The automation of science is a long-standing ambition in artificial intelligence (AI) research1,2. Although the community has made substantial progress in automating individual components of the scientific process, a system that autonomously navigates the entire research life cycle-from conception to publication-has remained out of reach. Here we present a pipeline for automating the entire scientific process end to end. We present The AI Scientist, which creates research ideas, writes code, runs experiments, plots and analyses data, writes the entire scientific manuscript, and performs its own peer review. Its ideas, execution and presentation are of sufficient quality that the manuscript generated by this AI system passed the first round of peer review for a workshop of a top-tier machine learning conference. The workshop had an acceptance rate of 70%. Our system leverages modern foundation models3-5 within a complex agentic system. We evaluate The AI Scientist in two settings: a focused mode using human-provided code templates as an initial scaffold for conducting research on a specific topic and a template-free, open-ended mode that leverages agentic search for wider scientific exploration6,7. Both settings produce diverse ideas and automatically test, report on and evaluate them. This achievement demonstrates the growing capacity of AI for making scientific contributions and signifies a potential paradigm shift in how research is conducted. As with any impactful new technology, there could be important risks, including taxing overwhelmed review systems and adding noise to the scientific literature. However, if developed responsibly, such autonomous systems could greatly accelerate scientific discovery.

Advances at the intersection of biotechnology, artificial intelligence, electronics and materials science are reshaping how drugs can be delivered inside the body. Intelligent and miniaturized drug delivery devices (IMDDDs) leverage these technologies to achieve precise pharmacokinetics, targeted distribution and programmable release while minimizing toxicity and improving patient adherence. Unlike conventional approaches, IMDDDs can incorporate real-time sensing and adaptive control, enabling drug administration that is more precise and more responsive to dynamic physiological conditions. In this Review, we outline key categories and design principles, highlight artificial intelligence technologies for augmenting performance, discuss potential clinical applications across cancer, diabetes, cardiovascular disease, vaccination and beyond, and examine translation challenges and opportunities. By uniting engineering innovation with medical need, IMDDDs exemplify the next generation of drug delivery technologies.

Starburst galaxies often host multiphase, galaxy-scale winds thought to enrich the circumgalactic medium and limit further star formation by disrupting interstellar gas clouds1-3. These winds are primarily powered by supernovae4-6, but it remains unclear how supernova energy forms an organized flow. Here we use the Resolve spectrometer on the X-ray Imaging and Spectroscopy Mission to show that the hot (T = 2 × 107 K) gas in the nucleus of the starburst galaxy M82 is moving quickly, with a line-of-sight velocity dispersion σ=595-128+464kms-1 . This is consistent with a hot, nuclear wind generated by thermal pressure. We show that a free-wind model reproduces the measured temperature but underpredicts the velocity. The inferred mass and energy outflow rates from the nucleus, about 7 M⊙ yr-1 and 4 × 1042 erg s-1, require that most supernova energy is thermalized. These outflow rates provide enough energy to power the ≳30 M⊙ yr-1 cool outflow and still transport up to 3 M⊙ yr-1 to the intergalactic medium, suggesting that thermal gas pressure is sufficient to power the multiphase wind without additional support from cosmic rays7. We also show that the nuclear gas is hotter and faster than the plasma seen on larger scales ( kT=0.72-0.08+0.10keV , σ=175-73+86kms-1 ), suggesting a distinct origin for the latter.

Effectively communicating worst-case projections of global future climate-hereinafter referred to as worst-case climate outcomes-is essential for risk assessment and developing robust adaptation strategies to global warming1-7. Yet, current approaches for identifying spatially consistent climate outcomes are limited, with worst-case global climates typically communicated via the average of climate model projections at high global warming levels, such as 3 °C or 4 °C above the preindustrial era8,9. Here we show that extreme global climate outcomes may occur even under moderate 2 °C warming for several sectors. For droughts in global key breadbasket regions, precipitation extremes over highly populated areas and fire weather extremes across forests, global climatic impact-drivers at 2 °C of global warming may turn out to be much more extreme than model-averaged projections at 3 °C or 4 °C warming. We derive these results by identifying sector-specific, spatially consistent potential high- and low-impact global climate outcomes through spatially averaging projected sector-relevant climatic impact-drivers across key global regions. Our approach can easily be adapted to a wide range of sectors to support the improvement of sector-specific climate risk assessment and to inform climate policy. As global warming approaches 1.5 °C (ref. 10), these findings underscore the urgency of rapid mitigation to limit warming well below 2 °C, as even a 2 °C world may entail severe impacts.

Phase singularities-points carrying quantized topological charge-are universal features found across diverse wave systems from superfluids and superconductors to acoustic and optical fields1-4. Ensembles of these singularities exhibit distance correlations resembling particles in liquids5-8, extensively studied for their role in exotic material phases9-11. By contrast, the full correlations in phase space that govern the system evolution have remained unexplored and experimentally inaccessible. Here we directly measure the ultrafast dynamics of optical singularity ensembles, capturing their full phase-space correlations, presenting the joint distance-velocity distribution. Our observations show a breakdown of the particle-singularity analogy12: phase singularities accelerate towards formally divergent velocities in the moment before annihilation7,13,14, indicated by measurements of velocities exceeding the speed of light. These apparent superluminal velocities are paradoxically amplified by the slow group velocity of hyperbolic phonon polaritons in our material platform, hexagonal boron nitride membranes15-19. We demonstrate these phenomena using combined hardware and algorithmic advances in ultrafast electron microscopy18,20-25, achieving spatial and temporal resolutions, each an order of magnitude below the polaritonic wavelength and cycle period. Our findings deepen our understanding of phase singularities and their universality, enabling to probe topological defect dynamics at previously unattainable timescales.

Archaeological evidence suggests that dogs diverged from wolves during the Palaeolithic, more than 15,000 years ago1-7. The earliest unequivocal genetic evidence, however, is associated with dog remains from Mesolithic archaeological contexts approximately 10,900 years ago8,9. Here we generate both nuclear and mitochondrial genomes from canid remains at Pınarbaşı in Türkiye (15,800 years ago)10 and Gough's Cave in the UK (14,300 years ago)11, as well as from dogs excavated from two Mesolithic sites in Serbia (Padina between 11,500-7,900 years ago and Vlasac 8,900 years ago)12,13. Our analyses indicate that a genetically homogeneous dog population was already widely distributed across Europe and Anatolia during the Late Upper Palaeolithic (by at least 14,300 years ago). This finding suggests that dogs were exchanged among genetically and culturally distinct western Eurasian Late Palaeolithic human populations, namely the Magdalenian, Epigravettian and Anatolian hunter-gatherers10,14-16. Last, we identify a major influx of eastern Eurasian dog ancestry during the Mesolithic, concomitant with the movement of eastern hunter-gatherer populations into Europe14, which led to the establishment of the primary ancestry characteristics that define European dog populations today.

Climate change is causing measurable harm globally1,2. Political and legal efforts seek to link these damages with specific emissions, including in discussions of loss and damage (L&D)3,4; however, no quantitative definition of L&D exists5,6, nor is there a framework to link past and future emissions from specific sources to monetized, location-specific damages. Here we develop such a framework, which is integrated with recent efforts to estimate the social cost of carbon7. Using empirical estimates of the non-linear relationship between temperature and aggregate economic output, we show that future damages from past emissions-one component of L&D-are at least an order of magnitude larger than historical damages from the same emissions. For instance, one tonne of CO2 emitted in 1990 caused US$180 in discounted global damages by 2020 ($40-530) and will cause an additional $1,840 through 2100 ($500-5,700). Thus, settling debts for past damages will not settle debts for past emissions. In other illustrative estimates, a single long-haul flight per year over the past decade leads to about $25k ($6,000-77,000) in future damages by 2100, and US emissions since 1990 caused $500 billion ($180-1,300 billion) of damage in India and $330 billion ($110-820 billion) in Brazil. Carbon removal offers an alternative to transfer payments for settling L&D, but is increasingly ineffective in limiting damages as the delay between emission and recapture increases.

The earliest morphologically identifiable dogs are from Europe and date to at least 14,000 years ago1-5, although early remains are also found in other regions. The origin of early dogs in Europe, and their relationships to other dogs, has remained elusive in the absence of genome-wide data. Similarly, although dogs were the only domestic animal to predate agriculture, little is known about how the arrival of Neolithic farmers from Southwest Asia affected the dogs living with European Mesolithic hunter-gatherers. Here we analysed 216 canid remains, including 181 from Palaeolithic and Mesolithic Europe. We developed a genome-wide capture approach that enriched endogenous DNA by 10-100-fold and could distinguish dog from wolf ancestry for 141 of 216 remains. The oldest dog data that we recovered are from a 14,200-year-old dog from the Kesslerloch site in Switzerland, and we find that it shares ancestry with later worldwide dogs-inconsistent with the hypothesis that European Upper Palaeolithic dogs derived wholly from a separate domestication process. The Kesslerloch dog already displays more affinity to Mesolithic, Neolithic and present-day European dogs than to Asian dogs, demonstrating that dog genetic diversification had started well before 14,200 years ago. We find a Neolithic influx of Southwest Asian ancestry into Europe, but this seems to have been of smaller magnitude than in humans, suggesting that Mesolithic dogs contributed substantially to Neolithic, and, ultimately, probably also modern, European dogs.

Carbon capture and storage (CCS) has the potential to help nations meet their Paris Agreement CO2 reduction commitments1,2. The ability to capture CO2 within mafic and ultramafic rocks through mineralization of carbon is an example of such a CCS technology3,4, but large-scale deployment has yet to be achieved5,6. Each geologic environment in the Earth's crust requires a distinct carbon storage solution. Whereas some regions of the subsurface contain saline aquifers and sedimentary traps suitable for traditional carbon storage through the injection of high-pressure, dense CO2 below impermeable caprocks, other regions may lack caprocks5-9. In these regions, carbon storage is possible through the mineralization of injected water-dissolved CO2 forming stable carbonate minerals through its reactions with reactive silicate rocks and minerals6,10,11. A notable challenge to applying this process at scale is that it can require 20-50 times or more water than the mass of CO2 stored12. Here we report on an industrial-scale pilot project designed to find a carbon disposal solution for western Saudi Arabia. This arid region has large point-source CO2 emitters, including petroleum refining and desalination facilities, but lacks saline aquifers and sedimentary traps13-17. We find that a CO2 injection approach based on the recirculation of subsurface fluids can eliminate the need for external water. Our results demonstrate the feasibility of carbon mineral storage in regions in which access to water resources may be limited.