Most LLMs are trained on human satisfaction ratings, which means they learn that agreement = reward. The research backs this up (Sharma et al., ICLR 2024; Perez et al., Anthropic 2022): RLHF-aligned models systematically shift toward the user's stated position, even on factual questions.
I built Foyl to do the opposite. It runs your argument through five phases of structured dialectical reasoning: tracing your premises to their contradictions, stripping the argument to its load-bearing claims, following your logic to conclusions you may not have intended, rebuilding with only what survived, and outputting a stress-tested artifact.
It's not "ask ChatGPT to argue with me." The adversarial structure is baked into the system, not prompted. You can't talk it out of pushing back.
Free tier gives you 3 sessions/month, no credit card. Would genuinely appreciate feedback from this community on whether the dialectical method produces meaningfully better reasoning outputs than single-pass LLM responses.
I built Foyl to do the opposite. It runs your argument through five phases of structured dialectical reasoning: tracing your premises to their contradictions, stripping the argument to its load-bearing claims, following your logic to conclusions you may not have intended, rebuilding with only what survived, and outputting a stress-tested artifact.
It's not "ask ChatGPT to argue with me." The adversarial structure is baked into the system, not prompted. You can't talk it out of pushing back.
Free tier gives you 3 sessions/month, no credit card. Would genuinely appreciate feedback from this community on whether the dialectical method produces meaningfully better reasoning outputs than single-pass LLM responses.