Superintelligence: so near, and yet… so far
What if the recent market correction in AI-related stocks may be more than just a correction? What if the market volatility of the so-called “hyperscalers” – Meta, Amazon, Alphabet, Microsoft and Oracle – are the first whisperings of a more permanent and disturbing realisation regarding the entire AI hyperbole? A realisation that while the use of brute-force systems – made up of quadrillions of integrated circuits using backpropagation and other fandangled mathematical methods – to attempt to mimic the human brain, may provide us with some useful machinery, it may not, in fact, yield what is touted as the imminent horizon moment of “superintelligence”.
The Turing Award, begun in 1966 is often described as the “Nobel Prize of Computing”. It is named in honour of British mathematician Alan Turing, whose groundbreaking work in cryptography during World War II and foundational contributions to computer science and artificial intelligence transformed the course of modern technology. In 2018, the award was bestowed upon French-American computer scientist Yann LeCun in recognition of his seminal role in establishing deep neural networks as a cornerstone of contemporary computing. LeCun’s influential work began in the 1980s at AT&T Bell Labs, where he led a research team responsible for pioneering the application of backpropagation algorithms to convolutional neural networks for image recognition tasks, laying the technical groundwork for the progress in machine learning that followed.
In 2013, LeCun was hired by Facebook, now Meta, to head its newly formed Facebook AI Research (Fair) laboratory. But after more than a decade in this role, he and Meta CEO Mark Zuckerberg have reached a fundamental divergence in their visions for the company’s future. Rather than pursuing the large language models (LLMs) that form the centrepiece of most Big Tech companies’ AI strategies, LeCun – having long argued that these models cannot achieve general artificial intelligence – is emphasising the development of alternative AI systems known as “world models”. According to LeCun, these emerging models will require another decade of research and refinement before becoming truly viable – a timeline that does not align with Zuckerberg’s ambitions for more rapid progress toward superintelligence.
As Meta strives to keep pace with its competitors, Zuckerberg has initiated an urgent push to accelerate the rollout of new AI-driven products, following the underwhelming debut of Meta’s Llama 4 model, which was surpassed by offerings from Google, OpenAI, and Anthropic, as well as the muted consumer response to Meta’s AI chatbot. In his drive to accelerate Meta’s AI ambitions and make his company “the leading AI company in the world”, Zuckerberg signalled late last month that Meta’s AI spending will continue to surge into next year, potentially exceeding $100 billion in 2026. This spending will be funnelled not into Meta’s existing Fair lab but into the company’s newly formed “superintelligence” group, which will be tasked with advancing Meta’s LLMs.
A central component of Zuckerberg’s strategy is the high-profile recruitment of Alexandr Wang, the founder of the data-labelling firm Scale AI, to head Meta’s superintelligence team, in a move that cost Meta $14.3 billion. Wang is a maths prodigy who co-founded Scale at the age of 19 but is not widely recognised as a technical AI expert. He is, however, known to be well-connected within the industry, which Zuckerberg is hoping will help to attract top talent to Meta.
As Meta intensifies its focus on the development of commercial AI products, the resulting corporate reshuffle has placed the seasoned LeCun in the position of reporting to 28-year-old Wang. In response, LeCun has announced his departure from Meta to launch his own start-up. His exit reflects, however, a more fundamental fracturing that has begun to occur in the AI narrative – one that had, until recently, appeared to be largely unified in its conviction that large language models were the primary path toward achieving artificial general intelligence.
LLMs aim to mimic the human capability to produce grammatical language, by breaking language down to its smallest components, into words or parts of words, which are called tokens. These tokens are then given “meaning” by vectorising them through immense textual training to create, for each token, a vector of its particular mathematical relationship with other words. The introduction of the transformer architecture by Google researchers in 2017 dramatically improved this process, enabling these models to analyse entire sequences of words to capture complex patterns, dependencies, and relationships across longer stretches of text. This breakthrough ultimately led LLMs to being able to accurately generate text themselves.
LeCun argues that LLMs are limited in their “thinking” ability, relying on human intervention to be trained to reach certain conclusions, rather than reasoning in the way humans do. LLMs, though useful, he argues, may not be the key to achieving artificial general intelligence, whereas world models learn not only from language but from videos and spatial data as well with a view to creating AI with a more comprehensive understanding of the world. LeCun is not alone in his analysis of LLMs, with Google DeepMind and Nvidia also investigating alternative AI methods.
The fracturing of the AI narrative is also evident in the massive shudder that recently reverberated through the markets as developments in LLMs have slowed and investors have begun to suspect that there may, in fact, be significant overinvestment in data centres to serve AI purposes. This year alone, the “hyperscalers” have laid out plans with a combined capex spend of $390 billion, with an additional $1 trillion expected over the next two years. OpenAI has meanwhile indicated a spend of $1.4 trillion on AI infrastructure over the next eight years.
Over the past month, the market has suddenly turned and become uncomfortable with this wave of multibillion-dollar pledges and agreements. Oracle alone has lost $250 billion in market value, with investors particularly uneasy about the mounting debt burden tied to the company’s ambitions to become a leading AI infrastructure provider. On aggregate, investors are beginning to reassess the once-unshakeable optimism underpinning the AI narrative that has helped propel confidence in tech stocks to record highs. The S&P 500 has been trading at a price-to-earnings ratio of roughly 40, far above its long-term average of around 20, underscoring how overvalued Big Tech firms have become.
Concern over hyperscalers’ spending on AI has also spilled into the US bond market this month, with the spread that investors demand to buy their debt over Treasuries climbing to 0.78 percentage points on 11 November, up from 0.5 points in September. The word “bubble” has entered the lexicon among stock market analysts who are increasingly asking whether the AI business model can generate returns commensurate with the enormous capital flowing into it, and what levels of performance will be required over the next three to five years to justify such staggering investment.
Once the market realises that Yann LeCun may be right – that the methodologies driving today’s AI race contain fundamental flaws, and that the vast investment in computing power for large language models may ultimately deliver underwhelming results – then the recent downturn may prove to be only the beginning. Such a realisation could auger in not just a correction like the one seen over the past month, but an almighty crash.
By David Buckham
David Buckham is the Founder and CEO of international consultancy Monocle Solutions. He is co-author, alongside Robin Wilkinson, of the recently released The Spell: A Story of Human Progress and How the West Lost its Soul.
Loading ...