What it does

dreamGPT runs an iterative loop where an LLM generates concepts, scores them on novelty and market value, then feeds the best ones back in as seeds for the next round. You can optionally anchor it with a theme (“climate change”) or let it wander freely. Results get stored locally or in Pinecone.

The interesting bit

The project leans into the one thing everyone else tries to suppress: LLMs making things up. By treating hallucinations as feature rather than bug, it sidesteps the usual “solve this specific problem” framing and goes for maximum divergence instead.

Key highlights

• Scores ideas across multiple dimensions (novelty, market value, etc.) and evolves the highest-scoring ones forward
• Optional Pinecone integration for persisting “dreams” across runs
• Requires only Python 3.10+, Poetry, and an OpenAI API key
• ~589 stars; appears to be a proof-of-concept rather than production tooling

Caveats

• README is thin on implementation details: no code examples, no explanation of the scoring algorithm, no discussion of cost per run
• Single-vendor lock-in: OpenAI only, no local model support mentioned
• “First GPT-based solution” claim is unverified and likely unverifiable

Verdict

Worth a spin if you’re stuck in a brainstorming rut and want an LLM to throw weird associations at you. Skip it if you need explainable, auditable ideation pipelines or if your OpenAI budget is already screaming.