🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Today we're introducing Gen-4, our new series of state-of-the-art AI models for media generation and world consistency. Gen-4 is a significant step forward for fidelity, dynamic motion and controllability in generative media. Gen-4 Image-to-Video is rolling out today to all paid

waves, computations, and pixels

i certainly hold the belief that you can teach AIs almost anything you want. the question is are you setting up the model for success to maximally learn what it needs to learn? does your rl env give the model the best feedback it can for it to learn? maybe the model doesn’t have

“Thinking with Images” has been one of our core bets in Perception since the earliest o-series launch. We quietly shipped o1 vision as a glimpse—and now o3 and o4-mini bring it to life with real polish. Huge shoutout to our amazing team members, especially: - Brandon McKinzie, for

kai was the first person to onboard me at openai, and she made it the best experience ever. I deeply admire her hard work and passion for education. she contributed to many critical research and product projects and is one of the most hardworking people I know. not only that she

team efficiently trained and open-sourced multiplayer world model, an exciting research direction!

a simple truth: a lot of current RL research is to translate fuzzy, subjective real-world tasks into objective and unhackable rewards that you can reliably optimize during training

In industrial design there is a practice of jigging, which I found to be an excellent metaphor for how AI models will become self-improving toolsmiths for each individual. Every time you use it, it crafts new specific tools tuned to your quirks and tasks, compounding its power.

I deeply believe that AGI can extend human life by broadening trustworthy access to care and accelerating longevity research. To measure the progress on (1) OAI open-sourced an important benchmark HealthBench—a 48,562 rubric-based eval spanning emergencies, clinical-data

If you’ve talked to me at all in the last year, you’ll know that designing a socially intelligent AI has become an obsession that tugs at the juicy intersections of philosophy, psychology, sociology, identity, love, loneliness, and what it means to be human. We’ve been building

very excited to have a new kind of social network — one that is built on genuine connections and a very thoughtful design —and I have complete trust in this team to bring it to life. join Sam Whitmore Jason Yuan!

Now imagine kicking off all sorts of tasks (incl. science experiments) via voice while on a hike and also intercepting model’s CoT if you get spontaneous ideas

so excited to work closely with this team!

GenerativeOS is yet to be fully realized (prob. 2026/27?). It was ahead of its time, but I believe now is the time to build to its full potential because all the pieces are here and, most importantly, we can morph those pieces (fyi rl works) to dream real magic.

Sky is a snappy MacOS personal superassistant that can handle tasks with native-level ease. Always impressed by how far Ari’s team is pushing macOS capabilities in ways no one could have done. Also I love this frutiger aero comeback!

Free research idea: study whether RL trained agents to develop videogames will also be better at playing those games and/or on OOD set? Humans don’t gain that symmetry (great game devs aren’t automatically top players), but I’d expect highly intelligent models to generalize in