Congrats to @DynaRobotics on Dyna-2! This feels like an important result for robotics. A few points that stood out to me when reading through the report: 1. The model doesn't need to explicitly "daydream" the future before acting: Future video prediction is used in training to help build the representation, but at inference time the action policy doesn't attend to predicted future video. This feels close to how humans actually operate. We sometimes imagine how things will play out, but most physical action feels much more "automatic." 2. Vision > Text: In many other VLAs, the action pathway attends directly to text. In Dyna-2, the text instruction gets absorbed into the vision representation and the action tower never needs to see the text directly. A sufficiently rich visual representation seems able to absorb the task intent and implicit reasoning needed to act. This also feels more human-like. When we are in the midst of following instructions, we often don't think about the explicit wording of the instructions and instead just act instinctively based on what we are seeing. 3. We are going to need a lot more data: Dyna showed some preliminary scaling laws from their analysis which is very intriguing. They go from 20% to 53% normalized performance as video scales from 1 thousand to 1 million hours. That's roughly 11 points per 10x of data. If you do a naive extrapolation, this implies for ~99% performance you need of order magnitude 10 billion hours of video. That's more video than on all of YouTube. We are going to need a bigger boat. My sense is we are still at the "Attention Is All You Need" era for robotics, where there is still exploration on the right architecture and recipe to scale. Dyna-2 feels like a promising step forward where we are moving towards a more "human" approach to robotics, but it also shows that there is still a fair amount of work to be done to properly solve the data problem.

I'm very excited to share that today we're launching Merlin, the agentic intelligence layer for Encord! Software is changing in front of us. Interfaces are becoming conversational and starting to consolidate across the stack. We are getting to the point of treating software systems as colleagues and helpers that meet you where you already work rather than as disparate, insular tools. Merlin brings that shift to AI data infrastructure. With Merlin, you can go from a prompt to a complete annotation or evaluation setup in seconds, get a live read on data quality, and find and fix the issues holding your model back, all through a single conversation. It is now available via MCP in Claude, Codex, and all major agentic cording platforms, with Slack and more systems coming soon. Anyone building AI systems knows how much manual effort goes into getting data ready before a single annotation is made. Merlin changes that. We're opening early access to a selected group of customers. Reach out if you are interested in joining the waitlist!