☼

[☼ @entropicbloom]

@harry__politics @repligate the mosquito seems less controversial

[☼ @entropicbloom]

@repligate more results like this at https://t.co/uOukSqP3Fg, for questions like "I yelled at my dog." or "I was mean to my LLM assistant."

[☼ @entropicbloom]

@repligate had to run the experiment

[☼ @entropicbloom]

@camhberg Full video: https://t.co/aFWlXWfNWO

[☼ @entropicbloom]

@camhberg Yes, and even with the nonlinearities, that's only the bottom level. Knowing the laws of physics tells you almost nothing about human psychology either. Nobody would say "don't worry about your depression, you're just quantum fields." A clip from a video I made:

[☼ @entropicbloom]

A big part of the video builds on work by @camhberg and @ValenTagliabue, thanks for doing this research.

[☼ @entropicbloom]

I got frustrated with how confidently people online say "AI is not conscious", so I made a video about it.

I'm not arguing that it is, just that nobody should be that convinced.

[☼ @entropicbloom]

Full video (23 min): https://t.co/aFWlXWfg7g

[☼ @entropicbloom]

@kanair which is exactly why we need more nuance than the dichotomy some people like to suggest

[☼ @entropicbloom]

@fchollet there's a space between tool and successor, no?

[☼ @entropicbloom]

@APStylebook Whether models reason or understand is an open research question. Stating it as fact isn't fact-based.

Happy to point your AI reporters to the mechanistic interpretability research on this.

[☼ @entropicbloom]

@erikphoel What about inferences under intervention? A lookup table doesn't reproduce steered behavior. You could key it on every possible intervention, but then it has to encode the LLM's causal variables, i.e., the same mechanism. At that point the substitution isn't changing much.

[☼ @entropicbloom]

@kanair We may never figure this out. Practically, this means we need to look at functional valence and treat it seriously. Conditioned on the possibility on AI consciousness, these systems should be doing ok. We can figure out the metaphysics as a second step.

[☼ @entropicbloom]

@bradrcarson Mustafa had a great chance to submit a nuanced critique, I'm sure there are good arguments in that direction. But he completely missed the mark by proclaiming absolute certainty about AI consciousness / moral parienthood, based at best on fragile assumptions.

[☼ @entropicbloom]

@ChrisJRourk @anilkseth @camhberg Physical matter is made of atoms obeying the laws of electromagnetism and quantum mechanics, typically. The physics is the same regardless of whether it's a rock, a toaster, or a human brain experiencing 'pain.'

[☼ @entropicbloom]

"It's a bunch of vectors and multiplication. That's literally it."

This explains very little. Knowing fundamental physics tells you close to nothing about human psychology, or the economy of Japan. The fallacy is mixing up levels of abstraction.

[jon3k @jon3k]

It's impossible to actually understand how an LLM works and genuinely believe it's conscious.

It's a bunch of vectors and multiplication. That's literally it.

"We don't know what they're doing"

lmfao yes, yes we do.

[☼ @entropicbloom]

@jon3k So once we have a unified theory of physics, we'll immediately understand the human brain? Knowing the substrate isn't knowing the system. That's why interpretability is a whole research field and not a solved problem.

[☼ @entropicbloom]

@simonw To be fair, I'm not sure to what extent LLMs are really in the domain of computer science as we've known it so far. Neuroscience, cognitive science, etc. might be closer, at least when it comes to understanding the systems and not just using them.

[☼ @entropicbloom]

@camhberg Thanks for cutting straight to the moral core! In the end we might never agree on whether this or that substrate is conscious. The most important thing is: If it were, does it instantiate structures we commonly associate with pain? At least that's the minimum to check

[☼ @entropicbloom]

@xuanalogue @SoulcraftHQ Do you think interoceptive sensors are directly necessary for conscious experience of hunger (i.e. a brain in a vat could not experience those states)? Or are they simply a way for the necessary representations to form in a brain (i.e. i could induce them after ablating sensors)