of this is particularly novel, Dr Karpathy says. But he had missed them. “They stack up and actually improved Nanochat,” he says. Speed-ups of this kind are inevitable as models become more capable. Much of the work of building terabyte-size frontier models is less glamorous than the AI industry’s enormous salaries and fancy offices suggest. It involves plumbing together the layers of an infrastructure stack that are bought in from third parties, debugging hardware and software set-ups and tweaking “hyperparameters”, the initial set-up of a training run, until the outcome looks solid. An AI system can do much of that today, with little supervision. But even the more nuanced intellectual work is nearing automation, says Joe Spisak, a researcher at Reflection AI, a lab based in New York that is building frontier models that are open-weight (meaning their parameters are publicly released). Give a frontier system a rough sketch of an idea for efficiency gains, and it is increasingly capable of designing an experiment, running tests on a toy model, seeing what works and responding with a plan that is ready to implement at scale. AI models can carry out these sorts of tasks, which take humans hours, in around 30 minutes. Increasingly, humans play the role only of research director, steering the AI to run experiments, which the models code up,

debug, optimise and monitor themselves. The productivity boost is alluring, but also alarming. As the role that humans play in the production process shrinks, they may lose control. The end result could be models trained by models, to achieve goals set by models, whose safety is verified only by models. Some fear a disaster. Max Tegmark, a physicist and machine-learning researcher at the Massachusetts Institute of Technology who has devoted much of the past decade to campaigning for AI safety, likens it to a driver flooring the accelerator on the motorway with their eyes closed. The result would be certain doom, he told the The Economist’s “Inside Tech” video show, as long as the driver refuses to open their eyes. Powerful AI systems could outcompete humans as the decision makers in government and commerce, says Professor Tegmark, disempowering humanity; they could offer supreme power to whoever first builds them, ushering in global totalitarianism; or they could simply cease to care about humanity at all, and gradually squeeze people out to make room for more data centres and power generation. Three years ago, Professor Tegmark led a call for a pause in global AI development, arguing that the creation of the then-cutting edge GPT-4 was tantamount to that blindfolded journey. This year’s CSET report warned that the systems created by RSI “pose extreme risks. This warrants preparatory action now.” Anthropic, it seems, is close to agreeing with that idea. There are also several physical constraints that will, for now, impose limits on the speed at which models can improve themselves. The most important is access to compute. Despite efficiency gains, newer models continue to use more computing power to train than their predecessors, forcing progress to occur at the pace of data-centre development. Consumer use of AI may also slow down AI-powered research and development, says Helen Toner, interim executive director of CSET and a lead author of its recent report. The limited capacity in AI data centres needs to be carefully split between serving paying customers, training future models and carrying out open-ended R&D. The more demand there is in the first category, the less capacity, in the short term, there is for the other two.

Then there is the issue of training data. Much recent progress in AI has been in areas where models can teach themselves how to succeed thanks to “verifiable rewards”. A piece of software either runs or it does not; a mathematical proof is correct or it is not. In such cases synthetic data, generated by models purely to train other models, can be checked for accuracy and added to the training data without risking the degeneracy that normally comes with training an AI on its own output. It is trickier to make a model better at creative writing or legal judgment. If the models need to learn from the real world, that could also limit the reach of self- improvement. “Closing the loop” may be a step on the road to superintelligence and— depending on your disposition—utopia or doom. But it is not the only step required to produce exponential growth in AI’s capabilities. ■ Curious about the world? To enjoy our mind-expanding science coverage, sign up to Simply Science, our weekly subscriber-only newsletter. This article was downloaded by zlibrary from https://www.economist.com//science-and-technology/2026/06/07/how-artificial- intelligence-got-better-at-building-itself

Science & technology | Lung-range forecast New techniques can predict and prevent lung cancer A molecular signature can identify those most at risk June 11th 2026 Although cancer treatments are improving fast, prevention has so far been mostly about promoting a healthy lifestyle. But efforts to find preventive drugs and vaccines are starting to bear fruit. New research shows that existing anti-inflammatory drugs hold promise for preventing lung cancer. Despite dramatic declines in smoking, lung cancer is the most common cancer diagnosis globally. Smokers who quit face less danger but remain at higher lifetime risk than those who never smoked. People exposed at length to high levels of air pollution, for instance around busy roads in London, are at higher risk too.

Now, an international team of 80 scientists has identified a molecular signature that predicts a higher risk of lung cancer. The group examined blood samples from nearly 50,000 people in the UK Biobank, a repository of data, samples and scans taken periodically from the same cohort of people. They used machine learning to analyse thousands of proteins in the samples and found a group of 14 whose levels increased five years or more before a lung cancer diagnosis. The team then confirmed that this 14-protein “signature” could predict lung cancer in eight other data sets from around the world, including a Taiwanese one in which few participants were smokers. The results were published in Cell on June 4th. In lab experiments on mice and cells, the scientists found that the 14 proteins were present in larger quantities when an inflammatory pathway that is linked to lung cancer was activated. This pathway was identified in 2023 by a team led by Charlie Swanton, a researcher at the Francis Crick Institute in London and co-author of the new paper. The group showed that air pollution triggers the release of a signalling molecule called interleukin-1 beta (IL- 1ß). When IL-1ß reaches lung cells that carry dormant cancer-causing mutations, it activates those mutations. The cells proliferate and grow into a tumour. Remarkably, Dr Swanton found that blocking IL-1ß in mice exposed to air pollution stopped tumours forming, offering the promise of preventive drugs. The new research provides the foundations for a blood test that could identify who will benefit from such treatments. Drugs that block IL-1ß in humans already exist. They are used to treat certain auto-inflammatory conditions, such as some forms of arthritis. As part of the new study, Dr Swanton’s team analysed data on cancer incidence among people treated with these drugs. In 2017 Novartis, a pharmaceutical company, reported the results from a big trial of canakinumab, an IL-1ß blocker, as a preventive therapy for heart attacks. The drug was ineffective against heart attacks, but as part of the safety checks of the trial, Novartis collected data on cancer incidence among participants.

The patients treated with canakinumab ended up with lower rates of lung cancer than the placebo group. The effect was modest though. Dr Swanton’s team reanalysed the trial data using the 14-protein signature. In people with lower amounts of the 14 proteins, more than 1,500 would need to be treated to prevent one lung cancer. But in those with larger amounts of the proteins, canakinumab nearly halved the risk. In this group, treating 55 people prevented one case of disease. Cholesterol-lowering drugs known as statins, which are widely prescribed to prevent heart attacks, have a similar prevention rate. The next step is to develop a commercial test for detecting the 14-protein signature. The team is also investigating whether other anti-inflammatory drugs could be similarly effective. Fresh UK Biobank data is expected in the next two years. It could provide similar clues about other types of cancer. A long-running trial in England, for example, has shown that aspirin prevents some colorectal cancers in people with Lynch syndrome, a genetic condition that makes people highly predisposed to certain cancers. Cancer prevention may finally get the level of scientific attention devoted to finding new treatments. ■ Curious about the world? To enjoy our mind-expanding science coverage, sign up to Simply Science, our weekly subscriber-only newsletter. This article was downloaded by zlibrary from https://www.economist.com//science-and-technology/2026/06/10/new-techniques-can- predict-and-prevent-lung-cancer

Science & technology | Doctors with borders Too much Chinese science is ignored by the West A bad reputation and cultural ignorance are probably responsible June 11th 2026 Chinese authors published as many papers as American, British, German and Japanese researchers combined in 2025. Yet two recent analyses drawing on databases of tens of millions of English-language scientific articles suggest they were largely overlooked by Western researchers. That is not for lack of value: China leads the Nature index, a ranking of countries by number of papers published in a set of respected journals. Both new analyses tracked how academic papers were cited by others. The first, a working paper by Abhishek Nagaraj of the University of California, Berkeley, and Randol Yao of the Massachusetts Institute of Technology, was published in January. It found that between 1980 and 2022, only about one- third of citations of Chinese papers came from outside the country,