Just three months into the Trump administration's promised crackdown on immigration to the United States, Immigrations and Customs Enforcement now has a 30 million contract with Palantir to build a "near-real time" surveillance platform called ImmigrationOS that would track information about people self-deporting (electing to leave the US). Meanwhile, the Department of Homeland Security has been sending aggressive emails telling people with temporary legal status to leave the US. It is unclear who has actually been sent the messages, though, given that a number of people who are US-born citizens have reported receiving them. The US Cybersecurity and Infrastructure Security Agency briefly seemed poised this week to cancel funding for the critical software vulnerability tracking project known as the CVE Program. CISA eventually came through with the funding, but some members of the CVE Program's governing board are planning to make the project into an independent nonprofit.

The United States under President Donald Trump is ramping up use of surveillance systems and artificial intelligence (AI) to track and arrest immigrants, raising fears that risks to accuracy and privacy could put almost anyone in danger of getting caught up in the crackdown. The Department of Homeland Security (DHS) and other immigration control agencies are using a suite of AI tools -- such as facial recognition scanners in public areas and robotic dogs patrolling the southern border for human movement -- as part of the crackdown on alleged illegal immigration. Many of the AI tools that immigration agents are using have been in place for years and are a legacy of previous administrations, according to Saira Hussain, an attorney at the Electronic Frontier Foundation, a digital rights advocacy group.

Data-driven emulation of nonlinear dynamics is challenging due to long-range skill decay that often produces physically unrealistic outputs. Recent advances in generative modeling aim to address these issues by providing uncertainty quantification and correction. However, the quality of generated simulation remains heavily dependent on the choice of conditioning priors. In this work, we present an efficient generative framework for dynamics emulation, unifying principles of turbulence with diffusion-based modeling: Cohesion. Specifically, our method estimates large-scale coherent structure of the underlying dynamics as guidance during the denoising process, where small-scale fluctuation in the flow is then resolved. These coherent priors are efficiently approximated using reduced-order models, such as deep Koopman operators, that allow for rapid generation of long prior sequences while maintaining stability over extended forecasting horizon. With this gain, we can reframe forecasting as trajectory planning, a common task in reinforcement learning, where conditional denoising is performed once over entire sequences, minimizing the computational cost of autoregressive-based generative methods. Empirical evaluations on chaotic systems of increasing complexity, including Kolmogorov flow, shallow water equations, and subseasonal-to-seasonal climate dynamics, demonstrate Cohesion superior long-range forecasting skill that can efficiently generate physically-consistent simulations, even in the presence of partially-observed guidance.

Flow-based methods for sampling and generative modeling use continuous-time dynamical systems to represent a {transport map} that pushes forward a source measure to a target measure. The introduction of a time axis provides considerable design freedom, and a central question is how to exploit this freedom. Though many popular methods seek straight line (i.e., zero acceleration) trajectories, we show here that a specific class of ``curved'' trajectories can significantly improve approximation and learning. In particular, we consider the unit-time interpolation of any given transport map $T$ and seek the schedule $\tau: [0,1] \to [0,1]$ that minimizes the spatial Lipschitz constant of the corresponding velocity field over all times $t \in [0,1]$. This quantity is crucial as it allows for control of the approximation error when the velocity field is learned from data. We show that, for a broad class of source/target measures and transport maps $T$, the \emph{optimal schedule} can be computed in closed form, and that the resulting optimal Lipschitz constant is \emph{exponentially smaller} than that induced by an identity schedule (corresponding to, for instance, the Wasserstein geodesic). Our proof technique relies on the calculus of variations and $\Gamma$-convergence, allowing us to approximate the aforementioned degenerate objective by a family of smooth, tractable problems.

Sign up for our daily newsletter to get the best of The New Yorker in your in-box. Elon Musk, who's taking his chainsaw to the federal government, is not merely a chaos agent, as he is sometimes described. Jill Lepore, the best-selling author of "These Truths" and other books, says that Musk is animated by obsessions and a sense of mission he acquired through reading, and misreading, science fiction. "When he keeps saying, you know, 'We're at a fork in the road. The future of human civilization depends on this election,' he means SpaceX," she tells David Remnick.

It's one of the most profound questions in science – did life ever exist on Mars? Now, experts have unearthed evidence that the Red Planet was once habitable. Scientists have found carbon residue in Martian rocks, indicating that an ancient carbon cycle existed. And it means the Red Planet was likely once warm enough to sustain life. Researchers have long believed that, billions of years ago, Mars had a thick, carbon dioxide-rich atmosphere with liquid water on its surface.

Opposition parties in Italy have complained to the communications watchdog about a series of AI-generated images published on social media by deputy prime minister Matteo Salvini's far-right party, calling them "racist, Islamophobic and xenophobic", the Guardian has learned. The centre-left Democratic party (PD), with the Greens and Left Alliance, filed a complaint on Thursday with Agcom, the Italian communications regulatory authority, alleging the fake images used by the League contained "almost all categories of hate speech". Over the past month, dozens of apparently AI‑generated photos have appeared on the League's social channels, including on Facebook, Instagram and X. The images frequently depict men of colour, often armed with knives, attacking women or police officers. Antonio Nicita, a PD senator, said: "In the images published by Salvini's party and generated by AI there are almost all categories of hate speech, from racism and xenophobia to Islamophobia. They are using AI to target specific categories of people – immigrants, Arabs – who are portrayed as potential criminals, thieves and rapists. "These images are not only violent but also deceptive: by blurring the faces of the victims it is as if they want to protect the identity of the person attacked, misleading users into believing the photo is real.

For the second time in the last month, Microsoft employees disrupted high-level executives speaking at an event celebrating the company's 50th anniversary on 4 April, in protest against the company's role in Israel's ongoing siege on Gaza. The two were fired within days. The Microsoft president Brad Smith and the former CEO Steve Ballmer were shouted down at Seattle's Great Hall on 20 March by a current and former employee. The April event was preceded by a rally outside that also included current and former employees of the tech giant. Protesters projected a sign onto the hall's wall saying, "Microsoft powers genocide" – a reference to Israel's extensive use of the company's AI and cloud computing services since 7 October 2023, as "the IDF's insatiable demand for bombs was matched by its need for greater access to cloud computing services," the Guardian reported.

The most accurate clock in space launches within days and will begin building a highly synchronised network out of the best clocks on Earth. But the project, decades in preparation, will only operate for a few years before it burns up as the International Space Station deorbits at the end of the decade. NASA's most accurate atomic clock will be tested on a mission to Venus The Atomic Clock Ensemble in Space (ACES) is a European Space Agency (ESA) mission that will generate a time signal with unprecedented accuracy and then transmit it via laser to nine ground stations as it passes overhead at 27,000 kilometres per hour. This network of clocks will be in extremely close synchronisation and provide highly accurate timekeeping around the world. The result is that ACES will be able to test Einstein's theory of general relativity, which says that the passing of time is affected by the strength of gravity, with great accuracy.

Designing efficient and effective architectural backbones has been in the core of research efforts to enhance the capability of foundation models. Inspired by the human cognitive phenomenon of attentional bias-the natural tendency to prioritize certain events or stimuli-we reconceptualize neural architectures, including Transformers, Titans, and modern linear recurrent neural networks as associative memory modules that learn a mapping of keys and values using an internal objective, referred to as attentional bias. Surprisingly, we observed that most existing sequence models leverage either (1) dot-product similarity, or (2) L2 regression objectives as their attentional bias. Going beyond these objectives, we present a set of alternative attentional bias configurations along with their effective approximations to stabilize their training procedure. We then reinterpret forgetting mechanisms in modern deep learning architectures as a form of retention regularization, providing a novel set of forget gates for sequence models. Building upon these insights, we present Miras, a general framework to design deep learning architectures based on four choices of: (i) associative memory architecture, (ii) attentional bias objective, (iii) retention gate, and (iv) memory learning algorithm. We present three novel sequence models-Moneta, Yaad, and Memora-that go beyond the power of existing linear RNNs while maintaining a fast parallelizable training process. Our experiments show different design choices in Miras yield models with varying strengths. For example, certain instances of Miras achieve exceptional performance in special tasks such as language modeling, commonsense reasoning, and recall intensive tasks, even outperforming Transformers and other modern linear recurrent models.