It all started when I failed to get Gemini to be unhelpful.
Most standard LLM agents come with a system prompt such as “you are a helpful assistant”, and it is very hard to override that behavior. The agent cannot help but always be goal-oriented. If you instruct it, for example, to develop a dialogue with you over several turns, a back-and-forth “muddle”, including diversions that seem unhelpful at first but ultimately are key in reaching a productive conclusion, the agent fails to do that: every response needs to address a stated or implied goal directly, and if it cannot, return new prompts to guide you along. But that’s not how real conversations work.
Another aspect of LLMs is that their responses are mean-reverting. They take a bazillion of inputs and narrow them down to the one most probable response. An LLM is really good at reducing variety, at finding the one likely answer to what the user wants. But it’s really bad at creating variety, in the sense of directing the user towards something they didn’t know they wanted.
I was fascinated when I tried out Jack & Jill, an LLM-based recruiting platform. Their key insight lies in the fact that it’s impossible for one agent to work according to conflicting goals, so they split it up into two adversarial agents: Jack roots for the candidate, helps develop your profile, and finds roles that might match you. Jill operates from the employer’s perspective, making sure to find the best candidate for a role. The end result is a shared goal (a great match between worker and company), but to get there, we first need to address sub-goals independently.
A metalogue is a conversation where the structure of the talk itself is relevant to the subject. If we talk about “muddles”, the conversation gets into a muddle. In Gregory Bateson’s metalogues, a dad is talking to his daughter. But the daughter is not just a naive character asking questions; her questions point out flaws in dad’s logic, and this back-and-forth leads the reader to appreciate the topic.
A standard LLM struggles with this concept: it cannot create confusion. It is allergic to “muddle”. It wants to resolve the tension, whereas a metalogue wants to inhabit it. To write a good one, you have to be willing to be "unhelpful."
I want to create a different kind of agent, an “unhelpful” Bateson-like agent. But that alone would just lead us astray and never converge on something useful. What if we pair it with the “curious daughter”, an agent that makes us question our logic until we gradually develop a satisfying answer?
I believe that if we don’t embrace human variance and creativity, AI is destroying what makes us human. The result is corposlop, mediocrity, and loss of skills and connections to humans and the environment.
Metalogue forces you to think for yourself and to touch grass once in a while. Its goal isn’t only a more creative and meaningful output but also to highlight the human and ecological elements of AI.