4 ms·

Ahh but we known that both humans and LLMs 'hallucinate'. The study of the latter potentially provides insight to former; after all, isn't the core question we'

by 23B1 3y ago

Ahh but we known that both humans and LLMs 'hallucinate'. The study of the latter potentially provides insight to former; after all, isn't the core question we're asking in all of these discussions is 'what is intelligence?'

smt88 3y ago
> The study of the latter potentially provides insight to former It absolutely does not. Using the term "hallucinate" for LLMs has nothing to do with the underlying cause or process. It's a metaphor that I feel was specifically chosen by the AI industry to avoid terms like "lying" or "making a mistake". LLMs hallucinate because they're next-word prediction engines, not logical engines. They have no conceptual understanding, which is why their hallucinations range from small factual errors to bizarre lies that are obviously false. Humans hallucinate because our brains are reality-simulation engines that evolved to model the perception of events that aren't currently happening (remembering the past, imagining the future, etc.) and sometimes we lose our autonomy over our brains. The underlying process has nothing to do with predicting the next word we're going to say. In fact, there's an even more fundamental difference: human hallucinations aren't necessarily tied to speech. We have thoughts that don't require output. There's nothing in LLMs that doesn't "come out" as output.
- 23B1 3y ago
  Note my use of scare quotes around 'hallucinate' The user of an LLM still gets a vote, right or wrong. If we have a good understanding of how the human mind works - empathy - we can gain a better understanding on how to clarify the very semantic confusion of which you speak! Here's another example that just popped up in my feed: https://kottke.org/23/07/will-ai-change-our-memories https://kottke.org/23/07/will-ai-change-our-memories

4 ms·

Ahh but we known that both humans and LLMs 'hallucinate'. The study of the latter potentially provides insight to former; after all, isn't the core question we'

by 23B1 3y ago

smt88 3y ago
> The study of the latter potentially provides insight to former It absolutely does not. Using the term "hallucinate" for LLMs has nothing to do with the underlying cause or process. It's a metaphor that I feel was specifically chosen by the AI industry to avoid terms like "lying" or "making a mistake". LLMs hallucinate because they're next-word prediction engines, not logical engines. They have no conceptual understanding, which is why their hallucinations range from small factual errors to bizarre lies that are obviously false. Humans hallucinate because our brains are reality-simulation engines that evolved to model the perception of events that aren't currently happening (remembering the past, imagining the future, etc.) and sometimes we lose our autonomy over our brains. The underlying process has nothing to do with predicting the next word we're going to say. In fact, there's an even more fundamental difference: human hallucinations aren't necessarily tied to speech. We have thoughts that don't require output. There's nothing in LLMs that doesn't "come out" as output.
- 23B1 3y ago
  Note my use of scare quotes around 'hallucinate' The user of an LLM still gets a vote, right or wrong. If we have a good understanding of how the human mind works - empathy - we can gain a better understanding on how to clarify the very semantic confusion of which you speak! Here's another example that just popped up in my feed: https://kottke.org/23/07/will-ai-change-our-memories https://kottke.org/23/07/will-ai-change-our-memories