Alex says his first company, Virtual, tried to sell customer service chatbots in 2002, before many people knew what a chatbot or live chat was. He prefers working on early ideas because his strengths are product and early market development, while founders focused on sales and marketing may need a market-ready product.
What Big Tech Missed And How Startups Can Still Win
Speaker
Alex, founder of Wit.ai and AMI Labs CEO, discusses world models, startup risk, and building ambitious companies with a Y Combinator host in Paris.
Alex says AMI Labs is building a foundational model that learns from direct experience, such as video, audio, and robotic interaction. He contrasts this with LLMs, which learn from text about the world. He expects LLMs to remain strong at language, math, and programming, while world models may suit noisy, high-dimensional, long-horizon tasks such as robotics.
Alex argues that large companies have resources and brand recognition, but startups can take risks that are difficult inside established organizations. He recalls a Meta chatbot that produced an inappropriate response, making release a legal concern. He says OpenAI was able to keep taking risks as it developed GPT models.
Alex says AMI Labs needs specialized talent, data, and compute, and that securing GPUs is difficult even with money. He describes raising €1.2 billion for the company and says the public expectations that come with that funding can be harder to manage than dilution. For early founders, he recommends choosing a narrow problem while holding a big vision within it.
Takeaways
- Match the timing of your market to your strengths in product or go-to-market.
- Use LLMs for language-first tasks and consider world models for tasks grounded in real-world experience.
- A startup can use its ability to take risks as an advantage over larger companies.
- Secure data and compute alongside talent when building foundational models.
- Start with a narrow customer problem and communicate the larger ambition behind it.
Mentioned in this talk
Search inside this talk
What was claimed 8
- uses
Waymo uses World modelsml“Some flavor of it.”
YouTube Verified3 sources· Talk, Job posting· archived
- acquired
“you actually sold Wit shortly after to Facebook”
YouTube Unverified1 source· Talk
- uses
AMI Labs uses World models“a big new approach to AI”
YouTube Unverified1 source· Talk
- featureWorld models: Learn directly from the real world through video, audio, sensory data, and interaction with objects
“We learn directly from the real world, like humans and animals, through video, audio, sensory data, and playing with objects.”
YouTube Unverified1 source· Talk
- featureWorld models: Expected to use fewer parameters and have less expensive inference than large language models
“You have fewer parameters, so inference should be less expensive.”
YouTube Unverified1 source· Talk
- featurelarge language models: Suited to language, mathematics, programming, and other sequences of discrete low-dimensional symbols
“LLMs are designed for that.”
YouTube Unverified1 source· Talk
- invested inYann LeCun invested in Nabla
“Yann was an investor and an advisor at Nabla”
YouTube Unverified1 source· Talk
- role atYann LeCun is executive chairman atAMI Labs
“Yann is the executive chairman of the company.”
YouTube Unverified2 sources· Talk
