Timeline for Do LLMs suffer from a kind of Dunning-Kruger effect, giving an inflated self-assessment in domains they lack expertise in?
Current License: CC BY-SA 4.0
31 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 20 at 5:41 | comment | added | pete | @howlger "confabulation" means it's inventing something with an imagination and/or purposely lying, which is much more wrong than "hallucination". A hallucination happens when one is unable to distinguish between fantasy and reality, which is much closer to what actually happens in an LLM. From the LLM's point of view, if you feed a news article that World War 4 started, it has no way of knowing whether it's real. It is, in my opinion, one of the most fitting terms ever chosen for AI and wrongly gets a lot of backlash. I cannot think of a more fitting word. | |
| S Jul 31 at 16:29 | history | suggested | Vikas Sharma | CC BY-SA 4.0 |
fixed grammar
|
| Jul 10 at 6:35 | review | Suggested edits | |||
| S Jul 31 at 16:29 | |||||
| Aug 4, 2023 at 16:26 | comment | added | JimmyJames | @user10186832 I can agree that both are anthropomorphic and if we could eliminate that entirely, it would be optimal. But I don't see that happening any time soon. The general public simply doesn't have the vocabulary. Anyway, both terms are analogies but one is far more accurate than the other. The process of how LLMs produce wrong answers is much like making stuff up (a.k.a: BS) than 'believed to be true or real but that is actually false or unreal.' I just think it's easier to understand how you can make stuff up without having beliefs (correct or otherwise.) | |
| Aug 4, 2023 at 15:43 | comment | added | pcpthm | This is stretching to call it "self-assessment", but "confidence" (output probability) can be fairy calibrated on the base model (no under/overestimation of ability) but a later fine-tuning may destroy the calibration, according to the GPT-4 report <cdn.openai.com/papers/gpt-4.pdf#page=12>. | |
| Aug 4, 2023 at 15:08 | answer | added | Barmar | timeline score: 3 | |
| Aug 4, 2023 at 15:03 | answer | added | Uk rain troll | timeline score: 1 | |
| Aug 4, 2023 at 12:24 | comment | added | barbecue | I see a lot of excessively literal-minded interpretation of human language in these discussions. OP DID NOT ASK if the actual Dunning-Kruger effect applies to LLMs. OP asked if there could be something analogous to DK with LLMs. A different question. Every answer that says Dunning-Kruger can't apply because LLMs aren't human is answering the wrong question. | |
| Aug 4, 2023 at 11:44 | comment | added | howlger | @user10186832 IMHO confabulation is better than hallucination, but both terms are anthropomorphic, and I agree with you that this should be better avoided. But for me it is not an error either, because the computer runs the model correctly and the model, does what it was trained to do. I would like to see a better and generally accepted term for this instead, which to my knowledge is not yet the case. So I will try to avoid those terms in the future. ;) | |
| Aug 4, 2023 at 11:02 | answer | added | NoAnswer | timeline score: 2 | |
| Aug 4, 2023 at 9:39 | comment | added | MT1 | @howlger confabulation is still anthropomorphic. It's an error, it is from a computer. | |
| Aug 4, 2023 at 9:30 | comment | added | MT1 | @howlger Nope ... dictionary.cambridge.org/dictionary/english/confabulation | |
| Aug 4, 2023 at 9:12 | comment | added | howlger | @JimmyJames Yes, I agree that confabulation is a better term than the more common term hallucination. | |
| Aug 4, 2023 at 8:46 | comment | added | howlger | @user10186832 Sure, as I said it might look similar, but the cause is quite different than the Dunning-Kruger effect. In your answer you said "LLMs aggregate references". But technically, there is no aggregation, but generalizing when training (or sometimes unfortunately overfitting) and interpolating when inferring. Bard has generalized that "How many" asks for a number, but Bard doesn't know how much data it was trained with as long as neither the prompt nor the training data contain that information as text. Bard just gave you a random number. | |
| Aug 4, 2023 at 6:07 | comment | added | MT1 | @howlger I asked Bard "How many references do you have to Python in your training data?" The answer it gave was 2,000. No DK effect there then! It was totally focussed on the computer language with no mention of snakes. | |
| Aug 3, 2023 at 18:37 | comment | added | JimmyJames | @howlger I really think we should start using the term 'confabulation' instead of hallucination. The latter is too anthropomorphic. | |
| Aug 3, 2023 at 12:45 | comment | added | howlger | LLMs do interpolate, not extrapolate, and in cases of too less training data, this leads to hallucination rather than the extrapolated conclusion "I know that I know nothing". The outcome might seem similar to the Dunning-Kruger effect, but it has quite different causes. | |
| Aug 3, 2023 at 10:17 | answer | added | OverLordGoldDragon | timeline score: 1 | |
| Aug 3, 2023 at 6:58 | history | edited | Rebecca J. Stones | CC BY-SA 4.0 |
added 1 character in body; edited title
|
| Aug 3, 2023 at 6:56 | comment | added | MT1 | @MikeWise you could just ask it directly. I asked Bard "Do you suffer from the Dunning-Kruger effect" and it gave an interesting answer, almost as though a human had intervened to help it! | |
| S Aug 3, 2023 at 2:05 | history | suggested | Wicket |
edited tags
|
|
| Aug 2, 2023 at 22:22 | comment | added | Mike Wise | You could test it by asking it to mimic someone under the DK effect and then "not do that". I think if enough people write enough text on a topic that suffers from the DK effect and the LLM is trained on, it, then it will mimic that effect to score higher on the training. | |
| Aug 2, 2023 at 21:55 | comment | added | Wicket | I moved the comments to genai.meta.stackexchange.com/a/147/12 as I realized that they might require to be extended, refining and probably restructured. | |
| Aug 2, 2023 at 21:37 | review | Suggested edits | |||
| S Aug 3, 2023 at 2:05 | |||||
| Aug 2, 2023 at 18:04 | answer | added | Gumpf | timeline score: 2 | |
| Aug 2, 2023 at 17:13 | answer | added | NoAnswer | timeline score: 24 | |
| Aug 2, 2023 at 14:30 | history | became hot network question | |||
| Aug 2, 2023 at 11:18 | answer | added | MT1 | timeline score: 6 | |
| Aug 2, 2023 at 6:44 | answer | added | Franck Dernoncourt | timeline score: 9 | |
| Aug 2, 2023 at 6:41 | history | edited | Rebecca J. Stones | CC BY-SA 4.0 |
added 20 characters in body; edited tags; edited title
|
| Aug 2, 2023 at 6:29 | history | asked | Rebecca J. Stones | CC BY-SA 4.0 |