The Quiet Acceleration

The Quiet Acceleration

There is a particular kind of silence that settles over a room when somebody who works inside a frontier artificial intelligence laboratory is asked, off the record, how worried they actually are. It is not the silence of someone searching for an answer. It is the silence of someone deciding how much of the answer they are allowed to give.

Over the past eighteen months, that silence has grown noticeably longer. The reason is not difficult to identify. The systems being built behind the security badges of San Francisco, London and Hangzhou are no longer merely larger versions of what came before. They are beginning, in measurable and reproducible ways, to participate in their own improvement. The question that once belonged to science fiction - namely whether a machine could meaningfully bootstrap its own intelligence - has quietly become an engineering problem with a budget line.

The word for what comes next, if anything comes next, is singularity. It is a term most people have heard, fewer can define, and almost nobody outside the field has been given an honest account of. Polling data from the Pew Research Center, the Reuters Institute and the Tony Blair Institute for Global Change consistently shows that public understanding of artificial intelligence has not kept pace with the systems themselves. People know the chatbots. They know the image generators. They have heard, vaguely, that something called AGI is supposed to arrive at some point.

What they have not been told, in plain language, is that the laboratories building these systems have begun publishing papers in which the models help design their successors, and that some of the most senior researchers in the field now treat a recursive self-improvement loop not as a hypothetical but as a near-term operational risk.

This article is an attempt to close that gap honestly. It is neither a prophecy of doom nor a sales pitch for inevitability. It is a stocktake, conducted in April 2026, of where the technology actually sits, what the people building it actually believe, and what the average person - the one who has never read an arXiv paper and never wishes to - ought to understand about the road ahead.

What the Singularity Actually Means

The term itself was popularised by the mathematician and science fiction writer Vernor Vinge in a 1993 essay delivered at a NASA symposium, in which he predicted that the creation of entities with greater than human intelligence would mark a point beyond which human affairs as currently understood could not continue.

Ray Kurzweil, the engineer and inventor now serving as a principal researcher at Google, took the idea and gave it a calendar. In his 2005 book The Singularity Is Near, and again in his 2024 follow-up The Singularity Is Nearer, Kurzweil placed the arrival of human-level machine intelligence at 2029 and the full singularity at 2045. Those dates, once treated as fringe optimism, now sit comfortably within the public timelines published by laboratories such as OpenAI, Anthropic and Google DeepMind.

The technical core of the idea is recursive self-improvement. An artificial intelligence capable of improving its own design, even slightly, can use the improved version to design a further improvement, and so on. The mathematician I. J. Good, who worked alongside Alan Turing at Bletchley Park, described this in a 1965 paper as an intelligence explosion. Good wrote that the first ultraintelligent machine would be the last invention humanity would ever need to make, provided the machine remained docile enough to tell us how to keep it under control. The caveat has aged considerably less well than the prediction.

For most of the intervening sixty years, the scenario remained theoretical because nobody could point to a concrete mechanism by which a machine might improve itself in any meaningful sense. That changed quietly, and then suddenly.

• In 2023, Google DeepMind published a paper titled FunSearch, in which a large language model was used to discover new mathematical results by iteratively proposing and evaluating its own programs.
• In 2024, the company followed with AlphaProof and AlphaGeometry 2, which together achieved a silver medal performance at the International Mathematical Olympiad.
• In 2025, Sakana AI, a Tokyo-based laboratory founded by former Google researchers David Ha and Llion Jones, published The AI Scientist, a system that the authors described as capable of conducting end-to-end machine learning research, including generating hypotheses, writing code, running experiments and drafting papers. The papers it produced were not, by the admission of the authors themselves, brilliant. They were, however, real.

The line between a system that does research and a system that improves itself is thinner than it sounds. Machine learning research is, in large part, the activity of designing better machine learning systems. A machine that can do machine learning research is, by definition, a machine that can participate in the design of its successor. The question is no longer whether such participation is possible. The question is how much of the work the machine is doing, and how quickly that share is growing.

What Is Actually Happening Inside the Labs

In June 2025, the consultancy METR - formerly known as the Model Evaluation and Threat Research group - published a study that has become one of the most cited pieces of empirical work in the alignment community. The researchers measured the length of software engineering tasks that frontier models could complete autonomously, and tracked how that length had changed over time. Their headline finding was that the time horizon of tasks completable by leading models had been doubling approximately every seven months since 2019. Extrapolated forwards, the trend suggested that by 2027 the best models would be able to complete tasks that take a human software engineer a full working week.

That extrapolation is, of course, only an extrapolation. Trends bend. Scaling laws break. The history of artificial intelligence is littered with curves that looked exponential until they did not.

Yann LeCun, the chief AI scientist at Meta and a recipient of the 2018 Turing Award, has spent the past several years arguing publicly that current large language models are a dead end for general intelligence and that the entire architecture will need to be replaced before anything resembling human-level cognition becomes possible. He is not a marginal figure. His view is shared, in various forms, by Gary Marcus, the cognitive scientist and author, and by a substantial minority of academic researchers who consider the scaling hypothesis to be a kind of expensive mysticism.

The other side of the argument is represented most prominently by Dario Amodei, the chief executive of Anthropic, whose October 2024 essay Machines of Loving Grace laid out a timeline in which powerful AI - defined as a system smarter than a Nobel laureate across most fields - could plausibly arrive as early as 2026. Demis Hassabis, the chief executive of Google DeepMind and a co-recipient of the 2024 Nobel Prize in Chemistry for his work on AlphaFold, has placed his own estimate for artificial general intelligence at somewhere between five and ten years from the present. Sam Altman, the chief executive of OpenAI, wrote in a January 2025 blog post that his company was now confident it knew how to build AGI in the traditional sense of the term, and was beginning to turn its attention to superintelligence.

These are not idle predictions made by outsiders. They are statements made by the people who control the budgets, the compute and the hiring decisions of the laboratories actually building the systems. Whether their predictions prove correct is a separate question from whether they are acting on them. They are acting on them.

The capital expenditure figures alone make that clear. According to the International Energy Agency, global investment in data centres reached approximately five hundred billion United States dollars in 2025, with the majority of new capacity dedicated to artificial intelligence workloads. The Stargate project, announced jointly by OpenAI, Oracle and SoftBank in January 2025, committed an initial one hundred billion dollars to a single American compute build-out, with a stated ambition of reaching five hundred billion over four years. Nobody spends that kind of money on a hunch.

The Self-Improvement Loop, As It Actually Exists

It is worth being precise about what self-improvement currently means in practice, because the popular imagination tends to conflate it with the science fiction version. There is no model in any laboratory that wakes up one morning, decides it wants to be smarter, and rewrites its own weights. What there is, instead, is a growing collection of techniques in which models contribute to specific stages of the pipeline that produces their successors.

The first of these is synthetic data generation. Training a frontier model requires trillions of tokens of high quality text, and the supply of human-written text on the open internet is, for practical purposes, exhausted. Epoch AI, a research organisation that tracks the resource economics of machine learning, published a paper in 2024 estimating that the stock of public human text would be fully utilised by frontier training runs somewhere between 2026 and 2032. The response from the laboratories has been to use existing models to generate training data for the next generation. This is not a marginal practice. It is now central to how reasoning models are trained. The o1 and o3 series from OpenAI, the R1 model from DeepSeek released in January 2025, and the Claude reasoning variants from Anthropic all rely heavily on training data produced by earlier models engaged in chain-of-thought reasoning, with the better traces selected and used as fuel for the next round of training.

The second is automated machine learning research. Beyond Sakana's AI Scientist, both Google DeepMind and Anthropic have published work in which models are used to propose, test and refine novel training techniques. In a March 2025 paper, researchers at Anthropic described using Claude to generate and evaluate new interpretability methods, with the model identifying features in its own internal representations that human researchers had missed. The work was framed as a safety contribution, which it is, but it is also a demonstration that the model was contributing materially to research about itself.

The third is code generation. The proportion of code inside the major laboratories that is now written by models, rather than typed by humans, has risen sharply. Sundar Pichai, the chief executive of Alphabet, told investors in October 2024 that more than a quarter of new code at Google was being generated by AI and reviewed by engineers. By mid-2025, that figure had reportedly climbed past forty percent at several frontier labs. The code being written includes the training infrastructure, the evaluation harnesses and the experimental scaffolding used to build the next generation of models.

The machines are not yet designing themselves. They are, however, increasingly building the tools used to build themselves. None of this constitutes an intelligence explosion in the strict sense that I. J. Good described. What it does constitute is the assembly of every component piece that such an explosion would require. The question is whether the components, once integrated and given sufficient compute, will produce the runaway dynamic that the theory predicts, or whether some bottleneck - physical, economic or cognitive - will intervene first.

The Bottleneck Argument

The most rigorous case against an imminent singularity does not rest on the inadequacy of current models. It rests on the structure of the resources required to scale them.

Training a frontier model in 2026 requires an investment of roughly one billion United States dollars per run, according to figures published by Epoch AI and corroborated by statements from Anthropic and OpenAI. The compute required doubles roughly every six months. The electricity required to power the data centres has begun to strain regional grids. In Virginia, which hosts the largest concentration of data centres in the world, Dominion Energy has warned that demand from artificial intelligence facilities could double the state's electricity consumption by 2030. In Ireland, data centres already consume more than twenty percent of national electricity. In the United Kingdom, the National Energy System Operator has begun publishing scenarios in which AI-driven demand becomes the single largest variable in long-term planning.

These are not trivial constraints. They imply that even if the algorithmic ingredients for recursive self-improvement existed, the physical substrate required to run the loop at meaningful speed might not.

The economist Tyler Cowen, writing on his blog Marginal Revolution throughout 2025, has been one of the more articulate exponents of this view. Cowen does not deny that the technology is improving rapidly. He argues, instead, that the rate of improvement is constrained by the rate at which human institutions can build power stations, train operators and lay fibre, and that these rates are not accelerating.

There is a counterargument, made most forcefully by researchers at the AI Futures Project, whose April 2025 scenario document AI 2027 has become something of a Rorschach test for the field. The authors, including Daniel Kokotajlo, a former OpenAI researcher who resigned in 2024 over disagreements about the company's safety practices, lay out a month-by-month projection in which a fictional laboratory achieves a fully automated AI research workforce by mid-2027 and a superintelligent system by the end of that year. The document is explicitly speculative. It is also, by the admission of its authors, based on extrapolations from real internal benchmarks.