The author suggests that this is a fundamentally hard problem in the current LLM scheme and proposes potential solutions. These include curating a fine-tune dataset of good brainstorm examples on non-conventional topics, using methods like RLAIF to iteratively critique LLM’s response in terms of creativity, and changing the training process to seek out knowledge, thinking, and deductive reasoning skills. The author also raises several big questions in the field, such as the need for a world model, the correct paradigm for AGI, and the need for real-world feedback to reach AGI.
Key takeaways:
- Language Learning Models (LLMs) like GPT-4 have some independent and creative thinking abilities, but they are not good tools for effective brainstorming, especially in cutting-edge scenarios.
- LLMs are trained to follow existing patterns in human-produced corpus and not natively taught to brainstorm, often converging to consensus in existing data.
- LLMs are only suitable for better-than-average level brainstorms and cannot provide much useful insights for truly frontier problems.
- The author suggests potential solutions such as curating a good fine-tune dataset of good brainstorm examples, using methods like RLAIF to critique LLM’s response in terms of creativity, and changing the training process to seek out knowledge, thinking and deductive reasoning skills.