lot of like
18:56
experimental research in there where you would put two people
19:00
in a room, give $100 to one of them.
19:04
And with the goal, with the task to share this $100 and then the
19:08
person that received $100 that can make an offer to the other
19:12
person, right? Can make can be anything.
19:14
And the other person can say yes.
19:16
And then they share that that cut or it says no and then they
19:21
both get nothing. Right now rational theory would
19:25
say, if let's say I've got $100, give me, let let me give you
19:28
$1.00. I'd rather have nothing then you
19:32
walk away with $99. Yeah, because the other person
19:37
knows that the that it's $100. Yeah, so there.
19:41
Yeah, so there's this sense of fairness like in negotiation and
19:45
then that sort of thing. And it turns out that a is are
19:47
actually really good in in in mimicking this.
19:50
I found a report that did this and I took it and I made it a
19:53
bit more complex to give him like a sort of extra reflective
19:56
brain thinking out their strategy.
19:59
I was experimenting with all this sort of like psychological
20:03
research, things that were actually done in humans to find
20:06
out. How to set up this arena in
20:09
order to get good outcomes? Because if you just let the A is
20:13
go there, they're going to find common ground with every other
20:17
AI. That's just how they're, they're
20:18
really nice. You know what I mean?
20:19
Oh, that's before you know it, you're, they're, they're
20:21
building 1000 startups, which is not really realistic, right?
20:25
And they're, they're also, it might be, let's say there's 100
20:28
attendees and there's a 99 super smart machine learning engineers
20:32
envelopes people and there is one investor and then there's
20:37
one, one person that's really desirable in that room because
20:40
the other ones might not have to add too much to each other.
20:43
So this is all given on context. So I thought like, why not use
20:47
the constraint of a game and apply some rules?
20:51
And I'm not going to go into detail, but you can imagine that
20:53
if there's rules and that's games are really good for this
20:55
type of research is they have to be they're constrained.
20:59
And then with the end goal of and this is maybe confusing but
21:05
taken from game theory to find Nash equilibria like that is the
21:09
position where 2 actors are placed in in in such a way that
21:14
if one of them moves, then everyone loses like that.
21:17
It's the optimal interdependent rewards that you're trying to
21:21
find between actors. And yeah, with giving that,
21:24
you're basically just going to give every attendee and advice
21:28
like, all right, go to this and that person and hear some
21:30
opening, opening words to start a conversation.
21:33
We think you might have something to talk about.
21:35
And because the whole idea I remember when you were walking
21:38
me through this is that for a meet up, wouldn't it be cool if
21:43
we had this information on the folks who signed up for the meet
21:46
up? And then we could simulate out a
21:49
few turns on this game and we could see, oh, our game is
21:55
suggesting you talk with these three people because you all
21:58
have these things in common. You might want to have a deeper
22:02
conversation about XYZ. And so it cuts through like you
22:06
were saying, it would be great if we could all talk to everyone
22:09
at a meet up, but sometimes you talk to someone and you don't
22:14
want to be in that conversation. Other times you just, sometimes
22:17
I've been in this situation where I don't want to talk to
22:19
anybody because I feel really shy and I feel like I don't have
22:22
anything to talk to anyone about.
22:24
And I imagine I'm not the only one who's felt like that in a
22:27
meet up. So that can be like two ways.
22:30
When I attend like a meet up that I've never been like, I
22:33
find it hard to maybe even get to speak to people.
22:37
And when I organized the meet up and I've been just like hosting,
22:40
it's like I find it hard to there's something that that
22:42
queuing up it's it's. Funny because we try and do this
22:46
in some of our meet ups by just saying, hey, we're going to have
22:49
a game at the beginning of the meet up and we put people into
22:53
teams by their birthday month. It's and so you don't really
22:58
have a choice. You're just going with everybody
22:59
that's in the same birthday month as you, and then you have
23:02
this opening. There's like this bond that
23:06
forms because you're on the same team.
23:07
Whether you win or lose the game, you have your team you
23:12
raise the rest of. The point there, because this
23:15
whole idea builds upon the assumption that there is such
23:21
thing as the perfect outcome. Now I I used to own a restaurant
23:26
and and when in back in Corona, we started building for the
23:31
restaurants AQR order application.
23:34
That was with the goal of having people spend as little time as
23:38
possible on their mobile phones. Because to me as a restaurant
23:41
owner, it was very clear what the benefits were to us.
23:46
But it was also as a visitor, I hated it to how to make this
23:51
experience better for people with what if we predict what
23:54
people want basically of, you know, they can order from the
23:57
menu, like they have access to the menu to their mobile phone,
24:00
but they can also, yeah, the, the the menu is ordered so that
24:03
that it's in the right conversation for that Long story
24:09
short, in the very beginning, I thought there was such thing as
24:12
the perfect match between a given menu item in given context
24:16
with with a user. And the more I learned about it,
24:19
the more I realized that. And I could have known this
24:22
because I've been on the sommelier and, and the cocktail
24:24
bartender side a lot that it's really in, it's really hang out
24:28
there until, until the match is made.
24:33
So to say, given that the entity that gives you the
24:36
recommendation, whether it be a nicely tied up sommelier or a
24:41
recommender system, digital recommender system, AI, the the
24:45
trust that you have in the system is affecting your
24:49
experience. And also known as selling ice to
24:54
Eskimos, right? You're a good someone.
24:56
Yeah, you can sell a cheap wine for a lot of money as long as
24:59
you gave the person that buys the wine social recognition.
25:03
Yeah, and actually, for some reason I'm thinking about
25:07
Schwinger's cat again, because it's you're not getting that
25:11
match until the moment. And so there isn't that perfect
25:15
match that you can show someone. It's really in the moment that I
25:21
decide that's when it's actually real.
25:24
Yeah, yeah. And then there's, there's so
25:26
many factors at play there and there's so much like the medium
25:32
is the message here. And also another presentation
25:36
that I saw lately was much more about the interface, UX, UI
25:39
building AI applications. And that's also fascinating
25:43
because AI Gen. AI is an interface and it needs
25:49
an interface because voice is an interface.
25:52
Like it's a way of communicating and chatting is it's an it's an
25:56
ability of this of like the communication part is what it
26:00
can do, but it's also very much affecting your experience,
26:05
whether you communicate through through chat or to voice or to
26:08
predefined buttons or buttons that are generated on the spot
26:13
for you but still have to be clicked.
26:15
And there's so much there and spoken to so many people that
26:18
are building products in the last few years.
26:20
And I think this is still one of the very much unsolved problems
26:25
is you can have it as amazing application, but how do you get
26:28
people to use it? Like one one of the first rough
26:31
lessons I've learned is that people don't give a shit about
26:34
chat bots. And ideally we do not want to
26:38
use them I'm. One of those people so to come
26:42
to to take it back to the the abstraction of the beginning of
26:45
the conversation. What we really want is something
26:49
that knows that we want. Now, Oodopus, exactly where are
26:57
you when we need you? And it's so funny because on
27:02
that, it goes back to what we were saying with the being able
27:06
to suggest something in the right time, in the right moment.
27:09
As opposed to us having to have that cognitive load and figure
27:14
out what it is that we are trying to do in this moment and
27:18
then type it out and type it in a way that AI is going to
27:23
understand. So maybe we don't get what we
27:26
want on that first time around and we think, is it because my
27:30
prompt isn't good enough? Do I have to rephrase this?
27:33
And then it's more cognitive load and it's just, it creates
27:36
this poor user experience. And so I've been all about that,
27:41
man. Like the the UI piece for this
27:46
feels like the most important part of everything, the way that
27:50
some folks are doing it where you can say, let's just click
27:53
around and you have. We're so used to that.
27:55
It has been around for a long time, yeah.
27:58
Yeah. And when you get like those, you
27:59
probably know what it's called. But those pictures where you
28:02
have two sidewalks and then there's the trail in between the
28:07
two sidewalks because it's the shortest path.
28:10
There's a whole paradox around that too.
28:13
But the that's what it feels like where we are still trying
28:17
to figure out what is that short path.
28:19
Right now we just have the sidewalks and we're trying to
28:22
figure out where's that path that we can just make a shortcut
28:26
and then. Let's do do do do you refer to
28:28
what we call in Dutch elephant elephant pets that when there's
28:32
a route that people will always find exactly that.
28:34
Poor design in the physical world, we see that.
28:37
Wow, that's not how we as humans want to use this.
28:42
It's a fork or whatever. It's Yeah, this form.
28:45
I was taking on what you just say, like to express what you
28:49
want. And still we're in in, in this
28:52
like in general, you know, as using AI is do use AI, don't
28:56
refrain from using it. But if you use it, don't expect
29:00
it's going to be like this miracle thing, like you have to
29:04
put in some hard work. And that hard work is often to
29:09
understand what you what it is that you want.
29:12
And yeah, good prompting is just basically explaining what you
29:19
want and and all the facets that are part of that.
29:22
But it's hard. And in the beginning we saw all
29:25
these this easy Reg systems where there's a user query
29:29
that's get embedded and then finds the nearest vector in a
29:35
database and then augments the response with this vector like
29:39
with this information. But of course, we know that the
29:42
query is not always semantically similar to the information that
29:47
is needed to answer that query. So then what do we expect?
29:52
Do we expect the user to ask good queries because forget it,
29:57
or yeah, or are we going to? No.
29:59
And that's, I think in 2024, we saw this huge explosion of
30:03
agentic RAG frameworks, like multi hop agentic RAG
30:07
frameworks, right? Where one very important part of
30:11
that is query decomposition, right?
30:13
Like to query understanding intent recognition, which says
30:17
old as as an op and everything, right?
30:21
And, and I've been building a lot of those kind of things as
30:24
well. And what occurs because there's
30:25
this query decomposition and then there's query routing.
30:28
Where are we going to send these queries?
30:30
What databases? Then there's also of course the
30:33
type of embeddings, because not everything should be or can be
30:38
done with the off the shelf embedding models.
30:40
But the complexity of a rec framework, I think is centred
30:44
there. Because this is true, right?
30:47
Like given that a decent LLM has access to the information
30:52
required to answer your question, right, you ask the
30:55
LLM, what did I have for breakfast yesterday?
30:59
Obviously that's not part of the training data, but if the
31:02
retrieved chunk gives gives the answer to that, then it's going
31:07
to answer correctly. And then this you can generalize
31:09
some more complex problems. So the problem of RAG is
31:12
retrieval, and retrieval is basically precision, a recall
31:18
like precision. Do we we retrieve the chunks
31:21
that we want, but are we also retrieving precision, a recall?
31:26
Precision being the percentage of the chunks that we retrieve
31:31
relevant to the question, Is there not too much noise and
31:34
recall? Did we retrieve all the chunks
31:37
required to answer a question? Which is actually the harder
31:39
problem because you don't really know what you don't know, right?
31:42
There might be some hidden chunks there that that if not
31:46
present, make your model hallucinate.
31:49
Now with all these genetic frameworks, I mean, it's useful,
31:52
but it gets so complex and it's, it's still prompt engineering.
31:55
Like you're not updating the model, like you're making layer
32:00
upon layer. And it's really hard to actually
32:04
check if you're doing a good job.
32:06
Like you have to the concepts just mentioned, precision in
32:10
recall, you have to measure them.
32:12
Otherwise you don't know what it what it is that you're doing.
32:16
I think also often people are trying to optimize direct system
32:19
by looking at the query user query input and the generated
32:22
output and then if that's not good the parameters starting
32:25
like prompt engineering. The real problem is retrieval
32:28
because giving that if the model has access to the right
32:33
information is probably going to be able to answer your question.
32:35
Now I was working on job or consultancy job for helping an
32:41
organization make an AI and like a virtual service engineer for
32:46
like the make industry complex machinery, but I think like
32:50
aircraft motors, that sort of thing.
32:51
You can imagine that those asset manuals like those technical
32:55
manuals are insanely complex. Yeah.
32:58
So it's it's the goal is to help an engineer find the right
33:02
information there. And so I start playing around
33:05
and like doing some multi hub retrieve another man super
33:08
complex and then DeepSeek just drop like a fucking bombshell.
33:15
Yeah, And we all had our sort of paradigm to look at it.
33:18
It like people with money started running, I mean from
33:23
NVIDIA stocks and stuff. But in the technical community,
33:26
I think like it's all been like everyone I speak to has been so
33:30
stoked about like this algorithm, the GRPO algorithm of
33:34
deep seeks. So a group relative policy
33:37
optimization, right? Really a real step up from the
33:40
PPO algorithm that was originally I think invented by
33:44
open AI in 2017 to train models with reinforcement learning used
33:50
to be done with human feedback, or at least a, a model, a reward
33:55
model. And now with this new algorithm,
33:58
we can do this without all of that, like without human
34:02
feedback, without having an external model to, to check our
34:06
results. And there was such a cool
34:10
insight that this was possible, right?
34:12
But this works it's reinforcement learning, right?
34:15
So it works with domains that have a definite answer like
34:20
mathematics coding. Does it execute the?
34:23
Is the answer correct? And then I thought, what if we
34:29
look at RAG as a closed domain, closed source problem?
34:34
Because the chunks like you have a question and there's an
34:41
answer. And to give that answer, let's
34:43
say that there is this database and the answer can only be
34:47
provided when Junk 1215 and 20 are being added to the context
34:54
of the response model. Then we have a close problem
34:59
because it's not about the answer, It's not about the
35:01
information on the chunks, It's just the chunk IDs.
35:04
And imagine that we would have a set of, I assume, synthetic
35:09
data. And there's a bunch of cool
35:10
pipelines out there that can help you create this sort of
35:13
synthetic data where you feed in this complex surface manual with
35:19
1000 pages. And you have it with
35:22
understanding. There's like images, tables,
35:25
headers, text that you make different chunks of all of them.
35:28
You add metadata, you map relation.
35:30
So you have a model do this, right?
35:32
And then you ask the same model with access to all of that
35:36
context, which is different from a rag system where the context
35:40
is hidden, right? It's masked, it's in the
35:42
database, it needs to be retrieved.
35:44
But in this case, the context is there.
35:46
It's like a puzzle that is already made, but it's still
35:50
pieces and the model has conceded.
35:53
And then you have a generate questions that can only be
35:58
answered giving a certain junk or a set of chunks.
36:02
Or maybe you like a set of chunks in a certain order.
36:04
And then you are creating this sort of truth that you can't.
36:08
You can use this to test your rag by blood because you have
36:13
this definite source of of truth and you can measure precision
36:17
Enrico, but you can also use it to train it.
36:21
And this is the idea like we we started talking about it.
36:23
We put some really cool people from different backgrounds.
36:26
There's there's a guy doing is building like models for Cancer
36:29
Research PhD. There's there's a professor in
36:32
computer science, all these different people from the
36:34
community, which is the cool thing of communities that we
36:38
have all these different expertises.
36:40
Yes, diverse. Yeah, yeah, yeah.
36:42
And we started thinking like, wow, man, like this is almost
36:44
imagine that you can, if you want, we can go into the
36:47
technical details of how this would work, but it basically
36:50
imagine that you can. This tested knowledge that a
36:54
domain expert has is often not directly related to what they
37:00
know, but their skill set of understanding the domain and
37:06
knowing where to find it. Right.
37:09
If you ask someone who has been working in a research lab
37:15
chemist and that thing to to do a certain test and they need
37:21
them, they made different. They may not even need that.
37:24
They may not even know yet what exact compounds they need and
37:28
where they are stored. But that's the good odds that
37:33
that person is going to be able to do this a lot quicker than
37:36
you need, right? Because, yeah, it has all these
37:40
sets of knowledge or to yeah, to ask the right questions to open
37:44
the right drawers without knowing that the answer is in
37:47
that drawer. But you get me, right.
37:50
A baseline understanding. Yeah, yeah.
37:54
And, and in, in, in the, in the case, for example, of a virtual
38:02
service engineer. And then a real human engineer
38:05
might have a question and that service engineer needs to
38:10
decompose that question into sub queries and needs to make a
38:14
decision where to find the data, how to interpret the state of
38:20
the machine giving like sensor data and that sort of thing.
38:24
And use and how to use that context to to make to make
38:27
queries to find out information. Now let's say that we have this
38:31
source of truth given this this machine, we have questions that
38:36
have correct answers and we have to the chunk IDs for yeah to get
38:41
to these answers. And these chunks are just to
38:45
give it more information that it can reason over or the answer is
38:50
inside of these chunks. I didn't quite get that, or it
38:53
doesn't matter either. So a a very a very simple
38:56
example. Let's say that the engineer once
38:59
it's, it's missing a part, and that's part A, and there's Part
39:03
B and there's A Sorry, it's missing a connector.
39:05
It's part A, there's Part B and it's, yeah.
39:08
The question is what part connects part A&B then?
39:13
And this is a simplified example, of course, but let's
39:17
say there is three other parts on the left connecting to A, and
39:22
there is three parts, unique parts on the right connecting to
39:26
B, and we're missing the part in the middle.
39:28
Then the question could be what parts do connect to part A and
39:33
what parts do connect to Part B? And if it is that all in the
39:38
retrieved information, we have only one common denominator now,
39:42
then we have the answer, right? That's like a logical deduction.
39:46
You can imagine way more complex, but this is the beauty
39:48
of what I think that can be. The beauty of reinforcement
39:51
learning is given that synthetic data set that we would have,
39:54
then we would give these queries to.
39:57
So we'd give these queries to A solid model saying like GW EN30,
40:04
2BU, you need some solid hardware to to be able to, to
40:07
run this. But there, and there's been
40:09
experiments out there of people with way smaller models as well,
40:11
like Tiny 0 is a really cool one.
40:14
I highly recommend checking that out.
40:15
But let's say we have a bit more body model and then we give it
40:18
these queries and then we each query we let it generate 10,000
40:23
reasoning paths and we give it access to.
40:28
A function function calling, we give it access to a database, it
40:32
can retrieve data in a cyclic manner and whatever do what you
40:36
need to do, The only thing that counts is then the answer at the
40:40
end. And then we have this GRPO
40:44
reward model that then given certain arbitrary functions.
40:49
But in this case we can imagine that our answer final answer
40:53
needs to be correct. But more importantly, our
40:57
retrieved chunks need to be the right chunks.
41:00
Then what We have this new subset with correct answers,
41:03
correct chunks. And then maybe we want to
41:07
further refine this subset by having a function that, I don't
41:12
know, like optimizes for the shortest reasoning path or the
41:15
longest, or there's different things you can do there.
41:19
But you'd maybe end up with the 10 best answers for each square.
41:24
And you do that, then it updates the parameters of the model
41:28
giving this algorithm. And you do that many times.
41:32
Then I would not mean maybe I am like way of the chart here, but
41:38
I'd love to hear if someone thinks so.
41:40
But yeah, can't it be so that we are squeezing this, it's hidden
41:45
somewhere in this model already, the ability to use the correct
41:50
reasoning path to find the right information.
41:54
Not necessarily return the correct answer to.
41:58
Optimize the reason box. So let me play this back for you
42:01
and make sure that I am understanding it correctly.
42:04
It's basically saying that we need to give the right context
42:11
to the model. What we don't necessarily need
42:14
to do is give it only the right context in one shot.
42:21
It can be various iterations of giving it context.
42:25
And since the model can reason, we can now say figure out if
42:30
that information is there because we've already done a bit
42:34
of training or fine tuning on it to show it what the right
42:41
context looks like and how to know if you are retrieving the
42:45
correct. Chunks in reasoning models and
42:49
mind you, we start here with a non reasoning model and this is
42:53
referred to in the DeepSeek paper as the aha moment.
42:56
All right, that the the model starts to reason and self
43:00
reflect on that reasoning and then take another path because
43:04
I'm not sure. And so it's a very, if you see
43:06
these reasoning models, like a one, but it doesn't review all
43:09
these reasoning tokens, but DeepSeek does.
43:11
If you see these reasoning models reason, it's a very
43:14
anthropomorphic, like human like way of reasoning.
43:17
And it shows a lot of doubt perhaps and maybe and, or and
43:22
and this makes it, well, first of all, very much understandable
43:25
for us. But it's also an important thing
43:29
because when there's too much doubt and our problem that we
43:33
want to solve is relying is like multi hop.
43:37
For example, we want to know how many atoms a compound has given
43:43
the combination of atom one first, sorry, compound one and
43:46
compound 2 and then do this and that to it.
43:49
So then we have to know, we have to be sure that we understand
43:55
the atom count of compound one and compound 2, right?
43:59
And so if we make a mistake there, a false assumption, then
44:02
our final answer is never going to be right.
44:06
So if you have a a reasoning model reason over this, you'll
44:09
see this a lot of maybe perhaps etcetera.
44:12
Now there's this great paper, it's called O1 Search.
44:15
It was, I think published just before the DeepSeek thing, or at
44:19
least it's not referring to that, but it's a way and it open
44:24
source the whole code as well to have reasoning models recognize
44:30
their own ambivalence, like if they're not sure enough and then
44:34
stop the reasoning process and basically perform a search For
44:41
more information like connect to a rack system does some other
44:44
things as well. And it's not function calling.
44:46
It's actually making a, it's formulating a question
44:49
encapsulated between string tags and edit with a special token
44:52
that gives the scripts the the call to stop this process, find
44:57
this new information. And it's actually sending it to
44:59
a parallel chain that has access to all the steps so far.
45:03
It's cleaning those junks for better position and then adding
45:06
it back to the main loop and then continuing and it can do
45:09
that and many times. So this is using a reasoning
45:13
model to be more sure and as you could imagine.
45:18
That you could. Make this ability like to
45:23
improve this ability a lot. And I think you could imagine
45:27
using that, that that from the mentioned reinforcement learning
45:30
way there as well. I'm trying to just make sure
45:35
that I understand this. Basically, the reasoning model
45:39
recognizes when it doesn't know if it is right or wrong or it
45:46
doesn't. It recognizes when it does not
45:49
have the right information or enough information to be able to
45:53
confidently make a decision. Yeah, and it can then go and
45:58
find that information. And you're saying instead of
46:01
like with the O1 search, the potential that we could have in
46:07
a RAG system is that instead of it going out and trying to find
46:11
that information on the web, we can just go and try and find
46:15
that information in our embeddings?
46:17
Yeah, in, in, in any, in any external source.
46:20
It could be the human, it could be web it it could be So it asks
46:25
the other question. It's the important part is in
46:27
any rag system like I've been going to like in very much into
46:31
the sort of hypothetical solutions.
46:33
But if you take a step back, is a rag system needs to know if it
46:40
needs to find more information is is it sure?
46:43
Is that the information that it has access to in the training
46:45
that is that enough? Do I need more information then
46:48
then I get more information. Is this enough?
46:50
Is this what I need to answer the question?
46:53
This is something that is you can't possibly do a rule based
46:56
in for real world scenarios, right?
46:59
Then part of that but more technical is that the system
47:03
needs to be able to understand the query of the user to
47:08
decompose that query in the optimal way and making that the
47:11
search queries optimal whether it be for web, human in the loop
47:15
or vector. But that's more a technical
47:18
thing. Like it boils down to how good
47:22
does the system understand its own limitations and how can it
47:27
be sure of things? And we cannot prompt, but you
47:31
can improve it. But like, how could you ever do
47:33
this in a sort of like rule based manner, right.
47:36
But what you could do is use reinforcement learning, given
47:40
that you have a set of correct answers.
47:43
Yeah, correct chunks and then just squeeze it out like blood
47:47
out of stone. I think the fascinating piece
47:50
here and one of the hardest pieces about this is that you're
47:54
trying to coerce the model into this unconfidence.
48:02
And as we all know, anyone who's worked with LLMS is their
48:07
confidence level is off the scale and so.
48:09
They they're definitely different with reasoning models,
48:12
right? I think what we see with scaling
48:14
inference time compute is this sort of not being so confident.
48:20
Yeah, to not think that it knows everything, right.
48:22
But still, I think that's the big unlock here is being able to
48:29
say, because of the way that this question was phrased, I can
48:37
only get this far. So maybe we need to go back and
48:39
and retry one of these steps so that like low confidence score
48:45
is huge. Yeah.
48:48
And I think because it's so hard to explain this in rules that if
48:54
this sort of approach would work, that it would probably be
48:59
limited to certain domains, just like in human domains where,
49:03
yeah, it's, you know, given this, given that good old
49:07
Bayesian, given this, given that, and that is information,
49:10
this would be the right question, right?
49:11
And this is what is like the best information looks like.
49:14
Now I can generalize that to maybe when I'm a bartender and I
49:18
know my way in the bar and I can maybe generalize that to the
49:21
kitchen, but I can't generalize that to the lab, right?
49:23
There's so much, so many smart people working on, on
49:27
reinforcement learning, on making better rack systems in
49:29
all sorts of ways at the moment. But we're going to see this like
49:33
we're not going to see AGI just yet in, in that sense, whatever
49:37
it means for another one. But this close source of smaller
49:43
domains where a combination of reasoning models, agentic
49:48
frameworks are working together to solve problems within a
49:52
closed source. Like Ethan Moloch said it, I
49:55
think he was that he said, give a good example.
49:58
We have now the open AI browser agent, which is like this.
50:01
Why it can do anything. Let's say that's the AGI kind of
50:05
thing. And we have a deep research,
50:07
right? And both are combinations of
50:11
agents and reasoning. And with browser agent, it's
50:16
nice, but it still really sucks. Like it's not useful.
50:19
Deep research is amazing. Not the Holy Grail grail yet,
50:23
but I'm using it like a lot. I'm not using browser agents.
50:28
That's what we're going to see whether it is in, in, in domain
50:31
specific rack intelligence or tasks like, yeah, it's, it's
50:36
fascinating. And I bet we're going to see a
50:38
lot of progress in the tactical. So take me back to what you were
50:41
doing with the manuals for all of these.
50:46
I think it was the airplane. Engineering.
50:50
Diagrams and all that stuff, machine diagrams that are very
50:53
complex. How does this play into that?
50:56
I know you gave the very simple example of there's a connector
51:00
and there's those two pieces, and we got to figure out which
51:03
one can connect the two. And you're assuming that there's
51:07
only one answer there. Yeah.
51:09
But there might be, there might be more.
51:10
Answers, of course. Yeah, what are the other
51:12
examples of this? How you build the rag system to
51:15
make it work? What does that all look like?
51:18
I guess is what I'm trying to figure.
51:20
Out I was asked in the in in this project to to look at like
51:24
how we could test the knowledge of engineers and there's like a
51:30
contradiction in terminus because try ask someone the
51:33
whole definition of testedology is that you can't really explain
51:36
it it's not it's just you do what you do because you know how
51:38
to do it because it feels right and.
51:41
Yeah, the idea. However complex it may be, and I
51:44
think we should really start very simple to to verify if it's
51:47
even we're chasing this this pot of gold at the end of the
51:50
rainbow. But that if there is an answer,
51:55
and you could indeed use a human knowledge for this as well.
51:58
But if there is an answer to a complex problem and there's
52:02
external information that that is that's needed to come to this
52:05
answer, then we should be able to use reinforcement learning to
52:12
find the right path. Because there is many possible
52:15
paths, right? And but try it, give it enough
52:19
tries, and you might just end up on a very good path and do that
52:25
many times. And then you have examples or
52:30
you train a model to to behave that way.
52:33
And I think what it comes down to, what is so fascinating is
52:36
that people say that models cannot answer questions outside
52:40
of their training data. Well, that's a truism, of
52:44
course. Yeah, of course.
52:45
But what is in their training data?
52:48
Is that a set of discrete words or is there emergent
52:53
capabilities of all this information stored in this
52:57
network? And one of these emerging
52:59
capabilities seems to be reasoning that is that aha
53:02
moment right of the DeepSeek paper and with reinforcement
53:06
learnings, if reasoning is already present in this set of
53:11
set of information in this set of data, what else could be
53:14
there? And I don't know, I don't find
53:17
it a weird idea that there's much more present.
53:20
So yes, models cannot answer questions outside of their
53:23
training data, but that's only really relevant for real time
53:26
facts as what did we have and what did I have for breakfast
53:30
yesterday. When it comes to capabilities, I
53:33
think there's much more possible than we've seen so far.
53:36
I like that you say that if there is an answer and we can
53:45
correctly make that answer be known, there's a yes or no
53:51
answer or a way to do it and task complete, task not complete
53:54
type of binary way of looking at it.
53:58
Then we can figure out how to get AI to try and do that.
54:05
I I love talking to you about this stuff.
54:07
I don't I don't know if they actually use this, but they
54:09
shout out to a crack camerat and he he he loves this the snake
54:13
game. So you can compare models and
54:15
they're playing snake against each other and a great way to
54:18
test the capabilities of the models, problem solving skills,
54:20
etcetera. And then someone in the Commons
54:23
was on LinkedIn said, I wonder how long it's going to take
54:26
before each game is over. Instantly.
54:30
Oh, wow. And it's just a roll of dice
54:35
because they all panned it out. They played it out.
54:39
And you can like in this sort of constraint space of a game like
54:44
that is a relatively simple game like that.
54:46
You can imagine a reinforcement can get to that.
54:48
Like I'd love, I'd love to hear people who know more about it to
54:53
to to say that this is possible or not.