AI Pulse · AI Trends Pulse

RedNote, the company behind the Chinese social app of the same name, just open sourced a new AI model called dots3 through its lab dots.ai. The headline feature isn’t the size, though 280 billion parameters with 16 billion active at any time is big. It’s a training method called TEMPO, which lets the model check its own progress on long tasks and update what it remembers as it goes, instead of just barreling ahead on a fixed plan.

Here’s why that matters if you run a business. Most AI agents today are good at short tasks. Answer this email, summarize this document, write this snippet of code. They fall apart on anything that takes many steps over a long stretch, because they lose track of what they’ve already tried and what actually worked. TEMPO is built specifically for that problem, self-critiquing along the way on tasks the model hasn’t seen before. If it holds up, that’s the difference between an assistant that can draft a paragraph and one you could actually trust to work through a multi-day research project or a complex ops task with less babysitting.

The team also released two new open benchmarks, VibeSearchBench and VibeLifeBench, built to test exactly this kind of long-horizon reasoning. That’s a useful signal on its own. Labs that publish their own test suite alongside a model are inviting scrutiny, not just hype.

Worth being clear-eyed here. This is the best-verified item in today’s roundup, but independent reproduction hasn’t happened yet. Treat the claims as promising, not proven, until other teams run their own tests on the model page on Hugging Face.

If long-running agent memory like this pans out, it’s exactly the kind of capability we like to fold into an AI command centre, the sort of setup we’re building with Omni.