james grieve

@ahguzelUK Amazing work @ahguzelUK - fantastic results already and I can't wait to see what comes next!

1/ Really proud of what the team pulled together on this one. Here's what we developed to make PROWL work 👇

@odysseyml So excited to see this work from the team starting to bring more of the tools that have worked so well with LLMs into the world model space.

Introducing PROWL! We’ve built RL agents that explore game environments, tasked with discovering failures in world models across physics, visuals, and actions. Those failures then become training data in an automated loop that advances world model performance.

Very excited to see PROWL out in the world. Using RL agents to improve world models is a fascinating new frontier, and brings more of the tools that have worked so well in LLMs into the world model space. This is just the beginning!

@odysseyml At @odysseyml we have always viewed our world models as a way of enabling conversations with a new form of intelligence. Starchild-1 takes that even further and we can't wait for others to experience it as well!

The next form of world intelligence. Coming soon.

Why We Must Build World Models

What made solving self-driving so crazy is that learning to drive really meant learning how the world works and interacts from visual observations of roadways. Imagine all the pretrained knowledge we humans have to know not to bump into things or how to negotiate with other humans. It’s why we can learn to drive with just 30 hours behind the wheel. To make a self-driving car, we had to find a way to give machines a similar base intelligence. After an insane decade of self-driving research, the problem was cracked. Models successfully learned the narrow world of driving from millions of visual observations, and thus became superhuman drivers. Self-driving cars now regularly demonstrate this emergent, really interesting understanding of the world—things like slowing down for occlusion, anticipating a pedestrian stepping out, or that driving a certain way would change another car’s behavior. Now I work on general world models, which push the concept of models learning the world much, much further. A general world model learns from visual observations of the entire world (universe?), not solely roadways. It’s difficult to comprehend just how diverse these observations are. It’s every ocean, glacier, mountain, forest, city, building, simulation, and scenario you can imagine. Modeling this diversity of pixels, sounds, and actions is dramatically harder, but what I believe will eventually emerge from general world models is a profound understanding of the world—of physics, cause-and-effect, and human behavior—at a depth we can’t really comprehend today. What knowledge or scientific breakthroughs or sci-fi applications will that intelligence enable? What will it contribute to our understanding of the universe and our place in it? What lives might it even be able to save? We don’t know yet, and we're so early, but I’m on a very rewarding journey with incredible humans at @odysseyml to find out. Our latest general world model—Odyssey-2 Max—is a humble step on this journey. Our scaling efforts have materially advanced the state of the art in physics-accuracy for world models, and in real-world use it demonstrates a fascinating understanding of the world. I couldn’t be prouder of this team for what they've delivered. We have much to do, and much we will share soon, but there’s no problem I can think of that’s more meaningful to work on than this.

@pgpetrakian Odytrain all the way @pgpetrakian!

Odyssey is cooking 🔥

Read more about it in our blog post here: odyssey.ml/introducing-od…

Another big day from the team at @odysseyml ! We've announced our biggest model ever, hitting leading scores in physics benchmarks across the industry. This is just the start of showing how increasing scale in real-time world models can unlock whole new applications in robotics, entertainment, simulation and more!

It’s time to go beyond language models. Introducing Odyssey-2 Max, our most powerful world model yet. It materially advances the SOTA in physical accuracy. This is a big step toward models that simulate and interact with the world in real time. A new intelligence entirely!

@spacecrafter3d @chrisnolet Congrats @spacecrafter3d, wishing you all the best!

I’ve started a new company with my best friend, @chrisnolet. We spent years at Apple working on Vision Pro together. Chris always impressed me with his extreme focus on technical excellence (always in service of a magical user experience!) and his relentless pursuit of big ideas. After I moved to Colorado and left Apple, we stayed close. We caught up every couple of weeks, often talking until 2 AM. Chris came to visit my family for Christmas and had snowball fights with my kids. I had the honor of being best man at his wedding. Now, we’re building @buttoncomputer together. Button is the wearable AI that can talk. You press it to get instant answers to your questions through a built-in speaker. Because we respect your privacy and agency, Button never listens to your private conversations—it can only hear you when you press the button. We're part of the current @ycombinator batch, and we’ll be presenting at Demo Day on March 24. We already have high-quality aluminum prototypes, and we're launching pre-orders next Tuesday, March 3 (link below). We can't wait to share more next week!

We’re excited to announce an investment from NVentures—NVIDIA’s venture capital arm—and Samsung Next to accelerate our research towards a general-purpose world simulator! x.com/odysseyml/stat…

@odysseyml Super excited to have both nVidia and Samsung on board - both great partners!

The internet is going to get even more weird and wonderful with world models from @odysseyml!

@olivercameron Love this - world models are going to fundamentally shift the structure of the internet.

What if every static image on the internet could come to life instantly?