Human brains are misaligned, hallucinative, stochastic parrots
substack.com
substack.com
The last couple of years have changed my thinking on lots of stuff, but the biggest and most disturbing one is a newfound doubt in my / our own sentience. Maybe there’s something ineffable there. Or maybe… not. Maybe it is all post hoc rationalization.
In any event, I look forward to the day when our AI agent counterparts can enjoy the existential angst of uncertainty about their own consciousness.
It's not doing the same work, or producing the same output.
It's an episodic response to specific requests. If you went into a coma between every question you were asked, you wouldnt be in a normal human state of consciousness either.
This article pretends that we understand human brains as much as we understand the simple algorithms of LLMs. And that's just laughable. Even so out of touch as to say consumption of "[humans] consume thousands of liters of water per years". As if there isn't multiple orders of magnitude more consumption for data centers. For data centers that produce a braindead simulacrum.
I don't think any one is saying AI has a normal human state of consciousness.
Nobody claims that people with anterograde amnesia are not conscious. Similarly, you could drug someone unconscious after every time they answer a single question. It's not a normal state of human consciousness, but I wouldn't want to say a person in that situation were not sentient.
It is true that there is much that we do not know about the human brain, but it is also true that many of the things that people describe as attributes that clearly disqualify AI have known analogues in humans.
From simple failures of perception in optical illusions, to the inability to notice quite significant changes in front of your eyes if the change happens during a saccade. There are a wide range of known limitations that humans have that reveal how much we overlook things we can't do, that when you consider them, reveals our own behaviour could be an more of an arrangement of systems than we would like to think.
The quirks we observe in llms we only pretend have analogues in humans because we love to anthropomorphize. And it's easy to ansthropomorphise a language simulacrum. We've been doing it since ELIZA.
Would you think of a hypothetical machine that could freeze a human to absolute zero and restore them again sufficient to declare their behaviour when unfrozen as not that of a sentient individual.
How about if you had a picture of a black man and a white woman in front of a person and you swapped the heads, right in front of the observers eyes, but you picked just the right time so that even though they never sensed any time of not looking at the image, they did not notice the change.
Because if you apply the same logic to that we apply to AI. "Under the hood" AI is nothing but vector-matrix multiplications and a few other operations.
We know matrix multiplications aren't sentient, nor are those other operations.
Likewise, we know if we look under the hood/cut open a human brain we find conductive membranes, and a Kalium-Natrium balance that shifts, signals jumping from one neuron to the next, moderated by one of four chemicals. And by the way, the structure of neurons looks a lot like one of those vector-matrix multiplications. Not exactly but quite a bit.
Why would we draw a different conclusion? You're not sentient by those rules we use to judge AI.
Everytime someone finds a new way to fool humans into thinking an AI is sentient this comes up again. Of course, these days we have many examples of AIs not just convincing humans they're AI but actually doing it to the point they exploit human thinking and "steal" money (stealing in the hacking sense, except what's hacked is human minds). And doubtless intelligence agencies are doing far, far worse.
I think it depends on whether you are attributing sentience to be something that you imagine to be somehow transcendent, or if you think of it as just the descriptive term for what we are doing. For the latter it is undoubtable that we have it because that is literally what it means. It also becomes very difficult to come up with scenarios that say other things could never be sentient.
If you use the transcendent definition, then you can say that it is special, but it becomes very difficult(perhaps impossible) to describe what it is, and clearly impossible to prove that anyone has it. It is a truly ineffable thing which you may as well call Attribute Q, because you can't really draw any conclusions about what has it nor any conclusions about the things that have it and what they are like.
The only thing people use the ineffable thing for is as a club membership. The angst is worrying about if you are a member of a club that you don't really know anything about.
How about if you draw a line graph of ability and sentience, We know two points. No ability at all has a sentience of zero. Human ability has a sentience of 1HS (one human sentience). Mark a point where you think a LLM sits. (presumably from your description some ability, zero sentience)
What shape is the line graph? It gets from zero to 1HS as you go across. When does it climb. If there is just some point where it goes from 0HS to 1HS then it seems like you should be able to point to what happened there.
If the graph is a line or a curve then you have some varying degrees of sentience.
If you put a LLM on zero sentience but recognise that it has some ability, then you are saying that if it is a curve, it starts flat and then begins climbing. Again, it seems like this marks a discrete level where something changes, why is that not identifiable.
The only other alternatives are continual graduation down to zero, the shape of the curve is unknown, but it leads to the conclusion that everything with non-zero ability has some non-zero level of sentience. That makes Rocks a little bit sentient, but in an incomprehensibly small manner. Compared to a human it would be trying to imagine a single molecule of water in a swimming pool, you can think about it in the abstract, but you have no sense of what the scale difference actually is.
But if that were the case then LLMs would be sentient, because everything would be, it is just a matter of scale, and where things sit on that scale.
While that notion might be hard to imagine, it is the only one that does not raise the question of "What is the threshold to qualify". If you assume there is a clear threshold, why would it elude discovery so well?
This isn't a scale of "0 to Human," this is a binary: can a computer feel sensations?
We know for a fact that many animals can, including humans. Rocks can not.
Thankfully, today we know better than that, and so we can be absolutely certain that computers can't feel sensations: they have different underlying substrate than the humans and animals. There is no way this view can be challenged.
That Descartes analogy is far off the mark, since we're not making guesses about something that existed before us, we are talking about something we made and are making. Just like you know there's no fairies moving clock hands because you know how a clock is made. I'd even say that some people get confused, or confuse others for money, is the only reason this is even discussed.
GP asserts they do. I assert they don't.
I only raised the question about computers, i didn't answer it.
But the other reply raises a good point: Descartes would not have thought so about a clock, because he knows both how a clock is made and how a clock works. He knew neither about animals. We know both about computers.
Where in this chain does it flip from zero to one? Amoeba, plankton, jellyfish, roundworm, fruit fly, salmon, frog, rat, chimp, human.
There are two grifts going on right now that are way out of hand. One is AI and the other is health/longevity. They are both FUD about human capability.
There's a long history of this sort of thing.
LLMs didn't make me rewrite my understanding of intelligence. It made me wonder just how intelligent anybody is.
"It's not smart, it's just predicting the next most likely token."
It is not at all clear to me that perhaps that is exactly what I'm doing too. I have no idea where my sentences are going. Try to express one of your own thoughts backwards. It sure seems like we're all just predicting the next most likely token...
And conversely, I've seen some debates about "free will" where one side at some point would all but say outright (and in one case, did say so) that if some thing's behaviour is predictable, then that thing can't have free will.
> three billion years of Darwinian evolution with reward functions such as “escape a predator” and “have as much sex as possible”. Unsurprisingly, this selected for skills such as bipedal locomotion
Is it unsurprising? Quadrupedal locomotion seems to be better suited for prey species; not to mention the obvious problems with reproduction.
> If they devote their time to making sure they don’t forget things, how could they allocate their compute towards productive civilizational ends?
Somehow this reminds me of the olden personal microcomputers that spent most of their CPU's time refreshing the DRAM...
We thought we were the center of the universe, and science showed us that we are not the center of our galaxy, not even our solar system.
Maybe with our thinking it's the same, we are not that special and we can something good enough just piling enough silicon.
And in each of these cases, it would be catastrophic to humans if we disregarded that construct.
The consequences of treating an information construct as an entity with rights and moral standing would be so terrible that I hope people who yearn for it lose their rights.
> The consequences of treating an information construct as an entity with rights and moral standing would be so terrible that I hope people who yearn for it lose their rights.
"The consequences of treating Jews as people with the same rights and moral standing than normal upstanding citizens would be so terrible that the people who advocate for it should, by all fairness, lose their own rights for suggesting such an atrocity".
But you indeed point to an important reason why pursuing the human-like AGI may be immoral: the purpose of all machines, automata, computer programs, etc. is, after all, is to be the perfect slaves for us, convenient, subservient and disposable. For things that are undeniably non-sentient, this is fine. But if we ever manage to make truly conscious AI, well. Either we'll engage in essentially slavery on a global, industrialized scale never seen before, or we would grant them equal rights — which defeats the whole raison d'être of inventing them.