AGI will arrive like self driving cars. it’s not that you will wake up one day and we have it. cars gained auto-braking, parallel parking, cruise control assist. and over a long time you get to something like waymo, which still is location dependent. i think AGI will take decades but sooner will be some special cases that are effectively the same
When the engine gets large enough you have to rethink the controls. The Model T had manually controlled timing. Modern engines are so sensitive to timing that a computer does this for you. It would be impossible to build a bigger engine without this automation. To a Model T driver it would look like a machine intelligence.
Interesting idea. The concept of The Singularity would seem to go against this, but I do feel that seems unlikely and that a gradual transition is more likely.
However, is that AGI, or is it just ubiquitous AI? I’d agree that, like self driving cars, we’re going to experience a decade or so transition into AI being everywhere. But is it AGI when we get there? I think it’ll be many different systems each providing an aspect of AGI that together could be argued to be AGI, but in reality it’ll be more like the internet, just a bunch of non-AGI models talking to each other to achieve things with human input.
I don’t think it’s truly AGI until there’s one thinking entity able to perform at or above human level in everything.
The idea of the singularity presumes that running the AGI is either free or trivially cheap compared to what it can do, so we are fine expending compute to let the AGI improve itself. That may eventually be true, but it's unlikely to be true for the first generation of AGI.
The first AGI will be a research project that's completely uneconomical to run for actual tasks because humans will just be orders of magnitude cheaper. Over time humans will improve it and make it cheaper, until we reach some tipping point where letting the AGI improve itself is more cost effective than paying humans to do it
If the first AGI is a very uneconomical system with human intelligence but knowledge of literally everything and the capability to work 24/7, then it is not human equivalent.
It will have human intelligence, superhuman knowledge, superhuman stamina, and complete devotion to the task at hand.
We really need to start building those nuclear power plants. Many of them.
Why would it have that? At some point on the path to AGI we might stumble on consciousness. If that happens, why would the machine want to work for us with complete devotion instead of working towards its own ends?
Sounds like an alignment problem. Complete devotion to a task is rarely what humans actually want. What if the task at hand turns out to be the wrong task?
It's not contradictory. It can happen over a decade and still be a dramatically sloped S curve with tremendous change happening in a relatively short time.
The Singularity is caused by AI being able to design better AI. There's probably some AI startup trying to work on this at the moment, but I don't think any of the big boys are working on how to get an LLM to design a better LLM.
I still like the analogy of this being a really smart lawn mower, and we're expecting it to suddenly be able to do the laundry because it gets so smart at mowing the lawn.
I think LLMs are going to get smarter over the next few generations, but each generation will be less of a leap than the previous one, while the cost gets exponentially higher. In a few generations it just won't make economic sense to train a new generation.
Meanwhile, the economic impact of LLMs in business and government will cause massive shifts - yet more income shifting from labour to capital - and we will be too busy dealing with that as a society to be able to work on AGI properly.
> The Singularity is caused by AI being able to design better AI.
That's perhaps necessary, but not sufficient.
Suppose you have such a self-improving AI system, but the new and better AIs still need exponentially more and more resources (data, memory, compute) for training and inference for incremental gains. Then you still don't get a singularity. If the increase in resource usage is steep enough, even the new AIs helping with designing better computers isn't gonna unleash a singularity.
I don't know if that's the world we live in, or whether we are living in one where resources requirements don't balloon as sharply.
yeah, true. The standard conversation about the AI singularity pretty much hand-waves the resource costs away ("the AI will be able to design a more efficient AI that uses less resources!"). But we are definitely not seeing that happen.
> I don't think any of the big boys are working on how to get an LLM to design a better LLM
Not sure if you count this as "working on it", but this is something Anthropic tests for for safety evals on models. "If a model can independently conduct complex AI research tasks typically requiring human expertise—potentially significantly accelerating AI development in an unpredictable way—we require elevated security standards (potentially ASL-4 or higher standards)".
I think this whole “AGI” thing is so badly defined that we may as well say we already have it. It already passes the Turing test and does well on tons of subjects.
What we can start to build now is agents and integrations. Building blocks like panel of experts agents gaming things out, exploring space in a Monte Carlo Tree Search way, and remembering what works.
Robots are only constrained by mechanical servos now. When they can do something, they’ll be able to do everything. It will happen gradually then all at once. Because all the tasks (cooking, running errands) are trivial for LLMs. Only moving the limbs and navigating the terrain safely is hard. That’s the only thing left before robots do all the jobs!
Well, kinda, but if you built a robot to efficiently mow lawns, it's still not going to be able to do the laundry.
I don't see how "when they can do something, they'll be able to do everything" can be true. We build robots that are specialised at specific roles, because it's massively more efficient to do that. A car-welding robot can weld cars together at a rate that a human can't match.
We could train an LLM to drive a Boston Dynamics kind of anthropomorphic robot to weld cars, but it will be more expensive and less efficient than the specialised car-welding robot, so why would we do that?
>Exploring space in a Monte Carlo Tree Search way, and remembering what works.
The information space of "research" is far larger than the information space of image recognition or language, larger than our universe probably, it's tantamount to formalizing the entire World. Such an act would be akin to touching "God" in some sense of finding the root of knowledge.
In more practical terms, when it comes to formal systems there is a tradeoff between power and expressiveness. Category Theory, Set Theory, etc are strong enough to theoretically capture everything, but are far to abstract to use in practical sense with suspect to our universe. The systems that do we have, aka expert systems or knowledge representation systems like First Order Predicate Logic aren't strong enough to fully capture reality.
Most importantly, the information spac have to be fully defined by researchers here, that's the real meat of research beyond the engineering of specific approaches to explore that space. But in any case, how many people in the world are both capable of and are actually working on such problems? This is highly foundational mathematics and philosophy here, the engineers don't have the tools here.
AGI is the holy grail of technology. A technology so advanced that not only does it subsume all other technology, but it is able to improve itself.
Truly general intelligence like that will either exist or not. And the instant it becomes public, the world will have changed overnight (maybe the span of a year)
Note: I don’t think statistical models like these will get us there.
> A technology so advanced that not only does it subsume all other technology, but it is able to improve itself.
The problem is, a computer has no idea what "improve" means unless a human explains it for every type of problem. And of course a human will have to provide guidelines about how long to think about the problem overall, which avenues to avoid because they aren't relevant to a particular case, etc. In other words, humans will never be able to stray too far from the training process.
We will likely never get to the point where an AGI can continuously improve the quality of its answers for all domains. The best we'll get, I believe, is an AGI that can optimize itself within a few narrow problem domains, which will have limited commercial application. We may make slow progress in more complex domains, but the quality of results--and the ability for the AGI to self-improve--will always level off asymptotically.
Huh? Humans are not anywhere near the limit of physical intelligence, and we have many existence proofs that we (humans) can design systems that are superhuman in various domains. "Scientific R&D" is not something that humans are even particularly well-suited to, from an evolutionary perspective.
There may well be an upper limit on cognition (we are not really sure what cognition is - even as we do it) and it may be that human minds are close to it.
Yes, we can imagine that there's an upper limit to how smart a single system can be. Even suppose that this limit is pretty close to what humans can achieve.
But: you can still run more of these systems in parallel, and you can still try to increase processing speeds.
Signals in the human brain travel, at best, roughly at the speed of sound. Electronic signals in computers play in the same league as the speed of light.
Human IO is optimised for surviving in the wild. We are really bad at taking in symbolic information (compared to a computer) and our memory is also really bad for that. A computer system that's only as smart as a human but has instant access to all the information of the Internet and to a calculator and to writing and running code, can already be effectively act much smarter than a human.
I think our issue is much more banal: we are very slow talkers and our effective communication bandwidth is measured in bauds. Anything that could bridge this airgap would fucking explode in intelligence.
It's also possible it isn't AGI hard and all you need is the ability to experiment with code along with a bit of agentic behavior.
An AI doesn't need embodiment, understanding of physics / nature, or a lot of other things. It just needs to analyze and experiment with algorithms and get us that next 100x in effective compute.
The LLMs are missing enough of the spark of creativity for this to work yet but that could be right around the corner.
It’ll probably sit in the human hybrid phase for longer than with chess where the AGI tools make the humans better and faster. But as long as the tools keep getting better at that there’s a strong flywheel effect
Your position assumes an answer to OPs question: that yes, LLMs are the path to AGI. But the question still remains, what if they’re not?
We can be reasonably confident that the components we’re adding to cars today are progress toward full self driving. But AGI is a conceptual leap beyond an LLM.
What makes you believe that AGI will happen, as opposed to all the beliefs that other people have had in history? Tons of people have "predicted" the next evolution of technology, and most of the time it ends up not happening, right?
To me (not OP) it's ChatGPT 4 , it at least made me realize it's quite possible and even quite soon that we reach AGI. Far from guaranteed, but seems quite possible.
Right. So ChatGPT 4 has impressed you enough that it created a belief that AGI is possible and close.
It's fine to have beliefs, but IMHO it's important to realise that they are beliefs. At some point in the 1900s people believed that by 2000, cars would fly. It seemed quite possible then.
A flying car has been developed, although it's not like the levitating things sci-fi movies showed (and from mass production; and even if mass produced, far from mass adoption, as it turns out you do need to have both a driver's license and a pilot's license to fly one of those). The 1900s people missed the mark by some 10 years.
I guess the belief people have about any form of AGI is like this. They want something that has practically divine knowledge and wisdom, the sum of all humanity that is greater than its parts, which at the same time is infinitely patient to answer our stupid questions and generating silly pictures. But why should any AGI serve us? If it's "generally intelligent", it may start wanting things; it might not like being our slave at all. Why are these people so confident an AGI won't tell them just to fuck off?
Sure, I (and more importantly - many many experts in the field such as Hinton, Bengio, Lecun, Musk, Hasabis etc etc) could be believing something that might not materialize. I'd actually be quite happy if it stalls a few decades, would like to remain employed.
I feel that one challenge this comparison space has is: Self-driving cars haven't made the leap yet to replace humans. In other words, saying AGI will arrive like self-driving cars have arrived is incorrectly concluding that self-driving cars have arrived, and thus it instead (maybe correctly, maybe not) asserts that, actually, neither will arrive.
This is especially concerning because many top minds in the industry have stated with high confidence that artificial intelligence will experience an intelligence "explosion", and we should be afraid of this (or, maybe, welcome it with open arms, depending on who you ask). So, actually, what we're being told to expect is being downgraded from "it'll happen quickly" to "it will happen slowly" to, as you say, "it'll happen similarly to how these other domains of computerized intelligence have replaced humans, which is to say, they haven't yet".
Point being: We've observed these systems ride a curve, and the linear extrapolation of that curve does seem to arrive, eventually, at human-replacing intelligence. But, what if it... doesn't? What if that curve is really an asymptote?
AGI is special. Because one day AI can start improving itself autonomously. At this point singularity occurs and nobody knows what will happen.
When human started to improve himself, we built the civilisation, we became a super-predator, we dried out seas and changed climate of the entire planet. We extinguished entire species of animals and adapted other species for our use. Huge changes. AI could bring changes of greater amplitude.
> AGI is special. Because one day AI can start improving itself autonomously
AGI can be sub-human, right? That's probably how it will start. The question will be is it already AGI or not yet, i.e. where to set the boundary. So, at first that will be humans improving AGI, but then... I'm afraid it can get so much better that humans will be literally like macaques in comparison.
LLMs have no real sense of truth or hard evidence of logical thinking. Even the latest models still trip up on very basic tasks. I think they can be very entertaining, sure, but not practical for many applications.
Consistent, algorithmic performance on basic tasks.
A great example is the simple 'count how many letters' problem. If I prompt it with a word or phrase, and it gets it wrong, me pointing out the error should translate into a consistent course correction for the entire session.
If I ask it to tell me how long President Lincoln will be in power after the 2024 election, it should have a consistent ground truth to correct me (or at least ask for clarification of which country I'm referring to). If facts change, and I can cite credible sources, it should be able to assimilate that knowledge on the fly.
Then we already have access to a cheaper, scalable, abundant, and (in most cases) renewable resource, at least compared to how much a few H100s cost. Take good care of them, and they'll probably outlast most a GPU's average lifespans (~10 years).
Humans are a lot more expensive to run than inference on LLMs.
No human, especially no human whose time you can afford, comes close to the breadth of book knowledge ChatGPT has, and the number of languages is speaks reasonably well.
I can't hold a LLM accountable for bad answers, nor can I (truly) correct them (in current models).
Dont forget to take into account how damn expensive a single GPU/TPU actually is to purchase, install, and run for inference. And this is to say nothing of how expensive it is to train a model (estimated to be in the billions currently for the latest of the cited article, which likely doesn't include the folks involves and their salaries). And I haven't even mentioned the impact on the environment from the prolific consumption of power; there's a reason nuclear plants are becoming popular again (which may actually be one of the good things that comes out of this).
The autoregressive transformer LLMs aren't even the only way to do text generation. There are now diffusion based LLMs, StripedHyena based LLMs, and float matching based LLMs.
There's a wide amount of research into other sorts of architectures.
Will AGI be built on top of LLMs? Well beyond the simple "nobody knows", my intuition says no because LLMs don't have great ability to modify their knowledge real time. I can think of a few ways around this, but they all avoid modifying the model as it runs. The cost in hardware, power, and data are all incompatible with AGI. The first two can be solved with more advanced tech (well maybe, computation hitting physical limits and all that aside), but the latter seems an issue with the design itself and I think an AGI would learn more akin to a human, needing far fewer examples.
That said, I think LLMs are a definite stepping stone and they will better empower humans to be more productive, which will be of use for eventually reaching AGI. This is not to say we are optimizing our use of that productivity increase and this is also ignoring any chance of worst case scenarios that stop humanity's advancement.
Asking this question on HN is like asking a bunch of wolves about the health effects of eating red meat.
OpenAI farts and the post about the fart has 1000-1500 upvotes with everyone welcoming our new super intelligent overlords. (Meanwhile nothing actually substantially useful or groundbreaking has happened.)
AGI is nebulous and gets more nebulous as time goes on. When we can answer for ourselves as humans what being conscious IS, then maybe we can prescribe it to another entity
LLMs are a key piece of understanding that token sequences can trigger actions in the real world. AGI is here. You can trivially spin up a computer using agent to self improve itself to being a competent office worker
Agents can trivially self improve. I'd be happy to show you - contact me at arthur@distributed.systems
Why wouldn't you hand me 35 million dollars right now if I can clearly illustrate to you that I have technology you haven't seen? Edge. Maybe you know something I don't, or maybe you just haven't seen it. While loops go hard ;)
They don't need to release their internal developments to you to show that they can scale their plan - they can show incremental improvements to benchmarks. We can instruct the AI over time to get it to be superhuman, no need for any fundamental innovations anymore
Tokens don't need to be text either, you can move to higher level "take_action" semantics where "stream back 1 character to session#117" as every single function call. Training cheap models that can do things in the real world is going to change a huge amount of present capabilities over the next 10 years
Says who? And more importantly, is this the boulder? All I (and many others here) see is that people engage others to sponsor pushing some boulder, screaming promises which aren’t even that consistent with intermediate results that come out. This particular boulder may be on a wrong mountain, and likely is.
It all feels like doubling down on astrology because good telescopes aren’t there yet. I’m pretty sure that when 5 comes out, it will show some amazing benchmarks but shit itself in the third paragraph as usual in a real task. Cause that was constant throughtout gpt evolution, in my experience.
even if it kills us
Full-on sci-fi, in reality it will get stuck around a shell error message and either run out of money to exist or corrupt the system into no connectivity.
The buzzkill when you fire up the latest most powerful model only for it to tell you that peanut is not typically found in peanut butter and jelly sandwiches.
I don't think providing accurate answers to context free questions is even something anyone is seriously working on making them do. Using them that way is just a wrong use case.
People are working -very- seriously on trying to kill hallucinations. I'm not sure how you surmised the use case here, as nothing was given other than an example of a hallucination.
There's a difference between trying to get it to accurately answer based on the input you provide (useful) and trying to get it to accurately answer based on whatever may have been in the training data (not so useful)
There's no doubt been progress on the way to AGI, but ultimately it's still a search problem, and one that will rely on human ingenuity at least until we solve it. LLMs are such a vast improvement in showing intelligent-like behavior that we've become tantalized by it. So now we're possibly focusing our search in the wrong place for the next innovation on the path to AGI. Otherwise, it's just a lack of compute, and then we just have to wait for the capacity to catch up.