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

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.

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.

https://pmc.ncbi.nlm.nih.gov/articles/PMC7484344/

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.)

It's small enough it can run on a gaming GPU locally. It's got a GGUF checkpoint on hugging face and is available on ollama. You can pull it and run scenarios against it in minutes: https://ollama.com/izzie/sycofact

The synthetic training data is also public, you can train other models over the data or reproduce my results. The labels were all generated by Gemma 3 27B with activation steering based on generated contrastive data. A write-up is planned at a later date, feel free to get in touch if curious.

Comments URL: https://news.ycombinator.com/item?id=47573491

Points: 2

# Comments: 0

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.

It has a pretty high false positive rate though, but it reliably highlights AI generated spam websites and saves me from having to read them.

  tabChar=$'\t'
  function prompt_command {
      echo "$(date +%Y-%m-%d--%H-%M-%S)$tabChar$(hostname)$tabChar$PWD$tabChar$(history 1)" >> ~/.full_history
  }
  export PROMPT_COMMAND=prompt_command
  
  function c {
      while read -r _ dir
      do
          if [[ -e "$dir" ]]
          then
              echo "$dir"
              cd "$dir"
              break
          fi
      done < <(cat ~/.full_history | tail -n 10000 \
          | cut -f 3 | sort | uniq -dc | sort -hr | grep "/$1$")
  }
  
  function _c {
      local IFS=$'\n'
      COMPREPLY=( $(cat ~/.full_history | tail -n 10000 \
          | cut -f 3 | sort | uniq -dc | sort -hr \
          | sed 's/.*\///g' | grep "^$2") )
  }
  complete -F _c c

I have a GitHub CI job which is able to automate the process for each release. https://github.com/jellyfin/jellyfin-mpv-shim/blob/master/.g...

https://gist.github.com/iwalton3/3f9bfce510f959782404be25cab...

You can also get a certificate through the Let's Encrypt DNS challenge without having to expose a server to the Internet, but you'll still need ownership of a domain name and either an internet connection or a local DNS server to support HTTPS using that certificate.

There is always the option of creating a local certificate authority for your devices, but this is kind of a pain. There are some new applications that aim to make this easier [2], but there is no easy way around having to install the root certificate on each device.

[1] https://blog.filippo.io/how-plex-is-doing-https-for-all-its-... [2] https://github.com/smallstep/certificates

> To avoid getting security warnings each time you launch the application, right click and select "Properties". Click "Unblock" towards the bottom of the page, and click "OK".

The trick would be finding a download that is still served over http.