Today, every major AI model — OpenAI, Anthropic, Google — is a single agent. But humans don’t reason alone, neither should AI. Progress emerges through debate, dialogue, and synthesis. Dialectical AI builds the orchestration layer for multi-agent intelligence — the middleware between foundation models and collaborative reasoning — enabling agents to interact freely without rigid, verbose graph-based rules.

This idea started from my pain of always feeling AI models skewed to become toady and lame. I feel that today, everyone is copying and pasting the result from one LLM to another, to cross check ideas and validate the results. I even found out that my Dialectical prompt (asking one single model to perform a full thesis, antithesis and synthesis chain of tought) was somehow very common as well. The question here are the following: Why can't we put them all on a single space, so they can interact, learn and debate with each other? Isn't the context itself and the Agent description enought to decide who should speak next? How can we evolve and deploy workflows that are emergent and non mappable to a graph/state machine? How can AI models overcome the User <-> Assistant ping pong bias and interact in trully multi stakeholder environments.

Seems interesting, would like to see if this would diminish the prevalence of AI hallucinations, good luck!

@kapesh That is one of the main ideas behind it, collective chain of tought with models being able to counter ballance the other.
This paper touches some of it and was one of my main inspirations to move forward with the idea:
https://arxiv.org/abs/2305.14325
I've built a completions endpoint so anyone can use a Group-As-A-Model, my goal is to test it against the main benchmarcks.

Seems interesting, would like to see if this would diminish the prevalence of AI hallucinations, good luck!

Dialetica AI

Humans don't reason alone, neither should AI.

Humans don't reason alone, neither should AI.