That is not at all the intention of the ARC team. By ARC teams definition, passing any single ARC-AGI benchmark does not mean that AGI has been achieved. Instead, AGI would be considered achieved when we are no longer able to come up with new benchmarks that the AI systems do not immediately do well on.
People are trying to solve it with software too, even if you don't hear about it.
The most high-profile example is the latest set of Qwen models, which replace most of the attention mechanisms with Gated DeltaNet (which uses constant memory with respect to sequenc length).
Test-time training architectures are also getting a lot of attention, and have shown great performance in the acedemic setting. It's only a matter of time before we start getting open TTT models.
Yes and no, it's not just text, it's images, video, etc, and it's not just the pages of content, it's also all the "thinking" as well. Plus the models tend to work better earlier on in the context.
I regularly get close to filling up context windows and have to compact the context. I can do this several times in one human session of me working on a problem, which you could argue is roughly my own context window.
My point though was that almost nothing of the model's knowledge is in the context, it's all in the training. We have no functional long term memory for LLMs beyond training.
The KV cache isn't memory, it's the extent of the process saved so the inference can start where the last generated output is concatenated with the next input. It's entirely about saving compute and has nothing to do with memory.
This really confuses how stupid LLMs are: they're just text logs as output and text logs as input; hence the goblins are just tokens that seem to problematically be more probable in the output.
But the KV cache is a thing made to keep a session from having to run through the entire inference. The only thing you can call "memory" is there's no random perturbations in the KV cache while there may be in re=running chat which ends up being non-deterministic. You can think of it as a deterministic seed to prevent a random conversation from it's normal non-deterministic output
Either someone hard-coded it in a system prompt to the reward model (similar to how they hard-coded it out), or the reward model mixed up some kind of correlation/causation in the human preference data (goblins are often found in good responses != goblins make responses good). It's also possible that human data labellers really did think responses with goblins were better (in small doses).
For one, everything humans' "intelligence" knows about solving the problem is contained within the finite brain size for the particular person and life. Unless the memory contents of the brain are being saved to storage and reloaded later, it won't "remember" that it solved the problem and save its work somewhere to be easily referenced in a later life.
I did trained some research models using the existing PyTorch/XLA on TPUs, and it was a mess of undocumented behavior and bugs (silently hanging after 8 hours of training!).
If anyone is trying to use PyTorch on TPU before TorchTPU is released, you can check out the training pipeline that I ended up building to support my research: https://github.com/aklein4/easy-torch-tpu
From my observations, there are generally four camps in the machine consciousness discussion:
1. People who haven't really thought about it, and assume they're conscious because they talk like a human.
2. People who haven't really thought about it, and assume they can't be conscious because humans are obviously somehow special. This appears to be the largest group, and is linked to our religiously rooted culture in which human exceptionalism is the default.
Those first two groups comprise the majority of people, and are not worth engaging with.
3. People who have thought about it, and came to the conclusion that they might be conscious, usually for computationalism/functionalism reasons. This is the group that I place myself in.
4. People who have thought about it, and came to the conclusion that they can't be conscious, usually for biological naturalist reasons. This seems to be the predominant group on Hacker News (among those who discuss it).
I'm not sure I'd agree that people in groups 1 and 2 aren't worth engaging with.
The interesting bit to do for both cases is look at the 'they talk like a human' and 'are obviously somehow special' parts, separate the ideas of language, intelligence (memory, fluidity, abstract reasoning), _aliveness_ (as a biological process) and finally ideas about metacognition and theory of mind, and see whether their idea of consciousness as a super-bundle of the above (which is how I assume a lot of default ideas about consciousness are) actually sticks, or whether it falls apart when beings can have a subset of those properties but not all.
Also, I nominate myself to be in the 'People who have thought about it and are becoming more doubtful that I myself am conscious, and the question might be moot.' group.
I'm curious about your doubting your own consciousness statement, given that "we humans are conscious" is pretty axiomic to its definition and one of the few pieces that most agree with.
If you're looking for one of the genuine angles on this:
Consciousness is horrendously under-defined, to the point some people go something like "you know, at this point I figure we'd be better off not having this word at all. "
You'd have to define those terms operationally first, somehow, before I could give you an honest reply. Most people can't -and those who do disagree- which suggests something structural.
I would place myself in 3, with the caveat that I don't think any current llms or other programs/dataset/relationships are close to conscious. It's certainly possible in the future, though.
Atoms arranged into a brain generate consciousness. There's no reason to think atoms in other arrangements can't. Brains aren't magic, just well optimized.
One big reason I don't think LLMs are (currently) conscious is because they are static. They do not change in response to input. I think they need some kind of temporal awareness (not just a 5-minute cron job), and some mechanism for self-modification or active learning based on their input. If an experience flows through them and leaves them completely unchanged, are they actually conscious of the experience?
But, in fairness, we don't have a science of consciousness yet. Anybody that is 100% confident in their proclamations about this topic is too confident.
> One big reason I don't think LLMs are (currently) conscious is because they are static
It is true that the LLM itself is static. However it's context window is self-modifiable, based on its inputs and outputs.
> I think they need some kind of temporal awareness... and some mechanism for self-modification or active learning based on their input.
Why?! (besides, they do, see above)
I bring this example up, and it's clear evidence in humans that neither of these things are required for consciousness, and one that I deal with in my home. People with dementia that have no memory that are no longer able to learn suffer a different issue that not being conscious.
> If an experience flows through them and leaves them completely unchanged, are they actually conscious of the experience?
This line of thinking precludes dementia patients with no retention of memory are not conscious.
I agree having an experience, and being conscious of that experience are two different things, though.
Assuming 3. Maybe in order to reproduce human level consciousness one would need to treat at least most human cells as neurons, and reconstruct all the diversity of neuron types and their signalling mechanisms.
If human consciousness is reproducible, maybe we will long underestimate the depth and diversity it uses to model reality the way it does.
5. People who has a financial interest in making sure that any eventual AGI isn't granted any kind of rights and continue to be exploited as an inanimate "thing", not as a "being", no matter the actual characteristics of this hypothetical AGI entity.
I mean, take a look a this language in the paper [0]
> This realization pulls the field of AI safety out of the welfare trap. It allows us to focus entirely on the concrete risks of anthropomorphism, treating AGI as a powerful, but inherently non-sentient tool.
This reads as someone that started with this conclusion, and then built an argument to support it.
As someone who places themselves in #4, at some point the people in #3 need to accept a bit of scientific humility. The reason we are "biological naturalists" is that we can point to hundreds of thousands of conscious species on planet Earth which are not humans, and whose consciousness clearly has nothing to do with an ability to say "Forsooth, I am a conscious thinking being." AI folks have been ignoring this since Alan Turing! And it's not a coincidence that humanity has yet to build a robot which is convincingly smarter than a cockroach.
If you grant that humans are conscious, then surely domestic cats are as well. It is simply irrational to talk about Claude's "consciousness" without actually engaging with this: cats, humans, pigeons, fish, etc etc all share some common features we associate with consciousness (I don't mean sensory awareness, I mean the fuzzy cognitive concept). Claude really does not. In fact Claude doesn't even have much in common with uncontacted hunter-gatherers! Claude imitates the solipsism of formally educated human philosophers.
It is uncharitable and curmudgeonly but totally scientific to dismiss people in camp #3 as unserious and not worth engaging with: they ignore scientific criticism and don't provide any themselves, it's just a mishmash of sci-fi-adjacent philosophy. There's nothing "functional" about ignoring animals and there's nothing scientific about waving your hands and saying "computationalism." That's certainly how I feel. I know this isn't a very nice comment. But I am so sick of AI folks thinking they can ignore animals and still have an honest conversation about machine consciousness. It's just sci-fi ghost stories.
Oh dear, just a short while after me saying I was confused by the term too.
Are you sure you're a <biological naturalist>? [1] Which is to say, do you adhere to Searle's position about syntax not leading to semantics?
Or is it more like: You're scientifically inclined, and thus you accept Ethology[2] or Neuroscience[3] as being empirically rigorous studies of animal behavior and cognition respectively?
Incidentally, Alan Turing's 1950 imitation game paper was actually pretty Ethological if you look it up. He immediately replaces the question "can machines think" with a more practical operationalization: the famous imitation game.
I didn't say I was a formal biological naturalist according to Searle, I put myself in one of the four boxes the parent comment offered. Please read my comment in context.
Your response is too condescending to engage with. You should have assumed I know what neuroscience is. Please don't ever email me about anything.
(ps. A quick search gives me the impression <biological naturalism> arguably rejects much of biology's findings on animal cognition. My mail is in my user description if you'd like me to dig up the relevant literature for you.)
What is the evidence that non-human animals have the "fuzzy cognitive concept" we call consciousness, but Claude "really does not"?
I personally have not been ignoring animal consciousness in how I think about the possibility of AI consciousness and I don't see how animals having consciousness means that AI can't.
I could have phrased it better, but the emphasis is on "fuzzy"! There isn't any evidence for any of this, it's pre-scientific.
My statement is an opinionated position on how we should direct our research efforts and ascertain what is plausible: the behavioral similarities between humans and cats are much more relevant to the question of consciousness than the behavior similarities between humans and Claude, because cats are obviously conscious and that's not true for Claude. The fact that there are almost no behavioral similarities between cats and Claude suggests to me that "Claude might be conscious" is just a ridiculous statement not worth engaging with, even at the level of pre-science. At the very least, the burden is on Amanda Askell and Dario Amodei to explain why nonhuman animals are irrelevant to the question of Claude's consciousness. They have not offered anything like that; instead they seem fully ELIZAed by the chatbot, high on their own supply.
> because cats are obviously conscious and that's not true for Claude
I'm not sure that I agree that's true, and I think that's the crux of the debate here: how do you define consciousness such that it's obvious that a cat is conscious, and why would that definition not include Claude being conscious?
What about robots? Not necessarily humanoid robots, but the classic RL demonstrations that can scurry around and achieve simple goals?
In the computational functionalist argument, the thing that we share with cats, pigeons, and robots (and in some ways Claude) is the fact that we react to our environment in a way that requires computation.
I myself lean (without confidence) towards weak panpsychism, where a lot of things down from humans to cats to fish to trees to bacteria are in some way sentient. We all have in common a computationally driven sense/"think"/act cycle, and that is where it derives from.
The problem with robots is, again, humanity has yet to build a robot with the intelligence of a cockroach, or apparent conscious agentic behavior of a nematode. If I see such a robot I will update my views on machine consciousness. I don't think either of us will live that long.
The problem with the "computational functionalist" argument is that a) there's ZERO evidence other animals brains are computational, that is begging the question; and b) pretty much any embedded system is a device that reacts to its environment in a way that requires computation, and none of them have anything close to the psuedoconsciousness of a bacteria. let alone an insect. Point a) is the more important one: only humans have meaningfully Turing-complete brains. Other animals might be hardware-capable but they'll never be trained to correctly execute a program, nor does their own intelligence seem especially amenable to being described by a classical symbolic algorithm - e.g. animals are very good at object identification, quantity discrimination, causal reasoning, and we don't have anything close to a symbolic algorithm for any of these[1]. Computation is linked to the ability to communicate symbolically, and most animals do not regardless of intelligence. The idea that "the brain is a computer" has always been a poetic description, not a scientific fact. It is more correct to say humans have the ability to think computationally because we think symbolically. Again, maybe someone can identify that animals do think symbolically even if they don't communicate that way, or (somehow) we will have a non-symbolic theory of computation. Perhaps a beautiful symphony. Absent either of these two things, "the chimpanzee's brain is like a computer" is simply not scientific.
The supposed "sense/think/act cycle" is just you begging the question again, applying a computational aesthetic in place of understanding; this time it's blatantly false. Animals do not have a "cycle": sensing is an act and processing senses is a thought. Thinking is an act and many animals can perceive themselves thinking (demonstrated in crows and chimps). Dogs think very deeply while they smell, and the manner in which they sniff (tentative whiff versus greedy huffs) is itself an act requiring thought. Most importantly: even in animals, thoughts can be totally disconnected from actions and senses. Actually this might be the most major difference between a pigeon and Claude: their thoughts and actions are not directly tied to environmental stimulus, whereas Claude can only think and act according to a short-term context provided by a human. You can fake an agentic loop with a prompt, but it's not convincing agency the way a nematode has convincing agency. It's just a chatbot in a loop. If you expose it to real sensory data like a webcam, the agentic behavior becomes even more brittle and unconvincing. It's just nothing like an animal.
[1] I know there's work being done on formal causal reasoning, I thought this monograph was interesting: https://direct.mit.edu/books/oa-monograph/3451/Actual-Causal.... I am not convinced by it. The funny thing about these causal theories... they don't have a causal explanation :) :) :) The argument works by going through cases until you agree it works, empirically, possibly after complicating things further by patching out oversights and inadequacies. Very amusing. Causality is a tough nut to crack!
I don't feel like I am in either 4 of those camps or that I'm part of camp 4, but the camp 4 problem is not the important one. My thought on this is that intelligence != consciousness and even brain != consciousness. Consciousness is the experience, consciousness is what you see, hear, feel, in the moment of it. It's the experience. It does not require any thought. In fact, if you look at Buddhist teachings, they teach the very opposite, they teach that the thinking mind is in fact an obstacle to experiencing consciousnesses fully, that it's only a sense, a tool (like smell, touch, vision, hearing). My bet is that a cat, or a dog has the experience of awareness the same way we have (although you can't never be too sure, even about another human being - look up "philosophical zombie").
Obviously, language-driven thought is not a requirement for consciousness, not just in other animals, but even in humans. The thinking mind takes a secondary role in ordinary daily human life. The truth is that a human being behaves the way they do is not because of thoughts, but because of conditioning (the thoughts are not the primary driver of decisions, actions and behavior). The 99% of the action and responses are trained, the thoughts that we have are also part of this conditioning (most thoughts are unconscious and they are inter-wired with the behaviors, even a seemingly conscious self-reflection thought can be an automated pavlovian trigger). For example, one may think that they get up and go to work because they have a thought "I have to get up, now I am going to go to work", this is an illusion and complete misunderstanding of what consciousness is. Or one can have a psychological insight about oneself, if it's repeated and follows a behavior consistently, the very thought is just the equivalent of whistle-salivation. The thinking mind gives us that 1% to self-reflect, adjust our behavior, learn, predict the future and that differentiates us from other mammals, it's a powerful tool, but just a tool, but it should not be confused with consciousness and it should not be confused with the mind as a whole (in the materialistic sense). The way our brain functions is anything but like an AI agent. And what is consciousness? It's not the thinking mind. It's the experience. It's the direct perception of the senses. The consciousness is what is seen, heard, smelled, touched, thought (the experience of having a thought) in the moment. When you practice meditation, you get to discover the consciousness directly by becoming separated from the thinking thread. The thinking thread becomes more like an external tool, like a computer inside you and you realize directly that it's just a part of the cognitive faculty that makes you navigate your life, not the entire thing.
The LLM (and the harnesses) as built right now merely simulate the tool (the thinking mind). It's not that because this is some code ran on a beefy, but regular piece of tech invented in 20th century you may have at your desk that it does not have awareness (that's also a good argument), but because the way they function and operate is nothing like human (or mammalian) brain, then why would you think that regular code running on a regular PC could gain awareness? My point is that there's no similarity argument, LLMs, despite all their incredible capabilities (to threaten our jobs), are not remotely similar to the way our brain works.
Secondly, even if someone built an artificial brain made of whatever that simulates the biological structure, because of the philosophical zombie problem (the fact that there's no way to scientifically observe consciousness), you could never be too sure if a key ingredient was not missing and you are looking at an NPC. The consciousness is not a property of the physical brain, it's literally immaterial, it's the direct experience of the senses. You can make an optimistic assumption that every person and animal experiences consciousness the same way you do, but there's no way to rationally accept this assumption for anything created artificially.
reply