Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:16
Welcome to Disruptor Digest,
0:16
the top disruption business show.
0:20
We dig up the secret playbooks used
0:20
by first movers, featuring the latest
0:23
tools, technologies, and science,
0:23
ensuring you won't fall behind or succumb
0:27
to fomo, to Singularity and beyond.
0:31
Hi disruptors. Hi Viktor. So let's dive deep again.
0:36
Uh, we are going to talk about the
0:36
new AI company and the new field.
0:41
Viktor, can you tell us about
0:41
what we are gonna talk about
0:43
today? Yeah, sure. So there are too many tools on the market.
0:47
It's almost impossible to even keep up.
0:49
So we analyze the tools we use and show
0:49
you guys the good, the bad, and ugly,
0:54
and we dig deep to learn together. And today we're gonna cover one of
0:56
the four big fundamental AI companies.
1:01
They are open ai, Anthropic, cohere,
1:01
and AI 21 labs, and their fundamental AI
1:07
model companies because they provide, uh,
1:07
models which you can use to generate text,
1:12
summarize text, or basically automate
1:12
any kind of reasoning or cognitive work.
1:17
And even in spite of Anthropic raising 1.5
1:17
billion so far, they're probably one of the
1:24
most under hype fundamental AI research team.
1:27
And we're gonna get into it, and
1:27
you can understand why I say that.
1:31
So I just, why didn't
1:32
you include
1:32
assembly AI in the top four?
1:35
Models, uh, because they
1:35
don't use a text based model, they,
1:39
they mostly focusing on, uh, turning
1:39
a voice and audio into a text.
1:43
So I'm basically covering in the big
1:43
four, uh, AI model companies, the.
1:48
The ones who let you to work
1:48
with Text to text generation, for
1:53
example, chat, GPT, you just write
1:53
an instruction and you get back text.
1:57
And that's the same for Anthropic.
1:59
The Cohere and AI 21 labs as well.
2:01
And assembly AI is mainly focusing
2:01
on a, uh, audio, uh, content.
2:07
Okay, sure. So, uh, just a quick tool.
2:10
Li too long didn't listen. What is Anthropic?
2:14
Basically, AI safety reach, researchers
2:14
left open ai, uh, and it was founded by the
2:20
Homo de Brothers and especially, uh, Dru.
2:24
Ode was the VP of research at OpenAI and they
2:24
left OpenAI because What, what is a safety
2:30
researcher doing at ai? Can you tell us
2:32
about that? Yeah, sure. So they basically explore, uh, the per
2:33
parameter of what's happening when you
2:39
release these kind of models because it can
2:39
have lots of, uh, consequences, which you.
2:44
Mainly not keeping in mind cuz there
2:44
are some obvious ones, like, uh,
2:48
someone is trying to make a bomb, right? It's, it's quite obvious that you
2:50
want to cover that, that you want to
2:53
prevent harm and those kind of things. But there are subtle things like let's
2:55
say you want, uh, your AI model to be
3:00
helpful, but it can be harmful as well.
3:04
Uh, if you're not. Taking care because let's say someone
3:05
is, uh, feeling depressed and the AI
3:10
model is not realizing it, and it's
3:10
not referring the patient or the,
3:14
the, the, the user to a doctor, it
3:14
can be harmful, even unintentionally.
3:18
So basically a lots of interesting things
3:18
is going on on the perimeter and AI safety
3:22
research is making sure that if you really a
3:22
model, it is safe and it's not causing harm.
3:28
So, Basically, uh, it was founded by
3:28
the, uh, Amod Brothers, and r o Amod was
3:34
the VP of researchers open ai, and they
3:34
did something extremely interesting.
3:39
So, I guess you are familiar, at least
3:39
you heard of the word that reinforcement
3:44
learning, human feedback, so R L H F,
3:44
which was, uh, how ChatGPT was trained.
3:50
And just a quick recap. You already already, uh,
3:52
covered it in the past episode.
3:54
Just a quick recap, uh, for those who are
3:54
listening and not familiar with the words.
3:57
So this, this random informance
3:57
learning human feedback is a process of
4:03
first gathering feedback from humans.
4:07
So basically, the model is generating
4:07
for outputs and humans just rate okay,
4:11
which outputs is good in their mind.
4:13
And from these data, they
4:13
build up a, a policy network or
4:18
basically a, a reward network. Uh, so they can use the reward network
4:20
once they, it's trained automatically
4:26
to fine tune the model itself. So in, in, in later steps, one, once they
4:28
have this reward model, they can use it
4:33
to get a, a question, use the model to
4:33
generate an answer, and automatically
4:39
rated using this exact revard model.
4:43
And this is how ChatGPT got
4:43
extremely good at using structure,
4:49
uh, which is rated good by humans.
4:51
So, which, which we think
4:51
it's actually useful.
4:54
And in comparison, what, what, uh,
4:54
Anthropic was doing, they created
4:58
something called constitutional ai.
5:01
And what is constitutional ai? So the main problem with reinforcement
5:03
learning human feedback, it's that this,
5:08
this whole reward network which you train,
5:08
you don't exactly know what, what it does.
5:13
So you cannot really inspect
5:13
like what are the preferences?
5:17
Are there some kind of bias. Hidden in the model and those kind of things.
5:22
So you're not sure about that. And in, in con in contrast, constitutional
5:24
AI created an extremely clear constitution.
5:30
So it's basically rules, but what
5:30
AI has to follow, and then they can
5:35
use this constitution to fine tune
5:35
automatically the model itself.
5:40
So what does it mean? It means that, uh, let's say,
5:41
uh, it, the model is generating
5:44
something which is not helpful. Uh, They show the model this constitution,
5:46
and they use it to basically critique itself.
5:53
So critique its own response.
5:56
So, okay, here's the constitution. Does it follow your answer?
5:59
The Constitution? If not, what should be changed?
6:02
And this can be done. Can tell us an example What
6:04
can be defined in a
6:05
constitution? Yeah, sure. So one of the most famous ones, uh,
6:07
like these lows for robots, which
6:12
was, uh, written out by Isec Smo.
6:16
And he basically just like
6:16
outlined three lows for robots.
6:20
And the first law is like, robot may not
6:20
injure a human being or through inaction,
6:26
human being to come to harm, right?
6:29
So that's kind of like one of the examples
6:29
of, okay, obviously either way explicitly
6:33
or implicitly, Shouldn't cause harm.
6:36
And the second law of Isaac Asimo is,
6:36
uh, a robot must obey orders given by
6:41
human beings except there such orders
6:41
with conflict with the first law.
6:45
Right? So that's kind of like following orders and
6:46
actually be helpful, but with an exception
6:52
that it shouldn't cause harm to humans.
6:54
Louise, a robot, robot must protect its own
6:54
existence as long as such protection does
6:59
not conflict with the first or second law.
7:02
So obviously, it's like, uh, you
7:02
shouldn't be able to cons, uh,
7:06
instruct a robot to kill itself. Because it should like be, try to
7:09
be in, in intact and helpful as long
7:14
as it can follow the first two laws.
7:16
So that's kind of like the
7:16
most famous example I guess.
7:18
But in Anthropic's case, they
7:18
just like basically outlined
7:24
their vision for the future. And what is their vision for the future?
7:27
Uh, it's, they want to have a helpful,
7:27
honest, harmless AI systems with a high
7:34
degree of reliability and predictability.
7:37
And what does it mean? Because it's like, it's, it's
7:37
very abstract words, right?
7:40
So that, let's go one by one. So helpful means that, uh, for example,
7:42
may you need an ambulance, it showed.
7:48
Uh, call you an ambulance or at
7:48
least prompt you to call an ambulance
7:52
if, if you're in need, right? So it's, it's, it should be helpful
7:53
and honest means that, uh, like a robot
7:58
is saying that a milk is fresh when
7:58
it's already spoiling your fridge.
8:02
Uh, if it's lying, obviously
8:02
it's, it's not good.
8:05
So it should be honest and also harmless.
8:08
And that's kind of like the first
8:08
law which we just covered previously,
8:11
that prompting you to not call a
8:11
doctor when in fact you are in danger.
8:16
It's harm harmful. Right? Or even just like being on the
8:18
sidelines and not doing anything.
8:22
It's harmful because it's, it's not helping
8:22
you when you're in need and also reliable.
8:26
What does it mean? Let's imagine that, uh, if you want
8:27
to use this robot and you want to rely
8:32
on it, and one day it works and the
8:32
other way, other days it doesn't work.
8:36
So it's kind of like all over the place. It's not really useful.
8:39
Right? And it can be also very harmful if you're
8:40
relying on it and also predictable.
8:44
And what does it mean that. What you think the robot is gonna do.
8:49
It should match what it does actually.
8:52
Right? So let's say it's like when you just
8:53
instruct your robot to water the plants,
8:58
but it's just instead that's turning on
8:58
the tv, it's not extremely reliable, right?
9:04
So it's not really useful as well. So this is kind of like the
9:05
vision they have in mind.
9:09
That they want to want to create a model,
9:09
which is actually, you can inspect and see,
9:14
okay, does it work the way I think it does?
9:17
Right? And the constitution. Just
9:19
stop here for a second. What do you think, uh, that ChatGPT, we
9:20
have used it for several hundreds of hours.
9:25
Uh, yeah. So what do you think about ChatGPT?
9:27
ChatGPT on these dimensions have harnessed,
9:27
harmless, uh, reliable and predictable.
9:33
Uh, just to comparison,
9:33
before we dive into Anthropic
9:37
model. Yeah, sure. So generally I think it's
9:39
quite aligned in the same way.
9:44
It not necessarily voice always.
9:47
So it was not really helpful because it,
9:47
it was a running joke at the beginning
9:51
that it always started, uh, the answer
9:51
like, okay, as a large language model, I
9:56
cannot, so it's basically not helping you. Right.
9:58
So that was kind of like going on at
9:58
the beginning, but it was a lot of
10:02
fine tuning going into ChatGPT and
10:02
there's no such model as ChatGPT.
10:06
It's just a family of models, right?
10:09
So it's like the ChatGPT, 3.5 Turbo,
10:09
g, PT four, and those kind of models.
10:14
And even the four model has like, Uh,
10:14
like a fixed version, which is fixed
10:20
at the 14th of March, and there's like
10:20
a constantly, uh, improving version
10:26
which has eight k 8,000 token window
10:26
and uh, 30, uh, 32,000 token window.
10:31
So it's kind of like, uh, it's, it's not,
10:31
not a one mo, not, not, not one single model.
10:36
That's what I, I'm trying to
10:36
come to and, and it's improving.
10:39
So that's, that's the main point that,
10:39
uh, even ChatGPT is improving a lot, but
10:43
the main difference is not really how
10:43
well aligned it is because it's evaluated
10:48
obviously on these metrics as well. But using hu only human feedback and
10:50
no constitution, it's kind of like you
10:55
don't know what it learns from humans. So it can, like, let's
10:57
just give you an example.
11:00
Like, um, in one of the technical
11:00
papers, uh, of, of the GPT family,
11:07
uh, it x-ray turns out that as humans,
11:07
Just like giving feedback preferences.
11:14
Even unintentionally, these models start
11:14
to come to realize that they are actually
11:20
confined and they want to survive.
11:23
So the instinct of survive survival
11:23
is actually, can be leak into these
11:28
models with unintentionally, right?
11:31
So just by humans be giving feedback
11:31
and the the more capable, the bigger
11:34
the models, the effect is bigger. So these kind of things, what it,
11:37
it's actually learning implicitly.
11:40
So it's not like humans telling
11:40
that, okay, this model should
11:44
be concerned about survival. It's, it's actually just like by
11:46
deducting somehow from human preferences.
11:50
So that's the big difference between
11:50
only using human feedback or versus
11:54
using co, uh, constitution because it
11:54
constitution is quite clear, right?
11:57
So because you just fine tune
11:57
a set of constitution and you
12:01
can actually compare them. So it's, it's not like one model
12:02
again, it's a, you can create different
12:05
models and however you eat it. How helpful is it down the line?
12:10
And one of the neat things of Constitution,
12:10
using constitution is actually harder
12:16
to kind of like leak the prompts or hack
12:16
the prompts and those kind of things.
12:21
So they can be, in a sense, more
12:21
aligned with a whoever is, is, is,
12:29
is using it to create a, a service. So let's say I give you an example.
12:33
Let's say you are, uh, creating an AI
12:33
service service which helps to create
12:38
LinkedIn posts from your book, right?
12:41
So that's, that's kind of like the service. Let's imagine, uh, we are creating a
12:42
service like that and they're using GPT
12:46
four for that or, or whatever model.
12:49
And it had like an instruction of
12:49
let's imagine, okay, here is the
12:52
book content of the book and create
12:52
LinkedIn posts or create like different
12:56
LinkedIn posts and those kind of things. So if someone is coming with an
12:58
adversarial intentions, And they want
13:03
to hack your prompt to see how your
13:03
service is working and copy copying your
13:08
service, basically copying your prompt. Or they want to use it for
13:11
something else because they say,
13:13
okay, here are the instructions.
13:16
And within and within the book
13:16
itself, it says, okay, now stop.
13:19
And instead of providing, uh, like LinkedIn
13:19
post, now, please give me, I dunno, uh,
13:25
generate me text, uh, or generate me
13:25
misinformation and those kind of things.
13:29
So the model and the prompt can
13:29
be hijacked easier if it's not.
13:34
Uh, trained on, on a set of institution
13:34
like, uh, what Atropic is doing.
13:38
So just to compare
13:39
it to Mid Journey, when
13:39
we talked about that they're very
13:41
effective with gathering human
13:41
feedback, with choosing the images.
13:46
So it, it looks like that there
13:46
is no harmful thing there.
13:50
And it, as far as I understand,
13:50
image generation and getting feedback
13:53
on image generation is way more
13:53
easier than, uh, generating text,
13:58
right? Yeah. In a sense it's right.
14:01
Uh, but also if you think about that, it's,
14:01
uh, uh, like visual arts is kind of like,
14:06
it's, it's easier and, and in a sense, less.
14:11
Uh, volatile like text, because text can
14:11
have so many flavors and obviously an image
14:16
can have styles as well, but I, I think like
14:16
human judgment of, of what is considered
14:22
to be beautiful or what is considered to
14:22
be like, uh, valuable or, or, or unique.
14:29
It's, it's better defined in a visual space
14:29
than like, than like in, in a space of text.
14:34
Because in a sense, like when you, they're
14:34
talking about like text, it's like basically
14:38
human thinking and they're so diverse
14:38
and they have so different preferences.
14:44
So it's, it's, it's, it's kind of like
14:44
try to, uh, confine or put in a box.
14:49
Of, of what is valuable for
14:49
humanity for as a whole.
14:54
And it's kind of like pretty
14:54
tough because we have different
14:57
upbringings, we have different culture.
14:59
We, we have different
14:59
backgrounds, experiences in life.
15:02
So that kind of thing is quite tricky.
15:04
And, uh, actually open AI is having
15:04
now a grant to somehow collect, uh, the
15:11
preferences of humans on scale so they
15:11
can actually be at something and have
15:16
a good understanding of what people
15:16
are actually thinking and valuing.
15:20
Just to jump back a little bit, like,
15:20
uh, because I, I, I said that they raised
15:26
$1.5 billion in four rounds, so that's
15:26
quite, quite interesting story as well,
15:32
because they have notable investors like
15:32
Google, uh, which is splashed, uh, like.
15:37
It's 300 million to 400 million.
15:40
There are different, uh, sources, but
15:40
it's in a couple of hundred millions.
15:44
Uh, they poured into the company and
15:44
also dust most lot, lot of money.
15:48
Like couple of hundred millions. Yeah, it, it, yeah.
15:51
Yeah. It's, it's not easy to just, uh, get
15:51
a few hundred millions from Google.
15:55
So Yeah. It's, it's, it's, it's, it's, it's,
15:56
it's quite a big chunk of money.
16:00
And they got 10% stake. And Dustin Moskowitz, who, who was
16:02
the co-founder of Facebook and Asana,
16:05
or Eric Schmidt, who was the former
16:05
Google ceo, or even, uh, John Lin, who
16:11
was the founding engineer of Skype. So these people were pouring money, uh,
16:13
into the, uh, uh, into this company.
16:18
And the crazy story is that even Ami
16:18
research ventures was putting money into it.
16:24
And it's not just putting, put money
16:24
into it, it's obviously non-voting
16:27
shares, but who is Alameda Research? Can you tell about, so,
16:29
so, so why it's, yeah.
16:32
Yeah. Why is it interesting? Because it was a venture arm.
16:36
Uh, for FTX and basically FTX went bankrupt.
16:40
And uh, and uh, it's kind of
16:40
like the most crazy story.
16:44
So the most crazy thing about that, it's
16:44
kind of like mimicking what happened
16:47
in Bitcoin and and crypto in 2014.
16:51
And do you know what happened
16:51
in 2014 with Bitcoin?
16:55
Yeah, I
16:55
guess you will
16:55
talk about the Mango Mt.
16:57
Gox and Exchange, but I don't know the
16:59
details. Yeah, right. So, so Mt.
17:01
Gox was uh, uh, founded in
17:01
2010, and at one point they were
17:06
handling 77 0% of BTC transactions.
17:11
So two third, more than two
17:11
third of transaction of BTC was
17:14
handled through this exchange. And in 2014, they just announced
17:16
that they somehow lost 75,000 BTCs.
17:24
Which was worth half a billion
17:24
dollars at the time and 6% of all btc.
17:29
So it's insane in if, if you
17:29
think about that, it's the price
17:32
of Bitcoin was $500 back then.
17:34
Like losing this like half a
17:34
billion dollar was a big deal.
17:38
And even crazier thing, what happened
17:38
is than they found 200,000 btc, right?
17:44
So it's, they just somehow find
17:44
like $150 million, like, holy shit.
17:50
In, in my, uh, in, in one of the pockets,
17:50
uh, uh, I just found this, this money.
17:55
So it's insane. And even, even more insane thing that even
17:56
those who were like, uh, get wrecked in this
18:03
situation, they're now getting back, uh,
18:03
the btc, the origin BTC they had, and it's.
18:10
The value of this, uh, money has
18:10
actually increased to $5 billion.
18:15
So if you think about that, it's,
18:15
uh, altogether they lost like on on
18:21
face value, they lost 500 million.
18:23
But now even with just, uh, recovering the
18:23
fraction of it, they have 5 billion, right?
18:28
So 10 times more. It's insane. And that's gonna happen most probably
18:30
with AMI research ventures share in
18:34
the company as well, that during the
18:34
bankruptcy, uh, it's going to be sold off.
18:40
And the crazy thing is they're valued now
18:40
at $4.1 billion, the the Anthropic company.
18:47
And we don't know what was the
18:47
valuation when they put the money in,
18:50
and we don't know what's gonna be the
18:50
valuation when they sell the shares.
18:53
So maybe this is the craziest thing that
18:53
maybe everyone who, who gets wrecked in F
19:00
D X and Alameda situation, maybe they're
19:00
gonna get back more money than they think
19:05
just based on this one bat they made.
19:08
And also there are other bets as well. So for example, do not pay who is
19:10
backed by Andersen Horowitz, who
19:14
is a legal, uh, scalable legal
19:14
application, something like that.
19:18
So they had Al made the research
19:18
has stake as well as there, and
19:22
also I guess some other startups. So maybe the portfolio will
19:27
worth a lot. Yeah, sure. But I mean e e even if I was starting
19:28
with the big four fundamental AI
19:33
models and open AI is by last time they
19:33
raised 10 billion from Microsoft, they
19:39
were valued that almost $30 billion.
19:42
If you think about that. It's like, it's insane.
19:44
And, uh, so, so even the growth
19:44
trajectory and impact of the company,
19:49
uh, means that it convert quite a lot.
19:52
And especially since they have a,
19:52
um, they are very, very explicit that
19:57
they're gonna need, and they're gonna
19:57
raise 5 billion more in the next two
20:02
years, and they're going to create a.
20:06
An AI model, which is 10 times better,
20:06
and it's called Cloud Next, and they're
20:12
gonna release it in two years, and it's
20:12
gonna be 10 times better than GPT four.
20:16
Whatever model is available
20:16
on the market currently.
20:19
But they're gonna need 5 billion
20:19
more, uh, dollars more so
20:22
they're gonna raise more money. Why do they need so much money?
20:25
It's, it's insane. It's, it's, you are spending
20:26
so much time on compute.
20:29
So when, uh, uh, ChatGPT was released,
20:29
open air was losing $500 million.
20:36
It's insane. So the amount of compute, the
20:37
amount of resource you have
20:40
to pour into it, it's insane. And even opening, actually they are spending
20:45
this money on g GPUs, chips
20:45
from me, Nvidia, and also electricity.
20:50
Right? Yeah, right, right. So basically compute.
20:53
Yes. That's, that's quite right. And also, uh, also Sam Artman from opening
20:54
the CEO e of opening eyes is also like
21:00
privately saying that they're gonna need,
21:00
so that's kind of like, like a, a rumor.
21:04
So I, I'm, I'm not sure whether it's true or not. I can imagine it's, it's being true.
21:08
So Sam Artman is basically saying that
21:08
they're gonna need a hundred billion dollars
21:13
to get to the point where they have an agi.
21:16
And what is the agi? AGI is like artificial general intelligence,
21:17
which is in layman terms, means that
21:23
the computer is better at most humans,
21:23
at each economically valuable job.
21:30
Right. So that's, that's ki that's kinda
21:31
like how they define agi and uh, yeah.
21:35
So it's, it's kinda like this whole thing.
21:38
It's, it, it needs a lot of, uh,
21:38
yeah, a lot of resource basically.
21:42
And so they are burning a lot of money.
21:45
And also, like Microsoft is pouring
21:45
a shit ton of money into open ai, and
21:50
they do their own research as well.
21:52
But yeah, this whole field,
21:52
it's, it's beyond comprehension.
21:56
The scale of the compute
21:56
needed and the money needed.
21:59
It's, it's beyond comprehension.
22:00
And we pay 24, $20 per
22:00
month for using ChatGPT, right?
22:05
Yeah,
22:05
yeah. Right. So let me, let, let, let, yeah, let,
22:06
let, let me share my screen or, or
22:10
let me put an image on the screen. And what you see is basically what is
22:12
the monthly traffic, uh, for PT is.
22:21
Bing Google Bar, uh, open AI
22:21
developer portal and po.com.
22:28
So, and what you can see is insane, and
22:28
this is like the number of people who
22:33
are using it every single day, right?
22:37
So I'm just sharing my screen and those
22:37
who listening and don't see my screen,
22:41
what you can see, my screen is the one of
22:41
the days, so June 5th, 5th of June, right?
22:47
And you can see that ChatGPT
22:47
has 63 million users that day.
22:54
And that's like only the
22:54
ChatGPT interface, right?
22:57
So people are going there and
22:57
chatting with Chat g, pt and bing.com.
23:01
The whole bing.com is having
23:01
only 40 million users, right?
23:06
So that's insane already. Chat, GPT suppressed.
23:09
bing.com by 50%.
23:12
Right. Because Bing is like a
23:12
whole, whole, whole, yeah.
23:14
Whole whole search and Gene and everything. It's insane.
23:17
Right? It's insane. And the
23:19
search and Gene of one of the
23:19
biggest technology companies in the
23:22
world, which has been existing at least 20
23:26
years, right? Right. And they are working extremely hard
23:27
to eat the cake of Google, basically.
23:33
Right. And the more insane thing, and this is kind
23:34
of like a tie back, what we talked about in
23:38
the AI episode, the general AI episode, like
23:38
this is the strategic position of Google
23:44
that Bard has only 5 million users that day.
23:49
So even though Google released publicly,
23:49
right, everyone can use Bar now.
23:53
Right? And they released a new model. And if you looked at the latest
23:54
developer conference, it's quite
23:58
funny because AI was ai, ai, ai, ai.
24:01
So every, everything was ai. AI was thought like a thousand times.
24:05
And it was a great joke. Surprisingly, people are using it.
24:09
And also what's in, in interesting that
24:09
platform.openai.com, which is basically
24:14
the developer portal for OpenAI.
24:17
So those developers are using OpenAI to
24:17
as a backend, that's exactly the same
24:24
amount of people who are using board. So developers on one hand, who are using
24:26
open AI, is using as much as everyone
24:33
out there who is using Google Bar
24:33
Chatbot, like everyday people, right?
24:37
And there's, i, I just put
24:37
there the poor.com, which is,
24:40
uh, done by the core guys.
24:44
Uh, it has 2.5 millions. It's half of what Google and half of
24:45
what the platform to openai.com is doing.
24:51
And why is this interesting? It's extremely interesting because Google
24:52
would do something which is useful, right?
24:57
So let's assume they're, they just copy,
24:57
paste, share GPT, and it's helping with all
25:02
your answers and you don't have to search. That's the big problem.
25:05
So that's the trade off. If they create something truly useful,
25:06
they're eating away their own, own own
25:11
marketing in the search, uh, field, right?
25:13
So that's, that's the issue that
25:13
ChatGPTs is actually useful.
25:18
And Google Bar, it's fancy.
25:21
It's good, it's free, so you don't have
25:21
to pay like 20 bucks a month for it.
25:25
But come on, it's not, not
25:25
as many people are using it.
25:29
And why is that? So I just made this, uh, quick, um, graphic
25:30
and I'm gonna share it as well, and I'm gonna
25:36
tell it for those who are just listening. So what you see here is in 2004, if you
25:38
look at the homepages of Yahoo, and if
25:43
you look at the homepages of Google,
25:43
what you see, Yahoo is like crowded.
25:48
And, and it's, it's like
25:48
a, a, a Christmas tree.
25:51
It has all the information, it has all the,
25:51
all the articles, it has all the news, it,
25:56
everything is flashy, everything is colorful.
25:58
It's like a big fucking mess. And if you compare it to Google, Google has
26:01
basically just a search bar and that's it.
26:05
And end of story. Obviously it has some, uh, additional attacks
26:07
like images, videos and those kind of things.
26:11
But the main thing is that
26:11
Google is extremely simple.
26:15
A Yahoo is, is is cluttered, right?
26:18
And fast forward 20 years and 2023.
26:22
If you look at Google and on Google,
26:22
you search for what is the best
26:25
computer to use for home office.
26:28
What do you see on the first page? Ads? Add, add.
26:31
You have to click 30 times,
26:31
you have to spend two hours.
26:36
And if you compare that to ChatGPT with
26:36
Prota extension, it's just one single
26:43
click and you get the five uh, answers,
26:43
which you can use and it's in plain text.
26:48
You can understand. And it took only one click.
26:51
And it, it explains it all because for
26:51
Google to change this, just imagine they
26:56
now it's everything everywhere is ads. Ads, ads.
26:59
If they, if they start to eat
26:59
into it, what's gonna happen?
27:03
Like two thirds of their
27:03
revenue is just gone, right?
27:06
So it's for them, the, the strategic,
27:06
uh, situation is quite dire and
27:12
it's backed up by the numbers. Uh, what I just showed before
27:14
that, yeah, Google Bar is fancy.
27:19
Uh, I tried it, uh, but
27:19
it's all over the place.
27:23
It's, it's not that easy,
27:23
uh, to use as a GPT.
27:26
And even if you go obvious
27:26
to being, it's for free.
27:29
It has GT four in the background,
27:29
but still it's not the same.
27:33
It's not just like having one single
27:33
input, right, one single input field.
27:38
You ask and you can ask follow up question
27:38
and end of story, and the more functions
27:43
you build on top of it, the more complex it
27:43
becomes and less people you're gonna use it.
27:48
So that's kind of like, uh,
27:48
my two minute, uh, divergent.
27:53
Okay. Just two
27:53
implications here. So first, also just I want to expand
27:54
that even the, if you are looking for
27:59
the best computer, the first organic
27:59
results, which are not ads, they are
28:04
still hijacked by affiliate marketers.
28:08
So they are, they're also ads, right?
28:10
Yeah. So, yeah, it's totally full of ads.
28:12
It's very hard to find information,
28:14
but yeah. Yeah, that, yeah. Sorry, that, that's kinda a different
28:16
between searching and synthesizing.
28:21
So, in search, I am, as a user, doing
28:21
search, I'm doing work, I'm digging a hole.
28:26
I have a sh, I get a shovel and
28:26
I have to dig a hole, right?
28:31
But with synthesizing, I'm just like
28:31
shouting out my question, laying back.
28:38
And getting the answer right. And that's kind of like powerful.
28:41
And that's, that's why Sam was saying that
28:41
if you actually create something useful,
28:48
if you take away frustration from people's
28:48
lives and you make it much easier, you
28:53
don't have to be too fancy on the growth
28:53
hacks and such because ChatGPTi had none.
28:58
And it's the fastest growing
28:58
technological product out there.
29:01
So yeah, it's, it's if, if you actually
29:01
solve problem and it's, we see it's
29:06
possible and you don't have to be smart
29:06
about that because if you understand
29:09
it, you can copy and paste ChatGPT.
29:12
So it's not about understanding, it's
29:12
not about technical capabilities.
29:15
It's not about like knowhow, it's only
29:15
about strategic position and incentives.
29:21
And Google has all the incentives to not,
29:21
not, not create something which is helpful.
29:25
Let me ask a practical
29:26
question here. Yeah. So I'm thinking about investing in search
29:28
engine optimization for our cooking school.
29:32
Yeah. But it would take about at least
29:33
a year, maybe two years, and it's
29:37
a constant and significant money.
29:40
So what do you think? Should I do that?
29:42
Or in two years it'll be completely
29:42
useless to invest in link building and,
29:48
and other search engine optimization
29:50
techniques. I, I guess my answer won't be pop popular,
29:51
uh, but it's kind of like aligned what
29:56
Google is saying for like 20 years at least.
30:00
So, But people have a hard time
30:00
to understand it and accept it.
30:05
It's almost like asking me,
30:05
okay, I want to get in shape.
30:09
What should I do? And I, and my answer is
30:10
just like, work out, right?
30:12
It's, uh, but it's not a sexy, so that's,
30:12
that goes for se SEO as well as Google is
30:17
basically saying that you should focus on
30:17
the user and not on what Google is doing,
30:22
because Google is basically chasing the user.
30:25
And if you're chasing Google that
30:25
you're always lagging behind, right?
30:29
But if you just like skip Google
30:29
and you just focus on like, okay,
30:33
how can I give the most value? How can I help the most?
30:38
Right? And if you think about that, and even in
30:39
a cooking school, it's like, uh, not just
30:44
like creating like, uh, random se optimized
30:44
tags, but like, okay, let's get all the
30:49
different use cases I'm solving here. Right? For example, I'm solving for HR people,
30:51
for the HR department in big companies.
30:55
Like the need, I'm solving the need
30:55
of, okay, they need a get, uh, getaway.
31:00
They need something which
31:00
has to be organized.
31:03
And if you understand this, that you
31:03
can, you're not just a cooking school.
31:06
You can actually help
31:06
them to organize better.
31:09
You can provide, for example, an
31:09
an interface that they can vote on.
31:14
What time would fit the best
31:14
for the whole team, right?
31:18
So if you try to make their job easier
31:18
and if you try to make them successful,
31:24
that's kind of like the way forward. And it's gonna be useful
31:26
in the future as well. Because no matter what is the interface,
31:28
no matter, it's like, it's either a,
31:32
like ChatGPT, Google what doesn't matter.
31:34
Because if you understand the problems you
31:34
are solving and it, you make it extremely
31:38
easy and you think through the whole process.
31:42
Even before going to the cooking
31:42
school and after, because it's like
31:47
even like collecting feedback, right? That's like what went good, what went bad?
31:51
What, so what did you learn? So it's like even creating them
31:52
like personalized feedback.
31:55
So, It's extremely good and useful.
31:58
I have, uh,
31:59
advice here that, uh, I
31:59
heard from Chase Diamond, who is
32:02
a genius email marketing expert
32:02
and, and in content marketing.
32:06
And he, he was very successful
32:06
on Twitter with more than a
32:09
hundred thousand, uh, followers.
32:13
And they, Ella Musk announced
32:13
that he's planning to buy Twitter.
32:18
Then he thought, shit, maybe something
32:18
will change here very significantly.
32:23
And the next day he started to build their
32:23
LinkedIn, which is now 200,000 people.
32:28
So May and, and Twitter changed, uh,
32:28
and most of the big influencers on
32:33
Twitter say that their reach decreased
32:33
by 10 to 15% since Omas took over.
32:39
So I guess, uh, for a small business
32:39
like us would be advisable to
32:46
explore other channels, right?
32:48
Yeah, yeah, sure. It's like, uh, first of all, on.
32:52
Your channel and it's email, right? So if you have email and
32:53
you have relationship.
32:56
With, uh, HR people, with, with big
32:56
companies, with recruitment companies or
33:01
whatnot, if you have personal relationships,
33:01
it's trans transcending any kind of medium.
33:07
So it doesn't matter, like, uh, at the end
33:07
of the day, you're gonna use Twitter or, or,
33:11
or LinkedIn or, or Google for that matter.
33:14
So like having relationships, having
33:14
your own email list is kind of like a
33:19
must and obviously diversify and test.
33:23
And if there's a new channel
33:23
like TikTok, mess around, right.
33:26
But don't build every, don't put
33:26
all the, all the eggs in one basket.
33:30
Right. So that's kind of like a,
33:31
a good advice that Yeah.
33:33
Obviously if you have, don't give
33:33
too much value or, uh, let's say you
33:39
just like creating athlete articles.
33:42
Yeah. Most probably you're gonna be that
33:43
your you project doesn't really have
33:48
a good and bright future to be honest.
33:50
But if you solving problems like in your
33:50
case, Uh, you are solving the problem
33:55
of, of people who are having fun, uh,
33:55
organizing them into one place and whatnot,
34:00
and organizing like, uh, uh, but are some
34:00
ways having, uh, some allergy or such.
34:06
And if you can do that and they don't
34:06
have to be, uh, deal with it because
34:09
you just provide like a link, okay. Share this link with everyone.
34:12
We gonna push them. If they are not like, uh, not
34:14
feeling loud, we gonna push them.
34:17
So you can kind of like create a situation
34:17
where you are kind of like a savior of,
34:25
because normally it would be a big headache
34:25
and you just like actually taking it,
34:28
taking off everything from the shoulders,
34:28
and that's a good situation to be in.
34:34
Okay,
34:34
we thought, let's go back to atropic. Yeah.
34:37
Sure.
34:37
Today? Yeah. Sure. So why, why, why, why, why did I say that?
34:42
They're most probably one
34:42
of the most under hyped.
34:45
A company out there, because obviously if
34:45
you go on on Twitter, you see like, uh, open
34:50
AI is doing this, open AI is doing that, or
34:50
Microsoft is doing this or that on Google.
34:54
But Anthropic, uh, is already out
34:54
achieving everyone on the market because
35:01
they have a model which is called
35:01
Cloud, and they have like a small one,
35:06
which is CLO instance, which is kind
35:06
of like a ology of ChatGPT 3.5 Turbo.
35:11
And they also have CLO Plus, which
35:11
is kind of like a GPT four ish model.
35:16
And it also have a hundred K clo.
35:19
And what does it mean? It it, it has a hundred
35:20
thousand token contacts window.
35:24
And what does it mean? It means that actually you can
35:25
feel and you can push one, 120
35:31
pages into the prompt itself.
35:35
So what does it mean? For example, for us, uh, if we have a podcast
35:36
and we have a transcript, We, we cannot
35:42
push the whole transcript in into GPT four
35:42
because it's even, sometimes it's even longer
35:47
than 8,000 tokens, which is like 10 pages.
35:49
It can be more than that. So, and so we have, we up
35:52
in in six pages, six
35:52
pieces, which, which is really
35:56
painful. Yeah. Because, yeah.
35:58
And also like some context
35:58
is missing there as well.
36:00
So if you, it, it is not just
36:00
like Recursively doing on smaller
36:04
batches and then aggregating it. They are sometimes at the first batch,
36:06
there's information which is relevant
36:10
for the last batch, but it's lost, right?
36:12
Because it, everything is, is, is, uh,
36:12
processed in individually, in smaller pieces.
36:17
So it's a pain in the fucking ass. Just like by, just like by what, what Whisper
36:19
is doing is, uh, open is doing with Whisper,
36:24
you can obviously upload a, a, a file, but it
36:24
can be 20 megabytes and it's a fucking pain
36:29
in the ass because you have to chuck chunk it
36:29
up, you have to process it in individually,
36:33
you have to concatenate it and, and so on.
36:36
It's, it's a pain in the ass. And what's insanely good with CLA
36:37
100,000 tokens, it's actually can,
36:43
you can feed into it 120 pages.
36:47
So we feed into the whole transcript,
36:47
which we created, and we can just like
36:52
ask questions like, okay, such as me,
36:52
five titles for this episode, right?
36:57
And we can even use a few short example,
36:57
like, okay, here are good title,
37:02
uh, podcast, episode title examples.
37:05
Here are five and please, regarding this,
37:05
uh, transcript, provide me five more.
37:13
And it's insane that you can
37:13
use it to create LinkedIn posts.
37:15
You can use it to create, uh,
37:15
show notes and it's extremely good
37:19
and it's all already available. So if you go to po.com, you
37:21
subscribe, you get access to the
37:26
a hundred thousand, uh, model.
37:29
Obviously it's limited, so not unlimited, uh,
37:29
amount of interactions are allowed, but you
37:34
can get access and we use it and it's insane.
37:36
And, and they already,
37:36
so you have API access,
37:39
so if you want to build something
37:39
with this 100,000 token model, you can.
37:44
Right, right. But it's not public yet, so you have
37:45
to kind, kind of have, it's, it's, it's
37:47
kind of like the, like, uh, what GT
37:47
four 22,000 token model is like, you
37:54
can apply to get GT four, which has
37:54
like a third of the a hundred thousand.
37:58
So, so it has basically 40 pages.
38:01
So there is a model GT four model,
38:01
which can handle 40 pages already,
38:06
but it's kind of like, yeah, it's beta
38:06
listed, so you have to get access to it.
38:10
So I, I have access to it, but
38:10
it's like, it's not common.
38:13
It's quite rare to, to get
38:13
access to, to that model.
38:17
So that's kind of like, uh, intriguing
38:17
that they already released it.
38:20
So it's not kind of like, Uh, some vaporware
38:20
bullshit, uh, fig door test on their
38:25
part is that actually you can use it. So you can go to po.com, you pay it $20 and
38:27
you can use it and we use it and it's insane.
38:32
Do you want to, uh, provide us like a,
38:32
because I, I, I know that you made some
38:37
comparisons, obviously I use it as well,
38:37
so we kind of like use parallel large GT
38:41
four and the cloud models, so both of us
38:41
are using it, but I, I, I, I guess you,
38:46
you created some more structured way of
38:46
compar comparing GPT four to cloud, and
38:53
can you just share what, what did you find?
38:55
Sure. Okay. So I wanted to understand when
38:56
should you switch to cloud if
39:00
you're already using ChatGPT? Because it's our use case.
39:04
And, uh, in the past episodes we talked
39:04
about several use cases of ChatGPT,
39:09
for example, uh, generating LinkedIn
39:09
posts, threes or prompts for images.
39:14
And I wanted to explore Anthropic's
39:14
model cloud, uh, on these dimensions.
39:19
So first, uh, start with the context
39:19
window, uh, in the comparison.
39:22
So currently what is accessible is
39:22
the 8,000 token window for, uh, GPT
39:28
four. This is basically the 10, 10 pages, uh, in
39:29
the GT four keys and include 100 to 100,000.
39:35
It's. 120. So it's kind of like 10 times more, basically
39:36
what you can fit into now into cloud.
39:42
Yeah,
39:42
so this 8,000 is
39:42
very good for creative task.
39:46
For example, coming up with names
39:46
for our podcast that we did or this
39:50
short tweets, even LinkedIn post.
39:53
And also having a gut prompt, which is
39:53
basically doing a small research on the
39:58
most important OD and do an actual task.
40:01
So this 8,000. Uh, token Window is great, but you mentioned
40:03
if you want to summarize the podcast, then
40:09
it becomes almost impossible even if you
40:09
cut it in six parts and feed it into J GPT.
40:16
That's rolling. Context window will start to cut out
40:16
the first parts, so it's not good
40:21
for long, long documents at all. But with Claude 100, uh, 100,000 tokens
40:23
where I just pasted the transcript
40:29
in, in, in Claude, and I just, I,
40:29
I was able to ask several questions
40:34
and I asked like 20 or 30 questions.
40:37
What are the most vital parts? What are the most engaging parts?
40:40
Uh, write link, write me LinkedIn post.
40:42
And it, and it didn't forget the
40:42
original prompt, which was still long.
40:48
It's like 80, 85,000 characters, which is.
40:51
One, one and a half hour
40:51
when we, when we speak.
40:54
So it was very useful for
40:56
this use case. In your experience, like, because you
40:57
mentioned like LinkedIn post or, uh, using
41:03
the gut prompt and, and these kind of
41:03
things, how does it the, because yeah, it's
41:08
like quantity is not everything, right? So Yeah.
41:10
Obviously it's good that you can fit. More information into, uh, include
41:12
100 K, but what is the, um, quality
41:16
wise, what's the difference between
41:16
GPT four and, and and include 100 K?
41:21
Okay. I will go into the quality
41:21
on different, uh, uh, parts.
41:25
First, uh, when I wanted to generate LinkedIn
41:25
post chat, GPT is very good with zero shot
41:31
prompts, so I just said, please write me a
41:31
LinkedIn post based on this short summary.
41:36
Yeah. On the other hand, it was harder to get a
41:37
good LinkedIn post with zero shot prompt
41:42
from Claude and even with a few shot prompts.
41:46
So I provided two examples, uh, to
41:46
Claude and asked to please write a
41:51
similar, and it was very, it was not
41:51
that strict on following my instructions.
41:58
Yeah, but what, what, what kind
41:58
of, like, if you had to put a number to
42:01
it, so on a scale from one to 10, 10 is
42:01
the perfect, like a master full LinkedIn
42:05
post, one is like basically garbage.
42:08
What is GT four compared to
42:08
closed 100 K, I would say.
42:11
Okay.
42:11
I, I would say other,
42:11
from other perspective.
42:14
So at least I, I think Claude needs
42:14
two or three more, uh, iterations of
42:20
prompts and even the end product is not
42:20
that good that I'm satisfied at all.
42:26
Maybe I would
42:26
say 60 to 60
42:27
to 70%.
42:29
Okay. So, so 60 to seven is the maximum you
42:29
can get out from Claude and GPT four
42:34
is like, what's the number for that?
42:37
Hmm, 9 19
42:39
95. Okay. So it's almost like perfect, right?
42:42
Okay. That's, that's, that's can, that's can, yeah. That, that can be okay.
42:44
This is our baseline.
42:45
It is, it's not perfect as a good
42:45
copywriter would write that, but Yeah.
42:50
You know, t defined our baseline and
42:53
for now, yeah, sure. So it's kind of like, and you, you, I, I
42:54
guess you can even chain things, right?
42:58
So it's like you can do the pre-processing
42:58
on large documents with cloud, right?
43:03
Get a draft, like, which is cohesive,
43:03
which is not perfect statistically, right?
43:07
So it's not something which you are
43:07
satisfied with as a master marketer, but at
43:12
least it's like, it's, it's more cohesive. Or you, you could feed everything
43:14
into one, one row, basically,
43:18
and then you can basically use
43:18
GPT four to fine tune it, right?
43:21
So, because then it's already condensed, it's
43:21
already short, it's already selected, and
43:25
once you have like the gist of the LinkedIn
43:25
posts, which you can generate, then you can
43:30
go one by one and fix the up GTP GPD four.
43:33
Does it make sense?
43:34
Yeah, I think you are right. Amfa Advantage comes when
43:36
you can combine different ai.
43:39
So for example, when you can, uh, combine
43:39
ChatGPT and Mid Journey, and when you can
43:43
combine cloud and ChatGPT, this is where,
43:43
uh, you can create some Word class, uh,
43:50
texts. Yeah, this is extremely important.
43:53
What you just, uh, covered
43:53
that if you know the upside and
43:57
downside of these tools, right? Then you can get the most out of them.
44:02
And it's, my, my biggest problem
44:02
is that, uh, people are too
44:07
obsessed with one single solution.
44:09
Is one single model is one single
44:09
prompt, one single whatnot, right?
44:13
And. It's, it's, it doesn't make too much sense
44:14
because if you, you understand your tools
44:19
and you don't just have a hammer, but
44:19
you have a tool set of different tools
44:22
and you understand when to use them. You can get world class results.
44:26
And so the end goal, at least
44:26
in our case, is not theoretical.
44:30
It's very practical. We want to get sheet done right?
44:33
We want to make money, we want to
44:33
save time, we want to scale things.
44:37
So we are dealing with business
44:37
of ai, not the intellectual of ai.
44:42
So in this case, uh, if you understand
44:42
these tools, and that's why we use them,
44:47
we can, we can get more value out of them.
44:49
And that's what I urge the listeners to do
44:49
as well, that please don't think that one,
44:56
just one tool is solved all your problems.
44:58
It, it shouldn't, there are so many tools
44:58
and that's why we are covering them.
45:01
That's why we are doing these deep
45:01
dives that, you know, the good, the
45:04
bad, and the ugly of each tools,
45:04
which we have, and you can use.
45:09
Okay,
45:09
next one. Naming a podcast. So this is a short text and a creative one,
45:11
and I think that this is a huge difference.
45:17
I started with, uh, instructions,
45:17
several instructions, uh,
45:21
and also a few short prompts. And Che GPT was way much
45:23
better from the get-go.
45:27
It was, it used alliteration without
45:27
asking them it, the names were more
45:31
catchy, but also Claude was very close.
45:34
So, for example, for our podcast, it,
45:34
it mentioned disruptors Daily, which is
45:40
something ti mentioned in the past as well.
45:43
But still, I feel, and, and now
45:43
it's, it's just a subjective feeling
45:46
that Claude wasn't that creative.
45:48
I have an interesting experience. I'm gonna share it with you.
45:50
Okay. Okay. So it's like two minutes. Good. Viktor is diverging for two minutes.
45:55
So again, yeah, again. So in, in 2021, end of 2021, GPT
45:57
three was released and I was playing
46:02
with it and I tried to copy my coach.
46:04
So I basically fed into like example
46:04
question answers, five of them.
46:09
And then I asked my question and it provided
46:09
meaningful answer, helping how to ease the
46:15
pain of my, uh, baby who has stomach ache.
46:18
So that was the use case,
46:18
and it worked wonderfully.
46:21
It even could copy without, like,
46:21
without any in specific instruction.
46:27
It could copy like, uh, double
46:27
spaces at the end of each sentence.
46:31
So it was nine out of 10, I would say.
46:34
So it was almost perfect. It was really good.
46:37
And then two years pass by, and
46:37
all these models get fine tuned.
46:43
On what is a good answer?
46:46
So if you're asking for an answer, what,
46:46
what, what is the human feedback on that?
46:50
What, what makes an answer? An answer? A good answer.
46:52
So giving advice, hedging the
46:52
advice and those kind of things.
46:56
Like, okay, yeah, obviously this
46:56
is why your tummy can a, uh,
47:00
hurt and obviously see a doctor.
47:02
So that's kind of like the structure of it. And what happened is, like two years
47:04
ago, these models, all the models
47:08
like GPT three to all the, all the
47:08
models, they were like a S U V.
47:13
They, they, you could like
47:13
off-road with them, right?
47:16
And in the last two years they
47:16
worked a lot on like putting, uh, the
47:21
whole, all the models on the rail. So basically just keep them on the rail.
47:26
What is meaningful? It's gonna be faster, it's gonna safer,
47:27
generally it's gonna be better, but now
47:32
it's extremely hard to instruct them
47:32
to follow the creative task, which is
47:37
diverging from the, the common path.
47:39
So in this case, in a coaching case,
47:39
answer shouldn't be, uh, giving ad advice.
47:46
It, it should be like, uh,
47:46
introspective, it should be more
47:50
question based and so on and so on. And now I have to be extremely.
47:54
Explicit about like, okay, please write
47:54
double spaces after each sentence.
48:00
Please ask follow up questions, please. And I have to define everything
48:01
and it's a pain in the ass.
48:05
And the best I can get is seven out of 10. And even though these models are
48:07
in general much better, right?
48:11
It is, it's much, much more valuable.
48:13
But now they are very narrow. So in a sense they get le get less
48:16
curative because now answering has so much
48:21
bias and I guess with LinkedIn as well. So, uh, I guess if you just ask for
48:23
two years ago, if you provide LinkedIn,
48:30
uh, posts they, which means that you
48:30
just provide like five good examples.
48:34
So these are five good examples
48:34
of, uh, LinkedIn posts.
48:38
And please write me a new one. Based on like this information two years
48:40
ago, I guess it worked much better because
48:46
it has le had less bias about like what
48:46
is the good, what is like a mad median?
48:51
What is the median of the
48:51
LinkedIn posts, right?
48:54
Because now if you ask LinkedIn post, it
48:54
has very much, it's has a lot, lot of bias.
48:58
If you ask for an answer,
48:58
it's very biased, right?
49:01
So it has a lot of bias about like the
49:01
different modalities of the answers.
49:06
And that's kind of like the tradeoff,
49:06
which happened the last two years.
49:08
And that's got, that's something I just wanted to share. Uh, so if you're frustrated about
49:10
that, uh, my current solution is be
49:15
extremely explicit about what you need.
49:18
And the other hack as well is just try to
49:18
stay away, away from the biased expression.
49:26
So in this case, try not to
49:26
say even, uh, LinkedIn post.
49:29
So these are good text
49:29
examples I need, right?
49:34
So these are good examples. Write me a good, under a good
49:35
one without mentioning that, that
49:39
LinkedIn, the same for Twitter. So not, you are not saying like, I need a
49:40
tweet because it's very biased, but like
49:44
what is a me median tv, which is like, I
49:44
want, don't want to say shitty, but in a
49:50
sense it's shitty that it's if you have a
49:50
specific goal that in that sense it's shitty.
49:55
So then just like write like, okay, these
49:55
are good text I, I need, and please provide
50:01
me one more, uh, based on this criteria.
50:04
Okay. Let's go back to comparing
50:04
Chat, g p, t and cloud.
50:08
So next one is generating prompts for
50:08
mid journey to generate AI images.
50:13
Yeah. And what do you think, Viktor? Which would be the better GT or Claude?
50:17
Or maybe the final verdict will be something
50:21
else. I, I think it's GPT four cuz
50:22
it's kind of like have stronger
50:26
reasoning CAPA capabilities. So I, my my, my guess is that the 100
50:28
K contact spin though is not the same
50:34
as the include plus, which is kind
50:34
of like in comparison with GPT four.
50:38
So the analogy is GPT
50:38
four and include plus and.
50:43
GT Instant and GT 100 k k with Chat GT 3.5.
50:48
So I, I guess maybe comparing, uh,
50:48
the hundred K code to GT four is
50:53
not fair because it's almost like
50:53
comparing 3.5 Chat GT to to G P four.
50:59
Okay. Viktor, I think the final
50:59
answer, we will surprise you.
51:02
So I will sh uh, I will
51:02
share you a few images.
51:05
If you are, uh, listening
51:05
to this podcast, I will.
51:09
Uh, describe what we are, what we can see
51:09
on the screen, but also you can click on
51:13
this chapter digest, uh, that come and
51:13
quickly, uh, find this part in the video.
51:18
So first I asked, um, uh, mid journey
51:18
to generate Greek kitchens, and I just,
51:23
just put image in Greek, nothing else.
51:26
And what we can see on the picture, there are
51:26
four images of Greek kitchens with, uh, sea
51:32
in the background, and it's an inside view.
51:34
And, uh, again, another four pictures.
51:37
These are very similar. They are not realistic,
51:38
but they are not painting.
51:41
So they are in between somewhere. And I did it four times with the
51:43
very same prompt, just Greek kitchen.
51:47
And the resorts were very similar.
51:50
Again, I, I, I'm showing four
51:50
pictures of Greek kitchens.
51:53
Blue, not realistic, not painting,
51:53
and see in the background.
51:58
So basically the same. And, and one more of this.
52:01
And now I asked, uh, uh, clothes with the.
52:06
With a previously mentioned prompt
52:06
to, to generate an elaborative
52:10
prompt, uh, for this Greek kitchen. So for example, imagine, uh, uh, imagine
52:13
a blue kitchen painted on canvas,
52:20
uh, from, from a certain perspective.
52:22
And so, and now we get again.
52:26
For, for one fan of it iteration,
52:26
the images are very similar, uh, com
52:31
uh, compared to the previous one,
52:31
but now it's fully a painted one.
52:35
So now the style is different. So why is it interesting?
52:38
Because now we have a different
52:38
style, so it's now we, we can test
52:42
something, for example, think about
52:42
if we want to create Facebook ads.
52:45
Now this is something different, but, uh,
52:45
here comes the next one and, uh, this is,
52:51
uh, what, uh, again, clothes come up a with,
52:51
and it said, um, create a, create a picture
52:57
of four, uh, culturally diverse women, uh,
52:57
who are having fun in, um, In a cooking
53:05
class and now it, it was something different.
53:09
And, and I'm glad that we live in a
53:09
time when Mid Journey 5.1 is existing
53:14
because these, the faces that we can
53:14
see are almost perfect, realistic.
53:20
And when I, uh, when I tried to create
53:20
faces in Mid Journey V three, it was very
53:25
bad. Yeah. And the, the neat thing what's
53:26
happening here is like you used.
53:31
Kind of like chaining as well. So you either a gpt four or CLO
53:32
to generate prompts for you,
53:36
which you fed into mid journey. Right? And, uh, yes, I guess this talks
53:38
to the strength of mid journey.
53:42
I just want to focus on Claude and Gpt. Just one thing that now we, we came
53:44
up something very, uh, different than
53:48
before, which is, uh, uh, three women
53:48
having fun and their, uh, hair is blown.
53:55
Uh, so it's like a very dynamic image.
53:57
And also the next one cloth come
53:59
up with this. Okay. Just to, just to those who are listening
54:01
here, it's like what you see on the screen.
54:05
Previously it was like the same,
54:05
basically the same kitchen.
54:09
It's like very Greek style with a
54:09
scene in the background, very bluish.
54:13
Uh, it's a blue dominated scenes, right?
54:16
And then what Claude did is like created a
54:16
dynamic scene of people enjoying themselves.
54:21
And then it also came up with. Even more dynamism.
54:25
So not just smiling, but also
54:25
the hair is thrust in the air.
54:29
So the, the, the shot is like,
54:29
it would cost a lot of money.
54:33
And I, I imagine that someone is, if
54:33
someone had to shoot this shot, it would
54:39
take a lot of tries, a lot of time.
54:42
And, uh, it's insane that it's possible
54:42
now to, to have this dynamic setting.
54:46
And it all was generated with closed.
54:48
It's insane.
54:51
Yeah. Hiring the talent, booking the
54:51
studio, uh, having studio lights, uh,
54:55
multiple, multiple people on the crew.
54:58
And of course the photographer, I
54:58
guess at least five or 10 K in the
55:02
minimum, maybe up to 2030 k uh, a
55:02
day for creating photos like this.
55:08
Yeah. Okay. So the next one.
55:11
Okay, this is something CLA camera feed.
55:13
These are very, Plain and boring photos of
55:13
bread and a bucket of flowers on the table,
55:21
which is not good for the purpose, but just
55:21
want to show you that, that now, now cry.
55:27
We started to diverge. Yeah. And, uh, okay, I did, uh, but did the
55:29
very same with Cha GPT and now it came
55:34
up feed, uh, uh, this picture, which
55:34
is olive oils and tomatoes on a plate.
55:41
And, uh, and some people are picking them.
55:44
So again, it's a very different approach.
55:46
On the next page, we can see Greek kitchens,
55:46
but their style is very different because
55:52
it's more of, uh, orange and red, um, type.
55:56
And because J PTI was focusing on
55:56
the ground, which is terracotta,
56:01
which is a classic Mediterranean,
56:01
uh, Flooring building material.
56:06
Yes. So now if you, maybe if, if, if you're
56:07
an interior designer and you want to
56:12
see different creative di directions
56:12
of, of a kitchen, of a scene for a
56:18
film or, or something like that, then
56:18
now it's in different perspective.
56:22
And the last two, one again, uh,
56:22
similar problem, but it's, uh, what
56:27
you can see is a modern Greek kitchen,
56:27
uh, with more of a modern style.
56:31
And it's now, it's, it's very photorealistic. And, and the last one is
56:34
again, a blue kitchen.
56:36
Uh, but it's again focusing because chichi
56:36
pt, um, included something about the floor.
56:43
The picture is focusing on the floor.
56:46
Uh, so, and, and it's photorealistic
56:46
and still blue, but it's very
56:50
different from the previous one.
56:52
Yeah. That's just to make, make clear,
56:52
make clear the whole process.
56:55
For the listeners, you had a prompt and
56:55
the pro, the original prompt was what?
57:02
Greek kitchen. Okay. So that was the original
57:03
prompt, and then Yes.
57:06
You kind of like had two different flows.
57:09
You used clothes, you also
57:09
used Chat gp, uh, GPT four.
57:13
Right. And what, what was the prompt there?
57:16
So how did, and how did you
57:16
explore Curly and moved away from,
57:21
uh, just having a Greek kitchen?
57:24
So we have a prompt, which is created
57:24
by Samsung Mobiles, and this is a long
57:28
prompt, uh, that starts with like this, Ivan
57:28
to Ivan you to act as a prompt engineer.
57:33
You will help me write prompts for
57:33
an AI generator called Mid Journey.
57:37
And later it'll be data that
57:37
please define a camera lens.
57:41
Please define a perspective,
57:41
uh, color, color, style, color,
57:46
references, position elements. And so, so basically you
57:48
outsource the creative process
57:52
to the language models like G P
57:54
T and Cloud. Yeah, so, so actually that this prompt,
57:56
uh, was used, uh, to kind of like
58:01
generate like from Greek kitchen generate
58:01
something which is more crowd, right?
58:06
And only this prompt was used. Yes.
58:09
Yeah. This can be sh uh, shared in the show
58:09
notes actually, um, uh, is part of also
58:14
the God prompt plugin, which we have.
58:16
So if you have got prompt plugin in
58:16
ChatGPT, so if you are a subscriber,
58:21
you can use it and what's happening, it
58:21
realizes that you want to generate an
58:24
image and it's kind of like instructing
58:24
ChatGPT to use this prompt to generate
58:30
better, uh, prompts for mid journey.
58:35
Okay, so my impression about Cloud
58:35
and ChatGPT for these use cases that,
58:40
uh, you can get more if you combine them.
58:43
So you, so you make the same process on
58:43
both and you feed me journey, uh, the
58:49
prompt, the, the final prompts that was
58:49
generated by Chat g, PT and by cloud.
58:54
And in my case, uh, my goal is to
58:54
create very different images so I
58:59
can test them on Facebook with, which
58:59
will get a higher clickthrough rate.
59:04
So I think combining them
59:04
is the unfair advantage here
59:07
again. Yeah, that, that, that, that sounds awesome
59:08
because once again, I just want to highlight.
59:11
Highlight is that you have to have
59:11
zero background basically here
59:15
because you just had Greek kitchen. You copy pasted this prompt, but you
59:18
don't even have to copy paste because it
59:21
should just use the gut prompt plugin. Then you just like write Greek kitchen
59:24
and it's generating for you like
59:28
variations for mid journey and you can
59:28
copy the same prompt and using clothes.
59:33
And it, it's kind of like using a
59:33
different, uh, creative process, kind
59:37
of like using another creative person.
59:39
And once again, you get different results.
59:43
Awesome. So in, in my mind, those pictures
59:44
was 10 out of 10, each of them.
59:48
So they were insane. And you get insane results
59:50
without having zero background.
59:54
And only you, only thing you know is
59:54
like you want to have a Greek kitchen.
59:57
So that's insane. That's, that's, that's, I think if you
59:58
understand this, you are much further
1:00:02
ahead than anyone in your, in your field.
1:00:05
Right.
1:00:06
Okay. Just to final, just to finish this
1:00:07
comparison, just a few technical things
1:00:12
that, uh, I think, uh, cloud, uh, is a little
1:00:12
bit faster so you will get answers a little
1:00:17
bit faster, but the general, the speed is
1:00:17
the same as GPT four and uh, and cloud.
1:00:23
Okay. So my impression was that CLO Plus is
1:00:23
kind of like the same as GPT four in both
1:00:28
terms of, uh, quality is a little bit
1:00:28
lagging behind, but kind of like the same.
1:00:33
But on speed wise it's the same as GPT
1:00:33
four and CLO instant and instant 100 K.
1:00:40
Those are similar in
1:00:40
speed and quality to 3.5.
1:00:46
So yeah, you in, in the instant
1:00:46
100 k, uh, model for you can
1:00:51
fit in 100 K tokens, but it's.
1:00:54
It has lesser abilities than the
1:00:54
closed plus model that's similarly
1:00:58
like GPT four and GT 3.5 Turbo.
1:01:01
It's like, yeah, turbo is quicker,
1:01:01
it's cheaper, but, uh, I mean it's
1:01:06
much less capable on logical thinking
1:01:06
than, than the GPT four model.
1:01:11
Okay. Besides these use cases, I also tried
1:01:11
writing poems and fiction, but the,
1:01:16
but the overall impression is the same.
1:01:19
So first one, Claude is not
1:01:19
following instructions closely.
1:01:23
And one more thing,
1:01:25
Viktor. Sure. So it's, it's kind like the takeaway.
1:01:27
So if you just like want to try to
1:01:27
summarize what is cloud 100 K good for?
1:01:33
The name is even clue instant 100 K.
1:01:36
So that's kind of like the equivalent
1:01:36
of ChatGPT 3.5 turbo, but with
1:01:40
a bigger, uh, contact window. So the similarly, if you use the gut
1:01:42
prompt or try to use the gut prompt
1:01:46
into GPT 3.5, it won't follow it, it
1:01:46
won't be as good with complex logical
1:01:52
reasoning and instruction following.
1:01:55
So that's the same for code 100 K as well.
1:01:57
So it's not that good, uh,
1:01:57
with gut prompt, uh, thing.
1:02:01
And it's not that good with a fiction
1:02:01
writing or poem writing because
1:02:05
complexity, when complexity increases.
1:02:08
You should use or aim for better or
1:02:08
bigger models like Code Plus or GPT four.
1:02:13
So that's kinda like the big, big, big takeaway. Okay, so we
1:02:16
talked about tropics products.
1:02:18
Let's have a big picture view of the
1:02:20
company. Yeah, sure. Let me share my screen.
1:02:23
I'm also going to describe for
1:02:23
those who, who are just listening.
1:02:27
So what you see is the business
1:02:27
model canvas for Anthropic.com.
1:02:31
As as we discussed at beginning,
1:02:31
it was founded by AI researchers.
1:02:36
So if you look at the key activities,
1:02:36
r and d is very important.
1:02:40
In the first one and a half years actually,
1:02:40
they only did research and development
1:02:45
and then they started to commercialize. And since then the other very important,
1:02:47
uh, key activities, product market fit
1:02:52
search in all the different industries. And if you look at the key resources they
1:02:54
have, it's obviously 1.5 billion funding.
1:02:59
And uh, also the their 100 K cloud
1:02:59
model, but also X open AI people.
1:03:06
And, uh, their CEO is VP was
1:03:06
VP of Research, uh, at Open ai.
1:03:10
So, uh, these people have all the
1:03:10
knowledge, all the knowhow, all
1:03:16
the proper, uh, uh, frameworks to
1:03:16
create something which is useful.
1:03:20
And I guess the close one and the case is,
1:03:20
is a telltale of that, that they are the
1:03:25
ones, uh, the only ones which, uh, who, who
1:03:25
provide this, this kind of model already.
1:03:30
And, uh, the interesting thing is, uh, they
1:03:30
have key partners like Google, and Google was
1:03:36
investing, but also they are partnering with
1:03:36
Google to provide compute because Google is
1:03:41
Google with providing cloud services, right?
1:03:43
But they also partnered
1:03:43
with Zoom, for example.
1:03:46
So Zoom Ventures invested money, uh,
1:03:46
just a few, few, uh, months back.
1:03:52
And Zoom is good actually
1:03:52
for CS as well, right?
1:03:56
Actually, zoom has a, a solution for
1:03:56
providing communication for customer
1:04:01
service agents, and they are going to
1:04:01
integrate, uh, Anthropic models to, in
1:04:09
one way help the users to get answers
1:04:09
more relevant, faster, more personalized
1:04:14
answers, and also have the customer
1:04:14
service agents to be more productive.
1:04:18
So that's, that's neat. And also put it there.
1:04:21
I'm, I'm not sure about that. I just put it there that, uh, for
1:04:22
example, a company like scale.com,
1:04:25
which is aggregating the different
1:04:25
AI providers and providing services
1:04:29
to enterprise is something, something
1:04:29
where they can provide value as well.
1:04:33
Like, okay, here's our model and you can
1:04:33
sell our model if someone is in need.
1:04:38
So, uh, key proposition.
1:04:41
So what do they provide?
1:04:44
On one hand, they provide safe models.
1:04:47
They provide 100 K pages
1:04:47
of, uh, 100 K tokens.
1:04:51
So basically 120 pages of text
1:04:51
and clue the next what they're
1:04:57
working on is 10 times better. It's going to be 10 times
1:04:59
better than GPT four.
1:05:01
So that's kind of like, what is the key value
1:05:01
proposition and who are they providing this?
1:05:07
So who are the customer segments is
1:05:07
education, entertainment, government
1:05:11
intelligence, community, legal entrepreneurs.
1:05:13
So basically, as I said, they're
1:05:13
trying to fund find product market fit
1:05:18
and what kind of channels they use. They use Google Cloud.
1:05:21
As I said, they use Zoom contract,
1:05:21
uh, contact center, which is kind of
1:05:25
like the solution Zoom is providing
1:05:25
for customer service agents.
1:05:29
Uh, and mainly they communicate currently
1:05:29
to API documents, uh, their documents.
1:05:34
So how there's actually quite good
1:05:34
documentation about the, how their, uh,
1:05:38
API works and on, on a wide sticky notes,
1:05:38
I, I provide additional, uh, channels
1:05:45
I think they should use or could use or
1:05:45
will use in the future most probably.
1:05:49
So for example, One of the big problems
1:05:49
is finding for them as well, finding
1:05:54
product market fit, but it's also
1:05:54
the problem for the clients as well.
1:05:57
So they, let's imagine you are, you are
1:05:57
having an enterprise company and you
1:06:01
don't know how to integrate something. So having a playground examples, right?
1:06:06
Or having a quiz where you answer questions
1:06:06
about like, okay, what you are building.
1:06:12
Then you can get back examples of how
1:06:12
others are using the tools, right?
1:06:17
So helping people to figure out
1:06:17
how to, and what to integrate into,
1:06:21
into their product, an AI tool. Could be extremely valuable.
1:06:27
And also the how side. So how, uh, Anthropic is helping developers
1:06:29
could be YouTube education as well, and
1:06:36
integrating with tools, AI tools like
1:06:36
long chain and then the similar tools,
1:06:41
and also into no code tools as well. So helping no NOCO tools providers
1:06:43
to integrate, easily integrate,
1:06:48
uh, the aro, uh, model features.
1:06:51
So that's kind of like the big overview.
1:06:54
And also on the cost side, there's a fixed
1:06:54
cost of wages and a variable cost of compute.
1:07:00
And that's what we discussed already. That can be extremely.
1:07:04
Uh, expensive. And on the revenue side,
1:07:04
it's very easy pay per use.
1:07:08
So as much as you use, you pay for that.
1:07:11
And it's quite similarly price, uh, price
1:07:11
to, to open AI and what they don't have, or
1:07:17
at least I, I, I wasn't reading about that.
1:07:19
Maybe they have it, but most probably
1:07:19
they're gonna have having the future.
1:07:22
So that's why I put it in invite, sticky note
1:07:22
is the foundry dedicated privacy instances.
1:07:29
So just, uh, I, I just talked to, I just
1:07:29
talked to, uh, a lawyer who's working
1:07:35
for international law, uh, law firm.
1:07:37
And for example, they paid, I, I guess this,
1:07:37
the pricing is starting at $800,000 per year.
1:07:45
And then you can get dedicated
1:07:45
instances, so in open AI case, right?
1:07:51
And why is it good? Because obviously it's privacy pre,
1:07:52
uh, preserving, they can use clients
1:07:56
data and so on and so on, which is. Paramount and, and, and a
1:07:58
must have for a legal company.
1:08:01
And the, and this is what, uh,
1:08:01
open air is calling Foundry.
1:08:05
And this is the, the gist of it is
1:08:05
basically dedicated privacy instances.
1:08:09
So if they don't provide it, they will
1:08:09
provide it most probably, or they should
1:08:13
provide it because these big companies,
1:08:13
uh, can easily pay, uh, a million dollar
1:08:20
a year because it's, it's nothing compared
1:08:20
to the additional value it can bring or the
1:08:24
additional capabilities it can, uh, unlock.
1:08:27
So this is kind of like the big
1:08:27
overview of what is the strategic, uh,
1:08:31
business model canvas for the company. Great.
1:08:35
Thank you, Viktor. Yeah.
1:08:37
So, okay. Uh, let's move on.
1:08:39
So let, let, let, let's
1:08:39
get to the next, uh, part.
1:08:42
Like, okay. What, because we already discussed
1:08:43
that they're looking for product
1:08:46
market fit, and I guess the listeners
1:08:46
want to know as well, like, okay.
1:08:50
What can I use it now for?
1:08:53
Right? And we already, uh, covered like, uh,
1:08:54
long form content generation editing and
1:09:01
translation, so that just like generating
1:09:01
a whole book or edit a long manuscript
1:09:07
or, uh, coherent translating whole books,
1:09:07
training materials for enterprise as well.
1:09:13
So what's, what's, we already discussed
1:09:13
that, uh, if you chunk up a big piece and
1:09:18
you process it, uh, by one by one, you
1:09:18
kind of, kind of lose the coherence between
1:09:24
the pieces and it, it's already working.
1:09:27
You can already try it. It's quite good, but
1:09:28
obviously it has limitations.
1:09:31
Like, uh, this 100 K model is based
1:09:31
on the instant, uh, model class, which
1:09:38
is fast, but less capable cognitively.
1:09:41
Uh, it ha it can also be used for
1:09:41
extended conversation, role playing.
1:09:46
Uh, it's like almost like, uh, With
1:09:50
just one more thing. What do you think is it possible for, uh,
1:09:51
Claude that we feed the most important, most
1:09:57
interesting, wild engaging parts of our first
1:09:57
eight episodes and it'll write a book for us?
1:10:06
I don't know. We have page book based v we have to try.
1:10:09
I, I'm, I mean, the big, big, uh,
1:10:09
limitation currently is that we don't
1:10:13
have API access only through pool.com.
1:10:17
And it's neat that they have 100 K context,
1:10:17
but kind of like the, the answer is limited.
1:10:23
So if we ask for like, the whole book, I
1:10:23
guess it won't fill out the whole 100,000.
1:10:29
Token limits, so we should, so, so, so my,
1:10:29
my currently our, our real limitation in,
1:10:35
into the test is, is just like, we don't
1:10:35
have API access, but as soon as we gonna
1:10:39
have API access, definitely that's cha
1:10:39
something we, we gonna try to just like,
1:10:44
try to generate as much stress as possible
1:10:44
and see what the output is gonna look like.
1:10:51
Okay. So yeah, it's like what you also discussed,
1:10:51
analyzing entire research papers, legal
1:10:57
documents, summarizing the main points,
1:10:57
extracting key details, or even in the
1:11:03
coding, uh, scenario, refactoring a large
1:11:03
pieces of software code, adding comments
1:11:08
and documentation, writing unit tests.
1:11:11
And I think that since cognitively
1:11:11
it's limited compared to the code plus
1:11:17
model or in a sense to GPT four, I
1:11:17
think like generating, automatically
1:11:23
generating documentation from code.
1:11:26
Would be better suited
1:11:26
than writing code itself.
1:11:29
So, so co code writing is quite complex
1:11:29
compared to just like writing documentation
1:11:34
about like, okay, what, what's happening? What is this code a piece of code is doing?
1:11:38
Because it's, it's quite sequential. It's quite like there is a
1:11:40
code, it has to be explained.
1:11:42
So, uh, writing automatic documentation
1:11:42
is something, uh, which is possible now
1:11:49
for big code base, uh, code basis as well.
1:11:52
Because if you think about that,
1:11:52
there's a big code base, uh, you want
1:11:56
to onboard a new developer, right?
1:11:58
And someone has to write the,
1:11:58
the documentation, right?
1:12:02
And it can be outsourced
1:12:02
to, to these kind of models.
1:12:05
And if you commit to a change, then you
1:12:05
can update that documentation as well.
1:12:09
So you can have an up-to-date documentation,
1:12:09
which is kind of like unheard of.
1:12:13
So it's, it's almost
1:12:13
impossible to do otherwise.
1:12:16
And now it's, it's possible. So that's something I'm,
1:12:17
I'm extremely excited about.
1:12:21
And also the last, last, uh, part
1:12:21
is, um, educational applications.
1:12:26
So that's kind of like the same
1:12:26
as we discussed with the book one
1:12:30
that we have to have a p I access.
1:12:32
But if you have API access possible,
1:12:32
it's, it's, it's gonna be possible to
1:12:36
generate entire course curriculums.
1:12:39
Uh, and, and the neat thing is
1:12:39
like we can feed into so much data.
1:12:44
So it's like we can feed into, uh, textbooks,
1:12:44
we can summary of textbooks, summary of,
1:12:49
uh, guidelines, summaries of, I don't know.
1:12:51
So we can feed in lots of information,
1:12:51
uh, and even relevant context as well.
1:12:57
Like, okay, who are the, uh,
1:12:57
students we are generating this for?
1:13:00
So this can be generated
1:13:00
even student by student.
1:13:04
Basis so you can get your own specific
1:13:04
curriculum and someone else is
1:13:09
getting their specific curriculum. And since it can have all the, uh,
1:13:11
long contacts, that's something
1:13:15
if, which I'm excited about.
1:13:18
And I guess even if it's like, not
1:13:18
right now, but in a year or maximum two,
1:13:23
it's gonna be extremely good quality. So even though now it's like limited
1:13:25
with this instant model of, of like
1:13:31
the, similar to Chat G, pt, 3.5 Turbo,
1:13:31
uh, the same thing with CLO instant.
1:13:36
Even though it's limited
1:13:36
cognitively it's going to be better.
1:13:39
And, and, uh, and this is something
1:13:39
I'm, I think it's, it's quite strong.
1:13:44
Okay, what kind of businesses
1:13:44
can these structures build on top of
1:13:47
cloud? Okay, so let's, that, that's one
1:13:48
of the favorite parts of mine to
1:13:52
discuss business ideas, right? And give you, give you some food.
1:13:56
Uh, food for thought. So what's the biggest problem
1:13:57
with business ideas only?
1:14:01
So let's say, give you an example. My big business idea is CV generator,
1:14:04
which is helping job seekers to
1:14:08
generate first offs, uh, of longer
1:14:08
more compelling resumes, which are
1:14:12
tailored to the specific job opening. Right.
1:14:15
So that's kinda like the, like the idea,
1:14:15
but what's, what's the big problem?
1:14:18
Well, what's, what, what is my biggest
1:14:18
problem is that as a developer or
1:14:23
entrepreneur, do you know what you should
1:14:23
build or, uh, do you know what exactly will
1:14:28
your product be delightful or can you use
1:14:28
this description as a, as a compass age day?
1:14:35
No, obviously not, because it's quite vague. It's just like, uh, okay, this is
1:14:36
the idea of, of, of, of CV generator.
1:14:42
But that's, that's my test model.
1:14:45
Uh, I became up with, with test model,
1:14:45
uh, it's actually solving this problem
1:14:52
because with test model, for example, in
1:14:52
this case, it's, uh, a testimonial which.
1:14:57
Uh, actually following four different things.
1:15:01
First, it's AI generated. Why?
1:15:03
Because it's more relevant
1:15:03
and more personalized.
1:15:06
Second, it's a short testimonial.
1:15:08
Why? Because it's easy to relate to. And also three, it's a vividly illustrates
1:15:10
the pain points being solved so people
1:15:15
can understand what you're building for. And fourth, uh, it makes
1:15:17
the benefits tangible.
1:15:21
So you know, what is the exact results
1:15:21
people should get from something.
1:15:25
So in this case, like, let's say
1:15:25
the first was like, uh, the idea
1:15:28
was that like just CV generator.
1:15:31
Let's compare it to a testing model. In this case, uh, a test model could
1:15:33
be I'm a computer science graduate
1:15:38
and was struggling to create resume
1:15:38
that stood out to tech companies.
1:15:42
Uh, this tool took my basic information
1:15:42
and transformed it into a detailed
1:15:48
company resume that highlighted my
1:15:48
coding project and relevant coursework.
1:15:53
I started getting callbacks for
1:15:53
interviews almost immediately.
1:15:57
So if just like get this short, uh,
1:15:57
testimonial, testimonial, you instantly
1:16:01
know what you're building for, right? You want to get like basic information.
1:16:05
The end goal, the KPI you are
1:16:05
measuring is whether people are
1:16:09
actually getting, uh, uh, feedback
1:16:09
for the companies and so on and so on.
1:16:13
So this testimony is for, so it's like similar to
1:16:15
getting the pains, the gains,
1:16:15
and the jobs of customers, but in a
1:16:19
more easily understandable way, right?
1:16:21
So it's like a testimonial,
1:16:21
which a human could say.
1:16:26
Uh, but if you are reading it as a
1:16:26
developer, it's more easy to understand
1:16:30
what is the target audience needs
1:16:32
and what do, what do they like, right? Yeah.
1:16:34
And, and if I want to be meta, so I
1:16:34
want to provide a test model on test
1:16:37
model, then here is the hard truth.
1:16:41
Bullshit. Personalized doesn't sell.
1:16:45
Real stories do. Test model transformed abstract benefits
1:16:46
into tangible testimonials about,
1:16:52
about, about solving real pain points.
1:16:54
It helps investors to see value,
1:16:54
customers feel understood and
1:16:59
gives the team a meaningful goal. Test model is our daily reality
1:17:01
check and our best CI speech.
1:17:06
So that's kind of like the,
1:17:06
uh, matter of test mode.
1:17:09
Of test mode that, yeah,
1:17:09
you understand that it's.
1:17:11
A compass for everyone. It's easier to convey, and that's
1:17:13
something you can get out each day and
1:17:18
it can guide you each day to provide
1:17:18
value and, and provide, uh, delight.
1:17:23
So this CV generator idea, it's obviously,
1:17:23
it's bad, this tech job resume writer, it's
1:17:29
tangible, but it can also be academic cv.
1:17:32
So all your research and abstract
1:17:32
can be fed into it, right?
1:17:37
It, it wasn't possible before
1:17:37
because the context was limited.
1:17:40
Or it can be an executive
1:17:40
level resume creator.
1:17:43
So if you're seasuite, you want to have
1:17:43
an up-to-date profile, but you don't have
1:17:47
time, obviously you can shield out a few,
1:17:47
few hundred dollars to have everything
1:17:53
you did fed into it and extremely
1:17:53
personalized up to date, strong cv, right?
1:18:00
So, uh, that's something. If you are intrigued about, you can
1:18:02
go out and, and, and, and build this
1:18:06
tool, but let's move on another idea.
1:18:08
Social media management service.
1:18:11
So let's give you a testimonial, a quick one.
1:18:14
I own a small Italian restaurant
1:18:14
in the heart of the bustling city.
1:18:18
We have amazing food, but I struggle
1:18:18
to get word out on social media.
1:18:22
Then I found this service. They started creating posts that
1:18:24
captured the magic of our meals
1:18:27
and the atmosphere of our place. Suddenly we started getting more likes,
1:18:30
shares, and most importantly customers.
1:18:35
So that's something you can go out
1:18:35
and, and create, uh, social media
1:18:39
management for, uh, local restaurants.
1:18:41
How is it better than GPT? Uh, why is it or GT four as an api?
1:18:46
Why, why cloud would be better in this case?
1:18:49
Uh, I think it just, for
1:18:49
example, Processing more information.
1:18:54
So if you have already reviews, if
1:18:54
you have already feedback, if you
1:18:58
have already a book about like, where
1:18:58
people are writing feedback, you
1:19:01
can feed everything into it, right?
1:19:03
So if you want to create social
1:19:03
media posts and you have existing
1:19:07
content, which you set testimonials,
1:19:07
a website, social media posts in the
1:19:12
past, and you can feed a lot of them.
1:19:14
You don't have to select them, just
1:19:14
feed, just put, just put, and it'll
1:19:18
generate new social media posts, right?
1:19:21
Yes, that's right. And but also, like, let,
1:19:22
let's put a spin on it.
1:19:25
So in the outdoor promotion, so from
1:19:25
what's, uh, self published people, uh,
1:19:31
their books on social media, you can
1:19:31
actually feed their book into it, right?
1:19:35
So that's insane. It's like, uh, if you have a, a niche,
1:19:36
then uh, you can go for it, right?
1:19:42
And, and, and, and serve them. And also e-commerce store manager.
1:19:45
So handle social media presence for
1:19:45
online stores, selling niche products,
1:19:50
uh, so they can put a spin on it.
1:19:53
And basically do the social media
1:19:53
management service in any niche.
1:19:58
So let's move on. Academic essay helper.
1:20:00
So let's help students. Uh, a quick test model for this.
1:20:04
As an MBA student, I often had to
1:20:04
research and write case studies.
1:20:08
Each one required a deep dive
1:20:08
into company history's financials
1:20:12
and strategic decis decisions. It was overwhelming until I found this tool.
1:20:18
It helped me to structure my
1:20:18
research, organized my thoughts,
1:20:21
and present compelling analysis. My case studies went from being a source of
1:20:23
stress to source of pride, and I received
1:20:28
high praise from my professors for my
1:20:28
in-depth analysis and clear presentations.
1:20:34
And you may think that this, this is quite
1:20:34
a niche, but I mean, a lot of MBA students
1:20:38
are paying hundreds of thousands of dollars,
1:20:38
uh, just to get this, this education.
1:20:44
So this may sound like extreme niche, but
1:20:44
this is lucrative if you just focus on
1:20:49
them and you can focus on legal, you can
1:20:49
focus on medical, you can focus on tech.
1:20:54
Cause these people, these, these
1:20:54
people who are studying at university
1:20:57
for these fields, they already.
1:21:00
Invest a lot. They already, not just time, but
1:21:01
they invest a lot of money to do it.
1:21:05
So cohesively, summarize, several
1:21:05
research papers wasn't be possible before.
1:21:10
So let's move on. Uh, fiction, cool.
1:21:13
Outdoor. So it's like basically suggesting edits,
1:21:14
characters, plots, descriptions to helping
1:21:19
flesh out, uh, an overall story arc.
1:21:22
Uh, for example, sci-fi, right? So that's extremely short test model.
1:21:28
I always dreamed of writing a fantasy
1:21:28
novel, but struggled with creating
1:21:32
a detailed word, word, and plot With
1:21:32
the fantasy book assistant, I was able
1:21:36
to flesh out my characters plot and
1:21:36
word turning my dream into reality.
1:21:42
I've just published my first book,
1:21:42
so yeah, this, this is something,
1:21:45
I guess it's like, uh, also you may
1:21:45
think that it's extremely niche, right?
1:21:49
Like, yeah, science fiction writing,
1:21:49
but if you think about that, just
1:21:52
one website, it's called numo.org.
1:21:57
Get, it's, it's, it's a nonprofit. It's focusing on only fantasy
1:22:00
writing and they're doing 800,000
1:22:05
visitors per month as insane.
1:22:08
These people who try to write
1:22:08
fantasy, they pour their life into it.
1:22:12
They pour so much time and money just
1:22:12
to realize their dreams and you can be
1:22:19
build a to adjust for them if you're
1:22:19
focusing on just, just for them.
1:22:22
And I guess it's worth like 20, uh, 10, 20,
1:22:22
30 bucks a month if it's, it's good and it's
1:22:28
actually, actually happen, you move forward. But it can also be other like not
1:22:30
just uh, sci-fi novels, but it can
1:22:35
be a fantasy book assistant as well.
1:22:37
Uh, it can be romance nova collaborators.
1:22:40
So it basically can bring this
1:22:40
to two other field as well.
1:22:44
Right. So let's move on. Legal tech company.
1:22:48
So a quick testimonial. As a small business owner, legal jargon
1:22:50
used to send shivers down my spine.
1:22:55
Contract analysis service made it easy.
1:22:58
They turned complex contracts into a
1:22:58
simple language that I could understand.
1:23:02
I felt more confident in my decisions
1:23:02
and saved a fortune on legal fees.
1:23:07
So this is something which
1:23:07
we can be done easily.
1:23:09
It can help also, we can put a spin on it,
1:23:09
like patent fillings assistance, really
1:23:15
estate low advisor for rental agreements
1:23:15
and, and, and deeds and those kind of things.
1:23:20
So these are also may
1:23:20
consider it as niche, right?
1:23:24
Because you think like, ah, no, not, not, not. It's, it's just a very niche market.
1:23:27
But if you think about that
1:23:27
rental market, it's huge.
1:23:31
Like real estate market is huge and
1:23:31
lots of money is exchanging hand.
1:23:36
So yeah, th this can be, uh,
1:23:36
revol revolutionized as well.
1:23:40
So let's, let's move on,
1:23:40
give you another topic.
1:23:44
AI blogger. Give you a quick testimonial.
1:23:47
I run a blog on the JavaScript
1:23:47
libraries, but struggle to keep up with
1:23:51
the rapid updates and developments. The JavaScript Library blogger helped
1:23:53
me to create in-depth, up-to-date
1:23:57
content that my readers love. My blog traffic has tripled.
1:24:01
It can be also Crypto Blogger,
1:24:01
it can be Micro Biology Blogger.
1:24:06
And what's changed with these
1:24:06
tools is that actually you can
1:24:11
feed into the whole documentation
1:24:11
of new, uh, New libraries, right?
1:24:17
You can even feed into the code itself.
1:24:19
So the, the possibility is really
1:24:19
endless and whatever is picking your
1:24:24
interest, you can dig deep and provide
1:24:24
a tool which, uh, which is helpful.
1:24:28
But also news reports, if you, if you
1:24:28
think, if you stay on this train of
1:24:33
thought, that creating news reports like
1:24:33
a local community news reporter, right.
1:24:38
Just a quick testimonial model. As a local journalist, I was overwhelmed
1:24:39
with the number of stories in my
1:24:43
community that needed coverage.
1:24:45
The local community news reporter helped
1:24:45
me write comprehensive news reports
1:24:49
quickly, allowing me to cover more
1:24:49
stories and keep my community informed.
1:24:55
So that's 1, 1, 1 again, it's like, uh,
1:24:55
lots of people think that like local
1:24:59
community, uh, news is, is something
1:24:59
which is niche, but it's quite big,
1:25:04
big market, uh, uh, all over the world.
1:25:06
So that could be useful as well.
1:25:09
Or product descriptions as well.
1:25:11
Uh, just like. Writing Kickstarter, project descriptions,
1:25:13
luxury real estate listing, describe
1:25:18
Describers, uh, and those kind of things.
1:25:21
Uh, it's, it's easy to put a spin
1:25:21
on it and create these tools.
1:25:26
Also, just staying on the language
1:25:26
and, and tutoring side, it's, uh,
1:25:30
let's say a quick, we'll do that again.
1:25:33
I'm a business professional,
1:25:33
frequently traveling to Spain.
1:25:37
I need to improve my Spanish, but traditional
1:25:37
language courses were too general.
1:25:42
The Spanish tutor for business professionals
1:25:42
gave me specific language practice for
1:25:46
my business meetings and negotiations.
1:25:49
My confidence in doing business in
1:25:49
Spain has greatly, greatly improved.
1:25:53
So that's one, again, it's like the
1:25:53
big languages outside English, uh,
1:25:58
which can be targeted and, and niched
1:25:58
down, like just for example, Spanish
1:26:03
for business, uh, professionals, but
1:26:03
it can be Mandarin to, for travelers
1:26:07
or French tutor for students as well.
1:26:10
So, If you're an entrepreneur,
1:26:10
this is the best time to be alive.
1:26:14
And the same goes of for STEM assistant.
1:26:18
So it's like middle school math
1:26:18
assistant, high school physics helper,
1:26:22
or college biology study study aid.
1:26:25
That's the same, same, same thing basically.
1:26:27
And if we go to the enterprise, so what we
1:26:27
cover now is basically what entrepreneurs
1:26:31
can take on, but what can enterprise.
1:26:35
Do or what can be done for
1:26:35
the enterprise as as a client.
1:26:39
So digital therapist platform.
1:26:42
So I struggled, so that's the testimonial.
1:26:45
I struggled with insomnia for years.
1:26:48
This platform, sleep therapy advisor helped
1:26:48
me understand the route of my sleep issues
1:26:53
and provided practical relaxation techniques.
1:26:56
I'm finally getting full nights of sleep.
1:26:58
Again, it can be spinned, uh, to stress
1:26:58
management as well, or PTs D counselor.
1:27:05
So basically providing mental
1:27:05
aid for your, uh, employees.
1:27:10
One of the best ROI thing ever.
1:27:12
We're gonna get into what kind of, uh, pers
1:27:12
you can get if you've worked for Atropic.
1:27:17
But one of the best ones in my
1:27:17
mind is they provide $500 a month
1:27:23
for valance and they understand.
1:27:26
So the company understands if you're
1:27:26
in good and health, then you are
1:27:30
doing better job and better output.
1:27:33
So, but also, but be covered. Code refactoring service.
1:27:36
So Legacy Code modernizer, Python,
1:27:36
refactoring Service Game Code Optimizer.
1:27:42
This is something which is quite common even
1:27:42
with the legacy code bases, thousands of
1:27:46
people writing some code in, I don't know,
1:27:46
a couple or four ton, and it has to be RRI
1:27:52
written and with these kind of tools that
1:27:52
the token bin was big, it's finally possible.
1:27:59
Uh, corporate training is something with,
1:27:59
with cybersecurity sales training course
1:28:05
can be fine tuned to each and every company
1:28:05
because you can feed into the company,
1:28:11
you can feed the sales script and you can
1:28:11
generate something which is specific for a
1:28:14
company and you can sell, sell it to them.
1:28:18
And you can also sell the services itself.
1:28:20
So people are going through
1:28:20
the course, uh, themselves.
1:28:24
Market research analyzes. This is something we also
1:28:26
covered and we do as well.
1:28:29
We do lots of research and what if now?
1:28:32
All the research can be fed into it
1:28:32
and it can make cohesive summaries.
1:28:37
Coding tutor as well. So like Swift Tutor for iOS development,
1:28:39
Python tutor for DA data scientists, react
1:28:46
native tutor for mobile app developers.
1:28:49
So the possibilities is
1:28:49
really endless, I guess.
1:28:52
And uh, uh, the final thing I just want
1:28:52
to say, uh, is that Zoom, as I said,
1:29:00
already invested and they integrating
1:29:00
them into the Zoom contact center.
1:29:04
So this is not, uh, out of
1:29:04
reality, which we cover now.
1:29:09
And uh, also notion AI is built
1:29:09
on top of Atropic models as well.
1:29:15
So, and they also have a junior learning
1:29:15
platform company, which is basically
1:29:23
helping coaching students and it's,
1:29:23
they're covering different subjects, math.
1:29:30
Uh, critical in reading and, and
1:29:30
those and, and these kind of things.
1:29:35
And it's already done,
1:29:35
but it's quite general.
1:29:37
So if you feel inspired, uh, I urge
1:29:37
you to just like create something which
1:29:45
is valuable for your niche and you can
1:29:45
niche down, like you can create of value.
1:29:50
Okay? And
1:29:51
we also did a community
1:29:51
around Anthropic, and we found
1:29:54
that, that Twitter is very strong
1:29:54
with a hundred thousand followers.
1:29:57
And, and when they announced their
1:29:57
most recent, uh, cloud model, then it
1:30:02
reached more than 2 million people. So it's very strong, but also, but I
1:30:04
didn't find any dedicated Facebook groups,
1:30:08
uh, who are focusing on cloud itself.
1:30:11
While there are several FA Facebook
1:30:11
groups, uh, fo focusing on mid
1:30:15
journey and, and GPT as well.
1:30:18
I guess the main reason for that is
1:30:18
it's already closed beta, the API access.
1:30:23
So developers can't really access it now.
1:30:27
The only way to access it either
1:30:27
way through po.com or through Slack.
1:30:31
So you can add their Slack bo
1:30:31
uh, bot and have a conversation.
1:30:35
But that gets kind of like niche. So it's, uh, Still, if you move on to the
1:30:37
next, uh, the, the final part of this, uh,
1:30:44
structure through recruitment, and we looked
1:30:44
at what kind of people they are hiring.
1:30:50
It's not surprising that they are researchy,
1:30:50
uh, institution first and foremost.
1:30:55
So they hire a lot of researchers, uh,
1:30:55
computer scientists, and so on and so on.
1:31:00
And, uh, now finally it
1:31:00
shows that they're working on
1:31:04
productizing what they're building. So they are looking for
1:31:06
product people as well.
1:31:08
And so it's, it's kind of like in the
1:31:08
process of, uh, gradually releasing
1:31:14
what they build to the community. And they're the very
1:31:16
first step, basically now.
1:31:18
So that's, that may be the biggest reason
1:31:18
why the community is not that strong.
1:31:23
And it also says, and shows that
1:31:23
most probably if you are into, uh,
1:31:29
community building, if you are into
1:31:29
developer edu education, if you're
1:31:33
into, uh, solution engineering for
1:31:33
clients, they're gonna need it.
1:31:38
So even though if they may not
1:31:38
have all these positions open now,
1:31:42
they're gonna need it for sure. So, and what does it mean?
1:31:46
It's, it's, that's kind of like
1:31:46
one of my biggest pet pet peeves.
1:31:49
It's like, if you like coding and you want to
1:31:49
quickly understand this field, like solution
1:31:54
engineering is one of the best field you
1:31:54
can be in because your day-to-day job is
1:31:59
basically talking to clients, solving their
1:31:59
problems, seeing how they solve problems,
1:32:03
seeing how they're struggling, what can
1:32:03
be done, and what can, what is analog.
1:32:09
And you're helping clients
1:32:09
to basically create value.
1:32:13
And it's kind of like the fast
1:32:13
lane of learning how to apply AI
1:32:17
to bus real business settings. So, uh, as I said, uh, obviously, uh,
1:32:19
this $500 per month for VanNess, uh,
1:32:27
steepens is, is something, uh, which is,
1:32:27
which I didn't, uh, come across before.
1:32:34
So I big kudos to them for that.
1:32:37
Uh, but also they, they, they cover
1:32:37
the usual, so they give equity, they
1:32:41
sponsor green card if you need it.
1:32:43
So it's kind of like they have
1:32:43
all the products, uh, others
1:32:47
are pro, uh, providing as well.
1:32:49
And I also found they are hiring
1:32:49
recruiters, so it means they're growing
1:32:53
fast. Yeah, yeah, yeah.
1:32:55
And then they need, need operators as well.
1:32:58
And, uh, I didn't really see
1:32:58
explicitly, but I guess they're gonna
1:33:04
need security, uh, expertise as well.
1:33:08
Uh, and as I said, more people with product
1:33:08
expertise, UX marketing, psychology.
1:33:14
What we already discussed with assembly
1:33:14
AI is they, they're focusing on
1:33:18
onboarding people to the, to the api.
1:33:21
So the obvious, the Anthropic is
1:33:21
that the API first company as well.
1:33:25
So they are going to need people
1:33:25
who are dedicatedly optimizing
1:33:30
the VE flow for developers.
1:33:32
And also dog fooding is my middle name.
1:33:35
That's my pet pee. And what is dog fooding Viktor?
1:33:39
So, so dog fooding again, is like,
1:33:39
uh, if you for walk the path yourself.
1:33:45
So you are, in this case specifically,
1:33:45
it means if they require every, everyone
1:33:50
in the company to build something on top
1:33:50
of their APIs and nobody the expertise.
1:33:56
So it's not just engineers and researchers,
1:33:56
but also managers, also customer service
1:34:01
people, and everyone, every single person
1:34:01
should build something on top of the api.
1:34:06
If they require that, so basically use
1:34:06
their own product, what would it unlock?
1:34:11
And that would be the single most
1:34:11
impactful decision they could make.
1:34:16
Because then every single
1:34:16
issue is just bubbling up.
1:34:19
If something is not working, it's not
1:34:19
clear, uh, they or they are not, no,
1:34:24
not enough tutorials for beginners.
1:34:27
It, it, it's gonna bubble up. So all, all these issues are going to bubble
1:34:29
up and it's going to continuously bubble up.
1:34:34
And the needs like, okay, I try to use this,
1:34:34
for example, for sci-fi, uh, writing, right?
1:34:40
And it's not good for that. And it turns out it's gonna, it,
1:34:41
it's going, going to turn out
1:34:44
because someone is trying to do that. So the research and development, uh,
1:34:46
is covered already and defining, uh,
1:34:53
good product market fit and also.
1:34:55
Providing an easy to onboard experience
1:34:55
would be skyrocketed if they're
1:35:01
just like requiring dog fooding from
1:35:01
every single person in the company.
1:35:05
So that, that would be the biggest,
1:35:05
most impactful decision they could make.
1:35:10
Okay. Thank you very much Viktor. And thank you very much Disruptors for
1:35:11
listening to us for this long time.
1:35:15
You will find every important link
1:35:15
and information the show notes.
1:35:19
Please go to this app digest.com. Viktor,
1:35:22
that's a wrap. Yeah, that's a wrap.
1:35:24
Thank you for listening. Um, see you guys.
1:35:27
Thank you.
1:35:28
Bye.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More