4 ms·

> that don't hallucinate “Hallucination” is part of thought. Solving a new problem requires hallucinating new, non existing, possible outcomes and solutions, t

by nomel 3y ago

> that don't hallucinate “Hallucination” is part of thought. Solving a new problem requires hallucinating new, non existing, possible outcomes and solutions, to find one that will work. It seems that eliminating the ability to interpolate and extrapolate (hallucinations) would make intelligence impossible. It would eliminate creativity, tying together new concepts, creation, etc. Is the goal AI, or a nice database front end, to reference facts? Is intelligence facts, or is it the flexibility and the ability to handle and create the novel, things that are new? The ability to have confidence, and know and respond to it, seems important, but that’s surely different than the elimination of hallucinations. I’m probably misunderstanding something, and/or don’t know what I’m talking about.

Barrin92 3y ago
>Is the goal AI, or a nice database front end, to reference facts? The latter given the kind of products that are currently being built with it. You don't want your code completion or news aggregator to hallucinate for the same reason you don't want your wrench to hallucinate, it's a tool. And as for hallucinations, that's a PR friendly misnomer for "it made **** up". Using the same phrase doesn't mean it has functionally anything to do with the cognitive processes involved in human thought. In the same way a 'artificial' neural net is really a metaphorical neural net, it has very few things in common with biological neurons.
- mitchdoogle 3y ago
  "hallucination" just means the AI produced incorrect information. Humans produce incorrect information all the time. But we know we are fallible. I've seen lots of people treating these AIs as infallible, that they can't get anything wrong. That's a huge problem with people, not the AI. Obviously it's worth it to try and eliminate the incorrect information, but what grand-op is saying is we don't want to do that if it takes away some valuable emergent properties.
  - zuminator 3y ago
    I think AI hallucinations are more than just incorrect information. They're incorrect information coupled with an air of certitude. Sometimes with a whole backstory. And often oblivious to any inherent contradictions. It's like, if you ask a random person what years Lyndon B Johnson was president, they might say something like 1964-1968, shrugging with a bit of uncertainty. And both of those years would be incorrect but close. Whereas an AI (if hallucinating) would say something like Lyndon B Johnson was president from April 15, 1865 – March 4, 1869, after the shooting of John F. Kennedy in Ford's theater. The dates are precise but completely wrong, as are the accompanying details. Then if you ask the AI when John F. Kennedy died, it might correctly respond November 22, 1963, in Dallas, TX, totally unaware how this information doesn't match with the erroneous information it gave earlier. I've been thinking there's some parallels between how AIs hallucinate and how human toddlers do. If you ask a toddler/young child a question about a fact they don't know, they will usually say, "iuno" (even when they should), but depending on the child and the circumstances, they will sometimes just make up a story on the spot and sound as if they believe it. "Who invented ice cream?" "Santa Claus! Mommy left him milk and cookies and he turned it into ice cream." It doesn't make any real sense but it seems facially plausible in their universe. But somewhere between first learning to speak and around 7ish, kids become markedly more accurate how they model the world, and their responses become correspondingly less fanciful. And they continue to improve beyond that point. So how are kids doing what LLMs are currently incapable of? How do we teach ourselves not to hallucinate? Or do we, really? I mean, if I tell myself I'm going to make it through the intersection before the light turns red, but I end up running the red light, was I just mistaken, or was that a self-delusion, i.e., a mini-hallucination of sorts? Probably a self-driving car would be less likely to make that category of mistake, so maybe I shouldn't be so smug about being grounded in reality.
- Applejinx 3y ago
  You see, there's your problem. Now that you mention it, I absolutely do want my wrench to hallucinate. A more salient question would be, 'how do I know it ISN'T hallucinating'…
ChatGTP 3y ago
“Hallucination” is part of thought. Any evidence to support this claim or just commentary ?
- nomel 3y ago
  Most intro courses to perception/cognition will cover the "gap filling" and "extrapolation" that takes place in a "data sparse" context that is the human experience/speech, but here's a related article [1]. I think "speculative thinking" and "heuristic simulation" are more appropriate/technical terms [2] (and surely involved in AGI). I used "hallucination" a bit liberally, but with text being the only "sense" of an LLM, I think it's somewhat reasonable. Disclaimer: I know little of this field. [1] https://www.scientificamerican.com/article/perception-and-memory/ https://www.scientificamerican.com/article/perception-and-me... [2] https://www.frontiersin.org/articles/10.3389/fpsyg.2021.728946/full#:~:text=Speculative%20thinking%20refers%20to%20thinking,prefactual%20thinking%2C%20and%20other%20types https://www.frontiersin.org/articles/10.3389/fpsyg.2021.7289....
usrbinbash 3y ago
The problem is, what we call "hallucinating" in LMs isn't a way of creative thinking and coming up with novel solutions. It also has nothing to do with "interpolate and extrapolate". It's simply when the predicted probable sequence isn't grounded in reality. When I ask an LLM to summarize the great water wars of 1999, and how the Trade Union was ultimately defeated by the Antarctic Coalitions hovercraft-fleet under Vice Admiral Zagalow, it isn't "extrapolating" from knowledge of history, it is simply inventing a load of bollocks. But that bollocks will be dressed in fine language and probably mixed in with plausible-sounding references that have a somewhat-logical-sounding relation to the training data. The problem is, the LM doesn't and cannot know when it produces bollocks. All it can care about is if the sequences produced are probable according to it's model.
- geysersam 3y ago
  It's a great illusionist. But ultimately it cannot separate relevant information from simple word correlations. > What is heavier, a small floating passenger ferry or a two metric ton heavy rock that sinks to the bottom of the ocean. > A two metric ton heavy rock would be heavier than a small floating passenger ferry. The weight of the rock is two metric tons, which is equivalent to 2,000 kilograms or 4,409 pounds. The weight of the passenger ferry would depend on its specific design and construction materials, but it is unlikely to be heavier than two metric tons. Therefore, the heavy rock would have a greater weight than the small floating passenger ferry. It completely relies on surface information such as "small floating" and ignores the deeper "correlation" that all ferries are heavy.
  - j-a-a-p 3y ago
    GPT-4 answer: The weight of an object is determined by its mass, regardless of whether it floats or sinks. So, when you ask which is heavier, a small passenger ferry or a two metric ton heavy rock, it all comes down to the actual mass of the ferry. A two metric ton rock weighs two metric tons by definition (or 2000 kilograms). However, a small passenger ferry, while it may look small compared to large ferries or ships, can weigh much more than two metric tons. Even a small passenger ferry can weigh dozens or even hundreds of tons, due to the mass of the hull, the engine, and other equipment on board. So, without specific information about the ferry's mass, it's safe to assume that a "small" passenger ferry is likely heavier than a two metric ton rock. However, if the ferry is particularly small and lightweight, or the term "ferry" is being used to describe a very small watercraft (like a raft or dinghy), it's possible for it to be lighter. You would need the specific weight of the ferry to give a definitive answer.

by nomel 3y ago

Barrin92 3y ago

>Is the goal AI, or a nice database front end, to reference facts? The latter given the kind of products that are currently being built with it. You don't want your code completion or news aggregator to hallucinate for the same reason you don't want your wrench to hallucinate, it's a tool. And as for hallucinations, that's a PR friendly misnomer for "it made **** up". Using the same phrase doesn't mean it has functionally anything to do with the cognitive processes involved in human thought. In the same way a 'artificial' neural net is really a metaphorical neural net, it has very few things in common with biological neurons.

mitchdoogle 3y ago
"hallucination" just means the AI produced incorrect information. Humans produce incorrect information all the time. But we know we are fallible. I've seen lots of people treating these AIs as infallible, that they can't get anything wrong. That's a huge problem with people, not the AI. Obviously it's worth it to try and eliminate the incorrect information, but what grand-op is saying is we don't want to do that if it takes away some valuable emergent properties.
- zuminator 3y ago
  I think AI hallucinations are more than just incorrect information. They're incorrect information coupled with an air of certitude. Sometimes with a whole backstory. And often oblivious to any inherent contradictions. It's like, if you ask a random person what years Lyndon B Johnson was president, they might say something like 1964-1968, shrugging with a bit of uncertainty. And both of those years would be incorrect but close. Whereas an AI (if hallucinating) would say something like Lyndon B Johnson was president from April 15, 1865 – March 4, 1869, after the shooting of John F. Kennedy in Ford's theater. The dates are precise but completely wrong, as are the accompanying details. Then if you ask the AI when John F. Kennedy died, it might correctly respond November 22, 1963, in Dallas, TX, totally unaware how this information doesn't match with the erroneous information it gave earlier. I've been thinking there's some parallels between how AIs hallucinate and how human toddlers do. If you ask a toddler/young child a question about a fact they don't know, they will usually say, "iuno" (even when they should), but depending on the child and the circumstances, they will sometimes just make up a story on the spot and sound as if they believe it. "Who invented ice cream?" "Santa Claus! Mommy left him milk and cookies and he turned it into ice cream." It doesn't make any real sense but it seems facially plausible in their universe. But somewhere between first learning to speak and around 7ish, kids become markedly more accurate how they model the world, and their responses become correspondingly less fanciful. And they continue to improve beyond that point. So how are kids doing what LLMs are currently incapable of? How do we teach ourselves not to hallucinate? Or do we, really? I mean, if I tell myself I'm going to make it through the intersection before the light turns red, but I end up running the red light, was I just mistaken, or was that a self-delusion, i.e., a mini-hallucination of sorts? Probably a self-driving car would be less likely to make that category of mistake, so maybe I shouldn't be so smug about being grounded in reality.
Applejinx 3y ago
You see, there's your problem. Now that you mention it, I absolutely do want my wrench to hallucinate. A more salient question would be, 'how do I know it ISN'T hallucinating'…

ChatGTP 3y ago

“Hallucination” is part of thought. Any evidence to support this claim or just commentary ?

nomel 3y ago
Most intro courses to perception/cognition will cover the "gap filling" and "extrapolation" that takes place in a "data sparse" context that is the human experience/speech, but here's a related article [1]. I think "speculative thinking" and "heuristic simulation" are more appropriate/technical terms [2] (and surely involved in AGI). I used "hallucination" a bit liberally, but with text being the only "sense" of an LLM, I think it's somewhat reasonable. Disclaimer: I know little of this field. [1] https://www.scientificamerican.com/article/perception-and-memory/ https://www.scientificamerican.com/article/perception-and-me... [2] https://www.frontiersin.org/articles/10.3389/fpsyg.2021.728946/full#:~:text=Speculative%20thinking%20refers%20to%20thinking,prefactual%20thinking%2C%20and%20other%20types https://www.frontiersin.org/articles/10.3389/fpsyg.2021.7289....

usrbinbash 3y ago

The problem is, what we call "hallucinating" in LMs isn't a way of creative thinking and coming up with novel solutions. It also has nothing to do with "interpolate and extrapolate". It's simply when the predicted probable sequence isn't grounded in reality. When I ask an LLM to summarize the great water wars of 1999, and how the Trade Union was ultimately defeated by the Antarctic Coalitions hovercraft-fleet under Vice Admiral Zagalow, it isn't "extrapolating" from knowledge of history, it is simply inventing a load of bollocks. But that bollocks will be dressed in fine language and probably mixed in with plausible-sounding references that have a somewhat-logical-sounding relation to the training data. The problem is, the LM doesn't and cannot know when it produces bollocks. All it can care about is if the sequences produced are probable according to it's model.

geysersam 3y ago
It's a great illusionist. But ultimately it cannot separate relevant information from simple word correlations. > What is heavier, a small floating passenger ferry or a two metric ton heavy rock that sinks to the bottom of the ocean. > A two metric ton heavy rock would be heavier than a small floating passenger ferry. The weight of the rock is two metric tons, which is equivalent to 2,000 kilograms or 4,409 pounds. The weight of the passenger ferry would depend on its specific design and construction materials, but it is unlikely to be heavier than two metric tons. Therefore, the heavy rock would have a greater weight than the small floating passenger ferry. It completely relies on surface information such as "small floating" and ignores the deeper "correlation" that all ferries are heavy.
- j-a-a-p 3y ago
  GPT-4 answer: The weight of an object is determined by its mass, regardless of whether it floats or sinks. So, when you ask which is heavier, a small passenger ferry or a two metric ton heavy rock, it all comes down to the actual mass of the ferry. A two metric ton rock weighs two metric tons by definition (or 2000 kilograms). However, a small passenger ferry, while it may look small compared to large ferries or ships, can weigh much more than two metric tons. Even a small passenger ferry can weigh dozens or even hundreds of tons, due to the mass of the hull, the engine, and other equipment on board. So, without specific information about the ferry's mass, it's safe to assume that a "small" passenger ferry is likely heavier than a two metric ton rock. However, if the ferry is particularly small and lightweight, or the term "ferry" is being used to describe a very small watercraft (like a raft or dinghy), it's possible for it to be lighter. You would need the specific weight of the ferry to give a definitive answer.