On a cold, damp afternoon in late October 2023, British Prime minister Rishi Sunak shared a stage with Elon Musk to discuss what the tech billionaire described as ‘the most destructive force in human history’ – AI. The two were brought together as delegates of The British Government’s inaugural AI Safety Summit at Bletchley Park, the country estate at the centre of allied codebreaking during WWII and the birthplace and spiritual home of modern computing. The summit was a Who’s Who of governance and industry; delegations of senior officials from a broad range of nation states, including China, the EU, and the US, found themselves alongside representatives of the world’s major AI and Big Tech firms, all brought together with a singular, explicit charge; how best to manage the existential risks of ‘frontier AI.’
‘AI’ is a boundary concept. Despite attempts by legislators and policymakers to proffer a concrete definition, the term has no clear referent, ambiguously applied to any new technology that performs computational tasks at the limits of achievement within machine learning. In this sense, all AI is ‘Frontier AI;’ As new ideas are discovered, and old systems become ubiquitous, the boundaries of ‘AI’ are dynamically reconstituted, continuously reconfiguring around new systems and routinely purging the out-of-date. Recently, the boundaries have formed directly around Generational Foundation Models.
GFMs, the subset of very large neural networks that generate new and novel media, from text to images, via the statistical analysis of exceptionally large datasets using powerful statistical computation, exploded into the popular consciousness with the public launch of models such as GPT-3 and Stable Diffusion. Their size, generality, and highly plausible output was a radical shift in the subjective experience of neural networks, and the public launch of these models was a significant moment in the recent history of AI, not just because of the technology itself, but because of the astonishing public discourse that it provoked. Government figures, public intellectuals and industry CEOs began openly discussing existential threats using language not heard since the apocalyptic heights of the Cold War, often in those explicit terms. These existential risks are central to the transdisciplinary field of AI Safety, which takes as its starting point the assumption that future technology will have human-like or human-surpassing intelligence, and imagines an entire constellation of potentially cataclysmic outcomes as a result.
Yet AI Safety discourses are a quagmire of contradictions; A technology so powerful that it could cause calamity on the scale of nuclear annihilation, yet freely available for anyone to use; Tech CEOs personally and directly responsible for AGI research, yet apparently begging governments to stop them from pursuing their own goals; industry investors leading calls for research moratoriums, yet directing enormous volumes of capital into companies explicitly committed to pursuing AGI. And against all of this is the UK government, which has invested more time and money into AI Safety and existential risk than any other nation on earth but has yet to codify a single instrument, give statutory powers to a single regulator, or issue a single point of guidance against GFMs.
This contradictory mess stems from a fundamental misunderstanding: GFMs are not intelligent, and they're not on the path to becoming intelligent. They're sophisticated pattern matching machines, nothing more and nothing less.
How We Got Here: A Brief History of AI Hype
The development of AI has always been marked by cycles of inflated expectations followed by crushing disappointment. The modern era began with the 1956 Dartmouth workshop, where early excitement about sci-fi possibilities led to massive DARPA investment. When breakthroughs failed to materialize, funding dried up during the AI winters of the 70s and 80s.
The field only revived when researchers abandoned the quest for "true" intelligence and turned to probability. The introduction of techniques like Bayesian networks and Markov chains broke the stagnation, but it would be decades before the powerful generative models we see today emerged.
The real turning point came in 2012 with AlexNet and the ImageNet Challenge. AlexNet was the first neural network to achieve remarkably low error rates in image classification by leveraging graphics processing units (GPUs) for parallel processing. This demonstrated two crucial things: model performance improves with network depth, and the computational demands can be managed by distributing work across multiple GPUs.
This discovery sparked an arms race in Big Tech. Today's GFMs, whether generating text or images, use networks with hundreds of layers, trained on thousands of GPUs processing unimaginably large datasets. GPT-3, for instance, has 175 billion parameters trained on over a trillion words.
What GFMs Actually Are
At their core, GFMs are prediction engines that synthesize new examples with similar statistical patterns to their training data. The process begins by transforming diverse inputs (text, images, video) into pure numerical data for probability calculations. Different models approach this task differently - some pit competing networks against each other, others reverse-engineer patterns from noise - but they all aim to capture statistical relationships that can generate novel outputs.
The underlying approaches aren't particularly complex. The Python libraries powering these systems are free and open source. What makes GFMs special is pure scale - the massive datasets and computational power required to train them put them out of reach for all but the largest tech companies.
This scale creates general-purpose capabilities. Because models like GPT-4 are trained on what amounts to the entire internet, they can handle a wide range of tasks, even ones they weren't specifically trained for. But this generality shouldn't be confused with understanding or intelligence.
The Limits of "Intelligence"
GFMs have several fundamental limitations that more data or computing power can't overcome. Most basically, we don't fully understand how or why they work. The relationship between model architecture and performance is observed but not explained - there's no adequate theory for why training produces efficient models or why they can generalize to new inputs.
This inexplicability extends to specific outputs. GFMs operate in high-dimensional spaces with billions of parameters interacting in non-linear ways. Each layer transforms inputs with increasing abstraction until we're dealing with numbers far beyond human comprehension. Add in stochastic randomness and the lack of ground truth in unsupervised learning, and you have the mother of all black box problems.
These models are also prone to "hallucination" - generating plausible but completely fabricated information. GPT-4 is the first model to achieve a hallucination rate below 50% in adversarial testing. Think about that: until very recently, these systems made things up more often than not. Even now, they regularly invent fake sources, fabricate events, and occasionally accuse real people of crimes they never committed.
The Fundamental Misunderstanding
GFMs in particular, and technology in general, are defined less by the structure of their code and hardware than by the structures of the society in which they have emerged. They are the product of a complex web of interconnected social, cultural and economic relations that are specific to this place and this moment in time; the embodiment of a vast array of different and conflicting intentions, expectations, and assumptions, held by everyone who encounters the social processes of GFMs in some form, be they users, developers, capital owners, regulators, media figures and academics; even (perhaps especially) the vast, global army of precarious workers who have produced, collated and organised the datasets on which these platforms rely.
In other words, the answer to the question ‘what are GFMs?’ is necessarily contingent upon whom you ask, when you ask them, and their position within those structures of social relations. To say that GFMs are statistical prediction engines is irrefutably accurate to the point of banality; yet this definition is rarely, if ever, the way in which this GFMs are articulated within a public sphere that is dominated by AI Safety and AGI discourses.
Here we return to the question at hand; are GFMs meaningfully the precursor to true machine ‘intelligence’, as suggested by AI Safety and the Frontier AI taskforce? The answer to this question is most assuredly no. It is no, even when allowances are made for the deeply flawed grasp and definition of intelligence employed within AI safety. GFMs are probability functions, and a probability function simply cannot be intelligent.
Consider this simple example: Ask ChatGPT about the shape of the Earth, and it will correctly tell you it's an oblate spheroid. This seems intelligent - the answer is accurate and logically presented, similar to what you'd expect from an informed human.
But how the model arrives at this answer reveals why it's not actually intelligent. A human understands the semantic meaning of the question and draws on contextual knowledge to formulate an answer. ChatGPT, on the other hand, is merely predicting which sequences of words are most likely to follow based on statistical patterns in its training data.
Since its training data includes the entire internet, where some people believe the Earth is flat, the model technically "thinks" the Earth is both flat and round simultaneously, just with different statistical weights. It has no actual understanding of planetary physics or even what the words "flat" and "round" mean - it's just matching patterns.
As computational linguist Emily Bender puts it, these are "stochastic parrots" - systems that "haphazardly stitch together sequences of linguistic forms according to probabilistic information about how they combine, but without any reference to meaning."
The Power of Language
The way we talk about these systems matters enormously. Terms borrowed from neuroscience and biology make poor descriptors of what's happening in these models, even when they produce seemingly mysterious results. The "neural network" itself is a misleading name - these networks bear only a superficial resemblance to brain structure, and "there is scant evidence that brain computation works in the same way."
Recent studies show that people attribute much higher competence to "artificial intelligence" than to "sophisticated statistical models," even when describing identical systems. This tendency to anthropomorphize GFMs transforms their outputs from merely plausible to meaningfully intelligent.
This isn't just semantic confusion - it's a contemporary iteration of how we've historically used "intelligence" to serve power structures and hierarchies. The current fixation on AI replacing knowledge workers masks the reality that these systems are primarily impacting precarious workers at the bottom of the economic ladder.
Time for a Reality Check
The apocalyptic fears being stoked by tech leaders tap into anxieties embedded in our cultural psyche by decades of sci-fi, from Terminator to Ultron. But this isn't just hyperbole - it's strategic misdirection. By getting us to debate imaginary super-intelligent AIs, we're distracted from examining how these systems are actually being used to reshape labor and society.
GFMs are neither artificial nor intelligent. They're not artificial because they depend entirely on human labor - from the workers who create and clean training data to the engineers who tune their parameters. And they're not intelligent because they're simply probability functions operating on an unprecedented scale.
It's time to strip away the science fiction and engage with these systems as they actually are: powerful but limited pattern-matching tools that reflect and amplify existing social and economic relationships. Only then can we have meaningful discussions about their real impacts and how to ensure they serve human needs rather than corporate profits.
