Episode 20 of The Information Bottleneck brings Yann LeCun to the microphone for a focused conversation about why he left Meta to form Advanced Machine Intelligence (AMI), what AMI will build, and why he believes world models and representation-based prediction are the right route to practical, grounded intelligence. The episode was published on Dec. 15, 2025; based on LeCun’s on-record timeline about his remaining weeks at Meta and public announcements about his departure, the interview was conducted on 2025-12-10 (inferred). The hosts are Ravid Shwartz-Ziv and Allen Roush.
Why start AMI now and the company’s purpose
LeCun explains that AMI is more than a research lab: it is intended to build products informed by many years of world-model research. He emphasizes that investor willingness to fund multi-year research programs has created a unique opportunity to found a company where "the first couple years are essentially focused on research" and yet aim at long-term productization. As he put it, AMI’s ambition is to become "one of the main suppliers of intelligent systems down the line," concentrating on systems that can predict consequences of actions and use optimization to plan.
"We think the proper way to handle this ... is have world models that are capable of predicting what would be the consequence or the consequences of an action or sequence of actions ... That's planning."
Openness and publishing as the core of research
LeCun argues that upstream research should be open. He says you cannot call it research unless you publish, because internal secrecy breeds self-deception and internal hype. Publishing, he stresses, enforces methodological rigor, motivates researchers over long horizons, and avoids short-term, product-driven diversion from foundational progress.
"You cannot really call it research unless you publish what you do because otherwise you can get easily fooled by yourself."
World models versus LLMs: why language scaling is not enough
A repeated theme is that LLMs work well for language but fail on continuous, high-dimensional, noisy modalities such as video and embodied sensor streams. LeCun highlights that tokenizing such data and training generative, pixel-level models is ill-suited to capturing the abstractions needed for prediction and planning; instead, he advocates learning abstract representation spaces and making predictions there.
"If you want to handle data that is high dimensional, continuous and noisy, you cannot use generative models that tokenize your data into discrete symbols. It's just no way."
JEPA, representation learning, and the long arc of self-supervised methods
LeCun traces his work on unsupervised and self-supervised learning from auto-encoders, sparse representations, and contrastive (siamese) methods through modern joint-embedding predictive architectures (JEPA). He discusses challenges such as collapse when representations are trained end-to-end and the evolution of objective functions (contrastive terms, information-maximization efforts like Barlow Twins, and VICReg) that made high-quality abstract representations practical.
He stresses the key idea: learn a representation that filters out unpredictable details (noise) and predict in that reduced space — a move that allows tractable, useful prediction where pixel-level forecasting fails.
Simulation, abstraction and what a world model should be
LeCun rejects the notion that world models must simulate every physical detail. Instead, he shows how scientific modeling works by layered abstractions — from quantum mechanics up to thermodynamics and phenomenological laws — and argues that useful world models simulate only the relevant abstract dynamics. He uses computational fluid dynamics and planetary mechanics to illustrate how coarse abstract variables can be both sufficient and computationally tractable.
"World models don't have to be simulators at all... they simulate only the relevant part of reality, in abstract representation space."
Learning basic physical concepts from data (object permanence, gravity)
Using developmental examples, LeCun explains how concepts such as object permanence and gravity can be learned from many videos or abstract simulations rather than by innate programming. He notes infants learn such expectations through play and simple stories; analogously, agents can learn from adventure games, synthetic environments, or videos that expose structured, redundant patterns.
Game AI, planning under uncertainty, and why some games remain hard
LeCun contrasts fully observable combinatorial games (chess, Go) where tree search and value functions excel, with partially observable, stochastic adventure games (e.g., NetHack) that require planning under uncertainty and efficient move proposal mechanisms. He reiterates the classical recipe: a generator of promising moves and a value function to evaluate positions, learned by imitation, RL, or a hybrid.
Safety by construction and constraints vs filter-based moderation
On safety, LeCun favors objective-driven architectures with explicit constraints over post-hoc finetuning or expensive generate-and-filter pipelines. He argues for systems whose outputs result from constrained optimization that respects guardrails — e.g., low-level operational constraints on domestic robots — rather than relying solely on brittle fine-tuning that can be bypassed.
"You should use those objective-driven AI architectures ... the system is intrinsically safe because it has all those guardrails ... it's by construction."
Silicon Valley monoculture, open-source dynamics, and geopolitics
LeCun describes a herd effect in industry where competition pushes many labs to follow the same LLM-focused path. At the same time he notes a paradox: some of the best open-source models currently come from Chinese groups, which has driven adoption outside U.S. corporate openness. He also explains why he chose a global company footprint with an important Paris presence for AMI.
Timelines, definitions, and realistic expectations about AGI
LeCun disputes the notion of a single, unitary "general intelligence" and reframes the conversation around human-level capabilities. He predicts a progressive path: machines will reach human-level competence across many domains over time, but that is not a single event. In the most optimistic scenario (with no unforeseen obstacles), he suggests dog- to human-level capabilities could appear within a multi-year horizon, but he cautions that hard obstacles could extend timelines to decades.
"There is no such thing as general intelligence... We can talk about human-level intelligence. ... The most optimistic view is maybe 5 to 10 years to something close, but we may hit obstacles that make it 20 years or more."
Personal background, hobbies and why he keeps working
LeCun reflects on his mission to increase usable intelligence in the world, his family background, and a life of making things: sailing, home-built flying contraptions, astrophotography, electronics and electronic musical instruments. He says that despite awards and the chance to retire, he continues because his work — building assistive and amplifying intelligent systems — is a mission.
Career advice
To students entering AI today, LeCun recommends learning fundamentals with long shelf life: mathematics, modeling, signal processing, control theory and physics — subjects that help you learn to learn. He cautions that many short-lived tool skills are less valuable than deep conceptual foundations.
References
Episode page and transcript: EP20: Yann LeCun — The Information Bottleneck (Dec. 15, 2025).
LeCun’s departure and AMI coverage: Financial Times: Meta's Yann LeCun targets €3bn valuation for new AI start-up; Reuters: Meta's Yann LeCun targets $3.5 billion valuation for new AI startup.
Explore more exclusive insights at nextfin.ai.
