Introduction

Good evening everybody and thank you all for being here. I've been giving talks about AI for about two or three years now. Ten countries, none of them Madrid.

This is my first time in my hometown so I'm very excited about this one.

So what is it about AI? Of course, AI progress, what that means for jobs and for entrepreneurship, agentic AI and how I learned to stop worrying and love the claw.

so who am I why should you listen hi I'm Daniel Duma I have a PhD in AI I used to hang out at Google and eat the free food they called it a job they paid me X Alexa founder of many startups on the right hand side you can see my academic track record my citation count if that means anything to you but I think much

better credentials and these are selfies so the man on the left hand side this is Jeffrey Hinton the winner of the Nobel Prize two years ago for his contributions to artificial neural networks and the man on the right hand side will feature in slides later on is the ceo and founder of nvidia the world's most valuable company because they make the chips on which the ai runs and this guy might not need an introduction here if we're going

to talk about sam altman i met him in 2017 he came to google gave a talk we had a little chat afterwards but i was a bit shy back then or maybe i was starstruck you know from meeting him and i never asked him for a selfie. So this is the one that never happened.

Thank you AI for solving this big pain of mine.

AI progress and why it matters now

So if you talk about Sam, I want to talk about Chad GPT or as I call it, GP. Now GP is many things to many people.

It's your lawyer, your doctor, your friend, your cousin, your girlfriend, whatever you want it to be. But it is also one more thing for all of us and that is the

The canary in the coal mine: crossing the exponential curve

canary in the coal mine it is the thing that happened that tells us we have passed the elbow on the exponential curve of progress this is a comic from 2015 if this was 2026 and edited by ai it would look a bit like this and i know it sounds exaggerated i know this curve looks like i mean it's hand -drawn right it doesn't it cannot be real right well let's compare it to this real graph

of the capabilities of large language models at executing tasks that take humans a long time this is long running agents right so on the y -axis you have the time they would take a human to do this on the x -axis you have the date the time and these dots are different models and as you can see this is exponential

progress not only this it is accelerating it doubled recently in in terms of speed of progress, that line is becoming straighter and straighter up, and we expect month -long tasks, human equivalent time, by 2027.

Not only this, we can give these models the same IQ tests that we give humans to see how smart they are, and this is a bell curve distribution over human intelligence, and you can see that the frontier models are all pushing genius level.

From benchmarks to real-world computer use

Beyond that, they're gaining new capabilities. this is cloud computer use using a computer it's googling stuff it's clicking on buttons filling in boxes and doing tasks with the UI that a human would use to do a very mundane tasks but the point is this is becoming possible

AI and the future of work

now so should we all wonder then does this affect me the writing is on the wall so let me tell you a little story I used to hang out at Google as I

I mentioned before, I had the extreme privilege of getting up to that mic every Thursday and asking the founders of Google on stage annoying questions. And one of my annoying questions

was, as a software engineer, I see it as my job to automate all jobs, by which I mean human jobs. And I used to believe that being the one automating my job would be the last one to go.

But it has recently dawned on me that maybe we are better automating ourselves ourselves than we are at automating anybody else, whose job we don't actually understand. So are we putting ourselves out of a job?

Automating ourselves: from optimism to layoffs

And their answer was, don't worry, your jobs are safe for many, many years. Many was exactly six, because that's when the layoffs started. And they continue to this day, Block just fired 40 % of its staff.

Firing vs hiring: what the employment data suggests

Now, I do realize that we had a pandemic. and so that explains this curve over here on this graph which shows you a time when everybody was hiring and nobody was firing and so afterwards there was an overcompensation for sure that's definitely a big part of it people are expecting a recession another part of it but maybe another

part of it is not the firing but the hiring so this is a graph that came out two weeks ago the red line is the one that matters this is year -on -year change in u .s tech employment u .s only tech only the red line has only gone after zero sorry under zero so more

firing than hiring three times in history one was the dot -com bubble one was the Great Recession and one is now is this the AI apocalypse I don't know

How coding changes when you manage agents

but I do know if the video plays that this is what coding programming looks like this is what I used to do this is what I do now this is my actual screen

they're recorded with my phone these are three agents coding and I just watch them give them instructions manage them review their work how many of you have used any of the software's on this screen Oh amazing excellent almost everybody then I don't need to explain to you that coding is now talking to the software to the machine telling it what you want and iterating over that is it

Beyond software engineering: automation across professions

only coding according to anthropic the red bars is what they see as covered in terms of AI automating work and the blue bars are what's left so of course there's software engineering but there's business and finance and office and legal and everything else so according to them there's more disruption to jobs coming from the AI that we already have this is not future capability this is

right now and that leads me to this thought which is that soon this will be be true or maybe not soon but at some point and that is that there is nothing that a human employee can do that a machine can do better faster cheaper and more reliably and if that is true stop hiring humans this is a real ad

campaign they love their hype in San Francisco and your photo to take away with you from this presentation is a homeless person sitting on market street next to a billboard that says stop hiring humans now you might think that's a problem and I would agree with you but I also think that if you look deep enough there might be a big opportunity inside there and I think that

The opportunity inside disruption: entrepreneurship

opportunity is entrepreneurship it's always been hard to be an entrepreneur I don't know if any of you in the room entrepreneurs around okay you guys know as well as I do and potentially better that it's been really really hard to do

this job because you have to do everything you have to hire people if you don't have to do the job yourself you have to constantly be motivated activated etc it might not be for everybody but maybe AI is here to help us by becoming our co -founder if you remember the guy from slide number four

The one-person billion-dollar company idea

Sam says soon it will become possible for one person to start a billion dollar company this is a tiny little example of what that might look like here's an

Example: an automated Etsy-style business

agency that made an AI automation an agent this is pre open claw where it And it would research products on its own. It would create these products, create these designs, upload them to a print -on -demand partner and sell them on Etsy.

And it made money. It made 42 % profit margin. So here's an automation that would make a business. Now this was 2025, I know, so long ago.

You would use something like NA10 or Make .com. I just found out yesterday that Celonis, where we are, owns Make .com. Anyway, you would use some kind of automation. But that's so last year.

OpenClaw and the rise of agentic AI

What is the new thing? well hot bot mold bot open claw thank you for checking how many of you have used this before

now for those who don't know what this is it's the most popular open source project probably in history the fastest growing and github stars for sure that's like a vertical line basically and it is also in every single headline that you read basically everything is about open

claw right now and the hype is unbearable honestly it's like thousands of people are queuing up in china to install this software on their phones because they think this agent is going to work

for them and do all kinds of things now so what is it what does open claw do how does it work so

How OpenClaw works: state, memory, tools, and an LLM brain

this is the open claw part the gray bit in the middle what is it it's a software you install on your virtual private server your own laptop your phone wherever you want that essentially keeps state memory remembers what you talk to it about it keeps configuration

for API services etc and the brain is of course an external LLM whether it's open AI Claude or maybe you self -host it with a llama it connects to external api's you could connect your Google Docs for example and read what's in your drive and it could use a browser this is not like a novel capability but it uses

what's already out there for browser use so you could pretend it's you and use websites on the internet for you and then you talk to it over slack telegram or whatsapp and

Heartbeat and the “it texts you” effect

it has this wonderful thing they call heartbeat which is approximately every 30 minutes there's a process that runs where they run a prompt with an lm and says check what we're doing is there anything we should notify the user about so basically it texts you which makes it feel amazing and agentic and everybody loves it and that's why it's this big hype is that it texts

you first we could talk about the architecture but i don't think we have the time for that

Common use cases: personal admin, content, and dev workflows

But what can you use it for? This is what everybody keeps asking me about.

I think the main use case is this side. It's personal admin. So it gives you daily briefings on news, on your upcoming events, on whatever is interesting to you, triages your emails,

tells you what to reply to, what's important, maybe gives you some draft of it.

People say it helps them with their flight check -ins. I doubt that, but okay.

You could use it to generate slop and fill the Internet with rubbish. Please don't do that.

or you could use it to handle your building your development your deployments with AI I found a

different use case or I thought I will try it anyway which is what if I go to it and say make me a start -up what make it yourself pseudo make me a start -up okay then so installing open claw

Setup in practice: skills, browser automation, and chat interfaces

and you want a computer whether it's your laptop whether it's a VPS whether it's a Raspberry Pi as i mentioned once you've installed it there there's a web ui that you can connect to directly but ideally you want to talk to it on discord and telegram on whatsapp and then you want to give

it skills capabilities you want to give it brave search xiai parallel something to search the internet and scrape pages then you want to connect it to a browser so you can do browser automation stuff so this is the basic which is just like runs a local browser but the amazing thing is laptop.

So it pretends it's you. Even if it's running on a remote server, it's actually using your Chrome on your laptop remotely.

And then yes, you talk to it over WhatsApp. There's a lot to the setup of it. I won't go in depth.

All you need to know is you can copy and paste the prompt for Claude and he will set it up for you. So we have a gentic AI to do that for us.

Resources: prompts and community documentation

And if you are going to take away one QR code from this presentation, this is probably the one.

There's this incredible GitHub repo with everything you need. Every question you have, how to set it up, how to configure tools how to do the browser use anything everything's in there the official

documentation is rubbish use this instead so when you're all done

Experiment: using an agent to find and build a startup idea

anyway so i went and told it okay your mission is to search the web every morning give me a list of great business opportunities that i could make as a sas app or as a mobile app this is all a conversation we've been having i picked one this one that i liked which is

The concept: improving restaurant delivery photos

imagine you're a restaurant that does food delivery and you need to take photos of your dishes to put them on the delivery app and maybe you're not great at taking photos and maybe they look like crap and maybe nobody buys from you or you have low reviews because of that so it wouldn't be great if you had software that would improve the photos of your restaurant automatically

and how do i sell this to you i will find your socials well i won't open claw will find your socials on the internet message you directly with an improved photo saying i love this dish look look, here's an improved photo. Would you like me to do your whole menu for 50 bucks? And then you make a SaaS app out of that.

I was like, I like it. This is a cool idea. All right, let's do it. So how did that go?

Well, installed it. Internet access. Here's your mission. Cool.

What worked: tracking, research, and scraping with APIs

Proposed options. I picked one. I said, create a database to keep track of all the things that we've done. And it did. And I thought that was amazing. It created a local SQLite, you know, like real access. Awesome.

Then it searched the internet for market size and stuff. and then he said but actually to really search the internet i need a proper api there's this thing called apify give me an api key to that all right boom i went signed up gave it an api key so i went and scraped google places and scraped instagram and other websites because that's what this

What failed: poor leads, existing competition, and premature pivots

provider does so amazing right but the restaurants that i found were tourist traps so i took the first five of the list or something not actually delivery places that even do delivery okay i checked delivery apps myself every single photo was beautiful every photo

was amazing already which means either delivery apps are doing the photos in -house or they're improving them with a as well but it's already in the in the final result in any of a check for competitors it never occurred to it that it might exist already it does there's dozens of exactly this that's out there

so probably just rehashing an existing idea and when I told it look there's all of these things that you should be aware of he said oh okay let's drop it let's pivot away let's do something entirely different which is later hold on hold on i haven't fully um uh invalidated

Where agents fit: from messy exploration to repeatable automation

this idea so what did i learn well if na10 or make is your factory that does something repeatable uh measurable you know that's you can run you know on a conveyor belt open close your plumber that does unscalable non -scalable things bespoke solutions to very specific problems so for that is great

and i thought well where does it fit in the startup process startup process starts as a squiggle it ends as a less squiggly line right well you could start ideating with gp or gemini whichever you prefer and come up with some ideas you could then try the concept as i was just doing with openclaw see what apis exist how you could connect it what it would look like and once you

1Once you have something repeatable that works again and again, then you build an automation with something like an HTML make.

Then of course, question mark, profit! There's your company right there. Some of you might have got that reference.

Reality check: reliability, security, and the cost of supervision

So yes, we like to model processes as a straight line, but the reality is spaghetti. Everything is spaghetti and OpenClaw is also spaghetti.

Why today’s agent stacks still feel like spaghetti

The memory management is terrible. the conversation session management is terrible you forget stuff I once told it do this thing and he said yeah I'll do it and notify you nothing happens I'm like hey did you do it do what like that was literally a conversation I had all

right then browser automation doesn't work like it just you watch it's you know headbutt a wall for hours while wasting your tokens everything is terrible configuration is bad so why is it cool then I think it's really cool

still because this is not a software per se it's a platform that you can tinker with you can create your own thing but I think beyond that the most important part is it's a pattern it's a pattern that now everybody else is forking you take this thing you create your own or you start from scratch you take the ideas and say I'm going to make it better and you implement in a different

language for a different platform people are doing it for the meta glasses whatever basically the idea of your own assistant that has local memory that you manage blah blah this is now going to stay this is now here for us so much so that jensen from slide three is just announced that two days ago that they have their own fork of open claw nemo claw so this is one that i found that i really like and i think this is what it will look like

eventually you still have a ui you still have an interface that a human can use but it has the agent built in so this is a crm that messages people for you on linkedin whatever but it also

Security and trust: when agents delete the wrong things

has the agent in it now two words of caution number one security this is Meta's head of AI alignment who recently went viral on all the media because she gave her inbox to opencloth to manage it proceeded to delete everything and then she was like no stop opencloth stop and it didn't and so she ran out of mail and

then cloud code recently of course deleted one of the many databases in production that it tends to do and in summary that means you cannot trust these models.

AI burnouts: why agentic workflows can intensify work

1They need continuous supervision, intervention, and review of their work. That means you won't have a single moment to breathe.

That means your agents are always hungry. They're always waiting for your input, for your instructions, for your review.

So AI doesn't reduce work, it intensifies it. Welcome to 2026, the year of AI burnouts.

How to adapt: the human skills that remain essential

Now, whatever we talk about we said a progress exponential maybe it's accelerating humans need to move up the value chain from doing from implementation to orchestration it's never been easier to be an entrepreneur which means we'll have a lot more of those and the open claw pattern shows us the way

Judgment, taste, and agency in an agentic software world

that software is going to go all software will become agentic ai agents can't be trusted though within a decision and so human modes remain your taste your judgment your trust your agency

You know what is good, what is appropriate. You can be trusted and you want to do stuff. You're the one who initiates stuff.

But yes, your work has intensified, not gone down.

Five practical recommendations

And what would I recommend to anyone worried about their jobs and how AI will impact that? I think these five main things.

One, play with it. Learn and tinker. It's easy. It's accessible to anybody.

Definitely worth trying that. Develop taste and judgment basically means become a senior. If you're a junior, find a way to become a senior quickly in whatever domain you're in.

Build domain depth, become an expert at something specific, find a niche, own it, and build a reputation about it. Talk about it. Talk about it on the internet.

Not with like AI slop. Just share your experience, what you learned, who you've met, whatever it is.

Conclusion

And finally, build your network. And so add me on LinkedIn and I'll build my network. Thank you all for your attention.