I can certainly see an argument for that, but also for thinking that NSAI would be a better way to get to alignment. (and i don’t think a pause is actually at all politically realistic though possibly desirable.)
I agree with your conclusion but I would reserve some funding for non-hybrid, non-LLM approaches. Deep Learning definitely has been successful but it is still essentially a statistical modeling technique. I think it is likely that evolution created brain mechanisms that go way beyond simple statistical analysis of input. We should spend more time and money looking for those non-statistical mechanisms. My guess is that, once we discover them, we will find that AGI does not require we turn the world into AI server farms.
"Still, as I wrote yesterday (and also six years ago"
I value your voice, it's been an amazing antidote for the PR-engineered AI hype of the last few years, but it's also extremely difficult to take you seriously where pretty much everything you post has those numerous "I told you so, but nobody listened" undertones (and they aren't exactly subtle...) such repeated emotional responses rarely go hand in hand with objective reasoning, so I think you're doing yourself a disservice and effectively diluting the important and knowledgeable remarks that you're making.
I’d rather we spend a trillion dollars on climate science or public health, if someone has a trillion lying around somewhere. We won’t even need a trillion to end tuberculosis!
"As a society we could spend another trillion dollars pursue LLMs, or a trillion dollars seeking new, hybrid approaches. I for one know which I would pick."
Hmm... or, building a decent, wholistic education that equips young people with the ability to reason, to explore, to imagine.
Apparently Larry has it right. No need to plan for or concern ourselves with actual human needs like education, jobs, etc if we can just achieve AGI and eliminate humans altogether, right?????????
Gary—this is a thoughtful piece, and I agree with the core diagnosis: current systems don’t truly reason, and scaling alone won’t get us there.
Where I think the conversation is still incomplete is in the proposed remedy.
Neurosymbolic systems add structure, but they don’t solve what appears to be the deeper issue: admissibility.
Both LLMs and symbolic systems share a common property—they can always produce an answer. There is no mechanism that determines whether an answer is allowed to be asserted in the first place.
In scientific practice, that constraint is fundamental. Hypotheses are not just generated—they are exposed to reality, filtered, and only retained if they survive. Critically, the system must also be able to withhold assertion when evidence is insufficient.
That full loop—exposure, evidence accumulation, survival filtering, and abstention—is what’s missing from both paradigms.
LLMs fail because they interpolate without constraint.
Symbolic systems fail because they assert without exposure.
Adding symbols doesn’t fix admissibility.
The gap is not reasoning machinery—it’s the absence of a mechanism that allows reality to reject structure before it is asserted.
In that sense, what’s needed is not just a hybrid architecture, but something closer to an executable version of the scientific method itself—where only structures that survive repeated exposure are permitted to participate in inference.
That shift—toward evidence-licensed assertion—may be the difference between systems that appear to reason and systems that can be trusted in high-stakes domains.
Thanks, this is excellent thinking. My biggest fear about AI is that the entire environment is essentially a big answering machine instead of a big questioning machine. Most people have a terrible grasp over what facts and answers are, so it’s hardly surprising that machines fare no better!
What I've learned over the past few years is this: the men who are hellbent on forcing an AI-centric future on all of us are mad, bad, and dangerous to share a planet with. In light of this, do we want neurosymbolic AI? Apart from machine learning technologies specifically developed for narrow purposes with an obvious potential to benefit humanity (e.g., medical research), do we want any sort of AI at all?
In light of the deficiencies in LLM machines, should that influence financing for the data centers now being built across the U.S.? Are investors pouring billions of dollars into data centers that will simply not do what the likes of Sam Altman proclaim they can do?
I 100% agree with the analysis and hybrid direction, but aren't there already public LLMs that generate, test, and execute code in order to answer such formal questions?
Never in the history of any subject has something that is so touted yielded so little utility, than neurosymbolic AI. Sure, one can ascribe a lookup table being scaffolded to an LLM as some kind of neurosymbolic system (though this is highly technically questionable), but the response to this is always the same. If it’s so effective, so power efficient, and so easily done, then where is it?
Needs to go flat ass broke first. The state of mind of hypers and hypees is very far from "let's start over". And even if it wasn't, they would never be able to know what "start over" even means.
I sent a copy of an AI qualifier printed at the bottom of an online science magazine (Science X Newsletter) to an academic friend in Denmark and realized as I was sending it that I will have to add a note to all of my online papers (academia.edu) that were written before the AI revolution that there was no AI writing in them. Think of any book you have on your shelves--or in any library--we really have entered a new world so different from the one before.
Markets have been pricing LLMs like they found the unified theory of everything. A 95% vs 34% success rate on a problem humans solve with a napkin and twenty minutes suggests the unified theory still has a few pages missing. The trillion dollars chasing scale may eventually hit the same wall the Tower of Hanoi keeps building
The Tufts result tracks with what DeepMind's AlphaGeometry team quietly conceded in 2024: their IMO-level solver was ~53% pure LLM, but the symbolic deduction engine did the actual proving — the neural net just proposed auxiliary constructions. Every headline "LLM solves math" system in the last two years has a classical solver hiding in the basement. The neurosymbolic future isn't coming; it already shipped, just with the symbolic half airbrushed out of the press release.
Thanks for sharing! It seems neurosymbolic module primarily acts as a temporal scaffold for the LLM?
In brains, all neurons have endogenous clocks, and evolutionarily we know this came about as multicellular life evolved, and before cells specialized into “neurons” to handle some computations. All neurons thus do their thing against a robust circadian clock that each neuron maintains and synchronizes to external time signals.
As I understand it, LLMs are stateless between tasks, and while a lot of complexity is encoded in its structure, its dynamics are severely constrained.
To me, this absence of “time sense” at a fundamental level seems to be the reason LLMs can’t reason or plan well. We try to bolt on physical time on top, but that’s pretty inefficient and can mean you “compute” something that’s no longer relevant to the changed environment you’re computing in.
I can certainly see an argument for that, but also for thinking that NSAI would be a better way to get to alignment. (and i don’t think a pause is actually at all politically realistic though possibly desirable.)
I agree with your conclusion but I would reserve some funding for non-hybrid, non-LLM approaches. Deep Learning definitely has been successful but it is still essentially a statistical modeling technique. I think it is likely that evolution created brain mechanisms that go way beyond simple statistical analysis of input. We should spend more time and money looking for those non-statistical mechanisms. My guess is that, once we discover them, we will find that AGI does not require we turn the world into AI server farms.
agreed
"Still, as I wrote yesterday (and also six years ago"
I value your voice, it's been an amazing antidote for the PR-engineered AI hype of the last few years, but it's also extremely difficult to take you seriously where pretty much everything you post has those numerous "I told you so, but nobody listened" undertones (and they aren't exactly subtle...) such repeated emotional responses rarely go hand in hand with objective reasoning, so I think you're doing yourself a disservice and effectively diluting the important and knowledgeable remarks that you're making.
I’d rather we spend a trillion dollars on climate science or public health, if someone has a trillion lying around somewhere. We won’t even need a trillion to end tuberculosis!
"As a society we could spend another trillion dollars pursue LLMs, or a trillion dollars seeking new, hybrid approaches. I for one know which I would pick."
Hmm... or, building a decent, wholistic education that equips young people with the ability to reason, to explore, to imagine.
Soon, young people won’t have any need to reason, explore or imagine.
AI will obviate all that.
Then we will have reached Nvidia..I mean nirvana.
Apparently Larry has it right. No need to plan for or concern ourselves with actual human needs like education, jobs, etc if we can just achieve AGI and eliminate humans altogether, right?????????
The future looks empty
Gary—this is a thoughtful piece, and I agree with the core diagnosis: current systems don’t truly reason, and scaling alone won’t get us there.
Where I think the conversation is still incomplete is in the proposed remedy.
Neurosymbolic systems add structure, but they don’t solve what appears to be the deeper issue: admissibility.
Both LLMs and symbolic systems share a common property—they can always produce an answer. There is no mechanism that determines whether an answer is allowed to be asserted in the first place.
In scientific practice, that constraint is fundamental. Hypotheses are not just generated—they are exposed to reality, filtered, and only retained if they survive. Critically, the system must also be able to withhold assertion when evidence is insufficient.
That full loop—exposure, evidence accumulation, survival filtering, and abstention—is what’s missing from both paradigms.
LLMs fail because they interpolate without constraint.
Symbolic systems fail because they assert without exposure.
Adding symbols doesn’t fix admissibility.
The gap is not reasoning machinery—it’s the absence of a mechanism that allows reality to reject structure before it is asserted.
In that sense, what’s needed is not just a hybrid architecture, but something closer to an executable version of the scientific method itself—where only structures that survive repeated exposure are permitted to participate in inference.
That shift—toward evidence-licensed assertion—may be the difference between systems that appear to reason and systems that can be trusted in high-stakes domains.
LLM copypasta.
Oh , AI is executing the scientific method, all right, just not in the way a computer executes but like a hangman does.
Just ask the editors of scientific journals who are now being inundated with AI generated “science”.
Thanks, this is excellent thinking. My biggest fear about AI is that the entire environment is essentially a big answering machine instead of a big questioning machine. Most people have a terrible grasp over what facts and answers are, so it’s hardly surprising that machines fare no better!
What I've learned over the past few years is this: the men who are hellbent on forcing an AI-centric future on all of us are mad, bad, and dangerous to share a planet with. In light of this, do we want neurosymbolic AI? Apart from machine learning technologies specifically developed for narrow purposes with an obvious potential to benefit humanity (e.g., medical research), do we want any sort of AI at all?
In light of the deficiencies in LLM machines, should that influence financing for the data centers now being built across the U.S.? Are investors pouring billions of dollars into data centers that will simply not do what the likes of Sam Altman proclaim they can do?
I 100% agree with the analysis and hybrid direction, but aren't there already public LLMs that generate, test, and execute code in order to answer such formal questions?
Without alignment though, surely we should pause.
Never in the history of any subject has something that is so touted yielded so little utility, than neurosymbolic AI. Sure, one can ascribe a lookup table being scaffolded to an LLM as some kind of neurosymbolic system (though this is highly technically questionable), but the response to this is always the same. If it’s so effective, so power efficient, and so easily done, then where is it?
Needs to go flat ass broke first. The state of mind of hypers and hypees is very far from "let's start over". And even if it wasn't, they would never be able to know what "start over" even means.
I sent a copy of an AI qualifier printed at the bottom of an online science magazine (Science X Newsletter) to an academic friend in Denmark and realized as I was sending it that I will have to add a note to all of my online papers (academia.edu) that were written before the AI revolution that there was no AI writing in them. Think of any book you have on your shelves--or in any library--we really have entered a new world so different from the one before.
Markets have been pricing LLMs like they found the unified theory of everything. A 95% vs 34% success rate on a problem humans solve with a napkin and twenty minutes suggests the unified theory still has a few pages missing. The trillion dollars chasing scale may eventually hit the same wall the Tower of Hanoi keeps building
The Tufts result tracks with what DeepMind's AlphaGeometry team quietly conceded in 2024: their IMO-level solver was ~53% pure LLM, but the symbolic deduction engine did the actual proving — the neural net just proposed auxiliary constructions. Every headline "LLM solves math" system in the last two years has a classical solver hiding in the basement. The neurosymbolic future isn't coming; it already shipped, just with the symbolic half airbrushed out of the press release.
Thanks for sharing! It seems neurosymbolic module primarily acts as a temporal scaffold for the LLM?
In brains, all neurons have endogenous clocks, and evolutionarily we know this came about as multicellular life evolved, and before cells specialized into “neurons” to handle some computations. All neurons thus do their thing against a robust circadian clock that each neuron maintains and synchronizes to external time signals.
As I understand it, LLMs are stateless between tasks, and while a lot of complexity is encoded in its structure, its dynamics are severely constrained.
To me, this absence of “time sense” at a fundamental level seems to be the reason LLMs can’t reason or plan well. We try to bolt on physical time on top, but that’s pretty inefficient and can mean you “compute” something that’s no longer relevant to the changed environment you’re computing in.