99
GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it.
(garymarcus.substack.com)
this post was submitted on 10 Aug 2025
99 points (99.0% liked)
technology
24027 readers
196 users here now
On the road to fully automated luxury gay space communism.
Spreading Linux propaganda since 2020
- Ways to run Microsoft/Adobe and more on Linux
- The Ultimate FOSS Guide For Android
- Great libre software on Windows
- Hey you, the lib still using Chrome. Read this post!
Rules:
- 1. Obviously abide by the sitewide code of conduct. Bigotry will be met with an immediate ban
- 2. This community is about technology. Offtopic is permitted as long as it is kept in the comment sections
- 3. Although this is not /c/libre, FOSS related posting is tolerated, and even welcome in the case of effort posts
- 4. We believe technology should be liberating. As such, avoid promoting proprietary and/or bourgeois technology
- 5. Explanatory posts to correct the potential mistakes a comrade made in a post of their own are allowed, as long as they remain respectful
- 6. No crypto (Bitcoin, NFT, etc.) speculation, unless it is purely informative and not too cringe
- 7. Absolutely no tech bro shit. If you have a good opinion of Silicon Valley billionaires please manifest yourself so we can ban you.
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I don't think an AI necessarily has to have needs or wants, but it does need to have a world model. That's the shared context we all have and what informs our use of language. We don't just string tokens together when we think. We have a model of the world around us in our heads, and we reason about the world by simulating actions and outcomes within our internal world model. I suspect that the path to actual thinking machines will be through embodiment. Robots that interact with the world, and learn to model it will be able to reason about it in a meaningful sense.
The basic need/want all biological forms have is energy. The need to acquire energy and the need to use energy more efficiently in order to survive longer between energy acquisitions.
Everything else evolves from that starting point. It's just a matter of adding predators and complexity to the environment, as well as adding more "senses", sight, hearing, taste etc.
One thing I think we're not yet realising is that intelligence probably requires other intelligence in order to evolve. Your prey aren't going to improve naturally if your predators aren't also improving and evolving intelligently. Intelligent animal life came from the steady progression of ALL other animal life in competition or cooperation with one another. The creation and advancement of intelligence is an entire ecosystem and I don't think we will create artificial intelligence without also creating an entire ecosystem that it can evolve within alongside many other artificial intelligence in the environment.
Humans didn't magically pop into existence smart. They were creating by surviving against other things that were also surviving. The system holistically created the intelligence.
I've had this thought for 5 or so years. With any luck, maybe I'll put it into something publishable before I'm rounded up by the anti communist death squads that come for academics. I think intelligence is fundamentally a social/collective phenomenon, at least in a broad sense where we consider a predator/prey relationship some kind of sociality too. Humans just got really really really surprisingly good at it relative to other species because of our ability to communicate using language, and thus transfer knowledge gains very widely and very quickly compared to a non-verbal animal limited to its immediate social sphere or, slower yet, passing down changes via evolution and genetic advantage. Still, though, this is reliant on a social world and a kind of thinking that can be fundamentally collaborative. Our language and mental models of the world aren't just geared to get us what we as individual agents desire; they were developed by a process that required them to be understandable to other similarly intelligent creatures.
Thanks for sharing! My approach to this is from a philosophy background, less coding, but I'll see if I can puzzle my way through some of the technical language I don't use regularly. I just took a "Phil of mind and artificial intelligence" course waaaayyyy back in 2020 that first turned me onto these topics, and that may as well be ancient history as far as "AI" is concerned. I developed a strong conviction even then that LLMs were structurally incapable of "General Intelligence," whatever that is, and began trying to think about what was missing from the process or approach that yielded actual intelligence. I more or less came to the conclusion that the tech field was largely oversimplifying the problem (when has that ever happened before) and substituting a solvable problem, "How do we train better Machine Learning programs?" for a much harder problem, "How do we construct an intelligence?"
However, since then I've specialized in other areas of legal and political philosophy and have less regular interaction with the current discussion other than to hate on it and be a gadfly whenever less skeptical members of my community begin to become seduced by the allure of yet another tech con.
Learning about that AI winter is exactly the pattern I predicted we'd reach at the time. Who cares what an up-jumped undergrad cares about auditing a seminar though, right? (not really, they were very graceful and receptive, but like, I couldn't publish anything due to lack of professional skill at that time either). Again, this wasn't my field of expertise, but I did find it an interesting problem where had life put me another direction, I'd be more specialized within philosophy of mind instead.
Yes. Dialectics layered on dialectics layered on dialectics.
Increasingly more complex ecosystems of dialectics over time.
The problem I forsee with the future of machine learning is that it's easy to create one dialectic. It's easy to create 2 or even 100... But at a certain point you reach "ok, so what dialectic do we add now?".
We do not have an accurate model of all the dialectics that exist in our biological world. How are we supposed to recreate the conditions that create the same kind of intelligence we have if we can't create the same dialectics that formed it? If we miss some parts out it will not create the same kind of intelligence, it may not have the same morals or the same fundamental beliefs as we do because the environment that shapes it won't be the same environment.
The creation of human-like artificial intelligence will require the recreation of conditions that gave rise to human evolution.
This brings me back to that techbro theory that we might be living in a simulation. If we aim to create artificial intelligence like us then we would end up creating a simulation that looks like our environment.
incidentally, I had a few thoughts on the subject here, and I generally agree with what you're saying https://theunconductedchorus.com/#unveiling_the_origins_of_volition
Yeah, it makes sense that selection pressures from the environment would lead to evolution of needs on part of the agent. Volition and directed action gives a survival benefit.
This is one of those things that starts getting into the fuzzy area around the unanswered questions regarding what exactly qualifies as qualia and where that first appears. But having needs/wants probably is a necessary condition for actual AI if we're defining actual (general) AI as having self awareness. In addition to what @Awoo@hexbear.net said, here's another thing.
You mention how AI probably has to have a world model as a prerequisite for genuine self aware intelligence, and this is true. But part of that is that the world model has to be accurate at least in so far as it allows the AI to function. Like, maybe it can even have an inaccurate fantasy-world world model, but it still has to model a world close enough to reality that it's modeling a world that it can exist in; in other words the world model can't be random gibberish because intelligence would be meaningless in such a world, and it wouldn't even be a "world model." All of that is mostly beside the point except to point out that AI has to have a world model that approaches accuracy with the real world. So in that sense it already "wants" to have an accurate world model. But it's a bit of a chicken and egg problem: does the AI only "want" to have an accurate model of the world after it gains self-awareness, the only point where true "wants" can exist? Or was that "want" built-in to it by its creators? That directionality towards accuracy for its world model is built into it. It has to be in order to get it to work. The accuracy-approaching world model would have to be part of the programming put into it long before it ever gains sentience (aka the ability to experience, self-awareness) and that directionality won't just disappear when the AI does gain sentience. That pre-awareness directionality that by necessity still exists can then be said to be a "want" in the post-awareness general AI.
An analogy of this same sort of thing but as it is with us bio-intelligence beings: We "want" to avoid death, to survive (setting aside edge cases that actually prove the rule like how extreme of an emotional state a person has to be in to be suicidal). That "want" is a result of evolution that has ingrained into us a desire (a "want") to survive. But evolution itself doesn't "want" anything. It just has directionality towards making better replicators. The appearance that replicators (like genes) "want" to survive enough to pass on their code (in other words: to replicate) is just an emergent property of the fact that things that are better able to replicate in a given environment will replicate more than things that are less able to replicate in that environment. When did that simple mathematical fact, how replication efficiency works, get turned into a genuine desire to survive? It happened somewhere along the ladder of evolutionary complexity where brains had evolved to the extent that self awareness and qualia emerged (they are emergent properties) from the complex interactions of the neurons that make up those brains. This is just one example, but a pretty good one imo that shows how the ability to experience "wanting" something is still rooted in a kind of directionality that exists independently of (and before) the ability to experience. And also how that experience wouldn't have come about if it weren't for that initial directionality.
Wants/needs almost certainly do have to be part of any actual intelligence. One of the reasons for that is because those wants/needs have to be there in some form for intelligence to even be able to arise in the first place.
It gets really hard to articulate this kind of thing, so I apologize for all the "quoted" words and shit in parentheses. I was trying to make it so that what I was attempting to convey with these weird sentences could be parsed better, but maybe I just made it worse.
That's called a bad replicator for the purpose of this discussion, because destroying the conditions that are required for its own replication to continue is not conducive to replication and therefore a replicator that does that is bad.
If good replicator is just being defined as personally producing a whole bunch of offspring, then I think it's just not a helpful term. A good replicator should be something that replicates effectively, not just a lot, and what you are describing as "less effective at replication" is clearly more effective at replication relatively speaking if its offspring are still around and its competitors are not. You would hardly say something is a good replicator if it produced an unfathomable amount of offspring and then just ate them all, right?
How is this relevant? No one was contradicting this idea, even implicitly, it's just not a meaningful factor in the discussion for the reason you go on to note.
Maybe I'm just reading it wrong, but it looks to me like it's all about how selection pressures produce traits seen in individuals because them having those traits is better for the survival of the species.
I don't think amoeba "want" to live, they just do things toward the end of surviving to replicate, with no awareness of anything. It's like machine learning, it's just a system of reactions that ended up being self-perpetuating via survival and reproduction. That's the essential element, and having any sort of "will" is far, far downstream of that.
Wanting to live is caused by replication because it was developed out of these systems in response to selection pressures.
I'm a total philistine, half of the words you said just passed over my head, I just don't see anything fundamentally different between amoebas and an electronic light sensor or a roomba or whatever. Certain inputs produce certain outputs, and things like whether it's chemical or mechanical or anything else is immaterial. You may as well tell me that every massive object "wants" to move toward other massive objects in proportion to the product of their masses and inversely proportional to the distance between them. The fact that one perpetuates an organism's existence and the other isn't is purely incidental and I think you're effectively projecting a teleology onto it by saying that these reactions by means of which an organism maintains itself are, by that very fact, evidence of a "want."
But that's a statement about the ultimate consequences of what happens, which doesn't tell us almost anything about the proximate nature of the actions. That's what I mean by "projecting a teleology." Consider a mutant amoeba that does something not conducive to self-maintenance, there is nothing inherently different about the nature of those actions as biological processes, it requires a zoomed-out view to explain normal amoebas as conforming to selection pressures and this mutated behavior as deviating from selection pressures. "But the amoeba dies!" Yes, but the event of the death later on is not useful for explaining the fundamental nature of the action itself, the death is a distant, emergent consequence of the action, an event distinct from the action (as is successful replication, and even more so with the "event" of surviving past that point in the future). You're using teleological reasoning to make some sort of metaphysical claim about events and organisms that fundamentally don't make sense from a materialist perspective.
This, as I have abridged it, is completely true. The issue is that the "want" is just a metaphysical, teleological complication that doesn't help us understand anything and just serves to mystify a mechanical/chemical/electrical process that is already entirely understandable.
Right, it is dialectics. The problem is that it's Hegelian dialectics, which is highly teleological and idealist, and not material dialectics.
You're conflating ultimate causes and proximate causes. It's just an extremely elementary mistake in understanding behavior. I didn't call it Hegelian on the basis that you mentioning Hegel was evidence (though that did help me make the connection), I called it Hegelian because it has the same ethos of the end existing inside each step of the process, drawing the process along intrinsically, which you cannot claim this theory of "want" is not.
It's no different than saying massive bodies want to be near each other (prioritized in terms of mass1 x mass2 / distance) because they keep exerting force that trends toward that outcome. It's no different than saying that liquids "want" to hold together, they just don't want it very strongly, or that rivers "want" to erode shorelines. With base organisms, it's just input -> output, and the function processing them was created by selection pressures, but there is nothing distinguishing the actual actions on a proximate level from ones that are ultimately self-destructive, because the self-destruction only happens later and on that basis, with no further information being required, cannot be used for establishing what was going on inside the base organism at a proximate level to cause the output.
Yes. And?
I'm not seeing how this contradicts anything I said. In fact it supports what I said by recognizing the necessity for a directionality that precedes (and is a prerequisite for) any kind of sentient desire or "wants."
@purpleworm@hexbear.net addressed this really well and gave a thoughtful, completely correct response. Not much more for me to say on it.
I think you're splitting hairs here between ever so slightly different aspects what I have been calling directionality. Desires or "wants" by definition require a mind capable of having a want or desire. Where you say "it really is a kind of want but not mentally because there is no mind yet" then that's simply not the kind of "want" we are talking about here, the thing that a self-aware (mind-possessing) AI would have if it were genuinely self aware and possessing of a mind. Everything else really is just an appearance of want and is a result of what I've been calling directionality. What you're talking about as the mindless "need to avoid death to continue" is still just the mindless non-intelligent and non-sentient directionality of evolution. And to specifically address this piece:
But it is part of the world (dialectics ftw!). There is a tension between inside and outside the individual cell (and also a tension between the "self" and "outside the self" of a sentient mind which is addressed further down, but this is not the same thing as the the tension between the cell and the world, as proven by the fact we aren't aware of all our cells and frequently kill them by doing such things as scratching) but the cell still isn't the most basic unit of replication in evolution, that would be the gene. Strands of RNA or DNA. Genes (often but not always) use cells as part of the vehicle for their replication, and either way they are still just chemicals reacting with the environment they exist within. There's no more intentionality behind what they do than there is behind, say, a magnet clinging to a fridge. That magnet does not "want" to cling to your fridge, like genes, it is reacting to it's environment and this will be true regardless of where you draw the boundary between the "self" of the magnet and "the outside world." To actually desire something the way we are talking about here requires the complexity of a brain capable of producing a mind.
Agreed. The emergent property of the mind and sentience comes out of the complexity of the interaction of the firing of neurons in a brain and the world they exist within, at least in all likelhood. We still don't know exactly what produces our ability to experience, where exactly qualia originate (i.e. why we aren't just philosophical zombies) but I think most neuroscientists (and philosophers who work on this stuff) would agree, as I do too, that without an outside non-self world for those neurons to interact with, there would be no actual mind. Even that the mind is a drawing of the distinction between self and non-self. But since that complex neural structure could never even begin to come about without that outside world and all the mechanisms of evolution (aside from a Boltzmann brain!), always having to include the phrase "and with the outside world" when describing the neurological origin of qualia and experience is some severe philosophical hair-splitting.
Um, yeah... that's pretty much what my argument was for the necessity of any genuine AI to have wants and desires, those "wants" necessarily would have had to have been there built in for it to even become AI.
Disagree. Again, if you want to split hairs on exactly where it is in that ladder of complexity that self-awareness arises, or where in the fuzzy chain we can draw a line between organisms capable of self-awareness vs those not, or even exactly what constitutes self-awareness then feel free. But a thing having an actual desire as something genuinely experienced, it requires some sense of selfhood for that experience to happen to.
I'd argue that the idea of self awareness or needs/wants is tangential to the notion of qualia. A system can be self aware, and develop needs without necessarily being conscious and having an internal experience. What needs and wants really boil down to is that the system is trying to maintain a particular state. To maintain homeostasis, the system needs to react to external inputs and take actions that keep it in the desired state. For example, a thermostat could be said to have a "need" to maintain a particular temperature, but it could hardly be argued that it has some sort of qualia.
Why sentience exists is a really interesting question all of itself in my opinion as it's not an obviously necessary quality within a self aware system. I suspect it may be related to having a theory of mind. When a system starts to model itself then perhaps you end up with some sort of a resonance where it thinks about its own thoughts and that's what creates internal experience.
We also have to define what we mean by intelligence here. My definition would be a system that has a model of a particular domain, and is able to make judgments regarding outcomes of different actions. I don't think mere intelligence requires self awareness or consciousness.
This is right. Having things like beliefs and desires is called "intentionality," and is orthogonal to both sentience/sapience and first-person subjectivity (qualia). You can have beliefs and desires without any accompanying qualitative experience and vice versa.
Machine learning requires needs or wants in order to evolve. If your model is going to learn how to utilise energy efficiently between recharging then it needs to desire energy (need/want). This is just "eat" and "collect water" process of learning. Then you give it predators to learn how to avoid being killed in the process of doing this so it learns survival methods. Add complexity to the environment over time and it'll learn more and more and more.
Reproduction probably needs some sort of social cues to learn, the ability to communicate with other models that they wish to reproduce, or the ability to start working in teams.
It all has the requirement of needs/wants. The basis of all animal intelligence evolving into more efficient methods of doing something is having needs.
I think "adding predators" is kind of cargo cult, and you can just give it obstacles that are actually relevant to problems it could conceivably need to solve, because we aren't making AI with the main hope that they can successfully survive a wolf attack, and it doesn't need to just completely recapitulate what humans did to evolve when it is fundamentally not an early human and we don't want to use it for those conditions.
It's going to have a completely different development path without any kind of predators. Almost everything in our world is either prey, predator, or both.
I struggle to believe that it will be adapted well for a human reality where predation exists (where humans are sometimes predator to other humans too, socially) without having a development path that adapts it for that reality.
We don't know how to skip straight to human level intelligence. You need to create wild animal level artificial intelligence before we can understand how to improve it into human level intelligence. If we can't make artificial monkeys or macaws in terms of emotional/social intelligence and problem-solving then we certainly can't make anything akin to humans or whatever the next level intelligence is that we're hoping to get when an AI can make better versions of itself iteratively progressing technology to levels we haven't imagined yet.
My point is that it's a different kind of thing. Role-playing that it's an animal in the wild instead of a cogitating tool is counterproductive. We aren't going to be sending it out into the jungle to survive there, we want it to deal with human situations, helping people, with its own "survival" being instrumental to that end. Even if it encounters a hostile human, it probably won't have anything that it can do about it because we aren't immediately talking about building androids, this AI will effectively be just a big bundle of computers in a warehouse. If you want to give an experimental AI control over the security systems of the building housing it, go off I guess, but the protocols to contain the person breaking in by . . . locking the doors with nothing else it can really do while calling the cops and the owner of the facility are not really "avoiding predators" on a level other than strained metaphor, and require no engagement in this "surviving in the wild" role-play, it's just identifying a threat and then doing something that, the threat being identified, any modern computer can do. If you want to say it needs to learn to recognize "threats" (be they some abstraction in a game simulation, a fire, a robber, or a plane falling out of the sky) sure, that's fair, that falls within obstacles it might actually encounter.
Nothing I'm saying bears on the level of intelligence it exhibits or that it is capable of. I'm not saying it needs to handle these things as well as a human would, just that it needs to be given applicable situations.
I feel like you've misunderstood me. You're talking consistently about one single AI.
Machine learning is not one AI. It is thousands of generations of AI that have iteratively improved over time by their successes and failures. The best performing of the generation go on to form the basis of the next generation, or you have survival mechanics that automatically form new generations.
This isn't training a model by giving something input data. Entire neural networks we do not understand are formed through an attempt at creating an artificial natural selection.
If your process isn't going to be similar to humans, you aren't going to produce something similar to humans. I honestly think that's dangerous in and of itself, you're creating something that might have a brain network that is fundamentally at odds with coexistence with humanity.
But you're able to designate what its goals are completely arbitrarily. It doesn't need to think like a human -- there are humans who have been at odds with coexistence with humanity -- it needs to be constructed based on the value of human benefit, and you can seriously just tell it that. That isn't changed by it cogitating in a structurally different way, which it also almost certainly would be doing anyway because the way we do is highly adapted to early humanity, but is structurally deeply based on random incidents of mutation before then. Something could think very differently and nonetheless be just as capable of flourishing in those circumstances. This difference is compounded by the fact that you probably aren't going to actually produce an accurate simulation of an early human environment because you can't just make a functional simulation of macroscopic reality like that. Even imagining your method made sense, it would still ultimately need to fall into aspects of what I'm saying about arbitrary stipulation because the model environment would be based on human heuristics.
But way more important than that is the part where you, again, can just tell it that human benefit based on human instructions is the primary goal and it will pursue that, handling things like energy acquisition and efficiency secondarily.
It requires needs or wants to be self directed. The needs can also be externalized the way we do with LLMs today. The user prompt can generate a goal for the system, and then it will work to accomplish it. That said, I entirely agree systems that are self directed are more interesting. If a system has needs that result in it wanting to maintain homeostasis, such as maintaining an optimal energy level, then it can act and learn autonomously.
Ok but how is it getting intelligent before the user prompt?
The AI isn't useful until it is grown and evolved. I'm talking about the earlier stages.
We can look at examples of video generating models. I'd argue they have to have a meaningful and persistent representation of the world internally. Consider something like Genie as an example https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/
It doesn't have volition, but it does have intelligence in the domain of creating consistent simulations. So, it does seem like you can get a domain specific intelligence through reinforcement training.