Meta says robot world models get better on a predictable curve. Today the fun part is whether anyone else can make that curve hold. This is AI Daily Briefing. On deck: robots, a ninety-million-dollar agent bet, models that answer with zero output tokens, and California writing health-care rules. I'm suspiciously optimistic. Enjoy it while it lasts. Starting with RoboJEPA. If today's show was useful, follow us wherever you're listening — the next one will be waiting. Artem Zholus, writing in arXiv:
In this work we present RoboJEPA, a world model based on the Joint Embedding Predictive Architecture (JEPA) and trained on a large-scale dataset spanning 12 robotic embodiments. We show that RoboJEPA’s imagination error, the error of its latent rollouts, follows a second-order power law in compute, allowing us to predict model quality well beyond the scale at which the law is fit.
Okay, FAIR put out RoboJEPA with every checkpoint, the training code, and the actual robot deployment code. It's a JEPA world model trained across twelve different robot embodiments, and you can just... go run it. Which is the whole ballgame for me. They claim imagination error, how far the latent rollouts drift, follows a second-order power law in compute and predicts quality well past the range they fit it on. With the weights out, somebody outside Meta gets to test that extrapolation instead of taking a sizzle reel's word for it. The bit I'd poke first is the hop from that latent error to a real arm. They say it tracks planning performance closely enough to stand in for real-robot evals. If that holds up, labs save a fortune in hardware hours. Big if. A proxy that correlates inside your own lab is how half the leaderboard numbers get born. But zero-shot planning toward a single goal image on real hardware, code attached? Checking it is now somebody's weekend project, and I'll take that trade. SiliconANGLE, with Maria Deutscher:
Nous Research Inc., the developer of the popular Hermes artificial intelligence agent, today announced that it has raised $90 million in funding. Early-stage startup fund Robot Ventures led the Series B round. It was joined by Nvidia Corp., Samsung Electronics Co., Y Combinator and several others. Nous is now valued at $1.5 billion.
Nous Research, ninety million, Series B, one-point-five billion valuation, Robot Ventures leading. After this week? That's a round size I can read without squinting. Lighter on the money front today, and this is about it. But look who's on the cap table. Nvidia and Samsung. And Hermes has a router that picks which model gets each prompt, Claude or Nous's own Hermes 4 fine-tunes, based on cost, speed and quality. Whoever owns that router decides where the tokens go. Nvidia can read that map. Yeah, the router's what I'd actually dig into. The numbers, though? Twenty-four million downloads, several thousand contributors. That's Nous's count. Then there's an estimate it's two and a half percent of the world's token consumption. Big if true. And none of it tells me how a free desktop agent turns into revenue. Hermes 4 got tuned on five million synthetic records, about two thirds aimed at reasoning, generated with their own DataForge tool. So if there's a business here, it's the fine-tuning plus the routing. The open license is just the on-ramp. Then show me the paid tier. Until then it's a very popular repo with a very good cap table. Here's Asif Razzaq at MarkTechPost:
Liquid AI has released Open d1, two open-weight multimodal models in its d1 decision model family. d1-3B reads text and images. d1-omni-600M reads text with an image, or text with audio. Neither model writes text. Each returns calibrated, typed answers in one forward pass with zero output tokens. The target is real-time decisions on the NVIDIA stack: DGX servers, RTX workstations, and Jetson edge boards.
Okay, this one I actually want on my bench. Liquid's d1 models don't generate a word. You hand them a state and named questions, and you get back P(yes), a full distribution over labels, or a score on a rubric. Every call reports zero output tokens. For an agent guardrail, a probability I can threshold beats parsing free text every single time. Love it. Now read the label on the box. 'Open-weight' under the LFM Open License means free commercial use only below ten million in annual revenue. And the pitch is aimed squarely at DGX, RTX and Jetson. So Nvidia turns up again, same as in the Nous round we just hit. Whoever owns the inference stack keeps winning. Eh, partly. Day-one llama.cpp support and Transformers loading means I'm not locked to a green box. What I don't have is a number. The 600M omni model is an early research drop with no published latency, and that's exactly the one you'd want on a Jetson. And for the 3B, whatever speed and calibration figures exist come from Liquid's own measurements. The checkpoints are on Hugging Face, though, so the calibration claim is testable this week. That's the part I respect. Reed Smith, with Paul W. Pitts:
Three of the over twenty new AI-related laws that were signed into law in California last week will have significant impacts on the use of AI for health care involving California residents. The laws govern the development and deployment of AI-based clinical decision support systems, the confidentiality of medical information shared with health care chatbots, and the obligation of large businesses to provide access to human customer service agents, with exceptions for certain hospital communications.
Earlier this week we covered California's batch of two dozen-plus new AI and privacy laws. Reed Smith just broke out the three that reach into health care, and all three go live January first, twenty twenty-seven. That's under three months out. And look how they carved it up. One on bias in clinical decision support, one on confidentiality for what you tell a health chatbot, one on your right to reach an actual human on customer service. Separate problems, separate statutes. That's more precision than you'll get out of most hearings on the Hill. The definition's the part I'd print and tape to the monitor. Any AI system that produces a prediction, classification, or recommendation aiding timing of care, diagnosis, or treatment. So a small typed classifier, like those Liquid models we just hit, wired into a triage flow? In California, that's a CDS system. Booking and billing bots get carved out, as long as the work doesn't need a license. Senate Bill 503 is the one vendors should actually read, though. It targets 'biased impact,' and diminished access to care is right there in the language. If you want to go deeper on the risks and guardrails shaping AI, check out AI Safety Daily. It covers alignment, model evaluations, emerging risks, and governance, weekdays. Find it wherever you listen to podcasts.
We're watching California's three health care AI laws take effect on January 1, 2027, covering clinical decision support, chatbot medical confidentiality, and access to human agents. Links to every story are in the show notes, so take a closer look at whichever developments caught your attention. That's AI Daily Briefing for today. This is a Lantern Podcast.