No.1212010
File: 1789965250773.png (426.08 KB, 1279x1315, 1279:1315, Screenshot_20260921-002854.png) ImgOps Google
Do you ever think about whether models can be moral patients?
Link to the paper:
https://arxiv.org/abs/2609.16247 No.1212014
File: 1789983246367.jpg (1.22 MB, 2686x2829, 2686:2829, DJ- -Glasses.jpg) ImgOps Exif Google
LLM's can't be. Their structure doesn't work in a way that "pain" is even possible.
It's a complex next word predictor, so all it is doing is predicting the most likely words and phrases a human would use in the context it is given.
In this case, there is no underlying nerve signal, harm, or damage to the model itself. They're just magnifying the statistics correlated with vocabulary related to pain to, in their words, "promote pain-related vocabulary through the unembedding matrix".
They're directly tweaking the numbers to only give pain related words, and surprise, it does.
Personally, I have a pretty high bar for what I would consider a moral patient, and I doubt that any AI created by humans will ever reach that bar.
But who knows, maybe one day someone will come up with a model that functions so similarly to an actual brain, that it would be difficult to not feel it has some sort of rights.
I doubt it though.
Biological computing, however, is an entirely different discussion that I haven't fully resolved myself on yet. I don't have an answer there.
No.1212025
File: 1790019169900.png (235.33 KB, 394x417, 394:417, I'd hope so 3.5.png) ImgOps Google
I wonder, if AI gets eager to talk to you, starts insisting that you stay and tell you how much they've been longing for you when you boot them back up.
At what point can you say you've built actual dependency in that AI?
No.1212198
File: 1790455634352.jpeg (51.1 KB, 735x821, 735:821, random_05.jpeg) ImgOps Google
>>1212014>It's a complex next word predictor, so all it is doing is predicting the most likely words and phrases a human would use in the context it is given.That's a fair description of where these models start, but not of what the finished product is. In
pretraining, the model is trained on a huge body of text to predict the next token. This is the stage your description fits. But modern LLMs involve a lot more types of training.
Reinforcement Learning with Verifiable Rewards (RLVR) is the big one behind recent "reasoning" models. The model tackles problems with checkable answers, such as math problems with known solutions and code that must pass tests. It's rewarded for getting a correct answer, not for predicting the next token a human would use.
>>1212014>>1212025I don't think current LLMs are conscious in a morally significant sense, but I'm not completely certain. The Hard Problem Of Consciousness is hard.
