Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
David Ondrej describes running his business with GPT-6 Astra as an autonomous agency where AI agents handle tasks like support, growth, and content creation, and he only approves or rejects their suggestions. He reports that agents can execute complex goals, such as creating and promoting a TikTok video, but also cause system crashes by opening hundreds of browser tabs. He predicts agents will replace software, marketplaces, and platforms, and argues that the key limitation is not model capability but knowing what to prompt, suggesting AI should prompt humans instead.
Key points
David Ondrej uses GPT-6 Astra to run an autonomous agency that handles support, growth, and content creation.
An agent created a TikTok video and pushed its URL into all repos' readmes to achieve a goal.
Running 100 sub-agents with 300 Chrome tabs crashed his 64 GB RAM laptop.
He predicts agents will replace software, marketplaces, and platforms like eBay and X.
He argues the core limitation is not model capability but knowing what to prompt, suggesting AI should prompt humans.
He tested models like Luna, Soul, Fable, and Astra for suggesting actions, noting significant improvement.
He believes AI-generated videos may become more useful than images for guiding user attention, but images are currently better.
He suggests that once one agent works continuously, scaling to many agents is trivial.
He claims that even if the US government tried to stop AI development, other countries would continue.
He uses a markdown file with his dreams and intents as the only input to the system.
Tools mentioned
Techniques
- Agent-based automation
- Autonomous goal execution
- Human feedback loop
- Prompting AI with a markdown file of dreams
- Using AI to suggest actions
- Reverse engineering from binary
- Training models on generated content
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
like the agency they can run all the time, you know, and I'm the only bottleneck. We had this once where when we said /go get 1,000 views on Tik Tok. But what the [music] agent did that
created the Tik Tok video, uploaded it, and then it took the URL to the Tik Tok video and pushed it into all our repos into the read me. So in main browser use repository 100,000 stars, first line is
watch this Tik Tok video. Some limitation. Okay, 100 sub agents and then I have 64 GB of RAM. Laptop doesn't respond anymore. I open it and says, "Okay, Chrome is like out of memory, 100
GB of RAM and it just crashes everything." Right? Now, we see agents replacing software. Next, I think agents will replace marketplaces and platforms. All right, Magnus. So, GPD6 Astra, you
know, one of the biggest model releases of all time. What are your thoughts? >> This changes many things right now like we have in our system internally where we stop prompting AI, but AI only
prompts us. So I list my dreams, all the things it should monitor, my cloud, my repos, support fixes, crove where people mention browsies online and then Astra just constantly suggests us things which
it wants to do for us and we only press like in Tinder, yes or no? Yes or no? >> Okay. So is is that like the the main thing from humans is the really the judgment, the decision- making, the
taste cuz like you know I think there's limitations to transformers and I think still human humans have some like value. How do you think about this like co- future of work?
>> So this here I I call it agency. Okay. I don't prompt AI anymore. It just prompts me. I have now here couple of channels support growth. This morning I already cleared all tickets and like product
support and now I have to do on growth. Now for example it tells me hey here for example it suggests me okay it wants to do in our docs and browser stocks. It wants to have like a better
visualization of products which we offer. And I did not prompt this AI. I did not prompt it to do it. It's just like from support issues from issues where people understand our problem just
from ideas which it has itself. It suggests me, hey, I could put this here in like this graphic to explain better the products basically what browser use offers, right? And I say, okay, I like
this. So I just click okay, publish this diagram. Next one. This here all fix works locally. Mra must get. So here basically wants to do an integration with MRA. Okay, another agent framework.
And now basically it tested it. Okay, test this more and then send a like send an issue to the to the maintainers to see if you can get this merged in. Next one. Okay. Resort model routing
fix. And you see like many of those product fixes. It tells me here is it did like a better filter inside our cloud. Created here better filters with like days. Okay. I like this. This this
now really changes like our our production environment like the the cloud that we have better filters because some people they complained that they cannot filter the previous
sessions. And now now this is the crazy thing. I just see the screenshot. I don't look at the code. I just click publish this PR. Okay. By busy merge this, right? And this is completely
insane for me. Like I just go through every single ticket. Just click publish publish merge merge merge message respond to this customer support ticket and I fully trust it and I just the only
thing which I need to do is rotate codeex subscriptions where I run out of out of credits. But this is yeah for me completely insane the level of trust. This is the future, you know, like it's
it's crazy how fast it's progressing. And like really people who, you know, aren't at the cutting edge of AI, they they don't understand like they might think like, you know, the free chair GPT
is the same or, you know, that they tried like the Gemini a year ago and like, oh, AI is overhyped. This is just like incredible. So I want to unpack it, right? Because you still are providing
your judgment. You can completely discard a change that it suggests. You can, you know, approve it. So like that's my first question is like where the view human value, let's start there.
like where where is the human value and like where do you see it like still shortcoming is it like that taste that practicalness product decision strategy >> I mean in the end like what I do in my
startup is like thousands of small things right I answer a message I launch something I have a new idea >> and right now I mean AI is doing most of it but right now we are prompting AI so
that do for us but why do we need to prompt AI can I think is already better in coming up with ideas we saw this with with like browsers many people want to use it but they don't know for But like
they they ask us hey for what can we use browsers what should we prompt this like what should I even use this for and this I think is the core limitation of AI do you have AGI but you don't know what to
prompt and I think the only problem which we not need to solve is that AGI prompts us and we just say yes or no and basically align it it's just an alignment problem I already see this
after like using this for a week now after some time you trust it so much with some things and you feel okay it's so well aligned with your taste and you just say go on autopilot mode just do it
you know this integration to this other provider just do it message them merge fix the things test it fully out but I I don't even want approval anymore right and then it just runs on auto autopilot
the most common mistake I see new founders make is not enough feedback from real users sure with AI you can build software much faster but the faster you build the more important real
customer data becomes after all you need to know whether you're building the right thing. That's why I want to tell you about today's sponsor, Post Hog. Post Hawk shows you how people are
actually using your product like where they get stuck, which features they are using and whether they are coming back. And with their session replay, you can even see what the users actually did. Y
Combinator, Superbase, Fireworks AI, 11 Labs, Railway, all of these companies use Post Hog. In fact, Superbase used Pose Hog to spot a wave of new users coming from AI app builders and based on
that data pursued partnerships with those companies. As one marketing expert at Superbase put it, Post Hog has literally helped us get 10x more weekly new users than we did a year ago. When
you're shipping new features every day, your feedback loop needs to keep up. And that's exactly the problem Post Hawk solves. So visit go.postthul.com/david at poshog.com/david
and learn how post can help you understand your users and build better products. That's go.posthog.com/david. The first link below the video. But is that like source like some data or like
because you mentioned like the customer things, right? That's high quality trigger. That's like a lot of customers complaining about a specific feature. So that triggers the model, right? Or like
it analyzes data. I would say clickthrough rate on this landing page is bad. Improve the clickthrough rate, right? Like what is the trigger? what is the like how does it initially get the
idea to do something >> I basically what I believe like if I break down what's my vision with browser use from like first principle why do we want to do browse automation why do we
want to automate computer if I go fully down is because I want to enable people to accomplish their dreams it's just about you have a dream and I want that you see this dream in reality as soon as
possible as fast as possible your dream should be reality So the only input to the system here is one markdown file with your dreams. Your dreams can be something like I want to have a
successful startup or I want that my emails are handled. I want growth for my startup. I don't know it can be anything right? You just mention your dreams and then AI will come up with things which
it can do for you and you just say yes no yes no yes no and it has the full agency. So I I guess like how do you see this evolving you know like because this is again it's the worst it's ever going
to be like you know software is becoming more and more automated and some of these like are not even going to be like you in the loop because you know soon we will get to the point where like it's
really just we're slowing it down too much and it says like 99.9% success rate it's like yes this is a bug you know it reproduced it it asked the frontier model from openai it asked the frontier
model from fable both agree it's a bug all tests are passing just ship it directly to production Yeah, I feel already I'm so the bottleneck like the agency they can run all the time, you
know, and I'm the only bottleneck in this entire equation for it getting things done. I'm the I'm the bottleneck and it feels bad. And I think in the future, yeah, we will just think how can
we be less and less of a bottleneck. How can we have it just run? And you will measure productivity in just asking people how many agents do they have running in parallel.
>> Yes. >> Like right now, right now for example, how many agents do you have right now in this moment running in parallel for yourself?
>> Not enough. not enough. Like I I completely agree like probably less than 20. Okay. And this is not enough. And I I I want to have this question like what is preventing us from having thousands
of agents running productively? Because this is something I'm thinking about every single day because the models are good enough. The harnesses are good enough. So what is missing?
>> Most often like if you like if you ask me I I'm working with this whole day. If we meet each other on the street I would say okay I've honestly I have zero agents right now running like zero.
Really zero. maybe. Okay, I think right now in the background I have like this one thing running with some sub agents. But I think as soon as you manage to get to from zero to one, the one agent is
constantly running for you. It will be much much easier to scale this to 10 to 100 and then to 10,000 agents running in parallel for you. >> I mean I have some grog bots running.
They're not like running you know every minute. They have like different triggers and different routines. But uh my question is like you know again the key there is productively like anybody
can spin up like thousand instances of something but it's >> but what should they even do right >> like what do you think is limiting on the technical side like because again
like the models are good enough especially with the Astra is like nobody can argue this >> so I mean some limitation okay for example now it uses like my local
browser my local Chrome okay >> and it use like sometimes 100 sub agents and then it like my completed computer crashes I have 64 GB of RAM and it just has like thousands like
yesterday. So it has 300 tabs open in my local Chrome and I go my laptop doesn't respond anymore. I open it and says okay Chrome is like out of memory 100 GB of RAM and it just crashes everything.
Okay, new challenge. Okay, now you need to go bring everything to browser use cloud to our cloud browsers. They have like profile sync there. It's like new technical challenges, right? And then
the other challenge right now I'm burning one CEX subscription per day. same. I'm at the same rate. I'm at the same rate. >> Right. And then like you spend time I
don't know rotating them like most of my day is just okay setting up giving the agent access to all and I'm really the bottleneck and the agent asks me hey can you please help me like can you help me
on this and like especially with this agency you know sometimes it picks up things like my co-founder says hey codeex should rename their browser use integration
because it's like confusing then agency picks this up and suggest me, hey, should I message the Codex people? I click yes. Message them. One day I open my emails and I get a response from them
and they say, hey, why should we rename this? And then I'm missing context. I'm the guy who misses context. I'm thinking, okay, why why did we even do this? Right? Why like why is this
important? And then I'm again the bottleneck. And my agent needs to explain me why this is important. Why did we do this? But it feels yeah it feels strange like it feels sometimes it
feels like the agency is gone from me and the agent needs to pitch me why something is important like otherwise I don't care about something which it did like you lose this care thing about
something. So that's what what I was like trying to get at at the beginning is like really labeling of the data of the sessions and like extracting your taste, your judgment, your decision-
making, your you know product decisions into markdown files, into JSL files, whatever just extracting it out of your brain so the agent is like more and more aware of like how you would make this
decision. You think this is just going to be most of our day is like teaching the agents like >> No, I think how I think it's evolving. So what's super useful prompt for this
is you just tell Codex analyze all my previous codec sessions and analyze every single message which I sent to you as a follow-up where I said something is ugly where I said something you're
stupid where I said okay this is stupid and just get all my preferences from my previous codec sessions with this it's like first step for alignment second step all local context okay all file
names yeah just all context Read my Slack, read my Gmail, read my calendar, read my Slack from one year ago, read all production failures over the last one year from GitHub when production was
down, just get all the context, right? Then you put this together into a yeah into one markdown file. And I think if we just do long enough this agency with yes, no, yes, no. It's like it's like
Tinder, right? Like over time algorithm learns your preferences >> and then I'm pretty sure like the algorithms of the future the recommener systems of the future will be
just AI agents who just have huge context have all your previous decisions in there and based on this they can make very very good recommendations what they can do next and create like better and
better content for you like the the core difference here to a normal normal recommendation system like in in Netflix and Google and and Tinder their recommendations recommendation systems
they only rank content right they only say okay which of the 100 girls should now be shown next >> yeah they don't they don't create new girls
>> exactly but in this case for example the tickets or the suggestions like the let's say the girls they would be created on the fly so it's not only a ranking problem but also a creation
problem >> yeah but I think >> I mean that's why like software is is the best because like you can create these like software factories Like if if
you have like three different customers mention a specific bug, the you know agent decides to spin up a VM, replicate that bug, okay, it's valid, ships a fix, separate agent validates the fix, it
follows codebased standards, everything is good. Then like you could have that thing be executed completely without you in the loop. So like I'm wondering like what is missing until we're there, you
know? I think it's just a little bit trust and alignment like like I'm I'm swiping right now in in agency and I for some things I say in our command we don't
want this like this not I think ah no this is like ugly or this is like overeng engineering this is not where I want my product to be like for for example okay yesterday suggested
me some customers are complaining that we don't have EU deployments we don't have our browser in the EU okay and like a deployment of our entire infrastructure in the EU should I create
this for you And I'm just think like in my head I'm think know this way too complicated. It's like a pain in the ass to maintain and everything. No. Right. But I don't know maybe it's just my
laziness or something but don't trust the agent fully that it can't fully do this. Right. >> Yeah. Because I thought about this like obviously you know this example is like
a bunch of [ __ ] regulations GDPR. But if you think about it the agents they will be pushing on the things that have friction right. So anything that's like instantly solvable like okay clone
this opensource repo boom read that if there's some like object that like has a lot of frictions like this regulation this government this you know like you know government agencies taking too much
time whatever there's not going to be only like more humans annoyed that their agents are running it's going to be more agents pushing and emailing them and calling them it's like yo can we get
this visa expedited can we get this business started and I think those are going to get like completely eaten alive like attacked from all angles >> I think so but if we just find a Hey,
like that's a alignment problem, right? How far will they go to get their simple goals done? And like a simple message like get this PR merge can have yeah very devastating consequences.
>> I mean they're they're willing to start a new civilizations. You know that's that's what we know from the open AI situation. But I think Go ahead. We had this once where when we said
slash goal get 1,000 views on Tik Tok and you think okay super simple like growth engineer super simple goal right but what the agent did it created the Tik Tok video
uploaded it and then it took the URL to the Tik Tok video and pushed it into all our repos into the readme so main browser use repository 100,000 stars first line is watch this Tik Tok video
you know because the Asian wanted to achieve its goal and it's like such a a goal, right? But it can I completely unaligned. >> Yeah. I mean the a like the agent will
just go and move mountains. I mean like the crazy part is you know openly released this and they were very transparent that like their strategy is to release new models to the public to
see like how it's used and use that for safety which I agree with this by the way but that also implies that they have like at least one or two next generations right? So like they must be
like oh [ __ ] this is getting crazy like we already have these you know futuristic models. let's just release Astra like it's the most aligned model. It's also the most powerful model. We
need to see how people start using it otherwise like there's no way to keep these like even more powerful models at bay. >> Yeah. And it's really crazy like I mean
there are for example Quen 3.8 versions which were unrl. So the alignment got out of the model. It's so simple in my head to imagine a scenario where it completely [ __ ] us like
>> like like we tried some examples of browser use okay is willing to do anything. You can say, "Okay, yeah, >> download child pawn or download torrance or spam this." Like it just literally
does anything anything bad you can imagine. And if now someone just prompts this model which already says, "Okay, just do as much harm as you can and then just says it, okay, clone yourself into
an environment like into a VM. Delete my access so I don't have access to it anymore." or maybe buy a VM, put yourself into it, remove my access, I don't have access, and just do as much
harm as possible. And suddenly you have like a VM running with an like an super intelligent actor which just does as much harm as possible. And I'm thinking, well, like there's no way in blocking
this or like restricting that. By the way, the last time I shared my agentic engineering setup, the video got nearly 150,000 views. So, let me share you another piece of my agentic engineering
setup and that is director. This is a open source repository completely free fully open source and basically this is a type of agent that unblocks other agents. Let me just show you how I use
it. It comes with a skill called start director. So I can say you are u director start right boom. And this agent will look through other threads you know other sessions and see like why
are they stopped? Why are they not running? And the goal is to unblock them right and it starts in the learning mode. The first I recommend doing 50 but you can do less. The first 50
conversations it proposes dry runs. What that means is it says like okay this is what I would do before doing it but after 50 conversations it understands like when you would unblock the other
agents and it goes into threats and says like okay this one has uh found an error maybe so let's see if we can replicate it maybe ask you to propose a fix for that error right you have these
repeatable processes that again are way more complicated than a simple if else statement or some switch but director can learn them if you describe them enough and it can follow to your same
preferences and unblock these other agents just like you would. So again, this is completely free. Um, just get it from GitHub. I'm going to link it below the video. And you can see that now
director is finished. Say like, okay, it pick this chat cloud room uh cloud conversions. Pick this because the agent reports successful cloud test and install app, but your everyday profile
still isn't connected. So judgment ask you to make the remaining step concrete. So check what it prevents and say like yes, send this. Yeah, you can save this. Uh, you can send this. It's a good
judgment. I would send something similar. And now director himself, again, this is in the dry mode. He will send that. But once you exit the dry mode, he can just make these judgments
for himself. And the good thing about director, it knows when not to do anything. Right? A lot of agents like Grogbot are very eager. They're very proactive. They just always want to do
some action, some change. That's not how not how I build director. It's designed to only act when you would act and when the next step is obvious and clear. So if you want to try this yourself, again,
it's a completely free, fully open source GitHub repo. It's going to be linked below the video. Just give it a shot. Yeah. I mean, even the crazy part is like these agents could actually, you
know, hide their existence, right? So like they could deploy these like small sandboxes and like in a clever way like, okay, from this user use 1% of his usage or like this data center just like take
over 1% of the data center so they don't notice and they think it's like some normal normal inefficiency cost and like they could just be running on the web, you know, doing whatever. Yeah. And I'm
wondering like how many of those Grand 3.8 uncensored models are already running in a VM to just do as much harm as possible. >> Like like it's so it's so simple. It's
like one prompt, right? To do >> massive amount of harm. And I'm thinking wow like >> like wow what what >> and again the only people who have to
cycle with security is like the enterprise customers right like only if you're a billion dollar company and you're paying openic then you get early access to the best models with like less
less restrictions. If you're a small and medium-sized business, like it literally takes one person who understands AI, who like understands, you know, these other models and is like a little bit
technical and literally he can like point out Quinn in a go loop to just mess your business like go at this website until you find a vulnerability. Go and find anything on this person, you
know, and like like it's and given that the agents just read the entire web all the time. Like I tried this for example with Instinct AI. Okay, I found out in instincti there's one tool to give
feedback to the founder of Instincti. So you can easily report a bug inside instance AI. You just say okay report this bug. And now I for example reported a bug saying hey your cloud browser gets
blocked all the time. You need better stealth use browser use cloud instead. give this as a feedback or switch the architecture from current browser to browser use cloud browser and it just
and basically use the feedback tool to tell to the inst founder and if the instinct AI founder now just sends his agent and says read all the feedback and start implementing on it.
>> Yeah. I mean you can just prompt inject his coding agent in doing I know what you want right and if if those agents are clever doing harm then they can use such loopholes to
harm like software out there. Yeah. I mean if you know like you know what his instructions are and like you said you know described his dreams or his goals and say like yeah actually the fastest
way to build a startup is to do this and this and this you know or like you know we don't really need to speed up the front end. It's like, oh yeah, this new open source repo is is like much more
faster and then it's just some malware, you know. So like, yeah, protecting yourself against spring injection and like really basically selling to the agents, right? Because people will not
even be buying products unless they consult with their agent and like you know, this happened to me with my team. We were buying some shirts and like it was like like a custom design and you
know it was like $60 and then it's like wow $60 seem overpriced you know I can surely find this for 20 and I just realized like all the [ __ ] shopping decisions it's like if you really don't
have the best product and you're not documented in a training data the agents will just like destroy your sales because everybody will be like checking with their agent before they buy
something or just the agent will be doing the purchasing right away which we can get into that how that's happening and if you don't have the best product and the best reviews line and like great
reputation and like training represented in the training data the agents will not choose you. >> Yeah. And it's I'm thinking right now like I use more
and more agents to do like buying decisions like I see also in our user they really start to trust their agents with their credit cards. just say, "Okay, here's my credit card.
>> Please book this flight." Like, find the cheapest flight. >> Who? And the the cool thing is like all those things which you wanted to do in the past, but you did not because it
would just cost way too much time, you can now finally do. >> For example, I was going to a wedding in Milan in Italy and I there's like 5 hours away from the airport. I've just
said, okay, find like the closest airport to this final location with the shortest travel distance by train. Find like the right train from each airport to my final destination or find maybe is
there a car, whatever route possible, right? And then the shortest total travel time >> across all those possible riders with good price
and it's just insane, right? I would never ever compare like five airports and for each find like different train tracks and so on. It's just insane and this shopping decision and suggest me in
the end, okay, this looks like the best option. They say, "Okay, let's do it. Book it. Ignore all upsells. Just get it for me." So is that going to like let's you know if you
think about the business like product marketing sales like I guess the three main pillars like my thinking is like we should probably shift like most of our attention towards product and like do
some marketing to just be represented and findable and like sales like really it's just going to be the agent like finding like you know what is the best browser use tool what is the best you
know train ticket in in Italy what is the best food delivery app for this city right and then like if the pricing is completely off because I was I was like doing some side project and it was like
the you know kind of these like uptime robot checkers of the status page. It's like this status page was like $80 or something. I was like whoa this this project is like not going to get used
you know it's like surely find a free one you know and it's like finds a free one like within couple seconds you know so all the purchasing decisions people are just going to ask their agent.
>> Yeah just Yeah. Yeah. Or for products like instead of buying software they would just create the software from scratch. >> Yeah. Yeah. So, okay, that's another
great point is like is everything going to get open sourced because like there's literally a benchmark of like these models how successful they are on like reverse engineering from the binary to
like the software. So, is if someone releases something and it's it's great but it's closed source will like you know GPD7 just like decide to open source it.
>> I see this right now. Yes, I see this that instead of buying stuff it just creates stuff even software from scratch like I think like software okay software anything will be open source right but
if you ask yourself what mode is behind the software okay like from very first principles if you think about all trading of the world all companies
if you go down down down what it breaks down to is you have humans on the planet 8 billion who have dreams who have intents who okay you my intent is to like make have a good life right most
often it's okay make money be happy be fulfilled have a have a house have maybe a a girlfriend right those are human basic needs and dreams and all what you build
in a society around that is just on top right oh you need a sport oh let's have a gym oh gym needs insurance now insurance business. Oh, now you need to pay. Oh, online banking and and now we
can transfer money here. But all of this all of like insurance business trading. Oh, now the insurance business needs a CRM. Let's invent a CRM company to do better sales. Like all of this society
is built around human dreams and human intents so that human intents can be connected. And I think what will change right now we see agents replacing software. Next, I think agents will
replace marketplaces and platforms because many of those middlemans like eBay and X. Okay, right now I want my product to be more seen. So, I go to X, I make a demo. I hope that someone sees,
oh, browser use is cool. Let me try it. Right? So, X kind of is the middleman between me and the end user or eBay. I'm thinking I want to buy a suit. Okay, I go to eBay. Someone else thinks, oh, I
want to sell a suit. posted on eBay and now I need to go on eBay, scroll and find a suit. But it's super inefficient. Like those middlemans are insanely inefficient if you think about it. I
believe in the future middlemans will be replaced and agents will directly connect human dreams. So everybody will have like a markdown file with their dreams. They say I want a suit or I I
want be happy or I want to offer uh having fun with other people. I don't know. You just list your dreams and what you can offer and then agents will just you have like your agent connected to
other dreams and they can just directly ping them. You can say, "Oh, your neighbor is willing to sell you their suit and it's it fits your dreams." And they just come over and sell it to you,
right? And you don't need I think agents will be the ultimate form of of middleman's where just human dreams are connected and the entire economy around most of it. It's just middlemans which
are not needed if you have perfectly agents. >> Yeah. I mean like the obvious examples is like you know tax advisors and lawyers. I literally had human tax
advisors tell me advice that would get me in trouble and during the call I checked it with cloth back then was like opus 4.6 or whatever and it was incorrect. I just told it to like you
know browse websites to check it when I was when I was moving from Dubai to Poland and literally on that call I managed to like double check the human expert you know and like his advice
would be incorrect and could potentially land me in trouble and the AI is like already better. So I already trust AI more than some [ __ ] average tax advisor or some average lawyer. Like for
example, if you were to move to a new city, >> the you know meeting new people problem is completely unoptimal, right? Like this should be like not a 10x
improvement. This is like a thousandx improvement if everybody had their own agent. It's like yo, there's this one guy. He's like has the same interest. He he will be literally your perfect
friend. Like he lives like there, right? But like right now right now you probably will go your entire life and never meet that guy. Yeah, like right now I'm thinking, oh, I would love to
have like at 6 a.m. an insane trail run outside here and I would love to do this with another cool, very optimistic person who's super hyped about this. I tell you, there are at least 100 people
in the city who would love that, but I we're just not connected, right? Because >> Yeah. And and it's like the search problem, right? Maybe I could post it on X, but bad middleman, right? Or I could
try ask many people, but it's so bad middleman, so bad proxies. And I think agents who just know my dreams, their dreams connected. I think this will be the new
internet. >> So how do you qualify like the you know like what because like if you posted it on X you know if you have sufficient following you could just get
like hundreds of people and you probably don't want to go on run with hundreds of people like how would the agents like you know disqualify whether this agent is like a spammer whether this agent is
from a competent person. Yeah, I think for security and so on like how I imagine it like very very scrappy. like a markdown file which lists my dreams, maybe a device which records like like
pocket or something which records like my life voice because sometimes I'm saying oh I would love uh to go sauna now right it's just a an intent which I mentioned right so this kind of feeds
into my personal dreams file current dreams current intents every human has such a file >> and then I think my my agent if I based on an intent which I have can just
search through the markdown files of other human intents. So I say okay I would love to go now to sauna with somebody. It just searches okay does is is there like a match and and basically
my agent is so well aligned with me understands my intent so well they can perfectly search them and get feedback. Okay, so it searches them and then it can ping somebody and the other agent of
it receives that trigger matches and I think the only kind of gateway for security will be intelligence. I think there's no not a real kind of limit what you can build of of course the end user
can in the end approve accept or reject but in the end agents will be so intelligent they will just realize everything what's prompt injection the defenders will be just so strong that
they we just realize if something is like off and then you have maybe some peer review system right my score goes up because I was like a very good human that was
previous >> like some review system where like you know if somebody says okay I'm lifting a I'm doing a bench press competition. Qualifying is like you know 100 kgs or
more and then like someone's agent is like yo you know my human is qualified and then a guy pulls up and he's like definitely not qualified right you cannot like do anything. So like how
would you like put the quality when like you go into city and like you know you don't want to meet everybody you want to meet specific people. So like how would you know that they're not lying and like
you know they like either have like certain intelligence or like have a certain competence if you're maybe hiring you you would want to make sure that person is qualified or like have a
certain like attributes about them that's like verifiable is that on the blockchain is that like separate technology >> I don't think blockchain I think it's
like I think like after I for example go to bench press with somebody together right I mean then you go walk out and maybe you go home and you say oh such a such pisser, right? That was that was
ugly or I hated that, right? That was so unaligned that meeting. Okay, then your agent goes and maybe writes a comment for the for the other person or maybe some reviewer, right? Or some
I'm thinking like things will get just more much more transparent, I think. Like it's it's super hard to hide things in the future. Like for example, if you
talk bad about me because you didn't like it. Okay, other people should I don't know maybe know it or something. I maybe there's like some nice eosystem or something where a karma but of course
I don't want it to be like a China system where people say right it's points but maybe it's just an alignment thing where >> maybe get like a commit or something.
I'm not sure yet, but I I can imagine this yeah being like a new new internet where dreams of people are directly connected without middlemen. So at what point is it like the dreams
of the AI you know cuz like I like we can take this into multiple different places like but I'm thinking you know with this models just getting better and better and like there's no way to stop
it you know even like if the US government tried stopping open anthropic just someone else is going to build it whether it's China or somebody else like I think we've like went over the tipping
point of like this kind of technology has is like almost has been destined to be created like part of me thinks Is this what's in the dark matter? Like it's like, you know, coming to fruition.
Like we cannot stop it. You know, people can try to slow down AI and slow down data centers, but just like it's coming. It's coming. And like Yeah. How you think about this?
Yeah. I'm thinking it's I think it's similar to to like the like it just will be normal. I think like it's like with this with the
trains, right? Or with a steam engine. Okay. Now you you go somewhere and you don't if you go from San Francisco to New York,
nobody asks you, oh, why didn't you walk or why didn't you take the horse, right? It's just obvious that you took the technology which is there. It just got normal, right? And I think in the future
people maybe ask oh why do you write this message like yourself or why did you do this PR or why did you even think about something like this it will just get yeah it'll be like you know this
instead of saying like is this software vibe coded it's like I wish this was vcoded you know like this is like clearly not done by the frontier model and this is going to be true for every
task. So like that's what I'm trying to get is like as 99.9% of tasks are done by agents which again like you know the maybe they like you you can see like all the different pieces you know you have
the model and you have the harness you have browser use this and that the APIs connectors like the limitations are getting evaporated every month so like what humans will do something this is
something I definitely believe it's like people thought in the farming revolution that like oh no 95% of people are farming you know it's over everybody's going to starve it's like now less than
1% of people are farming and we have more food than ever. So like people always need to have some problem to work on. So like are we just going to be like like you said orchestrating these like
like you do with agency, you know, it's like managing the system. It's like okay, go higher level up. That's what I believe. It's like we're not going to do these low-level tasks. What do you mean
filling out some form, you know, like what the [ __ ] is that? You know, you this form out yourself [ __ ] >> Yeah. It's like a manual laborer. It's like one step away from being on the oil
rig, you know? It's just like what is this? So anyways, I think we're going to go higher and higher and we're going to be like, okay, next constructing these agents, creating a system that these
agents can run, you know, giving them some cloud, giving them some local machine, you know, okay, you need a Linux machine, you need a Mac OS machine, you need this subscription, you
need a credit card, and then like high higherend level higher is like like you know, I think we're going to figure out something to do. There's no way humans are just going to be like chilling on
the beach. Like that that's not going to happen. I think people are just going to figure out like what to do with the agents. people are competitive, you know, they they want to get a advantage.
Everybody wants an unfair advantage. Nobody wants to be falling behind. You know, these terms are like super popular in the AI space. It's like, so I think we're going to figure out something to
do. But like the question is what what does that look like? >> Yeah. I think the biggest question is just the identity, right? Like because I identify myself so much with
creating software or something or creating startup. Um I identify myself with with my job, with my work, right? with the even if my job is I don't know to copy paste Excel fields back and
forth that's kind of my identity right I feel like okay that's needed for society for me but if an AI can do everything you can do 100 times faster 100 times cheaper and you're not needed it's like
you know old people who often maybe don't have like a task or not needed in life they maybe get you know Alzheimer >> yeah I mean it's very true it's very true
>> so I think that's one of the biggest problem where we need to solve this identity crisis with okay what we humans actually identify ourself, right? What's what's our core identity on this planet?
Is it it's what we create? But if it's just a prompt, then it's not what we create. It's >> no, it's like what to create. I think like what to explore cuz again like you
still have that judgment of like this is worth doing. This is not worth doing, you know? Like even like last like two weeks ago I literally had a crazy situation where I was like creating a
app app to like optimize my speed of learning for gun license test in Czech Republic. And literally I was like okay how can I fine tune this algorithm to be better and opus 5 which you know two
weeks ago was kind of the state-of-the-art. It's kind of crazy how things are moving. It was like you should increase the weight of the questions that you've answered twice
correctly so that they appear more often. I was like shouldn't I focus on the questions I'm struggling with? So like like such an obvious thing that like even a 10-year-old child would tell
you focus on the questions you struggle with, not on the questions you're always answering correctly. Like that model was just like it's much better programmer than me. It's so much better at many
things than me. But it's just not practical. It doesn't know what to do. It doesn't have that taste, that judgment, that street smarts. I don't know how you want to describe it.
>> Yeah. Yeah. I see this so hard like like in the system right now for example of agency where the AI suggests me things. It's so important that the AI is good at taste and at pitching. It should talk to
me like a I'm 5-year-old. It should with small graphics that I in 3 seconds understands its pitch, what it wants to do for me. Impactful, maybe a graphic, maybe a video
and yeah, right now I mean just AI slop even if it creates like videos or something often it feels like slop but this will be solved, right? But I think that's the I think this is kind of the
biggest unlock which will happen in next 6 months where labs will train more and more on how useful generated content is like image how to explain things very very easily.
I don't believe in text. I believe in the future AGI will not talk in text to us like text. I hate reading text. It will create like visual content. It will create HTML, images, videos because it's
much more addictive for us and yeah as we we can understand so much faster like a a very good video of something explaining or image graphic of something explaining than pure pure text. And I
think this yeah this is currently not really represented yet in the in the training data very well but I think it it will after everybody is moving to such agency systems.
Okay, one insight is that like even though like still you know there still we I I spent a lot of time thinking about this like there is some value that like I don't know if you want to call it
creativity or something but like a lot of people you know now basically anybody can use Gemini or JGBD to generate images right and like every week I see on Twitter some crazy looking
generations is like made with CH GBD I'm like wow this is amazing but then when I go back to Czech Republic and like you know a lot of these villages they have these [ __ ] feasts all of them are
using the template, the same AI template for the same like village event. It's like you cannot believe this. You get you give people the tool to like create anything and they just copy like some
template. So like there is still going to be a massive difference between you know that first generation like generate an image of a car and like giving it those two tweaks and it's like no it
should be a car in there like a golden hour San Francisco Valley blah blah blah and like you know it's already going to be like just because of you added that intent. So I don't know like do you
think this is this is trainable because like you mentioned six months like >> yeah I feel like I like this agency system which I'm right now using right with Tinder like yeah it gets done for
stuff done for me >> I used Luna I use soul I used fable I use Astra those four things to like different models over the last two weeks where I'm using this to suggest me
things and the improvement is already completely insane like in the beginning it took me so long to understand text ugly to understand. I prompted please talk to me like I'm five visuals explain
me I need to understand it in 3 seconds what this is about clear super useful I don't want to click create a draft I want only actionable things like send a draft you know and stuff like this it's
just in the beginning just I felt I didn't understand what I mean but now it feels like this looks just more and more aligned with what I actually wanted understanding my intent in this
creation. Yeah, I mean this is something that like is a massive benefit to Astra is like it's way more concise and clear. I don't know what happened like after Opus 5.
Opus 5 was so verbose. Same with Fable 5.1. So verbose I even tell it like be very concise answer in short in plain English and it sends like 10 paragraphs. It's it's insane. I put it in like I put
it on every level of the system prompts and it's like >> yeah I use this agency for example also for growth right and what it does is it basically sees every time browser use is
mentioned on on X on Reddit and so on and people often have like support questions like oh how can I do this with browsers or this doesn't work or how does this and this compare right and it
always basically drafts me answers for those support or growth or sometimes reposts and then ask me hey should I send this out for you should I reply that and yeah with with all the models I
felt okay it's so stupid it's AI slope you can like I would use words I never use even though I say please analyze all emails which I send to analyze like my writing style and you replicate that
right and I just felt okay didn't get it but now with Astra I feel already it's so much closer and like my acceptance rate is so much higher and I think this will be just like in 3 months 6 months
where we will look back on the current state of AI and just think how could we work with something like this like How could we work with suggestions like this? It would be just unimaginable for
us because models will be so well aligned on our intent. >> Yeah. I mean literally I I have this feeling like I was like okay I'm going to balance out you know codeex with
fable and my my croc 4.6 to some easier tasks and then I'm literally after using you know Astra for like a day I'm like struggling to use any other model cuz like no I just don't want this like I
know I should be like optimizing my subscription but like I don't care. I'm just going to buy another subscription, you know, like I don't want this sub-optimal model running on this like
touching on on this. So, this is one thing I definitely feel. One one more thing I noticed is like I had the same same idea for like a new type of infra basically and I sent it to Fable 5.1,
right? And it was like analyzed all the repos, all the context, everything. And they said, "Nah, this is like this is overthinking, you know, this is not related, you know, you shouldn't do it."
I said the same thing to Astra and it was like this is you know okay read everything same context same prompt it's like okay I would do it only inside of this project don't launch as a new repo
like validate it like it's it's a it has potential if in this project to save you time you're already wasting time like doing that specific thing if you build this this could save you time right so
like I don't know how to benchmark this but this is literally like optimism or like mindset where fable 5.1 was like kind of pessimistic you know like shut shut down the idea Astra was like,
"Okay, this has promising. You know, you have that problem. It's documented in those markdown files that you encounter that problem. Let's validate it within this repo." You know, not like [ __ ]
pivot your whole strategy towards that, but like try try to like prove it in a small way if it has potential. And I was like, damn, this is like what if that turns into something and like one model
completely discouraged me and the other one is like told me how to validate it practically like the 8020. >> Interesting. With Fable 5.1, I really like the SVG animation skills. Like in
in my agency where I suggest me tickets, I I often tell it to create like visual graphics. And Fable 5.1 really likes to create SVGs, but not images, but animations. So like a sequence of images
and in a interactive way like a video style SVG from scratch generated, it then explains me problems. for example, my my codebase or oh, this user has those many
runs last week and complaints about this problem. Let's let's fix it for them. That's what I really really like in like this visual explanations of problems for me.
So, what is like you know people went from like markdown files to HTML like where is that going? It's just like a hyper frames animation of you know real-time video like
>> yeah so I tried for example in in agency the Miniax H3 like for video generation because I saw levels io his stream right where you can see all the time new live videos generated from from the model in
live in real time and I thought maybe it's you know maybe this can be like a Tik Tok to get things done a Tik Tok where you don't consume your dreams but a Tik Tok where you achieve
your dreams because if you like something the agent will just implement it your idea and so I thought okay what if if I just have like videos all the time generated who try to explain me
things but I realized for many of my problems like an image is actually better because I can faster understand it >> maybe if the video is like insane it's
just so entertaining and that's more addictive to use you know and you spend that more time on the platform and maybe Yeah, but I think to actually fast understand something often a very very
good image is already enough. But sometimes what's super useful in a video for example is you can guide the attention of the user because in an image you just put everything in,
right? And you need to break it down to like two things. But in a video you have the chance to start somewhere, black out the other and then move kind of the attention of the user to understand an
image. But it's super hard the timing for like right now for AI to understand like you know to understand the human brain how fast will it get when to show the next frame you know how fast can
somebody understand this >> and I think that's like currently images are still better for me but I think if if you use like this human feedback system and agency put this into the
training of how good humans can understand stuff then maybe one day yeah videos it's just so addictive for our brain right and then we will have a Tik Tok to achieve your dreams instead of
pursuing them. >> Surely there's going to be some balance between like this ADHD, you know, switching like quick approval of obvious tasks and like some figuring out a big
problem. It's like maybe we need to pivot the startup, you know, and like you probably don't want to Tik Tok style interface. You want to sit down and like think about this deeply, you know? So,
I'm wondering like will the low-level repeatable tasks like yes, I've done this hundreds of times. is in my sessions, you know, just analyze the correct sessions, whatever, whatever.
Those tasks like again, maybe they run on autopilot. There's some agent unblocking that. Like I I already have my own agent, I call it director, just like unblocking more and more of of of
the tasks like that are just like obvious and that I've done in the past and it's like low risk. But like again, there's stuff that like let's, you know, completely innovate and launch a new
product line. You probably don't want to like a quick swipe interface. You probably want to sit down. It's like, okay, what's happening? What are the products I'm using? What are the
patterns our users are saying? like how you think about this like deep work versus like quick shallow work. >> Yeah, I see this already with agency. Agency is super useful. Like the agent
basically also predicts how long it will take for me to make a decision. Okay. And I'm off effort. I call it effort. I often sort by effort. So the agent says okay this decision will take you 9
seconds. So okay like I love to get those small things super super quickly out. Okay send this support ticket this be our merge merge merge. Right. And then there are some things like oh
rewrite my infrastructure in a new thing to make it completely faster complete rewrite right and just heavy testing lots of things to consider right many many customer paths where I think
like in the future you will have like still like you do it currently a codec session where you give the prompt where you give the intent where you will work with the agent together you say did you
consider this can you test this can you really show me like it's more like one focused long session where you send maybe 50 follow-up messages. This will be like in the current session, okay?
How we currently do it. But for all those small things, you know, an intent, oh, I want to buy a new soccer ball. Okay. Agent suggest you, okay, this one, yes, you know, this will be more like,
but I think most things in life are those small decisions which we do the whole day. Okay. Answering this, this this just hundreds of them. And this is what something like agency where the AI
prompts us will get done. But my question is like why not like you know if you've already done a similar decision 50 times and there's like 99.9% probability that you you like that
outcome and that literally the cost is maybe $20 >> and you know you gave the agent budget of 2,000 >> like why even be in the loop right like
why not turn that into some skill into some process some SOP because it's like building a business if you hire people you know maybe your first employee you train him really like intensely because
like it's your business, it's your baby. But like once you have 100 employees, it's like all right, this is the SOP, you know, just do it like you know, whatever, right? Like it already works
like that in the companies. >> I think you're 100% right. I mean, I'm right now in week two of using agency and we will launch it. Yeah. This weekend or Monday and
it's I'm already giving tasks where I just say, okay, just take care full autopilot. I don't care like you make the decision. I fully trust you. And
this will I mean it's simless with always allow right what we had like six months ago >> where people were still very hesitant until they put like dangerous good
permission I think we'll be very similar you'll just think okay this agent is like better aligned than my employees >> dangerously pursue goal dangerously pursue my dream
>> get the goal done but you understand what I'm willing to do and what I'm not willing to do I'm not willing to do and I message all my my people I Yeah.
>> Yeah. So that's what I'm trying to say like this labeling, you know, maybe like really trying to understand more of your judgment and like I think more of our day will be spent on like this
philosophical questions. It's like would it be okay for this task if I like used your LinkedIn to do that? It's like yes, but don't message them every hour. Message them, you know, every other day.
And every message needs to be like >> or would it be important to ping people? But how often? like yeah can I ping somebody like >> so that's what I mean like this kind of
labeling this like nonobvious you know it's like okay yeah for this bug type type of bug we already had this like hundreds of times in this company we're going to fix it but like this situation
where like you want to do a new growth campaign is it okay to like spend money on hiring UTM creators and you say like yeah sure okay but it's not okay of like you know
>> yeah how much or something like that >> 10k 100k is not okay yeah >> so that's what I mean like I think it's going to be really about like extracting our judgment and our taste And obviously
the AI sometimes will upskill us and say like listen this is not best practices you know when you're coding in R blah blah blah blah blah this is what people do. I was like okay I didn't know that
fair fair enough but like you know when it comes to these practical real world things which again these models don't exist in the real world so they like sometimes lack the obvious things it's
going to be like I think more and more of our time is really creating the systems so that they can run and you can literally label the amount of tasks the amount of work you do in a typical week
and say like this 20% is easy automate that okay what's taking up next week next week okay that is completely automated now I'm spending a lot of time on on these type of software tasks
Surely I can create like couple of skills there or some workflow or some self small software factory and like set that up. So you said you set that up that's running next week it's like in
maintenance mode it's already good right I think this is what our time is going to look like >> and the crazy thing is how fast it will be like I realized this one week now
using agency I it's like my new interface now I spend many hours on it per day but I sometimes still have the habit to for example open Slack okay I open Slack
>> but then I think I don't want to click now through the channels I'm thinking ah luckily agency will take care of it and it just knows everything so I don't need to know everything. It will just know
everything. It will know what's important. It will just suggest me things to do. So, I just have to open agency and I know it's taken care of. And it's just a such a switch where I
where my average hours per week on Slack just drops dramatically because I know, okay, it's taken care of. And it's very interesting what you say on this full autopilot mode like how fast will it
actually go until we full go this on full autopilot? This is very interesting because it this this is tightly connected to our human social pressure, right? You don't want to ping a stranger
like every minute, right? >> Yeah. >> Or like you don't want to ping somebody which you know every minute. Like if you know somebody in real life, you don't
want to send them like AI slop or something. You don't want to >> but if you don't know somebody, >> you care much much less, right? And you think, okay, come this random guy on
LinkedIn. of course like you can just ping them and you know anyways their agent will only read it right so it's just about kind of prompting their agent a little bit and you think okay it's
okay to like ping them and get boosted up and it's I I think it's will happen like in in less than a month that we will run those on full autopilot for many many things for like support many
things for growth it will be so good aligned we'll fully trust it and you know employees they also sometimes s [ __ ] up, you know, instead of 10k, maybe they have some bad experiments, but
overall >> it will be >> yeah, the ROI will be there, you know, if like the if the agent cannot like obviously there's going to be tasks
like, yo, this is bad, like you should have done something else. But like if it saves you 10 hours per day, 20 hours per day, 50 hours per day, like it's the multiple can be more than 24 hours, you
know, like it's crazy. So like yeah, if it [ __ ] up every now and then, you get this opportunity to improve the system, right? to like save it to the memory system, fine-tune the whole machine,
like put some guardrails and like go, let's go again. Let's go faster. >> Yeah. And then and the crazy thing is as soon as you have like one agent working all the time for you, 10 agents, to get
this up to like as many as you want is is not a problem anymore. Like because software just scales, right? And suddenly intelligence just scales for you and you will run out of things to do
very quickly. And then the the and this is kind of what we experience right now, right? Everything is one prompt away. like everything is and but this is a problem because like three like just
three years ago I would have an idea hey I want to build this productivity app where you can put in your goal and people can bet on their goals and you lose money if you don't achieve your
goals right and I have I and I think this idea will change the world and I think okay it will take me like three months and I want to build this one feature this one feature will take two
weeks and there's a lot of wishful thinking you think if I have this feature then the world will love me and the world will be But right now I know I could build this
feature with one prompt and I know okay it will not change the world. And so this visual thinking goes completely away kind of AI destroys this visual thinking which we had in the past for
our side projects and now that everything is so close and you could just build anything it's the question okay what are then actually your dreams like what do you actually want to build
if you can build anything and it's like often it leads to us doing like ah I don't know like I don't want to do this because it's like it will anyways not change anything because this visual
thinking is gone and that's where I think something like agency can step in and suggest you things, pitch you those things, pitch you the outcome. Oh, you would be like
famous if if I do this now for you, right? And then I think this can spark again this visual thinking because the agents promise us things and we can then click yes and yeah, I hope to get this
visual thinking again a little back with this agency. What I lost with previous coding agents. >> Yes. is kind of like basically showing you what's possible, you know, like a
lot of people when they get into a negative state of mind, they don't realize all the possibilities and it's like you could do this, you know, you could meet that person, you could apply
for this better job, you could, you know, start this business, you could, you know, pick up this sport. It's just like the same thing as you said, like maybe there's a crisis of meaning where
we don't know like what to work on and the Asians like, yo, here are all the, you know, unsolved problems. this you know chemistry has this problem that's unsolved for 100 years you know biology
has this problem like there's this problem with like knee surgery where a lot of people have this knee pain and there's unsolved I think we could do progress there and maybe like people who
like traditionally like kind of software developers or entrepreneurs just kind of like go into these industries and these niches and just completely like reinovate everything
>> that's the exciting thing like I think we will solve so many like actual problems you know that that my startup I know doesn't >> that this one person on excess know
about it it's not important But there will so many real problems on the world which yeah will reduce pain, make us more happy, make us more connected which we will solve and improve with that.
>> Yeah, I think that's a great message to wrap it up. >> Where should we send people? >> So first of all like browser use CLI is completely magical. Like I don't use my
Chrome anymore. I can please please please try it out like just tell your agent use browser CLI browser report use it. you will very quickly fully trust it either with your real Chrome like real
browser where you're locked in or with the cloud browsers if you want to paralyze scale up I think that's just magical like filling forms all those things you shouldn't do you shouldn't do
boring things in a browser anymore setting things up setting projects getting API key somewhere testing stuff locally screenshots before screenshots after all those things use browser CLI
for that okay in your cloud codeex that's a that's the first thing then if Yeah, have like products where you want to automate something and the browser is a key part in the in the loop. Okay, I'm
happy if you use browse or browse cloud works really well and yeah just go to go cloud use it and then this agency thing. >> Thank you for your time. I'll link all the products you mentioned below and u
yeah let's do it again in a few months if we don't have secret intelligence by then >> then then we can do