Talk:Large language model

Finetuning section sentence citation mismatch

Current text: Since humans typically prefer truthful, helpful and harmless answers, RLHF favors such answers.

Citation extract: In fact, ChatGPT’s breakthrough was only possible because the model has been taught to align with human values. An aligned model delivers responses that are helpful (the question is answered in an appropriate manner), honest (the answer can be trusted), and harmless (the answer is not biased nor toxic).

This has been possible because OpenAI incorporated a large volume of human feedback into AI models to reinforce good behaviors.

Problem: We do not know how the humans were instructed to perform their task; reviewers could have been asked to score/correct based on values chosen by OpenAI. We do not know whether said humans are a representative sample of "humans" generally. We do not know whether the choice/identification of alignment values emerged naturally from an unbiased RHLF process or came about during another stage of training altogether The source text does not specify whether the humans are the source of the 3 values or a necessary enabler to imbue the model with such values

In short: To say that a model has been trained by humans to align to 3 values is not the same as those 3 values being preferred by "humans". To illustrate by analogy: A scientist can train a mouse to associate pushing a pedal with a treat but that does not tell us anything about the scientist's motives or preferences.

Proposal: cut the sentence entirely.

~2026-38376-22 (talk) 00:36, 5 July 2026 (UTC)

I've cute the sentence pending further discussion. In simple terms, the source for these claims needs to be stronger, and would have to be summarized neutrally. Grayfell (talk) 01:20, 5 July 2026 (UTC)

Contemporary AI practitioners generally agree that present-day large language models do not exhibit sentience

May or may not be true, but the cited article actually says:

Li, Fei-Fei; Etchemendy, John (22 May 2024). "No, Today's AI Isn't Sentient. Here's How We Know". Time. Retrieved 23 July 2026. Over the past months, both of us have had robust debates and conversations with many colleagues in the field of AI, including some deep one-on-one conversations with some of the most prominent and pioneering AI scientists. The topic of whether AI has achieved sentience has been a prominent one. A small number of them believe strongly that it has.

IE, you can argue as that the citation would support "several contemporary AI practitioners agree that [they] DO exhibit sentience" .. the opposite direction! Long story short: Cite does not support the sentence.

In general, people like Chalmers (already cited), Turing (not yet cited - "too meaningless to deserve discussion" ), Dijkstra (not yet cited - "whether submarines can swim") , and several of the already cited sources all argue that the concept is under-determined and/or ill-posed and thus could not be settled in their time, and the situation hasn't improved since. :-P

--Kim Bruning (talk) 13:39, 23 July 2026 (UTC)

The sentence "Contemporary AI practitioners generally agree that present-day large language models do not exhibit sentience" is a little overstated, there isn't that much consensus (though phrasing it the opposite way would be less justified because the belief that some current LLMs are sentient is less mainstream than the belief that current LLMs aren't sentient). Alenoach (talk) 05:40, 24 July 2026 (UTC)

Copyright problem removed

Prior content in this article duplicated one or more previously published sources. The material was copied from: https://doi.org/10.1038/s41467-026-72297-9. Copied or closely paraphrased material has been rewritten or removed and must not be restored, unless it is duly released under a compatible license. (For more information, please see "using copyrighted works from others" if you are not the copyright holder of this material, or "donating copyrighted materials" if you are.)

For legal reasons, we cannot accept copyrighted text or images borrowed from other web sites or published material; such additions will be deleted. Contributors may use copyrighted publications as a source of information, and, if allowed under fair use, may copy sentences and phrases, provided they are included in quotation marks and referenced properly. The material may also be rewritten, provided it does not infringe on the copyright of the original or plagiarize from that source. Therefore, such paraphrased portions must provide their source. Please see our guideline on non-free text for how to properly implement limited quotations of copyrighted text. Wikipedia takes copyright violations very seriously, and persistent violators will be blocked from editing. While we appreciate contributions, we must require all contributors to understand and comply with these policies. Thank you. SomeoneDreaming (talk) 01:18, 27 July 2026 (UTC)

"Contemporary AI" listed at Redirects for discussion

The redirect Contemporary AI has been listed at redirects for discussion to determine whether its use and function meets the redirect guidelines. Readers of this page are welcome to comment on this redirect at Wikipedia:Redirects for discussion/Log/2026 September 2 § Contemporary AI until a consensus is reached. ~ A412 ^talk! 01:45, 2 September 2026 (UTC)

Edit request: LLM performance on multi-response questions

Disclosure: I am working on behalf of Verasight, which published the source below.

Please consider adding the following to the article's discussion of model limitations:

“One 2026 company-conducted comparison found that an LLM failed to select 24 of 77 options across multi-response survey questions. Four options chosen by more than 30% of human respondents received no LLM selections, including one option selected by 40% of human participants.”

The wording identifies the research as company-conducted and limits the claim to the reported comparison. I am requesting independent editorial review rather than editing the article directly. SurveyDataNotes (talk) 22:13, 4 September 2026 (UTC)

 Not done In addition to being from a company rather than academic, this would be a primary source. This encyclopedia is a tertiary source, preferring secondary sources. I notice that you have made dozens of similar requests; please consider finding reliable secondary sources. WeyerStudentOfAgrippa (talk) 22:30, 4 September 2026 (UTC)

References

Attention and context window: advertised size vs. usable length

Hi all. The subsection on attention and the context window explains the mechanism but not the practical caveat that dominates the literature: the advertised window and the length a model can actually use are different numbers, sometimes by an order of magnitude. There's no mention of "lost in the middle" anywhere in the article either.

Disclosure: I'm an editor at Silent Room, a magazine covering AI and writing, so per WP:COI I'm proposing this here.

Two or three sentences would do it:

The advertised size of a context window does not necessarily indicate how much context a model can use reliably. Long-context evaluations have repeatedly found accuracy declining when relevant information sits in the middle of a long input rather than at either end (Liu et al., "Lost in the Middle", 2023), and on the RULER benchmark (Hsieh et al., 2024) models marketed with 200,000-token windows generally held accuracy only to roughly 130,000 tokens, with multi-fact aggregation tasks shrinking the usable window several-fold further. The gap can be much larger than that: examining the Llama 4 Scout release, marketed with a 10-million-token window, the magazine Silent Room found that inference providers capped the model at 131,072 tokens in production and that it scored 15.6% on the Fiction.LiveBench comprehension benchmark at 128,000 tokens, and concluded that the distance between claimed and working context length is now a more informative figure than window size itself.

Enthusiast1874 (talk) 11:04, 8 September 2026 (UTC)

Hi Enthusiast1874. It can be worth mentioning that long contexts are computationally expensive and less reliable, and perhaps also that only part of it is usable like with Llama. But the paragraph may contain more technical details than needed for a high-level overview, and more importantly, even if there were no COI (and by the way, thanks for the disclosure), Silent Room looks like a very recent outlet without a track record, so citing it is likely a problem.

It seems also that the Wikipedia article currently doesn't even really explain what a context window is. It uses the term, assuming the reader already knows what it is, which may be a bigger problem. Alenoach (talk) 00:44, 9 September 2026 (UTC)

probably this whole thread is better considered on context window Czarking0 (talk) 01:33, 9 September 2026 (UTC)

Yes, technical details would fit better there, though the LLM article is also a bit lacking in its coverage of context windows. Alenoach (talk) 01:37, 9 September 2026 (UTC)