Today we are announcing JUWEL OS. While the world was building AI supercomputer assistance, we were re-thinking the entire infra stack. We realized early on, that AI was being wrapped around machines which simply were designed for this level of intelligence. As the world approaches closer and closer to AGI/ASI - we must completely rethink what it means to interact with machines from the ground up. This is why we've rebuilt the OPERATING SYSTEM KERNEL on SILICON, we've rebuilt the runtime of the future. - welcome! The next stop - full AI native services - done for you by your own assistant. Below is our first benchmarks, showing how our kernel - saves time/money/accuracy/security. - More to come. - by @VextLabs

Someone today said opus 5.5 is degrading. It animated this in 2 mins:

always insightful seeing how @poteto ships crazy amount of PRs at SpaceX. Some of the top priority lessons imo from this article that I already apply or will try out: - get agent to explain its understand to catch mistakes in understanding. Making sure the agent has the right/enough context on the right task helps prevent them going down the wrong path. I use a /new-task skill (link below) - /recall and gather context from previous conversations instead of always restating context. - architecting by designing core type definitions, function signatures, and through code instead of plan docs. I've seen this floating around before, but haven't given it a try. I think will be likely the biggest benefit in making sure the agents stop slopping around with Any types everywhere. Also another benefit is making sure you understand how the code works, and across a long horizon that's still the most important. - add some kind of /teach skill. Agents are very smart, horrible teachers by default. I use my /concisely skill and reinforce that human keeping in the loop is key but also a bottleneck. Teach me the most important points to keep me in the loop, i'll ask for follow ups to go deeper.

This is notable. DeepSeek, a lab usually first to pioneer novel algorithms and architectures, is saying that at this point, the ROI of improving data quality far exceeds that of working on novel post-training algorithms. I think this has already been true for some time for non-lab practitioners. If you're doing llm post-training, 80% of your effort should go into looking at your data. This means: - Hiring experts to dig through your RL tasks - Sifting through rollouts and sft data by hand to remove suspicious samples. Make sure all tasks are actually passable. - Making sure your data is diverse in both difficulty and category.

DeepSeek V4.1 Flash is cheap 50x cheaper being almost as good as Sol and Opus 👀

The future is open memory, open context.