CodeAct and Manus

Key Takeaways

Let's talk about Kodak, the framework that powers AI agent startups like Manis. Manis recently went a bit viral when it came out with an AI agent that can do a bunch of things for you by essentially interacting with your computer like a human. And then they got a $75 million investment from Benchmark at a $500 million valuation, but it's now being reviewed by the US government because Manis is apparently a Chinese company with interesting legal structures like being incorporated in the Cayman Islands and having offices in a bunch of different countries. Anyway, Kodak is interesting because it offers an alternative to function calling, which you're probably already familiar with. But as a quick refresher, function calling basically involves telling the LLM about potential tools, and then the LLM will generate structured commands, often in JSON, to invoke those predefined tools. Importantly, the LM doesn't actually do any of the execution of the functions. It just gives you what you should use to call that particular tool. With Kodak, you actually have the LLM just write code, execute it, and then work through problems in multiple turns like an AI agent. This agent can approach problems in a more flexible way and can also adjust things as it goes. The paper found that by using Kodact rather than more traditional methods like function calling, the authors were able to achieve better results on benchmarks. Now, the paper is almost a year old at this point, but it's an interesting alternative to function calling. I'm Diana and I break down texts and AI without the hype. Follow for more.