AMI Labs co-founder uses ICML keynote to challenge LLM dominance

The main conference program at ICML 2026 in Seoul opened on Monday with an invited talk that took direct aim at the dominant paradigm in artificial intelligence. Pascale Fung, co-founder and Chief Research and Innovation Officer of Advanced Machine Intelligence Labs, delivered a keynote titled "Towards AI Agents In the Real World," arguing that large language models fundamentally lack the capacity to understand physical reality.finance.biggo+1

A New Architecture for Intelligence

Fung, a fellow of the IEEE, ACL, ISCA, and AAAI who previously served as Senior Director of AI Research at Meta ↘ 1.78% FAIR, used the address at the COEX Convention & Exhibition Center to make the case that even frontier models like GPT and Claude perform well below human baselines on physical inference tasks. She called for a shift toward "world models" — AI systems that incorporate perception, prediction, planning, and memory to operate in the real world rather than merely predicting the next token in a sequence.builtin+2

The talk reflects the core mission of AMI Labs, the Paris-based startup she co-founded with Turing Award winner Yann LeCun after both departed Meta in late 2025. Officially launched in March 2026 with more than $1 billion in seed funding at a $3.5 billion valuation, AMI Labs is building AI that learns abstract representations of reality through raw sensory data — images, video, audio, and LiDAR — rather than human-written text.cathaycapital+3

The Scaling Debate Intensifies

The keynote lands amid an escalating intellectual dispute over whether scaling language models can lead to general intelligence. OpenAI CEO Sam Altman has pushed back against such skeptics. At a Stanford talk in June, Altman said "a whole generation of researchers" slowed the field "by underestimating what scaling could do". His comments came as OpenAI prepared to release GPT-5.6, a model family focused on science, coding, and biology capabilities that entered a limited government-approved preview on June 26.techblit+2

Two Visions for AI's Future

LeCun has long maintained that text-only training cannot yield genuine understanding. "It's kind of a slow way," he told the Observer in June, referring to the strategy of simply building larger language models. AMI Labs explicitly rejects the assumption that scaling LLMs will produce artificial general intelligence, instead pursuing systems designed to reason about cause and effect and spatial logic.observer+1

The conference, running through July 11, continues with workshops and additional invited talks as the field grapples with whether the next leap in machine intelligence will come from bigger models or fundamentally different architectures.icml