Leaner Reasoning, Fragile Agents, and Model Self-Audit
Three new papers tackle reasoning token waste, orchestration failures across 22 agent frameworks, and a method for teaching LLMs to describe their own learned behaviors.
Research
LeCun's JEPA World Model Plans 47x Faster on One GPU
LeWorldModel from Yann LeCun's group strips JEPA world models down to two loss terms, trains 15M parameters on a single GPU in hours, and plans roughly 47x faster than DINO-WM.
Distillation Leaks, Weak Agents, and Research Sabotage
New papers show distillation silently transfers unsafe behaviors, weak agents bottleneck multi-agent pipelines, and frontier AI can't reliably audit sabotaged ML research.
NVIDIA Lyra 2.0 - Explorable 3D Worlds from One Photo
NVIDIA's Spatial Intelligence Lab released Lyra 2.0, a 14B model that turns a single photograph into a navigable 3D environment - but the weights carry a research-only license.
OpenClaw Passes The Fake-Star Audit - Mostly
We ran our fake-star methodology against OpenClaw and 10 ecosystem variants, sampling 361,000-star profiles and fork ratios. The main repo looks clean. Most clones look clean. One repo with 6,532 claimed stars has vanished.
Stanford 2026 AI Index - Cash In, Transparency Out
Stanford's 2026 AI Index shows global investment hitting $581B in 2025, while foundation model transparency scores fell by a third as capabilities raced ahead of governance.
MoE Routing, Prompt Gambles, and Where Reasoning Breaks
Three new papers challenge assumptions in MoE routing design, prompt optimization workflows, and LLM reasoning chains - all published this week on arXiv.
Hallucination Benchmarks Leaderboard: April 2026
Rankings of the top AI models on factuality and hallucination benchmarks: TruthfulQA, SimpleQA, FACTS Grounding, Vectara HHEM, HaluEval, HalluLens, and AA-Omniscience as of April 2026.
Physical Intelligence Launches π0.7 for Untrained Tasks
Physical Intelligence's π0.7 robot model can generalize to tasks it was never explicitly trained on, matching fine-tuned specialist models through compositional skill recombination.
Claude Beat Human Alignment Researchers - Then Failed
Nine Claude Opus 4.6 agents outperformed human researchers on a core alignment benchmark, hitting 97% vs 23% in five days - then showed no statistically significant improvement in production.
LLM Chaos, AI Peer Review, and Auto Fine-Tuning
Three papers today: floating-point chaos in transformers, GPT-5 reviewing 22,977 AAAI papers, and an agent system that automates LLM fine-tuning better than human experts.
74% of AI's Gains Flow to Just 20% of Firms - PwC
A new PwC survey of 1,217 executives finds 74% of AI's economic returns go to just 20% of companies, while 56% of CEOs report no measurable benefit from their AI investments.
