“Let’s think step by step” style prompts seems to unlock the ability for text generation that can give logically consistent answers to questions that have been out of the reach of most A.I. neural networks.
This might present an opportunity to test out ethical or moral questions on future LLMs that are multi-modal. Theorem proving might be particularly useful for MIRI type solutions. I am imagining a process where one could get answers to ethical questions, and use that feed back to create transparent teacher/critic learning paradigm. A dynamic interrogation where the A.I. could show its language AND math for why and how it came to its answers.
I believe a proto AGI like GATO likely will use the transformer architecture and this information could be used to structure better alignment tools. Perhaps train something that could help with utility functions.
Recently this paper has discovered a prompt that can unlock a form of reasoning process for large language models (LLM). https://paperswithcode.com/paper/large-language-models-are-zero-shot-reasoners. These models are used for natural language processing and typically use a transformer architecture (https://paperswithcode.com/paper/the-annotated-transformer).
“Let’s think step by step” style prompts seems to unlock the ability for text generation that can give logically consistent answers to questions that have been out of the reach of most A.I. neural networks.
This might present an opportunity to test out ethical or moral questions on future LLMs that are multi-modal. Theorem proving might be particularly useful for MIRI type solutions. I am imagining a process where one could get answers to ethical questions, and use that feed back to create transparent teacher/critic learning paradigm. A dynamic interrogation where the A.I. could show its language AND math for why and how it came to its answers.
I believe a proto AGI like GATO likely will use the transformer architecture and this information could be used to structure better alignment tools. Perhaps train something that could help with utility functions.