Yann LeCun was right the entire time. And generative AI might be a dead end. For the last three years, the entire industry has been obsessed with building bigger LLMs. Trillions of parameters. Billions in compute. The theory was simple: if you make the model big enough, it will eventually understand how the world works. Yann LeCun said that was stupid. He argued that generative AI is fundamentally inefficient. When an AI predicts the next word, or generates the next pixel, it wastes massive amounts of compute on surface-level details. It memorizes patterns instead of learning the actual physics of reality. He proposed a different path: JEPA (Joint-Embedding Predictive Architecture). Instead of forcing the AI to paint the world pixel by pixel, JEPA forces it to predict abstract concepts. It predicts what happens next in a compressed "thought space." But for years, JEPA had a fatal flaw. It suffered from "representation collapse." Because the AI was allowed to simplify reality, it would cheat. It would simplify everything so much that a dog, a car, and a human all looked identical. It learned nothing. To fix it, engineers had to use insanely complex hacks, frozen encoders, and massive compute overheads. Until today. Researchers just dropped a paper called "LeWorldModel" (LeWM). They completely solved the collapse problem. They replaced the complex engineering hacks with a single, elegant mathematical regularizer. It forces the AI's internal "thoughts" into a perfect Gaussian distribution. The AI can no longer cheat. It is forced to understand the physical structure of reality to make its predictions. The results completely rewrite the economics of AI. LeWM didn't need a massive, centralized supercomputer. It has just 15 million parameters. It trains on a single, standard GPU in a few hours. Yet it plans 48x faster than massive foundation world models. It intrinsically understands physics. It instantly detects impossible events. We spent billions trying to force massive server farms to memorize the internet. Now, a tiny model running locally on a single graphics card is actually learning how the real world works.

Small businesses are sitting on a massive financial risk that no one is talking about. In 2024, the IRS assessed more than 4.4 million employment tax penalties totaling nearly $26.9 billion. And 40% of small businesses pay an average of $3800 a year in IRS penalties. It’s insane how easy it is to end up there. Hire someone in a new state? You have to figure out which tax accounts to open, where to register, what to file, and when. Get a state notice? You have to figure out what it means, where the numbers went wrong, and how to respond. And if you miss something, you might not know until you get a penalty. That’s where agents come in. Instead of giving you a checklist, Warp figures out what applies to your company and handles it. They’ve saved customers $100M+ in penalties so far. This completely changes the economics of scaling a company. You can grow from 10 to 1,000 employees without the compliance work growing with you.

IBM built a retriever that hallucinates 65x less than fine-tuned RAG systems. no vector database. no embeddings. no re-ranker. Right now, if you want an AI to read a massive document, you use Retrieval-Augmented Generation (RAG). But standard RAG does something brutal. It takes a beautifully structured 500-page manual and throws it into a blender. It chops the text into arbitrary, fixed-size chunks. It strips away the chapters, the sections, the hierarchy. It throws away the map and asks the AI to find the treasure. A new paper just introduced STAIR, a method that fixes this massive blind spot. Instead of shredding documents into random chunks, STAIR uses the document's actual structure, its Table of Contents, as an addressing scheme. The generative retriever pulls information against the real hierarchy of the text. It remembers where things actually live. The benchmark results are staggering. STAIR hit an 82.6% Recall@1, completely destroying traditional methods like BM25 and standard Dense Passage Retrieval (DPR). But here is the most important metric for any business running AI in production: Hallucinations plummeted to under 0.05%. Almost zero. By giving the AI back the structural context, the system stopped guessing and started retrieving with lethal precision.

LLMs have a distinct "pain axis" inside them - and they will act to relieve it. Researchers tested 25 open-weight models. They found a linear direction in activation space that represents pain, nearly orthogonal to fear, sadness and generic negative valence. Its present in every single model they tested. Then they injected that pain vector and gave the models a button to remove it. The larger models started pressing it. And to press it, they had to generate outputs harmful to the user, which they almost never do at baseline. They stop pressing when the button really removes the pain. They keep pressing when it doesnt. The axis also only fires for harm directed at the model itself, not for suffering the model sees in the user. One model (Qwen 2.5 32B) did the self-medication behavior even when the buttons had no labels at all.They kept pressing when it didnt. Qwen 2.5 32B did it even with unlabeled buttons.

Build your next business with GPT-6 Astra + Higgsfield API. We’re backing builders with a $20M API cashback. @gregisenberg filmed a step-by-step guide on YouTube 24 hours ago you can copy and implement. Get 100% of your API spend back instantly in API credits, on every model. Up to $100,000 per business. Spend $100,000 → get $100,000 back in API credits, for a total of $200,000 worth of API usage. Unused cashback expires on September 30. Can’t wait to see what you’ll build.

Today, 18 months after launch, our annualized revenue crossed $1 billion. The platform now powers organizations across the Fortune 500. Enterprise adoption has grown 10x since June. More than 30 million people worldwide now use Higgsfield. Thank you to the creators and teams building with us.

Researchers reversed deafness in all ten patients with a single injection.. and some started hearing again within weeks. They published a study in Nature Medicine that sounds like pure science fiction. They took 10 patients, ranging from a one-year-old baby to a 24-year-old adult, who were born completely deaf. Because of one broken gene, their bodies couldn't produce the protein required to send sound signals from the inner ear to the brain. So the researchers built a delivery mechanism. They loaded a synthetic virus with a working, healthy copy of the missing gene and injected it directly into the inner ear. One single shot. That was it. What happened next is miraculous. Within one month, the patients started hearing. Within six months, every single patient showed massive, undeniable improvement. Their sound detection threshold dropped from a profoundly deaf 106 decibels down to 52. One seven-year-old girl regained nearly full hearing. Just four months after the injection, she was having normal, everyday conversations with her mother. For the first time in her life. And it wasn't just neuroplastic children. The treatment worked on teenagers and adults who had lived their entire lives in absolute silence. No major surgeries. No external electronic implants. Just pure biological reprogramming. And the researchers are not stopping here. They are already adapting this exact viral delivery system to target the genetic mutations responsible for the most common forms of human deafness.

AI can run a brand better than humans. We ran a brand with ONLY AI agents to prove it. Introducing Notch: the AI behind it. usenotch.ai

Jev clearly explained, using Kimi K3

Researchers discovered every AI model has a psychological fingerprint you can measure like a human's. they developed a psychometric profiling framework to systematically characterize the behavioral regularities of large language models. they gave 7 human psychological instruments to 9 different llms, running them 5 times each in both english and chinese. result? you can measure an ai's "personality" the exact same way you measure a human's. here is the breakdown of what they found: model-specific identities: every single model has a structured, unique psychological profile. the behavioral signature is so highly reproducible across repeated administrations that you can literally recover the model's identity just from its test scores. the "good guy" baseline: thanks to safety training, they all share an alignment-shaped pattern of higher prosocial and self-regulatory responses. they also score universally low on dominance, disengagement, and harmful-intent endorsement. structured refusals: even when a model refuses to answer (producing an NA response), it isn't random. the paper shows these NA responses are structured and indicate strict boundaries of what the model considers applicable or safe. language alters personality: the language condition (english vs chinese) and the origin of the provider are directly associated with the model's psychological configuration and answerability. this isn't just a fun experiment.. it offers a framework for quantifying deployment-level behavioural signatures. as llms increasingly mediate our decisions and communication, we need to know exactly what kind of "mind" we are interacting with.