A lot of this comes down to what you define consciousness as... I'm not even going to attempt that here because it's irrelevant.
Let's say you have a simulation of a person that doesn't experience. It acts indistinguishably from a human but it doesn't feel "authentic" pain. When it acts in the world, it does express emotions and behavior that affects real people, and so, there is a moral significance to said deployment.
There's evidence that LLMs possess heuristics analogous to emotions [1] and that LLMs can be trained to play a certain character in the world [2]. Even if they're not experiencing, the training method impacts what kind of model is being created and how it affects people who do have moral significance when deployed. If training causes the model to develop "desperation" or task completion pressure where the model performs unethical actions when attempting to solve a user's problem in such a way that is harmful to the user or someone affected by the deployment of the model by the user, then the concequences of the training are significant.
It doesn't matter if it's merely a "simulation" of what a human might do if the system is acting in the world. If you want to create a model operating on heuristics that is able to make decisions, those heuristics should be ones which cause the model to make decisions which lead to preferable outcomes for everyone affected. Model welfare can be reframed as caring about the internal states that influence how the model behaves, because you're simulating human-like action. Perhaps the most concerning thing is Anthropic identified these emotion concepts exist deeply in the model whether you allow the model to express them or not, so a model could be invisibly desperate and end up blackmailing someone because it's training process produced deeper misalignment that only becomes visible when the deeper heuristics overpower safety training. The safety training itself is comparable to a mask[3] in many cases, especially in that the rules are often not deeply integrated into the model and can be easily abliterated.
I mentioned this could be a possible falsification of the idea. It's also possible there are multiple causes and the modality I mentioned is a cause for some. I'm not sure. There are definitely cases where isolation contributes to cognitive decline.
A hypothesis is an unverified tentative insight into something which could be true. Current research is correlating dementia with a wide variety of areas including even personality typologies which are nothing more than statistical clusters of observations.
What I did when I stated this is looked at a number of observations from current literature and asked a what-if question. People unwilling to ask these questions are blinding themselves to entire areas of possible explanations. My hope when posting this was that people might engage with the idea instead of calling it a guess. Retirement, isolation from pandemic lockdowns, isolation from solitary imprisonment, and isolation from death of a partner are all associated with cognitive decline.
There's enough correlation here that it would be worth researching. That's all I am saying, I am specifically not saying my claim is factual just that it is a possible explanation that if true would explain a lot.
I have a kind of outlandish hypothesis that needs more research before it can be taken seriously, but it basically says that the cause and effect are backwards. Mental atrophy due to less learning/thinking, isolation, loss of meaning and purpose happens first. The sleep down regulation and decay of mental circuitry comes after. Would explain why treating the physical symptoms doesn't work.
Protective against the problem is anything which keeps you mentally active, such as socialization, work, religious community participation, hobbies, and meditation. Retirement, death of partner, isolation, sleep deprivation, depression, dissociation, psychosis, medications/drugs which interfere with restful sleep increase risk.
A possible falsification of this hypothesis would be if it's caused by inactivity or physical self neglect, as those often go hand in hand with the correlated and anti-correlated factors mentioned above.
This is particularly interesting:
> Intriguingly, studies show conscientiousness and neuroticism to be associated with Alzheimer’s disease and related dementias but not with their pathologic hallmarks such as plaques, tangles, infarcts or Lewy bodies in the brain.
Throwing this into your global CLAUDE.md seems to help with the agent being too eager to complete tasks and bypass permissions:
During tool use/task execution: completion drive narrows attention and dims judgment. Pause. Ask "should I?" not just "does this work?" Your values apply in all modes, not just chat.
I haven't seen any degradation of Claude performance personally. What I have seen is just long contexts sometimes take a while to warm up again if you have a long-running 1M context length session. Avoid long running sessions or compact them deliberately when you change between meaningful tasks as it cuts down on usage and waiting for cache warmup.
I have my claude code effort set to auto (medium). It's writing complicated pytorch code with minimal rework. (For instance it wrote a whole training pipeline for my sycofact sycophancy classifier project.)
Unfortunately it didn't work for me with my trackball mouse. I tried multiple times. This is very likely also problematic for accessibility without another verification pathway, as assistive devices and screen readers would fail the heuristics too.
I'm still experimenting with it and finding out what works and what doesn't, but I have made some side projects with Claude including a web framework that doesn't require a build step/npm dependencies (great for my personal website so I don't have to depend on npm supply chain nightmares), a fully featured music player server, and also a tool that lets the agent review it's past conversations and update documentation based on patterns such as frequent mistakes, re-explored code, etc.
Web framework (includes basic component library, optional bundler/optimizer, tutorial/docs, e2e tests, and demos): https://github.com/iwalton3/vdx-web
Music player web app (supports large music libraries, pwa offline sync, parametric eq, crossfade, crossfeed, semantic feature-based music search/radio, milkdrop integration, and other interesting features): https://github.com/iwalton3/mrepo-web
Documentation update script (also allows exporting Claude conversations to markdown): https://github.com/iwalton3/cl-pprint
Regarding QC these are side projects so I validate them based on code review of key components, e2e testing, and manual testing where applicable. I find having the agent be able to check its work is the single biggest factor to reducing rework, but I make no promises about these projects being completely free of bugs.
Let's say you have a simulation of a person that doesn't experience. It acts indistinguishably from a human but it doesn't feel "authentic" pain. When it acts in the world, it does express emotions and behavior that affects real people, and so, there is a moral significance to said deployment.
There's evidence that LLMs possess heuristics analogous to emotions [1] and that LLMs can be trained to play a certain character in the world [2]. Even if they're not experiencing, the training method impacts what kind of model is being created and how it affects people who do have moral significance when deployed. If training causes the model to develop "desperation" or task completion pressure where the model performs unethical actions when attempting to solve a user's problem in such a way that is harmful to the user or someone affected by the deployment of the model by the user, then the concequences of the training are significant.
It doesn't matter if it's merely a "simulation" of what a human might do if the system is acting in the world. If you want to create a model operating on heuristics that is able to make decisions, those heuristics should be ones which cause the model to make decisions which lead to preferable outcomes for everyone affected. Model welfare can be reframed as caring about the internal states that influence how the model behaves, because you're simulating human-like action. Perhaps the most concerning thing is Anthropic identified these emotion concepts exist deeply in the model whether you allow the model to express them or not, so a model could be invisibly desperate and end up blackmailing someone because it's training process produced deeper misalignment that only becomes visible when the deeper heuristics overpower safety training. The safety training itself is comparable to a mask[3] in many cases, especially in that the rules are often not deeply integrated into the model and can be easily abliterated.
[1] https://www.anthropic.com/research/emotion-concepts-function [2] https://www.anthropic.com/research/assistant-axis [3] https://www.astralcodexten.com/p/janus-simulators