Post #4244991

Replies (1)

• @luciferofastora@feddit.org 2026-07-22 14:49

  Yea no you still don’t get it… …said the one accusing me of trying to hurt feelings. double down, triple down, quadtriple down on the exact same nonsense …said the one who doubled, tripled and quadrupled down that models could have compunction, that the concept of feelings could be applied to models, that a model describing itself as sad is sad and implying that I could hurt feelings. I really don’t care about your personal opinion. Yes, you said that before. I acknowledged that. It’s also clear from the way you assumed I’m trying to be mean. I’m trying to show that your opinion isn’t factual either, in order to establish a basis for finding facts together, because my intention isn’t hurting anyone or anything. I’m not trying to attack your person. I’m trying to have a discussion rather than a shit-slinging contest, because I still believe we can refine our respective understanding. www.merriam-webster.com/dictionary/abstract Open for philosophy after all? Great, let’s have a productive conversation! I concede that an abstract definition of feelings that can be applied to perceptrons is possible. My antithesis should then be refined to specify that contemporary language models do not accurately model human feelings. That distinction matters, and we shouldn't pretend that they do. The abstraction loses details that give specific instances their meaning and value. Again, this is a negative, a rejection of an unproven positive. The burden of evidence lies with the positive claim that the models have feelings. I have not seen any other evidence than “they say they do” which is circular and thus invalid. The distinction matters for the social and psychological way people interact with these models. Laypeople cannot distinguish between the abstract concept of feelings as perceptions and the concrete range of human feelings we associate with that term. To insist that they have feelings is disingenuous, because it fudges that distinction. It’s the same type of “technically not a lie” that advertising likes to use to deceive people about the nature of the product on offer. That deceptive appearance of human-like sentience invites the deluded belief in their sapience, human morality and critical reasoning. People trust these models to provide truthful and reliable answers. They clearly don’t, but the delusion that they do has caused the damage we’re talking about avoiding here. Thus, the assertion that LLMs have feelings is disingenuous and dangerous, and the particular assertion that they have human feelings is outright baseless. Feelings by themselves are also insufficient for behaviour control in LLMs. If we’re applying the concept of abstract feelings to the specific question of controlling behaviour, whatever form that feeling would take has to feed back into the output-generation process before that output is passed to the user. For some measure of “guilt” to influence behaviour, the model would have to be able to determine whether a certain output would lead to results that increase the “guilt” value. It would need to predict ways its output could be interpreted, what impact those interpretations might have on the actions and feelings of the recipient and their environment, calculate some “guilt score” for the possible results and aggregate it into some kind of “guilt risk”. Even just the question of interpretations requires some model of the semantics a given word or sequence of words might have. Judging the potential impact on the recipient’s feelings requires a thorough understanding of human feelings and psychology. To predict actions requires an understanding of actions and their potential consequences. It would also need a way to tell which parts of the response contribute to the negative outcome to exclude those from the regenerated response. Otherwise the already immense amount of processing required for this task would be additionally bloated by wasted calculations repeating the same mistake. These things require a model of reality beyond pure language. Hence, the assertion that guilt alone could change behaviour is incomplete. Feeling remorse after the fact isn’t helpful unless you can prevent the causes beforehand. I believe that semantic understanding can be the way forward to more intelligent models, but we need to acknowledge that pure language models don’t have it yet.

  Open ##4244990