Huge win for OpenAI - getting a whole team of top of the top vision researchers!

Introducing the first open-source optimized post-training losses in Liger Kernel with ~80% memory reduction, featuring DPO, CPO, ORPO, SimPO, JSD, and more, achieving up to 70% end-to-end speedup through larger batch size. Use it as any PyTorch module - Available today in Liger

We're excited to launch OpenAI's 4o image model in Playground today so you can make powerful edits to designs. We have a different spin on it: today we're launching 4 new tools for customers to change styles, colors, and put designs in mockups. In the coming weeks, we'll be

> GPT model stopped speaking Croatian > Nobody could figure out why. Turns out > Croatian users were much more prone downvote messages

You can say all you want, but this is the most impressive announcement of today. If it is legit.

People are finally starting to get that XML >> JSON.

Daily reminder that VLM-as-a-judge is non-existent even with frontier models.

You can think of it as sparks of AGI - instead of thinking and counting, the model takes mental shortcuts and answers based on its previous knowledge.

Is it that big of a problem for people to order stuff on DoorDash or Amazon? I thought that at least half of the experience is a dopamine hit when clicking the "Place order" button.