Tech Radar

Kimi K3 isn't just summarizing emails or writing Python scripts. It's autonomously designing GPU compilers and 45nm silicon chips from scratch. I've seen plenty of "agentic" claims lately that basically just mean the AI can click a few buttons in a browser, but building a Triton-like compiler with its own IR layer and a PTX code-generation pipeline is something else entirely. The "vision in loop" part is where it actually gets interesting. Instead of just guessing what the output looks like, it iterates between writing code and analyzing live screenshots. It sees the mistake, fixes the code, and checks the screen again. It's a tight feedback loop that makes the usual prompt-and-pray workflow feel prehistoric. Moonshot AI is pushing this toward a full-blown workspace with everything from IDE agents to browser extensions. It's a lot of surface area to cover, and I'm skeptical about whether one model can actually handle all these differ...

Moonshot AI just dropped K3, and it's a strange mix of a product update and a philosophical pivot. Most companies are still fighting over who has the biggest context window or the fastest token generation. Moonshot seems to have realized that having a massive memory is useless if the model can't actually do something with it. The new toolkit is a clutter of features. You've got scheduled tasks, specialized modes like Kimi Code, and a "Claw" tool. On the surface, it looks like they're just adding buttons to a chat interface. But if you look closer, they're trying to move the LLM from a passive oracle to an active agent that manages your workflow. I'm skeptical about whether "scheduled tasks" in a chatbot is a feature people actually need, or if it's just a way to keep users locked into their ecosystem. Still, the way they're handling multimodal inputs suggests they've found a way to make long-context reasoning feel less li...