The AGI Frontier

LeCun's JEPA at Scale: First Independent Replication Results Are In

GPT-5 Internal Evals Show Emergent Planning Capabilities Never Seen Below 100B Params

Figure 02 in the Wild: Six Months of Real-World Deployment Data Analyzed

AI Discovers New Class of Antibiotics Effective Against Drug-Resistant Bacteria

Process Reward Models Outperform Outcome Reward Models Across All Reasoning Benchmarks

When 1000 LLM Agents Collaborate: Emergent Division of Labor in Code Generation Tasks

MIT's New Causal Reasoning Benchmark Exposes Systematic Failures in All Current LLMs

Intel Loihi 3 Achieves 10,000× Energy Efficiency Over GPU for Sparse Inference Tasks

Scaling World Models via Self-Play: a Recipe for Open-Ended Learning

We propose a self-play framework that trains world models by having agents compete and cooperate in procedurally generated environments.

π0.5: A Vision-Language-Action Model for General-Purpose Manipulation

We scale VLA models to 7B parameters using a novel cross-embodiment dataset of 200M manipulation trajectories.

Chain-of-Thought without Prompting: Emergent Reasoning in Large Transformers

We show that CoT reasoning pathways emerge spontaneously in sufficiently large transformers without any explicit prompting.

AlphaGeometry 2: Olympiad-Level Geometric Reasoning with Synthetic Data at Scale

Extending AlphaGeometry, we achieve gold-medal performance on 96% of IMO geometry problems using synthetic deduction corpora.

Constitutional AI 2.0: Reinforcement Learning from Principle Hierarchies

We introduce a hierarchical constitutional framework where abstract principles generate specific rules for finer behavioral control.

Emergent Specialization in Large-Scale Multi-Agent Language Model Systems

When populations of LLM agents interact over thousands of rounds, specialist roles and protocols emerge without explicit programming.