Pathways to advancing AI
TL;DR: We simply don’t know if or when generative AI, an alternative approach, or some combination of both, will result in an artificial general or super intelligence that matches and exceeds all human intelligence. There is currently no universal consensus across experts in the field about how or whether we get there.
An article in the Wall Street Journal last week (14 November 2025), ‘He’s Been Right About AI for 40 Years. Now He Thinks Everyone Is Wrong,’ triggered another round of headlines about future advances in AI. It was an interview with Yann LeCun, one of the most significant contributors to this modern era of AI.
The article claimed most in the field of AI believe large language models will scale up and achieve artificial general intelligence (AGI) whilst LeCun thinks not.
Yet plenty of significant contributors in the field of AI have also voiced the need for breakthroughs beyond generative AI to make further progress. And generative AI is not just about large language models.
This post is an attempt to capture some of the different viewpoints.
The arrival of generative AI
AI as a term has been around since the 1950s, as have the two core approaches for building it - applying rules-based logic and teaching machines to learn.
The earlier advances in AI focused on applying rules-based logic, creating codified knowledge to build expert systems. Artificial neural networks were proposed as a form of machine learning but discarded by most as they failed to yield good results. However, the expert systems proved to be to brittle and failed to scale.
Fast forward to the 2000s, and new approaches to neural networks, nicknamed deep learning, combined with the availability of parallel compute using GPUs and an abundance of training data, meant everything changed. Every major breakthrough since 2012 has been based on deep learning using massive data sets.
These algorithms are referred to as generative AI because they learn how to generate novel outputs without explicit instructions. Whilst large language models are the most talked about, this applies to algorithms such as AlphaGo, that generated its own strategies to win at the game of Go, and AlphaFold, that learned how to fold proteins.
The question many are now asking is, what’s next? Are we on a direct path to achieving artificial general intelligence, or are further breakthroughs needed?
The image below is a simplification of the current approaches to AI.
The scaling hypothesis
The breakthrough with deep learning over classical machine learning (ML) was the scaling of training data. Classical ML algorithms outperform traditional statistical methods but reach a limit where more data no longer improves the model. With neural networks, so far, no limit has been found. As shown in the image below, deep learning specifically requires a lot of data. If you don’t have enough, a classical ML algorithm, or computational statistics, will perform better.
To give an indication of how much scaling occurred in just over a decade:
2012: AlexNet trained on 1.2 million images using 470 PFLOPs
2016: AlphaGo trained on 30 million moves using 1.9 million PFLOPs
2020: GPT-3 trained on approx. 200 billion words using 314 million PFLOPs
2023: GPT-4 trained on approx. 9 trillion words using 21 billion PFLOPs
PFLOP refers to one quadrillion (peta) floating-point computations.
The scaling hypothesis is that if we continue to scale the model size, data, and compute, along with minor algorithmic tweaks, then generative algorithms will eventually display general intelligence across most tasks.
The main proponents of this hypothesis have been people leading the companies building two of the most successful large language models - Sam Altman (OpenAI GPT) and Dario Amodei (Anthropic Claude). Ilya Sutskever, cofounder of OpenAI and coauthor of some of the key research contributions to generative AI, also supported it but has recently questioned whether we have enough data to continue scaling.
Two significant contributors to the field - Geoffrey Hinton and Yoshua Bengio - also support deep learning as the likely route to more powerful AI systems, and have been vocal in raising concerns that the current trajectory of AI could pose existential risks.
The hybrid ‘neuro-symbolic’ hypothesis
Interact with a large language model for a period of time and you will become familiar with its limitations. It often struggles with context and can make seemingly idiotic factual errors. This suggests that, whilst LLMs appear to have mastered much of human language, and are able to recombine information based on their training data which now seems to include most of the Internet, they lack other cognitive abilities that are the foundation of human intelligence.
Neuro-symbolic AI proposes to combine neural networks with logic-based inference - explicit rules and structured knowledge that would enable abstract reasoning and factual accuracy.
People proposing this approach include Gary Marcus and Judea Pearl. Pearl has been leading research in causal inference. Gary Marcus has been vocal in his criticism of large language models as a route to artificial general intelligence.
The world-model hypothesis
Similar to neuro-symbolic AI, the world-model hypothesis believes more is needed than just learning and generalising from massive amounts of data. That LLMs can interpolate from data, but not extrapolate. The core idea is that intelligence requires some form of internalised world model to be able to perceive, understand, and adapt to previously unseen situations. Under this hypothesis, an LLM will not achieve general intelligence without also having a world-model architecture.
People proposing this approach include Yann LeCun (founder of FAIR at Meta) and Demis Hassabis (cofounder of Google DeepMind). A counterargument is that LLMs have somehow created their own internal world model, or will in the future.
The embodied intelligence hypothesis
A further strand of thinking argues that you cannot get to general intelligence from disembodied systems. Under this view, intelligence emerges from being an agent in a world, perceiving, acting, getting feedback, and learning over time.
Embodied AI focuses on systems that learn through sensorimotor interaction, whether in physical robots or rich simulated environments. The claim is that many of the things we call “common sense” and “understanding” arise from this ongoing perception–action loop, not from language alone. On this view, large language models might become important components, but AGI would require agents that can move, manipulate, and experiment in their environment, developing skills over a lifetime rather than just being trained once on a static corpus.
People associated with this line of thinking include Rodney Brooks and the broader embodied / developmental robotics community.
The stochastic parrot argument
The substantial interest in AI has arisen from the breakthroughs with generative AI. Interacting with large language models can feel like being in a science fiction movie. It really is hard to overstate just how unexpected and exciting this development is. But it may prove to be premature to assume it has put us on a path to AGI.
A 2021 paper, ‘On the dangers of stochastic parrots’, challenged much of the excitement about large language models. The authors raised concerns that people were overstating the abilities of the models. The argument is that the models are just generating plausible-looking outputs, and carry significant near-term risks due to biases in the training data, potential for misuse, and environmental costs from building ever larger models.
People raising this concern include Emily Bender, Timnit Gebru, Margaret Mitchell and Melanie Mitchell. The argument is not about proposing an alternative or hybrid architecture to achieve AGI, and instead highlighting what’s missing in LLMs and what risks are created if we over-rely on flawed artificial intelligence.
Closing thoughts
There are many other approaches also being researched. The key message is that there is no universal consensus that LLMs are a route to AGI. Yann LeCun is not standing alone in his criticism of LLMs. It seems most people think something more is needed.
There is one critical aspect not covered by any of these hypotheses. What do we actually mean by general intelligence? Do humans even have it? Or is our intelligence just highly contextual and elastic? Is intelligence all that you need? There are other species on this planet that have superior sensory and navigation capabilities.
The classic line often used in movies involving an alien invasion is, ‘the most technologically superior species always wins.’ That doesn’t mean they are necessarily the most intelligent…
References
Computation used to train notable AI algorithms, Our World in Data
ImageNet Classification with Deep Convolutional Neural Networks, Krizhevsky, Sutskever and Hinton, 2012 (first mention of improvement through scaling)
Scaling laws for neural language models, OpenAI, 2020
The Scaling Hypothesis, Gwern Branwen, 2020
The Next Decade in AI, Gary Marcus, 2020
On the dangers of stochastic parrots, Emily Bender et al, 2021
Why neural net pioneer Geoffrey Hinton is sounding the alarm on AI, MIT, 2023
Embodied AI, Habitat, 2023
What Ilya Saw, Latent Space, 2024
Is deep learning actually hitting a wall? Garrison Lovely, 2024
Machines of Loving Grace, Dario Amodei, 2024
Neuro-symbolic AI, Greg Robison, 2025
Introducing the V-JEPA 2 World Model, Meta, 2025
Genie 3: A new frontier for world models, Google DeepMind, 2025
Embodied AI Agents, Pascale Fung et al, 2025
The Grand AGI Delusion, Gary Marcus, 2025
He’s been right about AI for 40 years, Washington Post, 2025
See related posts: The future of intelligence; A modern AI timeline.
Minor edits: adding to the references.