Alex on Funding, World Models vs LLMs, and Startup Ambition

Alex, a serial entrepreneur, shared insights from his extensive experience in AI and startups, including the founding of Wit.ai, Nabla, and his latest venture, Amie Labs. He discussed the challenges and opportunities in the AI landscape, particularly contrasting Large Language Models (LLMs) with "world models."

The Real Cost of Funding: Expectations

Reflecting on Amie Labs' record-breaking 1.2 billion euro seed round, Alex emphasized that the true cost isn't dilution but the immense expectations it creates. While investors understood the long-term vision, the external world's expectations are a significant pressure. He noted that raising substantial capital without visible progress within a couple of years makes survival incredibly difficult.

A Journey Through Early AI Ventures

Alex's entrepreneurial journey began with Wit.ai, a company he co-founded that was acquired by Facebook in 2015. He humorously recalled almost deleting an email from Mark Zuckerberg, mistaking it for spam, before realizing its authenticity. This led to a discussion about the Y Combinator playbook, which advises against engaging with corporate development unless actively seeking an acquisition, a rule that can be bent when a powerful figure like Zuckerberg is directly involved.

He highlighted his tendency to build companies that are "too early" for their time. His first company, Virtual, started in 2002, focused on chatbots for customer service. At the time, people didn't even understand what a chatbot or live chat was, making it a decade-long struggle before a nascent market emerged around 2010. Wit.ai, while also early, was not as prematurely launched as Virtual.

Alex believes his strength lies in taking very early technology and figuring out how to productize it and understand the nascent market. This contrasts with entrepreneurs whose strengths are in go-to-market strategies, for whom being too early can be detrimental.

Amie Labs: A Bet on World Models

Amie Labs represents Alex's most ambitious project to date, aiming to build foundational "world models." This endeavor is exceptionally capital-intensive, requiring billions of euros for GPUs to train these models. After years in applied AI, working upstream on foundational models that could impact millions is a deeply exciting prospect for him.

World Models vs. LLMs

Alex explained that LLMs learn by processing vast amounts of text, essentially learning about the world through human-written accounts. This makes an LLM akin to someone who has read every book but never experienced the world directly, leading to limitations in common sense and novel problem-solving.

In contrast, a world model learns directly from the real world, much like humans and animals. This involves processing sensory data such as video, audio, and touch, starting from scratch without relying on language. The training data primarily consists of video, but it can also incorporate audio and even touch data through robotics.

Use Cases and Advantages of World Models

In the short term, world models have significant applications in robotics. Current robots are highly specialized and struggle in open, unpredictable environments. World models aim to provide robots with a more generalized understanding of the world, making them safer and more useful in diverse settings.

Alex criticized current approaches like Visual Language Models (VLAs) as "bad hacks" that are inaccurate, slow, and expensive for real-time robotic applications. World models, while costly to train (comparable to LLMs), are expected to be more lightweight and less expensive for inference due to fewer parameters.

While LLMs excel at language-first tasks, mathematics, and programming (low-dimensional, discrete symbol sequences), world models are envisioned to be vastly superior for high-dimensional, noisy, and long-horizon problems that require a deep understanding of the physical world.

Current Status and Bottlenecks

While some ideas behind world models are in production (e.g., in autonomous driving like Waymo), the technology is still nascent. Alex drew a parallel to the Transformer paper in 2018, which laid the theoretical groundwork for LLMs but wasn't fully realized until OpenAI scaled it. He noted that large companies often struggle to take the necessary risks that startups can, citing an example from his time at Meta where a chatbot's inappropriate response highlighted the challenges of deploying such technology in a large, risk-averse organization.

The main bottlenecks for world models are: 1. Talent: A scarcity of experts who understand this specialized field. 2. Data: The need for vast amounts of real-world sensory data. 3. Compute: Securing sufficient GPUs, which is challenging even with significant funding.

The "Contrarian" Bet and Ambition

Alex doesn't view Amie Labs as a contrarian bet anymore, as the consensus is shifting away from LLMs as the sole path to Artificial General Intelligence (AGI). He believes that the definition of AGI has been conveniently altered by some who previously promised it through LLMs, indicating a recognition of their limitations.

He co-founded Amie Labs with Yann LeCun, who was an investor and advisor in Alex's previous company, Nabla. LeCun approached Alex with the idea of pursuing world models outside of Meta, believing it could accelerate progress. Alex, having spent years trying to teach machines human language, concluded that true intelligence requires grounding in real-world experience, not just textual knowledge.

Their roles are clearly defined: Yann LeCun, as executive chairman, leads the scientific vision and research direction, while Alex focuses on execution, ensuring the team has the necessary resources (people, electricity, GPUs, data, offices). He likened managing researchers to "herding cats," emphasizing the delicate balance between providing direction and allowing freedom to foster innovation.

Raising Capital and Geographic Strategy

Amie Labs' 1.2 billion euro seed round was the largest ever in Europe. Alex admitted he never imagined such a sum would be needed for a seed round, attributing it to the escalating cost of compute. He stressed the importance of a frugal culture in startups, even when dealing with large sums, to avoid complacency.

He noted that raising money becomes easier with successive successful ventures. While the first startup is always the hardest, subsequent successes build credibility.

Amie Labs operates out of Paris, New York, Montreal, and Singapore, deliberately avoiding a primary presence in San Francisco. This decision serves as a commitment test for new hires, ensuring they are genuinely dedicated to the project. Despite not having an SF office, Alex and his team maintain strong connections to the Silicon Valley ecosystem.

The Importance of Ambition

Alex urged aspiring entrepreneurs, particularly in Europe, to be more ambitious and take more risks. He recounted an anecdote from his time at Facebook where Mark Zuckerberg, upon reviewing Alex's proposal to hire 100 data annotators for Wit.ai, suggested hiring 10,000 instead, demonstrating a vastly different scale of ambition. This experience profoundly impacted Alex's understanding of what true ambition looks like.

He advised young founders to focus on solving a very narrow problem but to be incredibly ambitious within that scope. Instead of trying to build a product that does "everything for everyone," he suggested concentrating on one thing for one industry or customer segment, while still articulating a grand long-term vision.

A Glimpse into the Future with Amie Labs

If Amie Labs succeeds, Alex envisions a future with helpful robots that can assist with tasks, making many jobs less dangerous or arduous. He believes that machines equipped with common sense, derived from world models, will fundamentally change the world.