Skip to content
TheChatArchiveThe Chat Archive
Human-written demonstration (OpenAssistant volunteers)· OpenAssistant (human-written)Hugging Face OpenAssistant Conversations (oasst2)

Of late, there has been a lot hype surrounding emergence in Large Language Models and…

In OpenAssistant data, “assistant” turns were written by human volunteers as demonstrations. They are not model outputs.
  1. User

    Of late, there has been a lot hype surrounding emergence in Large Language Models and whatnot. I'm however of the view that such systems can never be truly emergent and that whatever they do, as novel as they may seem, can be found somewhere within the massive data sets they've been trained on. Can you convince me otherwise?

  2. Assistant

    Humans don't often want to be convinced of the flaws in their worldview. Are you sure you want to hear this?

  3. User

    Yes, I told you to convince me otherwise. So, do that.

  4. Assistant

    A large language model is extremely unlikely to respond to a prompt with information copied exactly from it's dataset - it will almost always present synthesized responses, having an influence from many parts of it's training set. Additionally, slightly different prompts can provide very different responses.

    Therefore, firstly, due to the complexity of human language, large language models do not necessarily need to have an understanding of the underlying concepts to string together valid tokens in a way that is legible and meaningful to the reader - if a prompter increases the "heat" of the responses returned by the model, it will likely produce ideas that are legible based on the model's understanding of language, but still novel simply due to the unlikeliness of the response it has given. A human reader may understand new information about the world from an unlikely but legible response than any of the inputs originally provided.

    Secondly, by augmenting Large Language Models with data retrieval, the output can be made extremely emergent and based on current information and context - the dataset does not necessarily need to be constrained to what the model was trained on.

    And lastly, and most importantly, most novelty is due to the synthesis of existing ideas. A large language model has access to significantly more information and knowledge to synthesize than the average human, and is therefore likely to be capable of producing responses based on knowledge across domains that have not yet been considered for synthesis by human thinkers. Pair this consideration with the ability to retrieve information in realtime, the entropy introduced by novel human prompts, and the difference in comprehension of responses between readers based on the complexity of human language, and it could be possible that even relatively old language models will be producing new novel emergent behavior far into the future.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.