Astra, can you write a Python script for the Fibonacci sequence i mean physically

so i was curious to see how opus 5.5 compares to astra in token usage when acting as a robotic policy TLDR: opus used ~6x more tokens but astra is still more expensive on average. in this particular example, astra got further and completed more tasks before giving up though was fun watching their approaches to failure recovery nobody beat the triangle

I asked Opus 5.5 to create an interactive 3D visualisation explaining how robot actuators work. zerotimedrift.github.io/insi… Repo in comments

Introducing FLUX 3 Action. An open weights 7B World Action Model that achieves first place on the RoboLab benchmark. It outperforms the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster.⁠⁠ FLUX 3 Action removes the usual trade-off between world action model performance and VLA speed: it still predicts video and actions together, but plans more than twice as far ahead and runs faster per second of robot motion than the strongest open VLA. Teams can fine-tune FLUX 3 Action on their own demonstrations to create policies for a particular robot and task. Together with @nvidia, we also integrated FLUX 3 Action natively into @huggingface's LeRobot, with fine-tuning recipes included and edge deployment on NVIDIA Jetson. Beyond robotics, we’re also seeing promising results training task-specific policies for acting in simulated environments like gaming, controlling a vehicle, computer use, and wherever else a model needs to understand a visual environment and then choose what to do next. FLUX 3 Action builds on the same image, video, and audio pretraining as FLUX 3, but uses a smaller architecture designed for practical deployment. In midtraining, we trained the model to predict actions and future frames together. We’re releasing the weights, code, fine-tuning recipe, benchmarks, and reproducible examples so researchers and developers can build on the model with their own robots, environments, and tasks (see below).

code: github.com/Physical-Intellig… paper: arxiv.org/abs/2504.16054

love this format, visualized a V-JEPA 2.1 training step based on public materials

compared GPT-6 Sol and Astra on my "robot drawing on the board" bench Sol used ~25k output tokens vs ~22k for Astra. interestingly, Sol's controller drew twice as fast to be honest, i expected stronger spatial reasoning from new Sol here, but it was way cheaper to run (almost 5x)

great paper, their model trained on 1k filtered RL tasks outperformed the one trained on 8k vanilla tasks (on the validation set) 8k vanilla tasks are sampled right after the test construction step 2 as i understand it

or these systems will implicitly emerge inside a single model with enough scale and proper data engineering #bitterlesson

right now, systems 1 and 2 in robotics communicate with each other mainly through text, one example is task planners prompting VLAs to execute subtasks i think this will be a major bottleneck, it’s hard to express richness of the physical intent with text, if you think about it e.g. when we try to grab a cup, our exact actions depend on whether it was hot/slippery/damaged somewhere/many other things the way our brains “handoff” this information from system 2 to system 1 during actions feels way more integrated

Jev reminds me of times when CLIP was used for image classification without task-specific training