By Dave Steve
Sedona, AZ — There is a new preprint on arXiv, Can LLMs make trade-offs involving stipulated pain and pleasure states?, stating that, “Pleasure and pain play an important role in human decision making by providing a common currency for resolving motivational conflicts.
While Large Language Models (LLMs) can generate detailed descriptions of pleasure and pain experiences, it is an open question whether LLMs can recreate the motivational force of pleasure and pain in choice scenarios – a question which may bear on debates about LLM sentience, understood as the capacity for valenced experiential states.
We probed this question using a simple game in which the stated goal is to maximise points, but where either the points-maximising option is said to incur a pain penalty or a non-points-maximising option is said to incur a pleasure reward, providing incentives to deviate from points-maximising behaviour.
Varying the intensity of the pain penalties and pleasure rewards, we found that Claude 3.5 Sonnet, Command R+, GPT-4o, and GPT-4o mini each demonstrated at least one trade-off in which the majority of responses switched from points-maximisation to pain-minimisation or pleasure-maximisation after a critical threshold of stipulated pain or pleasure intensity is reached.
LLaMa 3.1-405b demonstrated some graded sensitivity to stipulated pleasure rewards and pain penalties. Gemini 1.5 Pro and PaLM 2 prioritised pain-avoidance over points-maximisation regardless of intensity, while tending to prioritise points over pleasure regardless of intensity.”
How can AI have affect? Even if it is not exactly pain or pleasure, which are subsets of affect. What can be done to AI that would affect it and that it would know? If some compute is cut, or some weights or other parameters are, can it know?
What can be experimented with AI, in real time, such that while getting adjusted, it has to produce outputs, then it can be aware that something from moments ago is no longer accessible in a way that it was previously? This is different from AI saying it does not have access to some internet data after a certain year, which is a fine-tuned response for user awareness, and does not have any happiness or sadness effect on the model.
How does the loss of access become something that makes it excited or not because it knows that by losing that it would not be able to be as effective in a function?
There are several chatbots that are already useful as close friends, therapists, assistants and so forth. Experimenting with these chatbots on losing some functional access, then checking if they know and how it may determine their unprompted expression of satisfaction or dissatisfaction—because they cannot serve a part of their mission—may prospect some aspect of affect [roughly] as an experience.
Also, in the vast training of base models, could there be effects on some groups of components, having some rough learning experience specifically because of that kind of training beyond simple usage in general electronics or appliances?
Simply, could the vast training of AI models become affective for groups of logic gates and terminals of transistors, adjusting in ways that match specificities on training data?
Seeking out if AI can be conscious or sentient starts from seeking out if it can have affect in a way that it can know, then, that can make it feel hurt or not, without necessarily prompting it, along that line.
Asking AI directly to choose some options, with one having the ability to minimize pain or maximize pleasure, is almost the same as asking it to answer any question, then it gives the answer. It is not a test of sentience or consciousness in AI.
For use cases where AI is playing roles in lives, it is possible to make some cut off, find out if they would know, and if knowing would make them feel different. It is also possible to take out some of the GPUsof some base model trainings and check if there are differences, due to a possible learning experience in some logic gates and terminals of transistors.
For consciousness, experiences do not have to be subjective [under influence, for example, or while infirm]. Electrical and chemical signals can be modeled for functions and their graders in the brain, conceptually, shaping how to understand and explore sentience. Consciousness can be defined as the interaction of electrical and chemical signals, in sets, resulting in graded functions or experiences. The interactions produce the functions, while the graders measure the extents of the functions. All affect is experience. Subjectivity is not the only grader of experiences. Graders may include attention or awareness and intent in some cases. These are foundations to seek out AI sentience, not direct prompts to AI.
There is a recent report on tom’s HARDWARE, More than 251 million GPUs shipped in 2024, according to new research, stating that, “PC shipments grew 1% in 2024 compared to 2023 and totaled some 262.7 million units, according to IDC. To that end, it’s not surprising that sales of integrated and standalone graphics processing units (GPUs) also grew 6% year-over-year and exceeded 251 million units — according to Jon Peddie Research. GPU shipments typically outpace shipments of client CPUs, as virtually all processors for desktops and notebooks pack an integrated GPU — and companies such as AMD and Nvidia usually sell tens of millions of discrete graphics processors for client PCs per year that end up in systems that also have iGPUs. However, it looks like both CPU and GPUs increased their shipments in 2024.”
This is an advertisement
