AGI is poorly defined and thus is a science "problem", and a very low priority one at that.
No amount of engineering or model training is going to get us AGI until someone defines what properties are required and then researches what can be done to achieve them within our existing theories of computation which all computers being manufactured today are built upon.
I think that completely discounting the potential of new emergent capabilities at scale undermines this thesis significantly. We don't know until someone tries, and there is compelling evidence that there's still plenty of juice to squeeze out of both scale and engineering.
We don't know if AGI is even possible outside of a biological construct yet. This is key. Can we land on AGI without some clear indication of possibility (aka Chappie style)? Possibly, but the likelihood is low. Quite low. It's essentially groping in the dark.
A good contrast is quantum computing. We know that's possible, even feasible, and now are trying to overcome the engineering hurdles. And people still think that's vaporware.
It's not "key"; it's not even relevant ... the proof will be in the pudding. Proving a priori that some outcome is possible plays no role in achieving it. And you slid, motte-and-bailey-like, from "know" to "some clear indication of possibility" -- we have extremely clear indications that it's possible, since there's no reason other than a belief in magic to think that "biological" is a necessity.
Whether is feasible or practical or desirable to achieve AGI is another matter, but the OP lays out multiple problem areas to tackle.
Sometimes I think we’re like cats that learned how to make mirrors without really understanding them, and are so close to making one good enough that the other cat becomes sentient.
Nah,this sounds like a modern remix of Japan’s Fifth Generation Computing project. They thought that by building large databases and with Prolog they would bring upon an AI renaissance.
Just hand waving some “distributed architecture” and trying to duct tape modules together won’t get us any closer to AGI.
The building blocks themselves, the foundation, has to be much better.
Arguably the only building block that LLMs have contributed is that we have better user intent understanding now; a computer can just read text and extract intent from it much better than before. But besides that, the reasoning/search/“memory” are the same building blocks of old, they look very similar to techniques of the past, and that’s because they’re limited by information theory / computer science, not by today’s hardware or systems.
We can certainly get much more utility out of current architectures with better engineering, as "agents" have shown, but to claim that AGI is possible with engineering alone is wishful thinking. The hard part is building systems that showcase actual intelligence and reasoning, that are able to learn and discover on their own instead of requiring exorbitantly expensive training, that don't hallucinate, and so on. We still haven't cracked that nut, and it's becoming increasingly evident that the current approaches won't get us there. That will require groundbreaking compsci work, if it's possible at all.
The problem is that if it's an engineering problem then further advancement will rely on step function discoveries like the transformer. There's no telling when that next breakthrough will come or how many will be needed to achieve AGI.
In the meantime I guess all the AI companies will just keep burning compute to get marginal improvements. Sounds like a solid plan! The craziest thing about all of this is that ML researchers should know better!! Anyone with extensive experience training models small or large knows that additional training data offers asymptotic improvements.
I think the LLM businesses as-is are potentially fine businesses. Certainly the compute cost of running and using them is very high, not yet reflected in the prices companies like OpenAI and Anthropic are charging customers. It remains to be seen if people will pay the real costs.
But even if LLMs are going to tap out at some point, and are a local maximum, dead-end, when it comes to taking steps toward AGI, I would still pay for Claude Code until and unless there's something better. Maybe a company like Anthropic is going to lead that research and build it, or maybe (probably) it's some group or company that doesn't exist yet.
"AGI needs to update beliefs when contradicted by new evidence" is a great idea, however, the article's approach of building better memory databases (basically fancier RAG) doesn't seem enable this. Beliefs and facts are built into LLMs at a very low layer during training. I wonder how they think they can force an LLM to pull from the memory bank instead of the training data.
(Also, LLMs don't have beliefs or other mental states. As for facts, it's trivially easy to get an LLM to say that it was previously wrong ... but multiple contradictory claims cannot all be facts.)
If you believe the bitter lesson, all the handwavy "engineering" is better done with more data. Someone likely would have written the same thing as this 8 years ago about what it would take to get current LLM performance.
So I don't buy the engineering angle, I also don't think LLMs will scale up to AGI as imagined by Asimov or any of the usual sci-fi tropes. There is something more fundamental missing, as in missing science, not missing engineering.
What will it scale up to if not AGI? OpenAI has a synthetic data flywheel. What are the asymptotics of this flywheel assuming no qualitative additional breakthrough?
The missing science to engineer intelligence is composable program synthesis. Aloe (https://aloe.inc) recently released a GAIA score demonstrating how CPS dramatically outperforms other generalist agents (OpenAI's deep research, Manus, and Genspark) on tasks similar to those a knowledge worker would perform.
I'd argue it's because intelligence has been treated as a ML/NN engineering problem that we've had the hyper focus on improving LLMs rather than the approach articulated in the essay.
Intelligence must be built from a first principles theory of what intelligence actually is.
So the "Bitter Lesson" paper actually came up recently and I was surprised to discover that what it claimed was sensible and not at "all you need is data" or "data is inherently better"
The first line and the conclusion is: "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin." [1]
I don't necessary agree with it's examples or the direction it vaguely points at. But it's basic statement seems sound. And I would say that there's lot of opportunity for engineer, broadly speaking, in the process of creating "general methods that leverage computation" (IE, that scale). What the bitter lesson page was roughly/really about was earlier "AI" methods based on logic-programming and which including information on the problem domain in the code itself.
And finally, the "engineering" the paper talks about actually is pro-Bitter lesson as far as I can tell. It's taking data routing and architectural as "engineering" and here I agree this won't work - but for the opposite reason - specifically 'cause I don't just data routing/process will be enough.
Aye. Missing are self correction (world models/action and response observation), coherence over the long term, and self-scaling. The 3rd are what all the SV types are worried about, except maybe Yann LeCun who is worried about the first and second.
Hinton thinks the 3rd is inevitable/already here and humanity is doomed. It's an odd arena.
The counter argument is that we were working with thermodynamics before knowing the theory. Famously the steam engine came before the first law of thermodynamics. Sometimes engineering is like that. Using something that you don’t understand exactly how it works.
I am thinking we need a foundation, something that is concrete and explicit and doesn't do hallucination. But has very limited knowledge outside of absolute Maths and basic physics.
> If you believe the bitter lesson, all the handwavy "engineering" is better done with more data
Id say better model architechture than more data. A human can learn to do things more complex than an LLM with less data. I think modelling the world as a static system to be representation learned in an unsupervised fashion is blocked on the static assumption. The world is dynamical, that should be reflected in the base model
But yeah, definitely not an engineering problem. Thats like saying the reason a crow isnt as smart as a person is becauss they dont have the hands to type of keyboards. But its also not because they havent seen enough of the world like your saying. Its be ause their brain isnt complex enough
The bitter lesson was "general methods that leverage computation" win rather than more data. Like rather than just LLMs you could maybe try applying something like AlphaEvolve to finding better algorithms/systems (https://news.ycombinator.com/item?id=43985489).
> all the handwavy "engineering" is better done with more data.
How long until that gets more reliable than a simple database? How long until it can execute code faster than a CPU running a program?
A lot of the stuff humans accomplish is through technology, not due to growing a bigger brain. Even something seemingly basic like a math equation benefits drastically from being written down with pen&paper instead of being juggled in the human brain itself (see Extended mind thesis). And when it comes to something like running a 3D engine, there is pretty much no hope of doing it with just your brain.
Maybe we will get AIs smart enough that they can write their own tools, but for that to happen, we still need the infrastructure that allows them writing the tools in the first place. The way they can access Python is a start, but there is still a lack of persistence that lets them keep their accomplishments for future runs, be it in the form of a digital notepad or dynamic updating of weights.
Indeed. The Bitter Lesson has proved true so far. This sounds like going back to the 60s expert systems concept we're trying to get away from. The author also just describes RAG. That certainly isn't AGI, which probably isn't achievable at all.
8 years is a pretty short perspective. The current growth phase was unlocked by more engineering. We could’ve had some of these kinds of capabilities decades ago, as illustrated by how cutting edge AI research is now trickling down all the way to microcontrollers with 64MHz CPU and kBs of RAM.
Once we got the “Attention is all you need” paper I don’t remember anyone saying we couldn’t get better results by throwing more data and compute at it. But now we’ve pretty much thrown all the data and all (as much as we can reasonably manufacture) at it. So clearly we’re at the end of that phase.
I agree with your comment and the article. LLMs should be part of the answer, but the core of the progress should probably dive back into neural networks in general. Language is how we communicate as well as with other senses, but right now we're stuck at LLMs that just seem to be blown out elizas trained with other actual humans work. I remember early on, training of simple neural networks was done with rules in their environment and they evolved behavior according to criteria set, like genetic algorithms. I think the current LLMs are getting a "filtered" view of the environment, and that filter behaves like the average IQ of netizens lol
I think the gist of TFA is just that we need a new architecture/etc not scaling.
I suppose one can argue about whether designing a new AGI-capable architecture and learning algorithm(s) is a matter of engineering (applying what we already know) or research, but I think saying we need new scientific discoveries is going to far.
Neural nets seems to be the right technology, and we've now amassed a ton of knowledge and intuition about what neural nets can do and how to design with them. If there was any doubt, then LLMs, even if not brain-like, have proved the power of prediction as a learning technique - intelligence essentially is just successful prediction.
It seems pretty obvious that the rough requirements for an neural-net architecture for AGI are going to be something like our own neocortex and thalamo-cortical loop - something that learns to predict based on sensory feedback and prediction failure, including looping and working memory. Built-in "traits" like curiosity (prediction failure => focus) and boredom will be needed so that this sort of autonomous AGI puts itself into leaning situations and is capable of true innovation.
The major piece to be designed/discovered isn't so much the architecture as the incremental learning algorithm, and I think if someone like Google-DeepMind focused their money, talent and resources on this then they could fairly easily get something that worked and could then be refined.
Demis Hassabis has recently given an estimate of human-level AGI in 5 years, but has indicated that a pre-trained(?) LLM may still be one component of it, so not clear exactly what they are trying to build in that time frame. Having a built-in LLM is likely to prove to be a mistake where the bitter lesson applies - better to build something capable of self-learning and just let it learn.
Sometimes I think the fundamental thing could be as ‘simple’ as something like introducing l an attention/event loo, flush to memory, emotion driven motivation. There are quite a few fairly obvious things that llms don’t have that might be best not to add just in case.
Now that there’s a fundamental technical framework for producing something like coherence, the ability to make a reliable, persistent personality will require new insights into how our own minds take shape, not just ever more data in the firehose
the idea that you would somehow produce intelligence by feeding billions of reddit comments into a statistical text model is will go down as the biggest con in history
Am I the only one who feels that Claude Code is what they would have imagined basic AGI to be like 10 years ago?
It can plan and take actions towards arbitrary goals in a wide variety of mostly text-based domains. It can maintain basic "memory" in text files. It's not smart enough to work on a long time horizon yet, it's not embodied, and it has big gaps in understanding.
But this is basically what I would have expected v1 to look like.
I suspect most people envision AGI as at least having sentience. To borrow from Star Trek, the Enterprise's main computer is not at the level of AGI, but Data is.
The biggest thing that is missing (IMHO) is a discrete identity and notion of self. It'll readily assume a role given in a prompt, but lacks any permanence.
> Am I the only one who feels that Claude Code is what they would have imagined basic AGI to be like 10 years ago?
That wouldn't have occurred to me, to be honest. To me, AGI is Data from Star Trek. Or at the very least, Arnold Schwarzenegger's character from The Terminator.
I'm not sure that I'd make sentience a hard requirement for AGI, but I think my general mental fantasy of AGI even includes sentience.
Claude Code is amazing, but I would never mistake it for AGI.
Totally agree. It even (usually) gets subtle meanings from my often hastily written prompts to fix something.
What really occurs to me is that there is still so much can be done to leverage LLMs with tooling. Just small things in Claude Code (plan mode for example) make the system work so much better than (eg) the update from Sonnet 3.5 to 4.0 in my eyes.
No you are not the only one. I am continuously mystified by the discussion surrounding this. Clause is absolutely and unquestionably an artificial general intelligence. But what people mean by “AGI” is a constantly shifting, never defined goalpost moving at sonic speed.
The suggested requirements are not engineering problems. Conceiving of a model architecture that can represent all the systems described in the blog is a monumental task of computer science research.
I think the OP's point is that all those requirements are to be implemented outside the LLM layer, i.e. we don't need to conceive of any new model architecture. Even if LLMs don't progress any further beyond GPT-5 & Claude 4, we'll still get there.
Take memory for example: give LLM a persistent computer and ask it to jot down its long-term memory as hierarchical directories of markdown documents. Recalling a piece of memory means a bunch of `tree` and `grep` commands. It's very, very rudimentary, but it kinda works, today. We just have to think of incrementally smarter ways to query & maintain this type of memory repo, which is a pure engineering problem.
The reason people don't want to answer this question is because the value proposition from AI labs is slavery. If intelligence requires agency, they are worthless.
It's a research problem, a science problem. And then an engineering problem to industrialize it. How can we replicate intelligence if we don't even know how it emerges from our brains?
I think the author could have picked a better title. “<X> is an engineering problem” is a pretty common expression to describe something where the science is done, but the engineering remains. There’s an understanding that that could still mean a ton of work, but there isn’t some fundamental mystery about the basic natural principles of the thing.
Here, AGI is being described as an engineering problem, in contrast to a “model training” problem. That is, I think at least, he’s at least saying that more work needs to be done at an R&D level. I agree with those who are saying it is maybe not even an engineering problem yet, but should be noted that he’s at least pushing away from just running the existing programs harder, which seems to be the plan with trillions of dollars behind it.
108 comments
[ 3.4 ms ] story [ 73.0 ms ] threadAGI is poorly defined and thus is a science "problem", and a very low priority one at that.
No amount of engineering or model training is going to get us AGI until someone defines what properties are required and then researches what can be done to achieve them within our existing theories of computation which all computers being manufactured today are built upon.
An unfortunate tendency that many in high-tech suffer from is the idea that any problem can be solved with engineering.
We don’t even know how.
A good contrast is quantum computing. We know that's possible, even feasible, and now are trying to overcome the engineering hurdles. And people still think that's vaporware.
Whether is feasible or practical or desirable to achieve AGI is another matter, but the OP lays out multiple problem areas to tackle.
Of course it is. A brain is just a machine like any other.
Just hand waving some “distributed architecture” and trying to duct tape modules together won’t get us any closer to AGI.
The building blocks themselves, the foundation, has to be much better.
Arguably the only building block that LLMs have contributed is that we have better user intent understanding now; a computer can just read text and extract intent from it much better than before. But besides that, the reasoning/search/“memory” are the same building blocks of old, they look very similar to techniques of the past, and that’s because they’re limited by information theory / computer science, not by today’s hardware or systems.
Probably need another cycle of similar breakthrough in model engineering before this more complex neural network gets a step function better.
Moar data ain’t gonna help. The human brain is the proof: it doesnt need the internet’s worth of data to become good (nor all that much energy).
We can certainly get much more utility out of current architectures with better engineering, as "agents" have shown, but to claim that AGI is possible with engineering alone is wishful thinking. The hard part is building systems that showcase actual intelligence and reasoning, that are able to learn and discover on their own instead of requiring exorbitantly expensive training, that don't hallucinate, and so on. We still haven't cracked that nut, and it's becoming increasingly evident that the current approaches won't get us there. That will require groundbreaking compsci work, if it's possible at all.
In the meantime I guess all the AI companies will just keep burning compute to get marginal improvements. Sounds like a solid plan! The craziest thing about all of this is that ML researchers should know better!! Anyone with extensive experience training models small or large knows that additional training data offers asymptotic improvements.
But even if LLMs are going to tap out at some point, and are a local maximum, dead-end, when it comes to taking steps toward AGI, I would still pay for Claude Code until and unless there's something better. Maybe a company like Anthropic is going to lead that research and build it, or maybe (probably) it's some group or company that doesn't exist yet.
(Also, LLMs don't have beliefs or other mental states. As for facts, it's trivially easy to get an LLM to say that it was previously wrong ... but multiple contradictory claims cannot all be facts.)
So I don't buy the engineering angle, I also don't think LLMs will scale up to AGI as imagined by Asimov or any of the usual sci-fi tropes. There is something more fundamental missing, as in missing science, not missing engineering.
I'd argue it's because intelligence has been treated as a ML/NN engineering problem that we've had the hyper focus on improving LLMs rather than the approach articulated in the essay.
Intelligence must be built from a first principles theory of what intelligence actually is.
The first line and the conclusion is: "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin." [1]
I don't necessary agree with it's examples or the direction it vaguely points at. But it's basic statement seems sound. And I would say that there's lot of opportunity for engineer, broadly speaking, in the process of creating "general methods that leverage computation" (IE, that scale). What the bitter lesson page was roughly/really about was earlier "AI" methods based on logic-programming and which including information on the problem domain in the code itself.
And finally, the "engineering" the paper talks about actually is pro-Bitter lesson as far as I can tell. It's taking data routing and architectural as "engineering" and here I agree this won't work - but for the opposite reason - specifically 'cause I don't just data routing/process will be enough.
[1]https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson...
Hinton thinks the 3rd is inevitable/already here and humanity is doomed. It's an odd arena.
I am thinking we need a foundation, something that is concrete and explicit and doesn't do hallucination. But has very limited knowledge outside of absolute Maths and basic physics.
Id say better model architechture than more data. A human can learn to do things more complex than an LLM with less data. I think modelling the world as a static system to be representation learned in an unsupervised fashion is blocked on the static assumption. The world is dynamical, that should be reflected in the base model
But yeah, definitely not an engineering problem. Thats like saying the reason a crow isnt as smart as a person is becauss they dont have the hands to type of keyboards. But its also not because they havent seen enough of the world like your saying. Its be ause their brain isnt complex enough
How long until that gets more reliable than a simple database? How long until it can execute code faster than a CPU running a program?
A lot of the stuff humans accomplish is through technology, not due to growing a bigger brain. Even something seemingly basic like a math equation benefits drastically from being written down with pen&paper instead of being juggled in the human brain itself (see Extended mind thesis). And when it comes to something like running a 3D engine, there is pretty much no hope of doing it with just your brain.
Maybe we will get AIs smart enough that they can write their own tools, but for that to happen, we still need the infrastructure that allows them writing the tools in the first place. The way they can access Python is a start, but there is still a lack of persistence that lets them keep their accomplishments for future runs, be it in the form of a digital notepad or dynamic updating of weights.
Once we got the “Attention is all you need” paper I don’t remember anyone saying we couldn’t get better results by throwing more data and compute at it. But now we’ve pretty much thrown all the data and all (as much as we can reasonably manufacture) at it. So clearly we’re at the end of that phase.
I suppose one can argue about whether designing a new AGI-capable architecture and learning algorithm(s) is a matter of engineering (applying what we already know) or research, but I think saying we need new scientific discoveries is going to far.
Neural nets seems to be the right technology, and we've now amassed a ton of knowledge and intuition about what neural nets can do and how to design with them. If there was any doubt, then LLMs, even if not brain-like, have proved the power of prediction as a learning technique - intelligence essentially is just successful prediction.
It seems pretty obvious that the rough requirements for an neural-net architecture for AGI are going to be something like our own neocortex and thalamo-cortical loop - something that learns to predict based on sensory feedback and prediction failure, including looping and working memory. Built-in "traits" like curiosity (prediction failure => focus) and boredom will be needed so that this sort of autonomous AGI puts itself into leaning situations and is capable of true innovation.
The major piece to be designed/discovered isn't so much the architecture as the incremental learning algorithm, and I think if someone like Google-DeepMind focused their money, talent and resources on this then they could fairly easily get something that worked and could then be refined.
Demis Hassabis has recently given an estimate of human-level AGI in 5 years, but has indicated that a pre-trained(?) LLM may still be one component of it, so not clear exactly what they are trying to build in that time frame. Having a built-in LLM is likely to prove to be a mistake where the bitter lesson applies - better to build something capable of self-learning and just let it learn.
The underlying assumption is that it exists in the first place. Or rather, one must first accept an axiom.
In fermi, its that interstellar signals can be detected and further travel is possible.
In AGI, its that intelligence is a isolateable process which we can bootstrap in minimal time.
Both assumes human progress are templates of unlimited exponential growth.
the idea that you would somehow produce intelligence by feeding billions of reddit comments into a statistical text model is will go down as the biggest con in history
(so far)
It can plan and take actions towards arbitrary goals in a wide variety of mostly text-based domains. It can maintain basic "memory" in text files. It's not smart enough to work on a long time horizon yet, it's not embodied, and it has big gaps in understanding.
But this is basically what I would have expected v1 to look like.
I suspect most people envision AGI as at least having sentience. To borrow from Star Trek, the Enterprise's main computer is not at the level of AGI, but Data is.
The biggest thing that is missing (IMHO) is a discrete identity and notion of self. It'll readily assume a role given in a prompt, but lacks any permanence.
That wouldn't have occurred to me, to be honest. To me, AGI is Data from Star Trek. Or at the very least, Arnold Schwarzenegger's character from The Terminator.
I'm not sure that I'd make sentience a hard requirement for AGI, but I think my general mental fantasy of AGI even includes sentience.
Claude Code is amazing, but I would never mistake it for AGI.
What really occurs to me is that there is still so much can be done to leverage LLMs with tooling. Just small things in Claude Code (plan mode for example) make the system work so much better than (eg) the update from Sonnet 3.5 to 4.0 in my eyes.
Take memory for example: give LLM a persistent computer and ask it to jot down its long-term memory as hierarchical directories of markdown documents. Recalling a piece of memory means a bunch of `tree` and `grep` commands. It's very, very rudimentary, but it kinda works, today. We just have to think of incrementally smarter ways to query & maintain this type of memory repo, which is a pure engineering problem.
What if intelligence requires agency ?
Here, AGI is being described as an engineering problem, in contrast to a “model training” problem. That is, I think at least, he’s at least saying that more work needs to be done at an R&D level. I agree with those who are saying it is maybe not even an engineering problem yet, but should be noted that he’s at least pushing away from just running the existing programs harder, which seems to be the plan with trillions of dollars behind it.