Last week I wrote about the fundamental differences in the way that large language models and humans learn, and I noted that many of these differences speak to some integral limitations of text-based language models. This is particularly true when thinking about where AI may go from here and the potential of LLMs to form the basis for AGI. Yann LeCun (one of the so-called ‘godfathers’ of AI) got plenty of attention when he resigned from Meta recently, stating that Meta’s huge focus on LLMs was ultimately a dead end, and that the future lay in developing ‘World models’ that learned more as humans do, notably from visual learning. I should acknowledge before I go any further that LLMs are phenomenal tools. But LeCun’s reasons for wanting to take a divergent path on the future of AI led to me wanting to understand (in a non-technical way) why current models may have their limitations.
Let me remind you briefly of how LLMs and humans learn in fundamentally different ways. While LLMs excel at inductive reasoning, generalising from massive training datasets to recognise patterns and predict probable outcomes, humans use abductive reasoning to make creative leaps from incomplete information. Humans can learn efficiently from single examples through embodied, multisensory experience, whereas LLMs require hundreds of thousands of labeled instances and process only text-based, explicit knowledge (missing the 95% of human knowledge that’s tacit and experiential). Human memory deliberately degrades and updates, enabling adaptive generalisation, while LLMs’ perfect retention creates brittleness when facing novelty. Perhaps most critically, human intelligence emerges from complex evolutionary goals, intrinsic curiosity, and social learning within groups, whereas LLMs have only one objective (predicting the next token) with no genuine motivation, persistent goals, or capacity for collaborative intelligence. These are not mere performance gaps but are instead architectural differences that are rooted in how each system fundamentally works.
So my thought here is that these architectural differences will ultimately mean that we need a significantly different approach to get anywhere near AGI. Much of the (not-insignificant) hype around LLMs is based on the assumption that they can effectively develop in a straight line from where we are now to AGI. Yet we seem to be forgetting that language is not the same as intelligence. Probabilistic reasoning is not the same as intelligence. Mimicking human output without fundamental understanding is not the same as intelligence.
Technologist Mike Brock notes in his recent essay that whilst the current LLM-based approaches are amazingly useful tools, the claims that there is a straight-line path from where we are at to AGI are mistaken:
‘The AGI-from-LLMs thesis fails not because it’s too ambitious, but because it’s the wrong kind of ambition – an attempt to engineer a solution to what is fundamentally a philosophical confusion about the nature of intelligence itself.’
The constraints aren’t just engineering challenges, he says, they point to a fundamental misunderstanding about what kind of problem we’re trying to solve. Those that are hyping the future capabilities of LLMs have redefined the question from the original one, “How do we create systems that can understand, reason, and adapt the way humans do?”, to a new one, “How do we scale pattern-matching until it exhibits behaviors that look like understanding, reasoning, and adaptation?”. LLMs, Brock notes, separate learning and inference completely. Meaning that they learn once from historical data and this learning is then locked in place. They cannot observe and update their understanding based on what they are encountering now. They cannot hold multiple layers of meaning simultaneously. They cannot modify their understanding or processing strategies in the same way that human intelligence does continuously.
The fact that at their heart they are language models is, as Benjamin Riley said recently in The Verge, a fundamental limitation in itself (HT Tom Hopkins for the link):
‘The problem is that according to current neuroscience, human thinking is largely independent of human language…We use language to think, but that does not make language the same as thought.’
In other words, emulating the function of language is not the same as going through the cognitive process of reasoning which can then be expressed through language. In humans these functions literally use different parts of the brain. Put simply: ‘Our cognition improves because of language, but it’s not created or defined by it’. Humans develop intelligence through physical interaction with their environment which creates a grounded understanding. LLMs lack embodiment and operate solely in the symbolic domain of language which ultimately means that they can manipulate symbols (or tokens) but struggle to connect them to the real-world.
Given all this, it’s clear that we need something else if we want to take another leap forwards and ultimately if we want to achieve AGI (which is a big if). A 2025 AAAI report found that 76% of AI researchers believe that scaling up current AI approaches to achieve AGI is unlikely or very unlikely to succeed. Evidence demonstrates that simply scaling current approaches will lead to a plateauing of capability and diminishing returns from simply adding more data and computing power. Most researchers seem to agree that LLMs provide valuable components but they cannot achieve AGI alone. And there is growing consensus that AGI will not emerge from a single monolithic architecture but will instead necessitate ‘hybrid systems’ that integrate a diverse set of capabilities.
This could well mean that the path to AGI will be longer than many people seem to think. The most plausible path forwards appears to be evolutionary development of LLMs complemented by revolutionary leaps forward that address core limitations. Rather than evolving directly into AGI, LLMs will likely serve as sophisticated language processing models within larger hybrid systems. These systems will need to incorporate alternative architectures and capabilities including (but not limited to) embodied intelligence that can learn actively from its environment, understand causal effects, developed contextual understanding and adaptability, and the ability to update its learning continuously.
The path to AGI is likely a longer and more winding one than many have predicted which is probably no bad thing given how fast things have moved in the last few years.
A version of this post appeared on my weekly Substack of AI and digital trends, and transformation insights. To join our community of over thirteen thousand subscribers you can sign up to that here.
To get posts like this delivered straight to your inbox, drop your email into the box below.
Photo by Kajetan Sumila on Unsplash
Leave a Reply