Two separate campaigns target CVE-2025-8088, fixed last July, to conduct data theft and cyberespionage against military and government targets in Ukraine.

Discussion | Link

Gemini 3.5 Live Translate brings near real-time, natural speech translation to Google AI Studio, Google Translate and Google Meet.

Use-after-free bug can be exploited to evade sandbox defenses.

The FCC wants to legally force telecoms to collect new and renewing customers’ government issued identity number and physical address, impacting everyone from the privacy-conscious to domestic abuse survivors. “We never thought that would happen here.”

CPUs, GPUs, TPUs, and NPUs The post The Hardware That Makes AI Possible appeared first on Towards Data Science.

Training a speech AI model to correctly recognize or synthesize clinical terminology is surprisingly difficult. Drug names like Acetaminophen, Amlodipine,... Training a speech AI model to correctly recognize or synthesize clinical terminology is surprisingly difficult. Drug names like Acetaminophen, Amlodipine, Cefazolin, and Biktarvy are not part of everyday vocabulary. Procedure names, anatomy terms, and specialty-specific diagnoses introduce the same problem in a different form. Off-the-shelf speech systems can sound fluent and still miss the words… Source

Discussion | Link

When two AIs argue against each other, the legal system loses.

Individual gold atoms move around to form oxidation-proof structures.

The agreement follows AI infrastructure deals between Corning and Nvidia and Corning and Meta.

The world relies on a modest number of countries to keep watch over the ocean. That arrangement is starting to fail. Europe and Asia must now decide whether to let the system unravel, or to take it up together. Right now, in every ocean basin on Earth, a global network of instruments measures the state of the sea. Research ships steam along oceanographic transects from surface to seafloor. Anchored buoys watch the tropical oceans for the first signs of El Niño or tropical cyclones and take the pulse of the thermohaline circulation. Some four thousand autonomous floats sink every ten days to two thousand metres before rising to transmit temperature and salinity to ground stations via satellite. Underwater gliders patrol continental margins, and drifting buoys ride the surface in the most remote waters. Hundreds of elephant seals carry miniaturised sensors beneath the polar sea ice… Together, this network produces invaluable information that allows societies to anticipate and respond to a changing ocean and weather conditions, and protect the ocean in return. It is also far more fragile than most people, and most governments realise. A new study published in Nature Climate Change has measured for the first time just how fragile the ocean watch network is. The result is alarming. If observations from a single major contributor, the United States, were withdrawn from the Global Ocean Observing System (GOOS), the errors in our estimate of how fast the ocean is warming would jump by 163 percent. That is worse than randomly losing 80 percent of all global ocean data. The reason is geographical: US instruments cover every ocean basin and plug critical gaps that no other nation currently fills. This is not a theoretical concern. Proposed cuts to National Oceanic and Atmospheric Administration (NOAA) and the National Science Foundation in the United States now threaten exactly this contribution. And the situation is barely better on the other side of the Atlantic. The pressures are not confined to one side of the Atlantic, nor to the West. In China, scientists and policymakers are working to build a more resilient national contribution to ocean observation, but without the resources the moment requires. The marine monitoring system the world relies on is under strain almost everywhere. An observing system, not a programme Public conversations about ocean observations often focus on Argo floats. Designed to monitor ocean conditions, Argo floats are part of an international program that collects data on salinity, temperature and currents using a fleet of robotic instruments that drift and move up and down between the surface and mid-water level. D. Luquet IMEV, Fourni par l'auteur Each Argo is essentially a sealed cylinder of pressurised electronics with a clever buoyancy chamber: it floods with seawater to sink and is evacuated to rise again. These autonomous robots have transformed ocean science this century. However, Argo is just one component of GOOS and the complementarity of its parts matters. Argo profiles the upper two kilometres of the open ocean. Research vessels go deeper: GO-SHIP cruises survey from surface to seafloor along long repeated transects, providing the high-precision reference measurements that calibrate every other instrument and help validate climate models. Moored buoys deliver continuous time series critical for monitoring El Niño, the Atlantic Meridional Overturning Circulation, and the conditions in which tropical cyclones form. Underwater gliders target coastal currents, eddies and continental margins that floats cannot resolve. Elephant seals carry sensors into under-ice regions of the polar oceans that no other instrument can reach. Each platform answers questions the others cannot. Remove any one of these ocean watch components, and the observing system’s ability to deliver reliable information degrades not in proportion to the volume of data lost, but in proportion to where the gaps appear. What this network actually delivers The Global Ocean Observing System is too often described as “climate monitoring,” but it does so much more. Every operational weather forecast is built on these data. The numerical weather prediction systems run by the European Centre for Medium-Range Weather Forecasts, by Météo France, and by every other major weather service ingest ocean observations many times a day. Without them, forecasts drift quickly out of skill. The new artificial-intelligence-based forecast systems Pangu-Weather and GraphCast, despite their impressive performance, rely entirely on the same observational stream. AI does not replace observations; it depends on them. Sub-seasonal to seasonal forecasting, to help anticipate harvest seasons, energy demand and water availability weeks to months ahead, depends critically on knowledge about subsurface ocean heat and salinity. Tropical cyclone track and intensity forecasts, central to early warning and evacuation decisions, depend on ocean heat content beneath the surface, not just sea surface temperature, because hurricanes draw their explosive energy from the warm layers down to at least 200m depth. Marine heatwave warnings, now used routinely by fisheries managers worldwide, are impossible without sustained subsurface observation. Sea-level projections used to design coastal infrastructure require decades of consistent measurements, and salinity adds the density information essential for determining all ocean currents, including the AMOC. In short, GOOS underpins operational services from tomorrow’s storm warnings to next century’s adaptation plan. It is not a luxury but essential. Why models and artificial intelligence alone cannot save us There is a persistent misconception, amplified by the rise of AI, that sufficiently advanced models can substitute for direct observations. They cannot. Every forecast model, whether traditional or AI- based, relies on data assimilation: a continuous nudged adjustment of the simulation toward real-world measurements. An AI model trained on a richly observed past will perform poorly in a sparsely observed present. In a world of rising extremes and shifting ocean states, historical patterns become less reliable. An observation not made is lost forever. Satellite measurements of the sea surface cannot tell us what is happening hundreds or thousands of metres below, where heat accumulates, currents reorganise, and the precursors of the next season’s weather are already forming. To see beneath the surface, we need instruments in the water. The cheapest insurance we have The argument that ocean observation is too expensive collapses on contact with the numbers. The total annual cost of the global system, across all platforms and personnel, runs on the order of one billion euros worldwide. The European share is a fraction of that. Extreme weather events linked to ocean conditions caused tens of billions of euros in damage across Europe in 2024 alone. A single major North Atlantic hurricane season can cost the United States hundreds of billions of dollars. Marine heatwaves have collapsed fisheries worth billions and triggered mass coral bleaching on every reef system on Earth. Failed seasonal forecasts cascade through agriculture, energy and humanitarian response, with consequences rarely tallied. Every euro spent on ocean observation returns many times its value. It is one of the highest-return public investments available. Europe’s choice Europe needs to treat ocean observations as critical infrastructure, equivalent to satellite navigation or meteorological services. That means stable, multi-year funding for the operational backbone of the system: the floats, ships, moorings, gliders and data centres that process and deliver the data. France has the second-largest Exclusive Economic Zone in the world but contributes around 5 percent of global ocean temperature profile data. Present in the Atlantic, Pacific and Indian Oceans, France has five departments and regions and seven overseas collectivities, which are home to 2.7 million French citizens. Australia contributes more than three times as much. The European Union contributes about 12 percent, less than a quarter of the American share. Europe, and France in particular, should substantially increase its contribution. OceanObs’29, the decadal international conference set to be held this time in China, is an opportunity to negotiate a more balanced global system, reflecting economic capacity and maritime interests rather than historical accident. Europe-China scientific cooperation should increase, as they have largely complementary observing footprints. Together, they would cover much of the global ocean. The narrowing window The danger is a gradual erosion of the information on which a growing share of human activity and the blue economy now depends. Cyclone warnings become less reliable, seasonal forecasts less skilful, sea-level projections less precise. Each loss maybe individually tolerable. Together, they amount to flying blind into the most consequential transformation of the planet’s climate in human history. The ocean observing system is a planetary public service, built over decades by many nations. France and Europe possess the institutions, expertise and the maritime interest to play a far larger role. What is missing is the political decision to act, while the system can still be sustained. The loss of collaboration among nations would force a rebuild far more difficult and expensive than sustained investment in what already works. The window remains open but is narrower than it was. A weekly e-mail in English featuring expertise from scholars and researchers. It provides an introduction to the diversity of research coming out of the continent and considers some of the key issues facing European countries. Get the newsletter! Sabrina Speich a reçu des financements de l'ERC, EU Horizon 2030, CNES TOSCA et ANR. Sabrina Speich est présidente du comité d'experts du "Ocean Observations for Physics and Climate" des programmes UN GOOS et GCOS. John Abraham, Kevin Trenberth et Lijing Cheng, ne travaillent pas, ne conseillent pas, ne possèdent pas de parts, ne reçoivent pas de fonds d'une organisation qui pourrait tirer profit de cet article, et n'ont déclaré aucune autre affiliation que leur poste universitaire.

What’s holding organizations back from making the most of AI.

At night, the car projects its turn signals onto the road to alert other road users.

Luca Toni of Bayern Munich in action. Tsutomu Takasu/Wikimedia, CC BY Football fans everywhere are gearing up to celebrate the sport’s most skilled athletes as they prepare for the start of another Fifa World Cup. But few get to see how the next generation of Messis and Ronaldos are discovered. For decades, the beautiful game depended on the human eye: a scout on the sideline, attentively watching, waiting for that something special. That process, however, is becoming increasingly data-driven. Across elite football academies worldwide, technologies such as GPS trackers, automated video analysis, and AI-powered platforms are changing how players are identified and assessed. In a sport shaped by money and global competition, the appeal is obvious. Clubs want to make better decisions and they want to make them as early as possible. The quicker a talented player is identified, the greater the potential return. As my research shows, technology promises precisely this. Coaches and scouts are using new tools to make talent identification more efficient and, crucially, more objective. The question is whether talent can actually be reduced to numbers and whether technology and AI combat biases or create new ones. Football has always been unpredictable and full of promise. AMISOM/Flickr Early identification Modern football increasingly selects players at very young ages. Spanish footballer Lamine Yamal was scouted by Barcelona at just six years old; he made his first-team debut at 15. This is not a necessarily a new development or consequence of AI and technology. David Beckham famously joined Manchester United as a 16 year old. Nonetheless, not all talent develops or is identified early. US star Alex Morgan only began playing seriously in her teens, Italy’s Luca Toni did not reach the top level until his early twenties and England’s Ian Wright was nearly 22 when he signed his first professional contract. So-called late bloomers are reminders that development is unpredictable. Especially in childhood and adolescence, athletes grow at different rates physically, psychologically and socially. This makes it difficult to compare players directly, and even harder to predict who will succeed. AI systems, by contrast, rely on patterns. Neural networks analyse large datasets to identify what successful players have in common. The problem is that patterns are based on the past. If previous systems have favoured certain types of players, for example, those who mature early physically or come from particular backgrounds (such as regional accessibility, higher socioeconomic status, having pre-existing profitable networks or even just parents who are involved and interested in sport), then AI may simply learn to reproduce those patterns. What looks like objectivity can, in reality, be a new form of bias. Research has started to question whether AI can truly “solve” talent identification. While data-driven methods can support decision-making, they do not eliminate subjectivity. I have shown that decisions about what data to collect, how to analyse it, and what counts as “talent” are still shaped by human judgment and are, therefore, always subjective to some extent. A second problem stems from the fact that data does not capture context. Just about everything can be measured, of course, but player qualities are typically revealed through interaction, experience and context – not metrics alone. Studies have shown that coaches and staff struggle not with collecting data, but with interpreting it and turning it into meaningful insights. This clearly highlights that data is only useful if it is understood. Fairness and diversity Football has long struggled with unequal access and opportunity. Pathways into elite sport are shaped by social, economic and cultural factors. If AI systems are trained on historical data that reflects these inequalities, they risk reinforcing them. Players who are physically more developed compared to their peers typically get an advantage. But this does not necessarily mean they have greater longterm potential. Wright or Toni, who reached higher levels at a later stage in life, would have likely been overlooked in such systems. AI trained on such data may unintentionally prioritise early developers again and again. This would narrow the pool of talent rather than expanding it. There is also the question of how data affects young players themselves. Previous research has shown that increasing monitoring can create intense pressure and a sense of constant evaluation and surveillance from above, for both athletes and coaching and scouting staff alike. Football needs to decide what kind of talent it wants to recognise. Do we want to meet the new Ian Wrights or Luca Tonis? Or do we just want to copy and paste the same story, again and again? AI has clear potential. It can help clubs process large amounts of information, identify patterns that humans might miss, and even widen access by allowing players to be seen from anywhere in the world. Used thoughtfully, it could support more informed and inclusive decision-making. But we need to be clear about one thing: it is not a neutral tool and it is not increasing objectivity. It reflects the assumptions, values and biases embedded in the data and systems behind it. And it is chosen based on subjective ideas of what counts as valuable data. The challenge, then, is not simply to adopt new technologies, but to question how they are used. Clubs need to be aware of the limitations of data. They need to invest in education and expertise. And they need to ensure that technology complements rather than replaces human judgment. If the game becomes too focused on what can be measured, it risks overlooking what cannot. Without those safeguards, football risks losing what makes it special: its unpredictability, its diversity and that promise it has always proffered – that greatness can emerge from unexpected places. Leah Monsees does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

Amazon employees have a Slack channel for memes where the mock and commiserate about the company’s faulty AI coding product.

Discussion | Link

Stop re-computing the same context. Learn how to build a C++ runtime with copy-on-fork KV snapshots to eliminate redundant LLM prefills in multi-agent pipelines. The post Prefill Once, Fan Out: KV Snapshot Sharing for Multi-Agent LLM Pipelines appeared first on Towards Data Science.