Tool calling in LLMs promises to extend AI capabilities, but constrained decoding that forces models to use tools can actually hurt their performance. The video explores why this mismatch exists and what it means for building reliable AI agents.
There's a lot of misunderstandings about LLMs. For example, a really common mistake that LLM builders do is that they over rely on LLM tool calls. And like what we have observed is constrained decoding, which is really how people perform these tool calls, tends to negatively impact the ability of the LLM to perform. When I translate that to what that means is that whenever you use tool calling, you're really reducing the performance of your LLMs and its base knowledge. If you're a founder who's raised in this space, definitely think twice about how you would build that data set, how you would then productionize those agents.
The AI Agent Getting More Supplier Data in 5 Weeks Than a Team Did in 2 Years
