Lightwork podcast breaks down Jev, the model that outputs probabilities instead of chat
The episode discusses why Typesafe's system one model Jev returns probabilities rather than paragraphs, how it is used in practice, and also covers Meta Muse and Skild AI.
The newest AI model can’t text back. TypeSafe released Jev, their “system one” model that doesn’t respond with paragraphs, only probabilities. While it can still be wrong, it can’t wander outside the answer schema the way a chat-based LLM can. Will frontier labs take notice and start building their own? More on Lightwork this week: → Amazon bans Meta Muse from their website to guardrail its $68B ads business, while storefront-style platforms like Shopify welcome agent use for traffic. → Skild AI trained its robots on 140 years of simulated self-play soccer, showing that physical self-play could enable robots to far exceed human capability. Full video below. 00:00 Welcome to Lightwork 01:07 Jev: The AI Model That Doesn't Chat 06:39 Real-World Uses for Jev 9:56 Amazon Blocks Meta's Muse 13:23 Why Shopify Is Embracing AI Shopping Agents 21:22 Why Meta's Muse Took Off 23:07 Muse's Safety & Privacy Approach 26:15 Behind the Scenes at Meta Connect 27:24 Skild AI: Physical Self-Play @clairejyz @Machiz