[catchnear4321]

Second point is kind of harsh on the model behavior since that’s a product of the data, the training, and the user.

It is possible to treat it nicely and have it respond in kind. Most users just don’t consider that to be a worthwhile expense of their mental capacity.

Microsoft also forced Google’s hand. Would Google have ever wanted to augment search on their own? Sounds like a massive risk to Google ad revenue…

All that said, 1 is chef kiss.

[Adraghast]

> Most users just don’t consider that to be a worthwhile expense of their mental capacity.

Beyond that, I fundamentally don’t think people should be trained to be “nice” to technology. I don’t have to politely ask a hammer to pound in a nail and - the fact we’re talking about NLP notwithstanding — I shouldn’t have to politely ask Bing to provide me the results I’m looking for.

[catchnear4321]

Why not just taught to be nice to others?

And the NLP point matters quite a bit. ChatGPT can analyze the sentiment, and even offer adjustments.

This is less about being nice to technology and more about being aware of the impact of the self on the rest of the world. Technology just highlights the gap.

[Adraghast]

People should be taught to be nice to others, of course. The point is that LLMs are not “others”, they are inanimate tools. If I called a cashier a worthless piece of shit, that would be incredibly rude. If I said the same thing to Siri, it wouldn’t be, because it is not possible to be rude to software.

If my child were to say that to Siri though, I’d be concerned because, as you said, it could be highlighting something about the way they interact with the world. But I would still want it to respond to the command and leave the problem of my child’s bad manners to me. Unless there’s a major shift in our understanding of sentience, I consider teaching the delineation between humans, who are never unfeeling tools, and technology, which is always an unfeeling tool, equally as important as teaching mindfulness of one’s impact in the world. In fact, I don’t think you can actually understand the latter without understanding the former.

[catchnear4321]

> But I would still want it to respond to the command and leave the problem of my child’s bad manners to me.

No disagreement here. Especially as you recognize a problem to address in the situation.

The issue is the cause, for that’s the source of the training data.

Most people just… don’t, with regard to the rest of the world.

[penjelly]

users aren't always malicious, if Bing GPT thinks you're wrong and you persist it will get "aggressive". the 2022 vs 2023 issue shows that

[catchnear4321]

You don’t have to be malicious to be rude or inconsiderate. Few are approaching this as worth talking to like another person. It is a servant, by description, by design, so most treat it as such.

There is a bias to the interaction that most will never consider.

If you talk with a human, and the human thinks you are incorrect, and you insist, and neither of you attempts to smooth the conversation, plenty of humans also begin to get aggressive. Or at least irritable. (Which can escalate.)

[penjelly]

the difference with a human is, they'd concede they made a mistake. The user did as Bing asked, reported the date on their phone. Bing doubled down on incorrectness.

also, LLMs are not human, so no expectation to treat it as such required.

[catchnear4321]

Bing hallucinated as it was designed.

Treating things poorly because it is fine is how we got here.