Latest / Agentic Conversations (formally mlops.community) / AI, Marketing, and Human Decision Making // Fausto Albers // #313

Transcript

AI, Marketing, and Human Decision Making // Fausto Albers // #313 · Agentic Conversations (formally mlops.community)

0:00
So my name is Fausto, and I'm an absolute enthusiast in live
0:07
innovation and technology. I'd like to connect seemingly
0:11
despairingly ideas and most of the times I'm dead wrong.
0:14
But sometimes I'm not. And, and at least we can have an
0:19
interesting conversation about it.
0:20
And that is the very human quality that you'll see in this
0:24
podcast that really counts. We are back for another ML OPS
0:28
Community podcast. I'm your host, Dimitrios, and
0:31
today we're doing a session that I like to call Fridays with
0:35
Fausto. He has become one of my good
0:39
friends and every time I talk to him, I thoroughly enjoy it.
0:43
I love talking to him because he is very transparent about how
0:46
he's building, how he is trying to use all the new stuff that's
0:50
coming out in the world. We get into some of the topics
0:54
like the new image generation models, and of course,
0:57
everybody's favorite MCP. Let's jump into the conversation
1:03
now and as always, give us a review.
1:06
Leave some stars. Do what you can to help the
1:10
algorithm know. This is awesome.
1:18
Wild. The image Gen. stuff is
1:21
absolutely wild and we wanted to talk about this before even Open
1:27
AI dropped. And then I found Rev AI, or if
1:32
you put it all together, it kind of looks like it says reveal,
1:35
which I have no idea how that's working.
1:38
So first thing on my mind with the image use case is that I've
1:44
been seeing a lot of folks in the marketing department take
1:48
Facebook ads and then they say to their prompt is basically use
1:54
this photo, but put my product, which is another photo in the ad
2:00
and AD like this text and it will do it so well.
2:05
And that is highly valuable in itself, but is quite manual if
2:10
you think about it. So I'm almost extrapolating it a
2:14
few steps forward and thinking, you know what, what about when
2:20
you set up a pipeline that will just continuously look for the
2:26
best performing Facebook ads or the best performing ads on
2:31
different platforms? Exactly, this was just I was
2:34
thinking about that man. And then you're taking those,
2:37
you're just swapping out your product and then you are
2:40
throwing them in and testing them on different platforms and
2:44
seeing how well they work for your specific product.
2:48
Yeah, it's like real time generation.
2:52
I mean in anything that we see in Gen.
2:54
AI, I think an interesting thing is like not just on itself, but
3:00
how you know to combine it with existing technologies and in
3:04
this case AAB testing a scale, right.
3:08
And before we kind of had to make all the content ourselves
3:13
and then test it. And there is still sort of
3:17
aiming at user groups and this makes it possible to aim it at
3:22
the individual. But one step back and maybe from
3:26
a bit of like meta perspective, like what Gen.
3:29
I, what generative AI lets us do is to create the world's that we
3:34
are envisioning. This was literally the first
3:37
thing that I thought when I was introduced to GPT in GPTS in
3:42
back in 2022, I was taking a walk with my girlfriend to the
3:46
forest and I couldn't stop talking about like this.
3:48
I dare that we're entering a world where anything that you
3:52
can envision can be created in the digital world and therefore
3:56
has an I have an impact on the real world and so that it's
4:02
still hard, you know, even even when you know what you're doing
4:05
because to translate your creativity, do anything useful.
4:11
I mean, you're concept is as strong as your ability to to to
4:19
explain it, right? To to explain to someone else or
4:22
to an AI. Now, before we're going to
4:27
touch, I'm sure a lot of interesting topics, but one of
4:30
those things in the AI ecosystem that can really help there is
4:35
like Super whisper and Wisp flow and all these kind of things
4:38
that are based essentially helping you yeah to to give
4:44
words to your concept, right. So that's one I mean you could
4:48
do that iteratively with any kind of tool as well.
4:53
But yeah, I think that that's that's very interesting because
4:56
sometimes you sort of have this vision almost this this not just
5:00
urge like, oh, this is like the the app or the thing that I want
5:03
to create. But you know, you don't exactly
5:06
know how to put words with. And it gets more and more
5:08
important to do that. And whether it is to prompt a
5:13
image generation model or in a more complex workflow where this
5:17
gets integrated right, because, you know, making images is nice,
5:21
that's one thing. But to be able to integrate it
5:24
into a workflow, say you're building an app with a sick
5:30
front end in cursor, then first of all, you should have have
5:35
your, you know, your idea in in in in intricate project
5:40
description, which is where AI can help you.
5:44
But then it's extremely important because a is are
5:47
absolutely infallible right to feed the right context
5:52
information, what to do, what not to do, what libraries to use
5:57
at the right time. So at inference and yeah, with
6:01
with coding ID ES like cursor, like windsurf and and many
6:05
others like they let you granularly adjust this and then
6:11
image generation becomes one part of that workflow.
6:14
I mean, I didn't check, but the last MCPI was using for image
6:19
Gen. was stability II. But most likely there's already
6:22
an MCP out there, right? For the new open AI 4 O and
6:26
native image generation. So, yeah.
6:31
So to make a jump here at how I was using it, I was building
6:33
this this website and I wanted to have a check box between two
6:37
elements. And instead of going to Excel
6:41
draw and draw it and then point at it and then feed it to back
6:44
to the AII actually had it, you know, I made a print screen of
6:48
the current state. I fed it to GPTI said like I
6:51
want a chat box in between make it, you know, design it nicely
6:54
and it came back and all the text, everything in that whole
6:57
wireframe was exactly correct with the chat box added.
7:01
Then I added back to cursor. Then I call my 21st half magic
7:05
front end MCP 2 draws like this, element chooses by itself,
7:10
injects the code into the workflow and voila.
7:14
Well, let me tell you how I'm thinking about this because
7:16
you're during it. And again, this is incredible
7:21
how you're seeing like lift on your workflow in a way.
7:26
But I, I think about it as a platform engineer or an ML
7:30
engineer that is trying to create business value for their
7:35
company and they're looking for use cases that can add
7:40
significant value for the company, right?
7:43
And so I can imagine if you're a good ML engineer or engineer,
7:53
even product, it's probably more of the product folks that are
7:56
going out there and they're talking with the company and
7:58
saying, hey, how can we plug in AI to your workflow?
8:02
What is what are the things that you do on a daily basis that are
8:06
time consuming? And somebody in the marketing
8:09
department says, well, I'm constantly doing research on ads
8:14
and then I'm asking our designer to create different variations
8:19
of ads. And then I take those and I add
8:21
them to Facebook or I add them, I add them to XYZ.
8:24
And with Facebook, it's incredible because the creative,
8:28
the actual image is your targeting in a way.
8:34
These days, you don't have to be.
8:35
It used to be where you would have to try and figure out the
8:39
exact persona down to since their birthday, their
8:44
neighborhood, that type of thing.
8:47
But now it's not as much that because the ads algorithm is so
8:52
good that it can look at the creative and know, all right,
8:58
let's show this to these people and then let's see what kind of
9:01
signals we get and then we'll show it to similar people.
9:04
And so the marketers say the creative is the targeting.
9:10
And because of that, like we were saying, you can just AB
9:13
test so much more and you can create so many more variations.
9:18
Whereas when you were bottlenecked before and saying I
9:21
can only create 30 pieces of creative per day and we're
9:27
paying quite a lot for that whole workflow in time.
9:33
Now it's like I can just have the human looking at all of
9:38
these different creatives that are being churned out
9:40
continuously and making sure that the text and the words are
9:44
are correct or there's nothing funky in it.
9:47
And then you can imagine the workflow is just clicking a box
9:50
and pushing it right to Facebook.
9:54
There is still the human in the loop right in this in this
9:58
paradigm. And I think what is what is
10:02
interesting to me is that with every new innovation there is as
10:09
now like the the the talk about how is this going to be
10:12
implemented? It's going to be human loop.
10:14
Is the quality going to be good enough and and how is it going
10:18
to affect industry XYZ. Now those are valid discussions
10:24
to have, but I think that it's also interesting to look how
10:29
it's going to not influence the current state of the world.
10:33
Like, for example, take marketing content.
10:37
This post of this dude went viral on.
10:39
I saw it on LinkedIn, but but he created sort of like Don Draper
10:44
style advertisements and he said like marketing ads.
10:48
It's it's so over, right? And then I thought, yeah, it may
10:52
be in the short run, but in a sense, marketing is about
10:56
influencing humans on an emotional level in order to, you
11:02
know, in manipulate decision making, right.
11:07
Another interesting development in AI and actually Cass Sunstein
11:12
from one of the authors from the influential book Nudge Nudge
11:17
Theory published an interesting paper last year called
11:22
Patronalistic AI and Choice agents.
11:24
It was picked up by nature Google it.
11:26
It's interesting and where he argued that AIS are AI agents
11:32
are efficient decision makers, right?
11:34
And we're going into a world where AIS are making decisions
11:38
on behalf of the their users. Now, I think we can see some
11:42
some early stage examples from this, this already happening.
11:47
But that essentially means that marketing traditionally tries to
11:51
influence the end user, the buyer.
11:54
But what if the decision maker, maybe not the end decision
11:59
maker, but in large part is an AI or a set of AI agents?
12:03
And then how is the traditional marketing still useful if, you
12:10
know, people make their decisions on what their AI agent
12:15
is basically suggesting, right? Because we know that we're
12:19
manipulative. We know that we're that we do
12:23
that we do buy things that we should not, right?
12:25
So I think when marketing becomes so much more effective,
12:29
you're also going to see a backlash, especially when in the
12:32
younger crowd of Gen. Z that are saying like, all
12:36
right, I'm, I'm going to sort of like offload my decision making
12:39
because I'm making too many decisions.
12:40
I don't want to be manipulated, especially not when you know
12:45
more and more capable AI and people understand that stuff as
12:49
well, right. So then I think that the whole
12:52
sort of like Don Draper style advertisement, the advertisement
12:57
may be dead, yes, but in a different way that they are
13:00
suspecting now. Not because the images are
13:03
getting generated, but because it's going to be a whole nother
13:07
way. Like they probably need to find
13:09
a way to influence the descriptions and and and all the
13:16
information that a buyer's AI would spider to to sort of
13:21
influence that if you if it doesn't make sense.
13:23
It's almost like the tactics aren't going out of style
13:27
anytime soon. They're still going to be
13:29
useful, but the way that we implement those tactics are
13:32
going to be different. Those are fascinating things to
13:35
touch on, but I also wanted to steer the conversation in the
13:40
direction of MCP. And just because I've heard some
13:47
folks in the community talk about how useful it is, but also
13:53
how it's like, huh, I don't know if I really get it.
13:56
And then most recently I have been reading a thread in the
14:01
community from MEB who started this and shout out to MEB who's
14:06
like talking about using anthropic API and just having
14:12
the worst time ever. So I know they're having some
14:16
uptime issues over there at Anthropic.
14:20
Godspeed to them. Hopefully they get it sorted
14:22
out. But it's like you want to try
14:25
and, and build a service and there's this cool thing that now
14:29
has a lot of attention, attention like MCP.
14:33
And next thing you know, you're trying to integrate it in and
14:37
there's these bottlenecks in your stack where you're
14:40
recognizing like I, I can't have this as a service for my company
14:46
if we have specific SL as, and that's not a specific MCP thing.
14:51
It's more just like a general system thing that where MCP
14:57
might not be the point of failure, but then you're
15:01
recognizing like, oh man, I've got this.
15:03
I'm I'm relying on Cloud 3.7 and it's just not working.
15:07
No, but that of course Entropic came up with the multi context
15:14
protocol as an open, an open thing to use, right.
15:18
I've even heard that Open AI is looking to yeah, is looking to
15:21
adopt it as well. So those are two separate
15:24
things, right? You're talking about the uptime
15:26
of the model. I think that is something, I
15:29
mean, as developers and people that use AI for coding
15:34
assistance, we've been relying on a sonnet since that 3.5
15:40
release upgrade I think in autumn 2024.
15:43
It's just by some margin the best.
15:47
SO 3.7 was much anticipated 3.7 thinking 3.7 Max now in cursor
15:52
as well. So I feel it might just be sort
15:55
of overload use and and and and honestly, I have been, I've been
16:03
actually since I upgraded to a cursor dot .481.
16:10
I think it was like the last update, still a little buggy,
16:13
but they introduced an enhanced and improved model routing
16:19
system, which is on itself something to be seen in a in a
16:24
bigger scheme as well. Because again here, like making
16:30
choices is a hard thing to do. Making informed choices is a
16:34
harder thing to do. And to be using the most
16:39
efficient, most capable model for whatever your use case is,
16:43
that's, that's a hard thing to do.
16:45
So we probably as a whole tend to use too much force.
16:50
Yeah. And I'm sure that that that
16:55
cursor will improve this capability.
16:58
And also then sort of walks around the problem if models
17:03
have downtime, because there's a lot of models that we can use.
17:07
I mean, DeepSeek-V3, the new one, Gemini 2.5 Pro was, I just
17:14
saw it pop up in my model listing all benchmarks, right.
17:18
But yeah, yeah. And, and the context window, I
17:20
mean, because, you know, in the beginning, you know, we were
17:24
just using it for scripts then then a few scripts in your
17:28
project. But especially when you're like
17:29
integrating back end and front end, there's can be hundreds of
17:33
different files and, and, and all sorts of dependencies that
17:38
you're that you're relying on. So the context window becomes
17:40
really important. And I'm curious to see how our
17:43
Gemini is going to change that, if if if anything.
17:48
Yeah, because now we know that you might see it does so well on
17:53
all these different benchmarks, but that doesn't mean much.
17:57
It is a cool thing to use in your marketing material when you
18:00
come out with this and it's great for people to share
18:03
online. But until you actually get your
18:07
use case in front of it and see how it works and see how or even
18:12
your workflow and see how it integrates with what you're
18:15
trying to ask it, you don't have the best idea of what it's
18:21
actually going to do. Now you know, you know that I
18:25
that I used to own a restaurant, right?
18:27
And, and it was inspired on New Orleans, LA because I thought
18:33
that that that history was amazing because creativity was
18:37
born out of scarcity, right? The kitchen and the music.
18:42
It's, it was all from very little and that and, and, and
18:48
and and and scarcity is the model of innovation, right?
18:51
And, and at the moment with AI, we're in an age of abundance and
18:58
we're getting, we're really waiting for another crazy good
19:04
model to solve our problems instead of being really creative
19:08
and seeing what we can do. For example, like in what
19:13
intricate way you could use NCPS and, and, and, and rules
19:18
incursion I'm referring to. But in, in general, it's more
19:21
about making all these different calls and making sure that the
19:24
right information is being sent at the right moment.
19:27
And MCPS are a way that you could orchestra this rules are
19:32
another way. But and, and like we're, it
19:38
feels like we're sort of spoiled and it takes some focus of our
19:44
ability to creatively come up with problems and really think
19:47
hard and deeply and longer about a certain problem with a certain
19:50
model and. Because there's always, so it's
19:53
like the grass is always greener.
19:55
So you're like, well, maybe if this model can't do it, then
19:57
we'll try it, we'll swap it out. Yeah, And, and, and I think, you
20:02
know, there's a lot of voices that say like, oh, like people
20:05
are getting lazy. It's going to kill creativity.
20:08
I don't think so. I mean, there's, there's,
20:09
there's different. People have always responded
20:12
differently to innovation and you know, the the, the naturally
20:17
curious. We'll explore and and and we'll
20:20
do, I think, but. Is the.
20:26
Ever expanding model capability, ever going to come to and halt?
20:30
Well, there is a way to look at it that you know, scaling laws
20:33
and as in pure pre training. Yeah, I mean, it's going to stop
20:40
once right there, there's only so much energy and and money in
20:46
the world. But maybe I'm actually curious
20:50
to see maybe that's a good thing because that will force us both
20:53
the, the innovation labs, PPC, open AI, etcetera, to do all
20:59
these different things like latent latent reasoning without
21:05
tokens, inference time, compute different algorithms like the
21:11
ones that the DeepSeek release and, and at the application
21:15
layer as well. Like how do we make sure that my
21:18
AI gets, you know, these calls the right information at the
21:21
right time, etcetera. So while there's no real reason
21:24
to think that the raw capability of a is is going to go down
21:28
quickly, it is interesting to see if, if, if there's maybe a
21:33
little rest, maybe maybe this is me just just just sort of like
21:38
wishing for for a little Peace of Mind to figure out what I
21:43
should do to get the most out of what is now, you know, Well, the
21:47
other side of that is things that I've heard folks talk a lot
21:50
about, which in a way could be likened to the term forward
21:56
compatibility. Where you have backwards
21:58
compatibility, but then you also want to be thinking about
22:02
forward compatibility. And a perfect example of this is
22:05
when we had Flores on the podcast, I think a month or two
22:09
ago, and he was talking about how he did so many janky things
22:14
in the beginning to try and extend the context window.
22:18
And then three months later, he basically had too much context
22:23
window in all the models, and so he no longer had to do that
22:27
stuff. And he also was thinking, would
22:30
my time have been better spent trying to optimize other things
22:34
and then recognizing that the context window was going to get
22:38
better? And so because you don't have
22:41
this crystal ball, you don't really know which part of this
22:45
whole system is going to get better in the near future.
22:50
So I can let time be the one that optimizes that for me
22:56
instead of myself spending the time and brute forcing it.
23:02
Yeah. Well, I mean to answer the
23:04
question, one should really know where what it is that you're
23:08
optimizing for, right? I think a lot at the moment
23:14
feels, you know, the energy that we spend on, on, on on learning
23:20
about new things and and and and and, yeah, learning to work with
23:25
this new technology. And for myself, it's, it's
23:29
genuine curiosity and, and, and for marvelling.
23:35
But I have to admit there's also this sort of race, right?
23:38
I, I have to keep up with, with, with the beer, the beers in, in
23:43
the builders community and, and everyone is, is doing it.
23:48
And, but you have really, we don't really know yet what it is
23:52
that we're optimizing for to become the best.
23:54
What I mean, it's only a few years ago we said prompt
23:57
engineer and now prompt engineering is still important.
24:02
But I think it's important in a way that you are able to, to use
24:06
the tools to write the prompts for you Again, same as with the
24:11
cursor rules. I mean, a set up, a good cursor
24:15
rule set up for a, for a, for a large project would in my
24:19
opinion, require hundreds of rules like different NDC files.
24:24
And now there's ways to have the AI in cursor write your cursor
24:29
rules and update things that are good and, and, and and make note
24:34
of things that were that didn't work out.
24:36
So then it, it is maybe less of a skill of writing those rules,
24:41
but it is, it is and, and increasingly a skill of knowing
24:46
how to work with these little tools to, to do what's, yeah,
24:52
optimal. Yeah, and knowing what good
24:54
looks like, so that if it does break the rules or if it does go
24:58
off the rails, then you you have that.
25:00
But dude, how are you? How have you been playing with
25:03
MCP and where have you found it intriguing or absolutely mind
25:11
blowing? If we want to.
25:13
Get really out there on it. Yeah, man, it's, I mean,
25:18
essentially it's nothing new because we could already do it
25:22
with function calling, right? And when agents were introduced
25:27
to cursor, literally the first thing that I did was to create a
25:31
tools dot Pi file and then define the tools that I wanted
25:35
there and then instruct the agent to use the tools dot Pi
25:39
file. If so, that that was kind of
25:41
doing the same, but not as good. So what makes MCPS special?
25:47
That it is yet another abstraction around complexity.
25:53
It is a unified manner to connect server and client and
25:58
that makes it well. And there's a huge open source
26:01
community evolving. So if you have a need, then you
26:05
can just, you know, look in one of these forums or or on GitHub
26:10
and you'll find an MCP quite surely.
26:13
And that has been quite impactful because so my my game
26:20
is more back end than front end. And when designing apps, there's
26:26
now MCPS, for example, the one of 21st Def magic.
26:31
I highly recommend you check that one out.
26:33
What? What is that?
26:34
At 21st def is a platform where there is a huge library of
26:39
components, templates, basically pieces of code, front end code
26:44
that you could use in your project.
26:46
And the MCP 2 is you can call it by back slash UI or it
26:51
recognizes your intent. But mostly with MCPS if you
26:56
really want them to use then call them you know deliberately
27:00
and and it will communicate to 21st dev.
27:05
And based on your description. For example, I want a hero
27:09
section with nice gradient colours and the text must stream
27:14
in. Then it will search that
27:16
platform, find the code snippets and then communicates those back
27:19
to cursor, injects them and then and then just implements them
27:24
and it works really good. Then there's another MCP tool
27:27
called Browser Tools. It works with a Chrome extension
27:32
and you have to run a little server on your side and and it
27:36
will give the cursor AI, the cursor agent access to your
27:40
browser console, to your logs, to everything.
27:43
It can make print screens. So if there's an element or
27:46
something that you want to change before then you were like
27:49
trying to look, yeah, this element or you make a print
27:51
screen and then make a mark on it and then feed it back to
27:54
cursor. And now it can sort of see what
27:58
you're making in the front end when you're running it in a
28:01
local host. And that is pretty, that's been
28:03
pretty amazing. Like for someone that I mean,
28:07
I'm, I wouldn't say I'm, I'm, I'm able to make a production
28:11
ready apps for front end, but it is another way of letting people
28:18
see what it is that I'm trying to envision what I'm trying to
28:21
communicate. And then, you know, you can go
28:22
further on that. And something that I highly
28:26
recommend any source scraper, NCP, there's a hyper browser
28:31
that I really like. And so when you're building,
28:34
you're starting your project and often you're working with new
28:37
libraries. I mean, for example, the new
28:40
Open AI Agents SDK that also already has its own MCP by the
28:45
way. But then there is documentation
28:48
and you can with this MCP, you can literally just in the cursor
28:52
check, you can tell it to go fetch the new documentation of
28:58
Open AI Agents SDK, save it to individual markdown files in a
29:03
dedicated folder. And then you can have adapt to
29:06
create your cursor rule set of on that.
29:09
And so it can constantly refer to the new endpoint to the new
29:15
whatever instructions that it has.
29:17
So scraping front end tools and yeah, cool one just for best
29:23
practices, the GitHub one can in like search GitHub make all your
29:29
pushes and it that's also been in sometimes forget it.
29:33
And then I really regret it in the morning, you know, yeah, So
29:39
those, those and of course sequential thinking, but I think
29:41
everyone that works with has worked with MCPS knows the
29:45
sequential thinking ones. I'm I'm curious to see because
29:47
I'm using one, but surely there's different ones around.
29:49
And if, if people think what is the best and does it work for
29:55
them? When do they call it and?
29:56
Yeah, yeah, exactly. And really, it feels like it is
30:00
unlocking your workflow to help you, like you said, get that
30:07
idea that you have out from your head onto something tangible.
30:12
So then you can go from there. It's not necessarily that you're
30:15
looking for that final product, but you're really thinking about
30:19
how can I get to something as quickly as possible.
30:26
Yeah, yeah. And, and, and what we're seeing
30:29
is we have, you know, the developer or whatever your
30:34
position is at one side, then there's this huge ecosystem of
30:39
options, possibilities and emerging capabilities of AI
30:44
tools at the other side and what you need.
30:47
Is. A router in between and zoomed
30:50
out like this. The cursor is an interface of
30:54
such router, it's not just an IDE.
30:56
And I'm saying cursor because I work with it, but I don't want
30:59
to say that's the. Yeah, could be windsurf, could
31:03
be GitHub Copilot if you really weren't.
31:05
And. I think they have agent
31:09
capabilities now. I mean, I would expect, yeah, I
31:12
thought I saw something on that. So.
31:14
But yeah, maybe, maybe not that one, but for sure windsurf.
31:18
Yeah. And, and, and, and we're going
31:20
to see much more of that because just that just the way that a
31:26
recommender algorithms, quote UN quote helped us and, and having
31:32
the illusion of unlimited choice while only feeding as a subset
31:36
of it. Yeah.
31:38
And this this thing is is also like a router between all these
31:45
different tooling. Now I was actually having a was
31:48
giving a talk yesterday at a Congress of responsible AI,
31:52
which was very focused on environmental and, and, and, and
31:57
responsible use and the environment.
31:59
And I was sort of that cowboy that that they that they got
32:04
into a tell about the other side.
32:06
And I mean, I fully agree with the fact that we should take
32:11
good care of of our earth and all, but this was about like,
32:16
how do we have people use AI responsible?
32:21
For example, if you have a use case that requires text
32:24
classification, you could use a simple Burt model and that only
32:28
uses very little resources compared to the heaviest
32:33
thinking models. And I think I heard some Altman
32:36
saying something similar as well, that he didn't like the
32:38
interface of Jetty BT at the current time where there's so
32:41
many different model options where you have to make these
32:43
choices again and again. And yeah, you know, the, the,
32:48
the, the request can be routed to the right model.
32:51
But with every other abstraction that comes on top of this, that
32:56
takes away choices that we make, which surely programmers are
33:01
also saying about coding with AI is, you know, I want to make,
33:04
you know, these decisions myself.
33:06
And this is, I don't know man, I don't know what, what is the
33:10
best answer, but we are putting an awful lot of trust more and
33:17
more into decision making. And this is what I referred
33:20
earlier to that the the paper about choice engines and really
33:26
nicely touches upon well, this sunshine is a behavioural
33:29
psychologist, not necessarily a technical dude.
33:32
He was on on point there that we're yeah, on of loading our
33:41
decision making into a is and we're doing that all the time
33:46
and more and more. Yeah, which it, it almost like
33:50
atrophies that muscle of critical thinking on it, because
33:55
you, if you're going to trust that it can make the decision
33:59
for you better than you can, or at least as good as you can,
34:02
then you now no longer have to spend so much time in that
34:09
decision making process. And on one side, that's a great
34:14
thing because it frees you up to do other stuff.
34:16
But on the other side, when you actually have to figure out an
34:20
important decision or you have to walk back and look at why was
34:24
this decision made. Good luck.
34:26
Yeah. And.
34:27
It's in you know, I think a lot of people like non-technical
34:32
people as well, use AI for brainstorming, for generating
34:36
ideas. But that also is is is is
34:41
already slippery slope because if you don't use very specific
34:47
prompting, and here prompting is important, but again, you can
34:50
automate that as well, then then for example, an image generation
34:57
model will create sort of the same style images and so your
35:03
set of choice is already limited.
35:05
So you're in a illusion of having choice while you're only
35:10
seeing a very small subset of the possibilities.
35:13
And same with I don't know, I have seen a lot of AI generated
35:18
content on LinkedIn lately. It's so bad and LinkedIn almost
35:22
encourages it with those replies that I think they got rid of
35:27
that feature, but the reply is where it would just be already
35:31
generated. Like great job or wow this is
35:34
awesome. Now, I, I think one of the,
35:36
because most people are using Jet GPT, right?
35:38
And what I think that I'm seeing is that people have a certain
35:42
topic that they think is, you know, viral worthy or they just
35:46
find it generally interesting or they know something about it.
35:49
And maybe they let Jet GPT write the entire post of say 1000
35:54
words or they write their posts and have GPT sort of edited.
36:00
But either way, especially GPT 4 O, but also the 4.5 research
36:06
preview model as well, tend to do this really annoying thing
36:10
that is yada yada yada yada and then recap it.
36:14
So here and then. Here's the catch.
36:17
If I see that one more time and it's like a rhetorical question,
36:24
you know, like, so there there's a that's, that's not just a, a,
36:30
a new is, is this a new era? No, it's a new blah, blah, blah.
36:34
I don't know. Like it's, it's, it's sort of
36:37
like asking itself a question and then answering it.
36:40
And I, I didn't look, you know, I'm not a linguist expert, but
36:43
it doesn't take much AI used to recognize this pattern.
36:47
So whether it be with image generation or text generation,
36:52
you see that if you're not giving examples, which is the
36:58
stage before fine tuning, which I don't expect marketeers to do
37:02
right, or whatever people outside of the technical
37:05
community. And and there's of course like
37:08
spiral and that's sort of in in between, which is SAS is sort of
37:11
fine tuning, but it starts with giving it the input output
37:15
examples that you want to see. And that requires thinking, that
37:20
requires critical thinking. And so it's, it's, it's almost
37:24
like there's a group of people that don't use AI at all.
37:27
And there's a group of people that use AI, but they use it for
37:30
too much. And the, the, the, the, the, the
37:33
golden ratio is, is, is somewhere in between where you
37:36
do have to think like this is how it should look.
37:40
And with text, I think that's that that's, that's sort of easy
37:46
to do. I think with images and video,
37:49
it's harder and also not my my game.
37:51
But you know you only have this mental mental image then, right?
37:55
Yeah, but also with text, it's blaringly obvious for us.
38:00
And with images you can, you used to be able to really see,
38:05
oh, this they I generated. But now there's some of these
38:08
models like this Rev AI. I'm not sure how you pronounce
38:13
it. The realism in that model is I
38:17
can't tell. And I like pride myself in being
38:21
able to tell when it was AI generated.
38:23
But coming back to the idea that you're talking about here, it
38:27
reminds me of when I was just talking with Devanch on a
38:34
podcast a few days ago, and he mentioned this thing called the
38:40
Doorman Paradox. Have you heard of this?
38:43
No, but I love paradoxes so bring it on.
38:47
For some reason, I knew it was going to be important for you.
38:50
Basically a hotel said, oh, we're going to cut costs and we
38:54
have sliding doors right now, so why do we need somebody to stand
39:01
at the door and open the door for people?
39:04
We can get rid of that job and they got rid of that job, but
39:08
there were secondary and tertiary effects that kicked in
39:12
when they no longer had somebody at the door.
39:15
And I'm sure you can think of them right away.
39:17
You're thinking like, oh, well, people don't get greeted as they
39:19
walk into this hotel and that's a worse experience.
39:23
But then also potentially you have more people loitering
39:27
outside of the the entrance. So then walking into the
39:30
entrances, it's not as good of an experience and these things
39:36
were kept in check or they were done by the doorman, but it
39:41
wasn't necessarily their job. This is Murafak's paradox.
39:47
Oh no, I mean, not not the specific example of the doorman,
39:50
but Murafak AI scientist legendary thing back in the 70s.
39:56
I propose more effects paradox where he has said that in the
40:00
future jobs will not be entirely replaced by AI, but jobs
40:07
consists out of consists out of tasks and different.
40:11
I mean, that is even like an abstraction, but there's
40:14
different tasks in one job and some tasks can be economically
40:19
feasible replaced by AI and some cannot.
40:23
And then what might happen is that if.
40:26
Jobs get sort of like split up, then we can see the truly human
40:33
capabilities specialized on the remaining parts of the job
40:38
because you know, whatever realistic an AI can be.
40:45
And I'm going on dangerous territory here, but one of the
40:49
most important things that I have seen in the restaurant
40:53
industry is, you know, when is a customer happy?
40:56
And mind you, I actually developed an application for the
41:00
recent industry, and I learned in the hard way, and I was
41:02
wrong, you know, because one of the most important things is
41:10
social recognition. As the species that we are, we
41:14
are constantly. Yeah, looking for social
41:18
recognition. And we can only get that by
41:22
other humans. But I don't know if this stays
41:27
true. Right, so they're I might be
41:32
wrong here, but obviously when it comes to say the job of a
41:35
waiter and part of their job is to predict what product the
41:39
customer will be most happy with in experience wise, then an AI
41:46
could probably do a better job. But if I am sitting in a
41:52
restaurant and it's really nice, Somoye is saying this and that
41:58
wine is great. If I really look deep into my
42:01
thought process and my experience, then I trust this
42:07
entity. So let's say this woman has
42:11
knowledge I perceive to have have knowledge that I don't
42:14
have. So there's a game of trust and
42:16
there's also maybe I don't want to sort of like make her happy,
42:19
you know, I want to impress her that I choose the right wine and
42:23
and I don't think I have that with, with, with an AI, right?
42:27
There's a lot of complex decision making and experience
42:30
there. Yeah, exactly.
42:31
And the whole reason I brought this up was because of this
42:34
thought that we're outsourcing too many of our decisions
42:37
sometimes or we maybe it's not too many, maybe it's just the
42:41
right amount, but we're, we're definitely outsourcing more and
42:44
more decisions. But the, the, we do that because
42:46
we feel that we're both two things.
42:51
There's an awful lot of decisions to make.
42:55
You know, the, the, the downside of ultimate individual freedom
43:00
is that you have a lot of decisions to make.
43:02
I mean, you can't really see like we can the whole thing that
43:08
you see happening with like conservatives, the conservative
43:11
right wing populism and, and such is also like populism in
43:15
itself is in a simplification of complexity and and therefore the
43:20
diminishing of decision making. And so that there, there is a
43:26
need for for us humans to make less decisions.
43:32
And there, and then at the and the other end, there's also,
43:35
it's becoming increasingly more complex to determine what is
43:40
increasing my social status. And it's just very, it's very
43:45
hard to grasp. And yeah, I think and I think
43:50
Yuval Harari writes beautifully about this, about this affected
43:54
AI might have and coming back to image generation mimicking the
44:00
real world, take it all the way to a virtual reality.
44:04
But essentially what he says, we're the brim of an age where
44:08
people are able to create the world's that they want to have.
44:13
And therefore 1 point that is always being brought up when
44:18
it's when we're talking about image generation or video
44:20
generation, whatever. Like how realistic is it?
44:22
Well. I back to differ.
44:24
It doesn't really matter even when there is watermarks,
44:26
weather is it's it's it's clearly AI generated.
44:30
If men define situations as real, they are real in their
44:33
consequences. Thomas and Thomas sociology and
44:37
and and and if people want to believe stuff, they are
44:42
believing it. And and you know, all this sort
44:45
of like cognitive biases saying like, yeah, but it could be
44:47
real. Yeah, yeah.
44:49
I mean, this picture might not be real, but it could have
44:51
easily been real. And if it fits their world view.
44:56
So one being here like the, the of course, like the creator side
45:07
is faster than the response. The, the, the, the, for example,
45:11
the European Union almost has to respond to, with policy to new
45:16
innovation and they're notoriously lagging behind.
45:19
And there will always be the way.
45:21
But also it might not even be effective at all because I don't
45:25
think, yeah, I'm getting kind of lost my train of thought here,
45:31
but it is that when something is in abundance, it sort of loses
45:36
value. So if if AI generated content in
45:41
whatever form is abundant, it loses value.
45:45
So that is shining some light on this sort of dystopian negative
45:50
view that maybe we are perceiving value in the new
45:57
world of yeah, the source, the authority that is creating the
46:04
content. And yeah, and and I don't know
46:08
if it's going to happen, but it seems like a positive thought,
46:11
you know, to to think that if generation is sort of free, then
46:17
something else is going to inject value into that process.
46:22
Yeah, because that becomes the commodity and that is the piece
46:27
that is like the least human part about it.
46:33
And So what is going to be the valuable part about that and
46:37
where, where is that part going back to the doorman's paradox
46:41
that we were talking about, or as you called it, what was it?
46:45
And modavec. Modavec's paradox, you're
46:49
looking at the uniquely human pieces of this.
46:52
How can, how are we now extracting that?
46:55
And it potentially is just that when you walk into a hotel and
47:00
someone says hi to you, that is the most valuable part of it.
47:08
So maybe so that do you the human connection.
47:10
So maybe the world, the world oldest profession is going to be
47:14
the worst last. Yeah, but it it's funny to think
47:19
about that as you're as you're pondering it.
47:21
It's like something so quote, UN quote, simple as engaging
47:29
another human is actually the best part of the value chain.
47:36
Like it's the biggest thing out of all the stuff that that
47:40
person does in their job or going to the the sommelier.
47:44
It's not necessarily the wines they recommend or getting to
47:49
know you. It's that you're talking to
47:51
another human face to face. Yeah, it's it's, it's the the
47:55
the human. Experience and the social
47:56
recognition, like the the worst thing that can happen to a
48:01
human, to a customer in the service industry is to be
48:06
ignored relative to what they perceive other people are
48:10
getting for personal attention. If you ignore someone and give a
48:19
lot of attention, human quality attention to someone else,
48:23
they're going to notice and they're going to hate it.
48:28
So, so I mean, hopefully this, this opens because things are
48:36
going to get automated and yeah, it's like, you know, eating,
48:43
watching, what is it? Ray Dalio said, looking into the
48:48
crystal ball is anyone who's looking into the crystal ball is
48:52
bound to eat ground glass. So don't, don't, don't pin me on
48:57
it. But we're definitely seeing some
48:58
very interesting shifts and it's going so fast.
49:01
And I think in, in, in your community, people are, I think a
49:07
lot of people are very open to, you know, to, to, to innovation,
49:10
to change. But I think you have to realize
49:12
that there's a huge part of society that responds very
49:15
differently to innovation. And like, like, you know, like
49:19
this, I don't understand it. I, I will ignore it.
49:23
And yeah, it's it's an interesting dance for sure.