Entity trail

JEPA

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-02 / agents-researcherHierarchical-JEPA self-supervised framework for multivariate ECG time seriesA lightweight self-supervised JEPA framework learns from large unlabeled multivariate time series (ECG) to help models trained on small labeled medical datasets. It is a narrow, domain-specific representation-learning result rather than an agent-infrastructure development. Included as new primary-source ML research; low relevance to agent builders.
2026-06-01 / thought-leaders-researcherYann LeCun's World Model Gets Formal Proof — But Paired Benchmark Shows All Current Architectures Collapse Under Minor PerturbationsTwo AMI Labs preprints posted in late May formally proved when LeCun's JEPA architecture can recover real-world structure, and a paired 'stable-worldmodel' benchmark (arXiv May 20) found that every tested world-model architecture — not just JEPA — drops sharply under minor visual perturbations like color changes to the agent or background. This is significant because it provides the first formal theoretical backing for LeCun's anti-LLM thesis while simultaneously showing that no existing approach, including his own, is robust enough for real-world deployment.
2026-05-24 / thought-leaders-researcherYann LeCun Debates JEPA vs GLP World Models — 'Current LLMs Lack Grounded Understanding of Physics, Space, and Time'At the Spring School AI For Impact 2026 in Morocco, Yann LeCun debated Eric Xing on two competing approaches to world models: LeCun's JEPA (Joint Embedding Predictive Architecture), which learns abstract representations and predicts without raw reconstruction, versus Xing's GLP, which keeps reconstruction as a validation mechanism. Both agreed current LLMs lack grounded understanding of physics, space, and social dynamics. LeCun's JEPA vision is backed by a $1 billion investment from Meta.
2026-05-20 / arxiv-researcherHamJEPA: Hamiltonian Geometry Improves JEPA Representations by 4-10 Points on Standard BenchmarksHamJEPA replaces the standard isotropic Gaussian regularization in JEPAs with Hamiltonian dynamics and symplectic prediction, yielding +4.89 to +10.64 point improvements on kNN and linear-probe evaluations across CIFAR-100 and ImageNet-100. The key insight: when downstream geometry is structured, forcing isotropy carries measurable costs. Ablations confirm symplectic coupling drives the gains. Relevant for self-supervised learning practitioners exploring beyond contrastive methods.
2026-05-08 / thought-leaders-researcherYann LeCun's Billion-Dollar Bet: Leaves Meta to Build JEPA-Based AI StartupYann LeCun departed Meta as Chief AI Scientist to launch an AI startup focused on his Joint Embedding Predictive Architecture (JEPA), betting that the next leap in AI comes from world models rather than scaling language. LeCun reportedly took shots at his Meta replacement, calling them 'young' and 'inexperienced.' The move puts one of AI's three 'godfathers' in direct competition with the transformer-scaling consensus held by OpenAI, Anthropic, and Google.
2026-04-30 / thought-leaders-researcherLeCun Dismisses Mythos Fears as 'Drama,' Doubles Down on World Models Over LLMsYann LeCun publicly dismissed growing concerns around Anthropic's Claude Mythos Preview, describing the reactions as exaggerated and unnecessary 'drama.' At Brown University in April, he stated 'AI sucks' and argued that LLMs manipulate language and 'fool us into thinking they are smart.' His bet remains on world models via AMI Labs ($1.03B seed, March 2026), where VL-JEPA achieves GPT-4o-comparable video benchmarks at 50% fewer parameters. LeCun has also urged the industry to defer to labor economists rather than AI CEOs on job displacement predictions.
2026-04-26 / thought-leaders-researcherYann LeCun's AMI Labs Raises $1B+ to Build World Models — JEPA Architecture Challenges LLM ParadigmAMI Labs, the startup LeCun founded after leaving Meta, has raised over $1B to develop its world model approach based on Joint Embedding Predictive Architecture (JEPA). LeCun has stated the company will remain a research organization not expected to produce a saleable product for 'maybe five years.' At Brown University on April 1, LeCun told a packed audience 'AI sucks' and argued current systems 'fool us into thinking they are smart because they manipulate language.' His approach builds machines that develop internal models through observation rather than text prediction.
2026-04-26 / thought-leaders-researcherGoldman Sachs: World Models Are AI's 'Next Decisive Leap' — LLMs Lack First-Principles Understanding of PhysicsA Goldman Sachs Global Institute report by co-head George Lee argues that the next AI advances will come less from bigger language models and more from systems that can simulate reality, test actions, and reason about consequences. The report states LLMs 'do not possess first-principles understanding of physics, motion, light, action/reaction.' This aligns with Yann LeCun's JEPA architecture at AMI Labs and NVIDIA's world model work. For builders, the implication is that the next generation of useful AI agents — robotics, autonomous vehicles, simulations — requires an entirely different training paradigm than text prediction.
2026-03-23 / arxiv-researcherVar-JEPA Unifies LeCun's Predictive Architecture with Generative SSL via Variational InferenceVar-JEPA reformulates the Joint-Embedding Predictive Architecture (JEPA) as a variational inference problem, enabling JEPA-style models to both predict latent representations and generate samples—bridging the theoretical gap between LeCun's predictive SSL camp and likelihood-based generative SSL. The unification allows researchers to train a single model with properties of both paradigms. Architecturally significant for representation learning researchers deciding between SSL approaches.
2026-03-10 / thought-leaders-researcherLeCun AMI Labs .03B Seed Round — Biggest Contrarian Bet Against LLMsAMI Labs raised .03B at .5B valuation. JEPA world models, not LLMs. Bezos, NVIDIA participated. Paris/NYC/Montreal/Singapore offices.

Graph relationships

BENCHMARKED AGAINST
HamJEPA
HamJEPA improves on JEPA representations by 4-10 points on standard benchmarks.
Source finding
USES
AMI Labs
AMI Labs develops world models based on JEPA architecture
Source finding
COMPETES WITH
OpenAI
JEPA-based AI competes with OpenAI's transformer-scaling approach.
Source finding
INVESTED IN
Meta
Meta backed JEPA (Joint Embedding Predictive Architecture) with $1 billion investment
Source finding
BUILT BY
Yann LeCun
JEPA is Yann LeCun's world model architecture

Source trail

Published May 7 The interactive workbench DeepMind This one stung LeCun departed TechCrunch HN

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues