AMI Labs CEO Rejects AGI Label, Courts South Korea
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Alexandre LeBrun, CEO of AMI Labs (founded by Turing Award winner Yann LeCun after he left Meta), rejects the AI industry's preferred terminology, saying 'There's no good definition. What is superintelligence? I don't know. It's not a very useful word.'
- AMI Labs raised $1.03 billion in March at a $3.5 billion pre-money valuation, but remains pre-product with no committed timeline — 'We'll make a surprise when we're ready,' LeBrun said.
- LeBrun was in Seoul for the International Conference on Machine Learning (ICML) scouting industrial partners for world models — AI systems that predict the next state of the physical world rather than the next word.
- LeBrun frames world models and LLMs as 'complementary, not replaceable,' comparing LLMs' limits in medicine to 'a doctor trained only on textbooks and without a residency' — useful, but covering 'only 1% of healthcare.'
- LeBrun is targeting Korea for its robotics, semiconductors, and manufacturing base — sectors 'the first wave of AI barely touched' — and its track record as 'the fastest adopter of the internet 25 years ago,' calling the combination 'unique.'
- SBVA CEO JP Lee, one of AMI's Asia backers, pointed to Seoul's June plan to mobilize roughly $880 billion for chips, AI data centers, and physical AI as evidence Korea should 'keep investing in physical AI, too.'
- LeBrun says current robots remain 'really dumb in the physical world,' citing a public event where a dancing, kung-fu-performing robot approached and kicked a child — the hardware is 'incredible,' but 'there's no brain.'
Why it matters: AMI Labs is betting its $1.03 billion that useful AI runs through physical-world prediction rather than language — and that Korea's hardware-heavy base makes it the ideal launch partner. LeBrun's refusal to engage with hype vocabulary positions AMI against OpenAI, Meta, and xAI's branding wars while it seeks the real-world training data lab-based models can't produce.
