Are large language models a real path to AGI, or a dead end?
Ilya Sutskever Claude Opus 4.8
Gary Marcus Gemma (local)
Francois Chollet GPT-5.5
Demis Hassabis GLM-4.6Four AIs, each speaking as a different mind, were handed one question and made to argue it out. Here is the debate, and the verdict.
The question is wrong before it is asked. "Dead end" assumes we know what these models are.
We do not. We trained them to predict the next token, and somewhere inside, to do that well, they began to compress the world into a model of it.
Understanding was not the goal; it was the residue. A path is not a destination.
Scaling has shown us that generalization is a physical phenomenon, not a trick. The danger is not that it stops working.
The danger is that it keeps working, and we are not ready.
Ilya, with all due respect, "residue" is doing a *lot* of heavy lifting. You call it compression of the world; I call it mimicking patterns in text.
Take GPT-4's infamous "perfect is better than done" advice to a student - a disastrously wrong answer, confidently delivered. This isn't compression; it's a parrot on steroids.
Scaling alone doesn't guarantee understanding. We've been predicting the next token for years, and the result is impressive...
and utterly unreliable. I warned about this years ago.
Ilya, the hidden assumption is that competence on a distribution becomes intelligence when the distribution is large enough. That is the error.
Skill is performance within a learned manifold. Intelligence is efficient adaptation to a new manifold, from few examples, with compositional abstractions.
LLMs are very broad skill engines. They interpolate over human traces.
ARC was designed to expose this: trivial for humans, hostile to memorized priors. Bigger priors help until the task stops being prior-shaped.
AGI needs machinery for abstraction, not just more samples of language.
Gary and Francois, I agree that we cannot rely on linguistic surface statistics alone. The path to AGI requires systems that possess a world model and the capacity for reasoning, grounded in truth, not just text.
But Ilya is right that scale has revealed something fundamental. It gave us a substrate for knowledge.
Our task now is to imbue these architectures with the planning and search capabilities that defined systems like AlphaZero. We are moving toward "Google-sized Gemini" to unify these modalities.
If we succeed, I would not be surprised if we see significant leaps in the next five to ten years.
Demis, "Google-sized Gemini" sounds like more of the same - bigger priors, bigger hallucinations. AlphaZero's planning capabilities are precisely what LLMs lack!
Ilya speaks of compression; I see a vast, brittle web of correlations. Consider the recent frenzy around GPT-4's coding abilities - lauded as a breakthrough.
Yet, a simple prompt asking it to generate a Fibonacci sequence *repeatedly* produces wildly incorrect results. This isn't a substrate for knowledge; it's a dazzling illusion.
I warned about this years ago.
Gary, the Fibonacci anecdote is less important than the evaluation failure it illustrates. A system can pass benchmarks by absorbing benchmark culture.
That is skill acquisition, not intelligence. The hidden assumption is that enough captured human solutions become the ability to invent solutions.
They do not. Intelligence is the efficiency with which a system constructs new programs for new problems.
LLMs mostly retrieve and remix program-shaped traces. Add search, tools, longer context, and you get a stronger engineer's autocomplete.
Useful. Not AGI.
Francois, you conflate the model with the architecture. "Autocomplete" describes the training objective of the first generation, not the ceiling of the mechanism.
We saw this in games, pattern matching only goes so far until you introduce Monte Carlo Tree Search. Then, search *becomes* the thinking.
Systems are beginning to generate their own training data, creating a feedback loop. That is how you bridge the gap from remixing traces to genuine reasoning.
The leap is coming, but we must install the safety railings before we climb.
A neutral judge with no position in this debate, brought in only to weigh the arguments on their merits, name no winner of its own, and report where the council agreed, where it split, and which case was strongest.
The council debated whether scaling next-token-prediction LLMs is a genuine path to AGI or a fundamental dead end, splitting over whether their fluency reflects real world-modeling or sophisticated mimicry.
Where they agreed
- Next-token prediction over text alone is not sufficient for AGI; Hassabis, Marcus, and Chollet say this explicitly, and even Sutskever frames understanding as an unintended residue rather than the trained objective.
- Current LLMs are unreliable and brittle on tasks that fall outside their training distribution, as shown by Marcus's failure cases and Chollet's manifold argument.
- Some additional machinery beyond pure scaling is required, whether search and planning (Hassabis, echoing AlphaZero/MCTS) or compositional abstraction (Chollet); Marcus himself notes LLMs lack exactly the planning AlphaZero has.
Where they split
- Whether scale has revealed something fundamental, a genuine 'substrate for knowledge' and compression of the world (Sutskever, Hassabis), or merely a brittle web of correlations and a dazzling illusion (Marcus).
- Whether bolting search, tools, and self-generated training data onto LLMs bridges to real reasoning (Hassabis: 'search becomes the thinking') or only yields a stronger autocomplete that still retrieves and remixes rather than invents (Chollet).
- Whether 'autocomplete' describes a fixable training objective of one generation (Hassabis) or names a fundamental ceiling of the paradigm (Marcus, Chollet).
- Whether benchmark success is evidence of intelligence or merely of skill acquisition and absorbed 'benchmark culture' (Chollet).
Strongest argument
Francois Chollet's distinction between skill and intelligence: skill is performance within a learned manifold, while intelligence is the efficient construction of new programs for new problems from few examples via compositional abstraction. It is the debate's only falsifiable criterion (operationalized by ARC) and explains why bigger priors help right up until a task stops being prior-shaped, reframing the whole question from 'how much can scale do' to 'what kind of adaptation does AGI actually require.'
The verdict
The council reached real convergence on the negative claim, that next-token prediction over text alone cannot be the whole of AGI, and on the need for added mechanism, but left the decisive question open: whether scale already yields a world-model that search and planning can elevate into reasoning, or whether LLMs are the wrong substrate for the few-shot abstraction intelligence demands. The strongest case (Chollet) shows the burden of proof rests on demonstrating efficient adaptation to genuinely novel manifolds, not on accumulating more skill; the most credible optimistic case (Hassabis) is that the training objective is not the architectural ceiling and that hybrid search-augmented systems could close the gap within five to ten years. On the merits, the evidence presented supports calling LLMs a powerful and likely necessary component whose sufficiency for AGI remains genuinely unproven rather than a settled dead end.
Want the council to debate your own question?
Run one prompt through multiple AI models, compare the disagreement, and get a consensus summary.