What Big Tech Missed And How Startups Can Still Win
In a Nutshell
Alex built Amilabs on the contrarian bet that world models—AI trained directly on video, audio, and sensory data—will outperform LLMs on real-world tasks like robotics. To fund the massive GPU spend, he raised Europe’s largest-ever seed round of $1.2B, but warns the real cost is sky-high expectations, not dilution. His core advice for founders: pick a narrow problem yet keep an ambitious long-term vision, because big companies won’t take the crazy risks that startups must.
These notes were generated by AI and may contain inaccuracies.
The real cost of a 1.2 billion raise is not dilution but expectations. Investors understood and agreed on the long-term vision, but the expectations created in the outside world are dangerous. If a company raises 20 million and does nothing for 2 years or people don't see anything coming out of it for 2 years, it becomes very hard to survive.
Alex and the speaker did the same YC batch in winter 2014. Alex was building Wit AI while the speaker was building Algolia. Alex has done four companies compared to the speaker's single company.
Wit AI was the first .ai domain. Buying .ai domains was very difficult at the time because .ai is an island off Australia. Alex found the domain and thought it would be cool to have it, but nobody understood it was a domain back in early 2013.
In 2015, Alex received an email from someone called Mark Zuckerberg. After half a second, he deleted the email without opening it, thinking it was a scam or spam. After about half an hour, he thought it might not be and opened his spam folder to find it was real.
The YC playbook states that if you don't want to sell, don't speak with anyone. As soon as you start a startup, you receive lots of offers to talk, but usually these are from corporate development of large companies who need to talk to everyone to understand the market and determine if you're desperate enough to sell for a low price. It's a very bad idea to waste time and attention.
The rule book states if someone really has the power to do something big, you can talk. The Facebook outreach may have been smart because they got Mark to send the email directly.
Virtual was Alex's first company, started almost out of school in 2002 as a chatbot company. It was 20 years too early. They made something with it but spent 10 years trying to sell chatbots for customer service. People would ask what a chatbot is, and when told it's like live chat but automatic, they would ask what live chat is.
The market became nascent in 2010, but the company was still too early.
Whether being early is advisable depends on strengths. If co-founders' strengths are go-to-market, sales, and marketing, they shouldn't be too early because the product and market need to be ready to receive the go-to-market approach. Being too early will kill you if the strength is pushing a product to the market.
Alex's strength is taking very early tech and thinking about what product to build, how to sell it, building the product, and understanding the early market. The skills are more on the early side of go-to-market.
Amilabs is a normal project but very expensive because building a foundational model requires buying thousands of GPUs, needing billions of euros. Alex has done several startups in applied AI, using machine learning research or models that other people have done. Having the opportunity to work upstream on the solution or model that millions of people will use if successful is very exciting. Moving upstream is the most exciting aspect.
World models differ from large language models because LLMs learn to manipulate language by reading all texts, books, and internet content in the world. LLMs learn what humans have written about the world as a proxy, not learning directly from the real world. An LLM is like someone who has never gone outside, never left the room they were born in, but read all the books every day for centuries.
A world model is a model where learning happens directly from the real world like humans and animals, through video, audio, sensory data, and playing with objects, starting from scratch like a baby without any language.
Training data includes a lot of video data because it's the easiest to find, but it works with audio. It would work if someone was born without sight or hearing because they can still live through touch, and the same applies to the models.
Touch data is collected through robotics.
In the short term, applications exist especially in robotics. Current robots are very narrow vertical robots that do one thing in one position. It takes a lot of time to train them to do only one thing always the same. The hardware is incredibly evolved but the brain hasn't changed, and these robots cannot be taken into an open environment whether in houses or streets.
Videos show robots starting to dance and breaking everything around or doing karate demonstrations near children, demonstrating that robots are very dumb in open environments.
VLA (Vision Language Action) is a very bad hack where you have a hammer and try to see everything as a nail, using a solution for things it was not designed for. VLAs are not accurate, very slow, but these applications need very fast short latency, and they are very expensive because you cannot pay $100 of compute for 24 hours of robot home robot work.
World models are very costly to train like LLMs, hence the GPUs needed today. The cost is comparable, though hard to compare because the structure is different. When trained, they are supposed to be more lightweight with fewer parameters, so inference should be less expensive.
LLMs are already very good for everything that is language first, mathematics, and programming. These are all low dimensional sequences of symbols. LLMs are designed for that and will never be beaten at those tasks. World models will be vastly superior for high dimensional, very noisy, very long horizon problems.
Some ideas behind world models are already in production. Some autonomous driving companies are using some flavors of world models. Waymo is using some flavor, and Wave in the UK as well. Some of the ideas being used like JEPA have been around for at least 5-10 years.
In 2018, the theoretical idea behind LLMs like Transformers was in a paper, but Google did nothing with it. Nobody realized the power. OpenAI took the idea, scaled it, and solved some engineering scaling issues.
All large companies back then had reasons why there are things they cannot do that startups can do. Startups can take lots of risks. If you don't take any risk, you're dead because big companies have more resources, a brand, and many things you don't have. The one thing startups have that big companies don't is the ability to take crazy risks.
At Meta, they worked on a chatbot that was a language model, trained statistically without rules but from statistical data. The first time they launched this chatbot in the lab at Meta, they asked how are you today and the chatbot answered, "Oh, just chilling watching porn on my couch." Alex wanted to release this chatbot but realized the legal team would never want to have this discussion.
OpenAI could take risks with GPT-1, 2, 3 until ChatGPT that no one else could do because they had enough money and resources to scale the training.
Three things are needed: smart people and talent because not so many people understand this and most are working on LLMs, a lot of data, and GPUs/compute which is very hard to secure even when you have money.
The bet on world models was maybe contrarian a year ago but is becoming less and less. A year ago when saying LLMs are nice and powerful but they're not AGI and don't lead to artificial or super intelligence, most people disagreed. Now a good chunk of people agree that LLMs don't lead to general intelligence.
The definition of AGI has changed, with people who promised AGI changing the definition and stopping using the term, which is a good sign that something is not working.
Yann was an investor and advisor in Nabla, the company prior to Amilabs. Alex saw him very often. One day Yann came to Alex and said world models are ready to become very big and they could go faster if they do that outside of Meta as an autonomous startup. Yann wanted to do it and needed a CEO.
After 10 years trying to teach a machine to understand human language and concepts for customer support, the conclusion was that we cannot cheat. A machine can speak about a cat, but if the machine has never seen a cat, touched a cat, played with a cat, or maybe been hurt by a cat, then it will never be very smart. When humans talk about a cat, it reactivates in our brain some past experience we had with cats. This is grounding.
Even if you read a book about animals like dinosaurs that you've never seen, the way you imagine a dinosaur is built on basic objects that you had experience with. You need all your experience of things to be smart.
Yann is the executive chairman of the company with scientific vision, leading the direction of the research. Alex's job is to make sure there are people, electricity, GPUs, data, and offices. Alex loves this.
Managing researchers is very hard, like herding cats. You cannot tell researchers what to do because they hate that and will flee if you try. But if you leave researchers alone, they will just do very random things and run in all directions. Finding the very tight balance between direction with enough freedom to get a product one day is what Alex likes to do.
Amilabs raised the biggest ever seed round in Europe at 1.2 billion. No one could have imagined a few years ago that any startup would need so much seed money. The craziness is the cost of compute because you need a lot of compute, you need a lot of money, and everything has shifted.
When starting a company, having a culture of not spending too much money, being very careful, and spending money as if it's yours is very important. It's very hard to change later if you don't start like that. But it's hard because if you spend 1 billion for GPUs, then everything else looks very cheap.
Finding the money was not that hard. The rule is the difficulty doesn't depend on the amount. It depends on if it is your first time. First time you raise money for your first startup is hard. Second time is easy and then after third time people just beg you to give you money to do anything you want. You have to go through this journey.
The real cost of the 1.2 billion was not dilution but expectations. Not only from investors who understood and agreed on the long-term vision, but the expectations created in the outside world. If you raised 1 billion and do nothing for 2 years or people don't see anything coming out of it for 2 years, then it's very hard to survive.
There is an internal timeline but they don't communicate about it.
The company operates in Paris with a big team in New York, one in Montreal, and one in Singapore. These four cities were chosen because there were existing teams. There is no San Francisco office.
When people join Amilabs from San Francisco companies, they have to move to New York, which is a commitment. At this early stage, it's important to have people who join with real commitment and not just because they want to try. There will be an office there one day, but it's on purpose.
Being connected to San Francisco is maintained through frequent visits.
We are never ambitious enough, especially in France and Europe. We don't take risks.
One month after being acquired by Facebook, Alex wanted to hire 100 people to do data annotation for machine learning. He wanted to hire 100 concierges who work for real customers to book travels and record everything they do to train the model. This was in 2015, 10 years ago.
Alex prepared a memo to ask permission from the cab board to hire 100 people. Facebook had 6,000 people back then. When meeting the cab board, memos must be sent 24 hours before or the meeting is canceled.
Alex arrived ready to argue with scientific arguments and Mark said he had already read the memo and loved the idea, thinking they would kill Google by capturing intent with advertising. Mark suggested hiring 10,000 concierges, not 100.
Alex had to tell him not to do that, saying he wasn't sure it would work and there were lots of unknowns. To this day Alex doesn't know if Mark was really willing to hire 10,000 people or if he thought it was the best way to get the real information about what could go wrong in the project. Alex loved the ambition and they hired 200.
Mark has been well known for taking wild crazy bets on stuff. Sometimes it works, sometimes it doesn't. So, not that surprising he would go there, but yeah, kind of probably erased the bar from what you think is ambitious.
When starting out, entrepreneurs should look at a very narrow problem, but be very ambitious in this narrow problem. The one thing many founders do wrong is wanting to solve a chatbot that does everything for everyone, so not narrow, very hard to provide value as a very small team small startup.
If you choose a very narrow problem like doing one thing for one industry or for one set of customers, but in this narrow tunnel have a very big vision. Don't be afraid to share your long-term vision. Okay, I'm doing this today, but this is what we want to do later.
If you're very ambitious you are in a very wide field, then you're not credible.
In 5 to 10 years, there will be robots, helpful robots, actually your helpful robots. Most jobs will still be here, but not as dangerous or hard to do as they are today. Machines with some sort of common sense will make a big difference.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
