Ronald Skorobogat

@thsottiaux Access to GPT Pro in codex and better VSCode extension support for those of that still check code

What is something that you feel is surprising that Codex still can't do well and we should have gotten right a while ago?

@OfficialNathanY really cool breakdown. the table 6a number that jumped out was 4.00 vs 3.42 from-scratch on calvin abc to d, surprising that a pretrained rgb image prior transfers that well to a dense optical flow prior

What if you combined a world model and a VLA? DAWN is one of the coolest papers doing something similar. Instead of predicting future images, DAWN predicts pixel motion fields (2D vector field) telling pixels in the scene where it should move, and uses it to plan VLA actions

@chongyiz1 really cool construction. the river swim crossover is interesting: gcql catches q-learning as horizon grows, but loses on cliff walking. is that from the absorbing states turning dense rewards into a long-horizon success event, so gcrl wins when td horizon is the bottleneck?

1/ Reinforcement learning is usually framed as maximizing rewards. But can we cast it as reaching the right goals? New blog on bridging RL, goal-conditioned RL, and stochastic shortest path: iclr-blogposts.github.io/2026/blog/2026… Also #ICLR2026 Poster: Thu 10:30 AM–1:00 PM, P4 #4611. 🧵⬇️

@davidrmcall really cool work. the uniform copy beating backprop-through-chain (rows I vs J) leans on flow jacobians being roughly psd via ot straightness. did you try it on distilled few-step students or late-training checkpoints where flows get more rotational?

We developed a simple, sample-efficient online RL technique for post-training image generation models. We see it as a possible steerable alternative to CFG, driven by any scalar reward, including human preference.