AI

Understanding audio has always been the multimodal frontier that lags behind vision. While image-language models have rapidly scaled toward real-world deployment, building open models that robustly reason over speech, environmental sounds, and music — especially at length — has remained quite hard. NVIDIA and the University of Maryland researchers are now taking a direct swing at that gap. The research team have released Audio Flamingo Next (AF-Next), the most capable model in the Audio Flamingo series and a fully open Large Audio-Language Model (LALM) trained on internet-scale audio data. Audio Flamingo Next (AF-Next) comes in three specialized variants for different use cases. The release includes AF-Next-Instruct for general question answering, AF-Next-Think for advanced multi-step reasoning, and AF-Next-Captioner for detailed audio captioning. What is a Large Audio-Language Model (LALM)? A Large Audio-Language Model (LALM) pairs an audio encoder with a decoder-only language model to enable question answering, captioning, transcription, and reasoning directly over audio inputs. Think of it as the audio equivalent of a vision-language model like LLaVA or GPT-4V, but designed to handle speech, environmental sounds, and music simultaneously — within a single unified model. https://arxiv.org/pdf/2604.10905 The Architecture: Four Components Working in a Pipeline AF-Next is built around four main components: First is the AF-Whisper audio encoder, a custom Whisper-based encoder further pre-trained on a larger and more diverse corpus, including multilingual speech and multi-talker ASR data. Given an audio input, the model resamples it to 16 kHz mono and converts the waveform into a 128-channel log mel-spectrogram using a 25 ms window and 10 ms hop size. The spectrogram is processed in non-overlapping 30-second chunks through AF-Whisper, which outputs features at 50 Hz, after which a stride-2 pooling layer is applied. The hidden dimension is 1280. Second is the audio adaptor, a 2-layer MLP that maps AF-Whisper’s audio representations into the language model’s embedding space. Third is the LLM backbone: Qwen-2.5-7B, a decoder-only causal model with 7B parameters, 36 transformer layers, and 16 attention heads, with context length extended from 32k to 128k tokens through additional long-context training. A subtle but important architectural detail is Rotary Time Embeddings (RoTE). Standard positional encodings in transformers index a token by its discrete sequence position i. RoTE replaces this: instead of the standard RoPE rotation angle θ ← −i · 2π, RoTE uses θ ← −τi · 2π, where τi is each token’s absolute timestamp. For audio tokens produced at a fixed 40 ms stride, discrete time positions are interpolated before being fed into the RoTE module. This yields positional representations grounded in actual time rather than sequence order — a core design choice enabling the model’s temporal reasoning, particularly for long audio. Finally, a streaming TTS module enables voice-to-voice interaction. Temporal Audio Chain-of-Thought: The Key Reasoning Recipe Chain-of-Thought (CoT) prompting has improved reasoning across text and vision models, but prior audio CoT work showed only small gains because training datasets were limited to short clips with simple questions. AF-Next addresses this with Temporal Audio Chain-of-Thought, where the model explicitly anchors each intermediate reasoning step to a timestamp in the audio before producing an answer, encouraging faithful evidence aggregation and reducing hallucination over long recordings. To train this capability, the research team created AF-Think-Time, a dataset of question–answer–thinking-chain triplets curated from challenging audio sources including trailers, movie recaps, mystery stories, and long-form multi-party conversations. AF-Think-Time consists of approximately 43K training samples, with an average of 446.3 words per thinking chain. Training at Scale: 1 Million Hours, Four Stages The final training dataset comprises approximately 108 million samples and approximately 1 million hours of audio, drawn from both existing publicly released datasets and raw audio collected from the open internet and subsequently labeled synthetically. New data categories introduced include over 200K long videos spanning 5 to 30 minutes for long-form captioning and QA, multi-talker speech understanding data covering speaker identification, interruption identification, and target speaker ASR, approximately 1 million samples for multi-audio reasoning across multiple simultaneous audio inputs, and approximately 386K safety and instruction-following samples. Training follows a four-stage curriculum, each with distinct data mixtures and context lengths. Pre-training has two sub-stages: Stage 1 trains only the audio adaptor while keeping both AF-Whisper and the LLM frozen (max audio 30 seconds, 8K token context); Stage 2 additionally fine-tunes the audio encoder while still keeping the LLM frozen (max audio 1 minute, 8K token context). Mid-training also has two sub-stages: Stage 1 performs full fine-tuning of the entire model, adding AudioSkills-XL and newly curated data (max audio 10 minutes, 24K token context); Stage 2 introduces long-audio captioning and QA, down-sampling the Stage 1 mixture to half its original blend weights while expanding context to 128K tokens and audio to 30 minutes. The model resulting from mid-training is specifically released as AF-Next-Captioner. Post-training applies GRPO-based reinforcement learning focusing on multi-turn chat, safety, instruction following, and selected skill-specific datasets, producing AF-Next-Instruct. Finally, CoT-training starts from AF-Next-Instruct, applies SFT on AF-Think-Time, then GRPO using the post-training data mixture, producing AF-Next-Think. One notable contribution from the research team is hybrid sequence parallelism, which makes 128K-context training feasible on long audio. Without it, audio token expansion blows past standard context windows and the quadratic memory cost of self-attention becomes infeasible. The solution combines Ulysses attention — which uses all-to-all collectives to distribute sequence and head dimensions within nodes where high-bandwidth interconnects are available — with Ring attention, which circulates key-value blocks across nodes via point-to-point transfers. Ulysses handles intra-node communication efficiently; Ring scales across nodes. https://arxiv.org/pdf/2604.10905 Benchmark Results: Strong Across the Board On MMAU-v05.15.25, the most widely used audio reasoning benchmark, AF-Next-Instruct achieves an average accuracy of 74.20 vs. Audio Flamingo 3’s 72.42, with AF-Next-Think reaching 75.01 and AF-Next-Captioner pushing to 75.76 — with gains across all three subcategories: sound (79.87), music (75.3), and speech (72.13). On the more challenging MMAU-Pro benchmark, AF-Next-Think (58.7) surpasses the closed-source Gemini-2.5-Pro (57.4). Music understanding sees particularly strong gains. On Medley-Solos-DB instrument recognition, AF-Next reaches 92.13 vs. Audio Flamingo 2’s 85.80. On SongCaps music captioning, GPT5 coverage and correctness scores jump from

You’ve probably heard some version of this idea before: that many of us have an “inner Neanderthal.” That is to say, around 45,000 years ago, when Homo sapiens first arrived in Europe, they met members of a cousin species—the broad-browed, heavier-set Neanderthals—and, well, one thing led to another, which is why some people now carry a small amount of Neanderthal DNA.  This DNA is arguably the 21st century’s most celebrated discovery in human evolution. It has been connected to all kinds of traits and health conditions, and it helped win the Swedish geneticist Svante Pääbo a Nobel Prize. But in 2024, a pair of French population geneticists called into question the foundation of the popular and pervasive theory.  Lounès Chikhi and Rémi Tournebize, then colleagues at the Université de Toulouse, proposed an alternative explanation for the very same genomic patterns. The problem, they said, was that the original evidence for the inner Neanderthal was based on a statistical assumption: that humans, Neanderthals, and their ancestors all mated randomly in huge, continent-size populations. That meant a person in South Africa was just as likely to reproduce with a person in West Africa or East Africa as with someone from their own community.  Archaeological, genetic, and fossil evidence all shows, though, that Homo ­sapiens evolved in Africa in smaller groups, cut off from one another by deserts, mountains, and cultural divides. People sometimes crossed those barriers, but more often they partnered up within them.  In the terminology of the field, this dynamic is called population structure. Because of structure, genes do not spread evenly through a population but can concentrate in some places and be totally absent from others. The human gene pool is not so much an Olympic-size swimming pool as a complex network of tidal pools whose connectivity ebbs and flows over time. This dynamic greatly complicates the math at the heart of evolutionary biology, which long relied on assumptions like randomly mating populations to extract general principles from limited data. If you take structure into account, Chikhi told me recently, then there are other ways to explain the DNA that some living people share with Neanderthals—ways that don’t require any interspecies sex at all. “I believe most species are spatially organized and structured in different, complex ways,” says Chikhi, who has researched population structure for more than two decades and has also studied lemurs, orangutans, and island birds. “It’s a general failure of our field that we do not compare our results in a clear way with alternative scenarios.” (Pääbo did not respond to multiple requests for comment.) The inner Neanderthal became a story we could tell ourselves about our flaws and genetic destiny: Don’t blame me; blame the prognathic caveman hiding in my cells. Chikhi and Tournebize’s argument is about population structure, yes, but at heart, it is actually one about methods—how modern evolutionary science deploys computer models and statistical techniques to make sense of mountains upon mountains of genetic data.  They’re not the only scientists who are worried. “People think we really understand how genomes evolve and can write sophisticated algorithms for saying what happened,” says William Amos, a University of Cambridge population geneticist who has been critical of the “inner Neanderthal” theory. But, he adds, those models are “based on simple assumptions that are often wrong.”  And if they’re wrong, what’s at stake is far more than a single evolutionary mystery.  A captivating story of interspecies passion Back in 2010, Pääbo’s lab pulled off something of a miracle. The researchers were able to extract DNA from nuclei in the cells of 40,000-year-old Neanderthal bones. DNA breaks down quickly after death, but the group got enough of it from three different individuals to produce a draft sequence of the entire Neanderthal genome, with 4 billion base pairs.  As part of their study, they performed a statistical test comparing their Neanderthal genome with the genomes of five present-day people from different parts of the world. That’s how they discovered that modern humans of non-African ancestry had a small amount of DNA in common with Neanderthals, a species that diverged from the Homo sapiens line more than 400,000 years ago, that they did not share with either modern humans of African ancestry or our closest living relative, the chimpanzee.  This model of a Neanderthal man was exhibited in the “Prehistory Gallery” at London’s Wellcome Historical Medical Museum in the 1930s.WELLCOME COLLECTION Pääbo’s team interpreted this as evidence of sexual reproduction between ancient Homo sapiens and the Neanderthals they encountered after they expanded out of Africa. “Neanderthals are not totally extinct,” Pääbo said to the BBC in 2010. “In some of us, they live on a little bit.” The discovery was monumental on its own—but even more so because it reversed a previous consensus. More than a decade earlier, in 1997, Pääbo had sequenced a much smaller amount of Neanderthal DNA, in that case from a cell structure called a mitochondrion. It was different enough from Homo sapiens mitochondrial DNA for his team to cautiously conclude there had been “little or no interbreeding” between the two species.  After 2010, though, the idea of hybridization, also called admixture, effectively became canon. Top journals like Science and Nature published study after study on the inner Neanderthal. Some scientists have argued that Homo sapiens would never have adapted to colder habitats in Europe and Asia without an infusion of Neanderthal DNA. Other research teams used Pääbo’s techniques to find genetic traces of interbreeding with an extinct group of hominins in Asia, called the Denisovans, and a mysterious “ghost lineage” in Africa. Biologists used similar tests to find evidence of interbreeding between chimpanzees and bonobos, polar and brown bears, and all kinds of other animals.  The inner-Neanderthal hypothesis also took a turn for the personal. Various studies linked Neanderthal DNA to a head-spinning range of conditions: alcoholism, asthma, autism, ADHD, depression, diabetes, heart disease, skin cancer, and severe covid-19. Some researchers suggested that Neanderthal DNA had an impact on hair and skin color, while

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. Just before Artemis II began its historic slingshot around the moon, Jared Isaacman, the recently confirmed NASA administrator, made a flurry of announcements from the agency’s headquarters in Washington, DC. He said the US would soon undertake far more regular moon missions and establish the foundations for a base at the lunar south pole before the end of the decade. He also affirmed the space agency’s commitment to putting a nuclear reactor on the lunar surface. These goals were largely expected—but there was still one surprise. Isaacman also said NASA would build the first-ever nuclear reactor-powered interplanetary spacecraft and fly it to Mars by the end of 2028. It’s called the Space Reactor-1 Freedom, or SR-1 for short. “After decades of study, and billions spent on concepts that have never left Earth, America will finally get underway on nuclear power in space,” he said at the event. “We will launch the first-of-its-kind interplanetary mission.” A successful mission would herald a new era in spaceflight, one in which traveling between Earth, the moon, and Mars would—according to a range of experts—be faster and easier than ever. And it might just give the US the edge in the race against China—allowing the country to beat its greatest geopolitical rival to landing astronauts on another planet. While experts agree the timeline is extremely tight, they’re excited to see if America’s space agency and its industry partners can deliver an engineering miracle. “You wake up to that announcement, and it puts a big smile on your face,” says Simon Middleburgh, co-director of the Nuclear Futures Institute at Bangor University in Wales. Little detail on SR-1 is publicly available, and NASA’s own spaceflight researchers did not respond to requests for comment. But MIT Technology Review spoke to several nuclear power and propulsion experts to find out how the new nuclear-powered spacecraft might work. Nuclear propulsion 101 Traditionally, spaceflight has been powered by chemical propulsion. Liquefied hydrogen and liquefied oxygen are mixed, and then ignited, within a rocket; the searingly hot exhaust from this explosion is ejected through a nozzle, which propels the rocket forth. Chemical propulsion offers a significant amount of thrust and will, for the foreseeable future, still be used to launch spacecraft from Earth. But nuclear propulsion would enable spacecraft to fly through the solar system for far longer, and faster, than is currently possible.  “You get more bang per kilogram,” says Middleburgh. A nuclear fuel source is far more energy-dense than its conventional cousin, which means it’s orders of magnitude more efficient. “It’s really, really, really high efficiency,” says Lindsey Holmes, an expert in space nuclear technology and the vice president of advanced projects at Analytical Mechanics Associates, an aerospace company in Virginia.  The approach also removes one other element of the traditional power equation: solar. Spacecraft, including the Artemis II mission’s Orion space capsule, often rely on the sun for power. But this can be a problem, since it doesn’t always shine in space, particularly when a planet or moon gets in its way—and as you head toward the outer solar system, beyond Mars, there’s just less sunlight available.  To circumvent this issue, nuclear energy sources have been used in spacecraft plenty of times before—including on both Voyager missions and the Saturn-interrogating Cassini probe. Known as radioisotope thermoelectric generators, or RTGs, these use plutonium, which radioactively decays and generates heat in the process. That heat is then converted into electricity for the spacecraft to use. RTGs, however, aren’t the same as nuclear reactors; they are more akin to radioactive batteries—more rudimentary and considerably less powerful. So how will a nuclear-reactor-powered spacecraft work?  Despite operational differences, the fundamentals of running a nuclear reactor in space are much the same as they are on Earth. First, get some uranium fuel; then bombard it with neutrons. This ruptures the uranium’s unstable atomic nuclei, which expel a torrent of extra neutrons—and that rapidly escalates into a self-sustaining, roasting-hot nuclear fission reaction. Its prodigious heat output can then be used to produce electricity. Doing this in space may sound like an act of lunacy, but it’s not: The idea, and even a lot of the basic technology, has been around for decades. The Soviet Union sent dozens of nuclear reactors into orbit (often to power spy satellites), while the US deployed just one, known as SNAP-10A, back in 1965—a technological demonstration to see if it would operate normally in space. The aim was for the reactor to generate electricity for at least a year, but it ran for just over a month before a high-voltage failure in the spacecraft caused it to malfunction and shut down.  Now, more than half a century later, the US wants its second-ever space-based nuclear reactor to do something totally different: power an interplanetary spacecraft. To be clear, the US has started, and terminated, myriad programs looking into nuclear propulsion. The latest casualty was DRACO, a collaboration between NASA and the Department of Defense, which ended in 2025. Like several previous efforts, DRACO was canceled because of a mix of high experimentation costs, lower prices for conventional rocket propulsion, and the difficulty of ensuring that ground tests could be performed safely and effectively (they are creating an incredibly powerful nuclear reaction, after all). But now external considerations may be changing the calculus. The Artemis program has jump-started America’s return to the moon, and the new space race has palpable momentum behind it. The first nation to deploy nuclear propulsion would have a serious advantage navigating through deep space.  “I think it’s a very doable technology,” says Philip Metzger, a spaceflight engineering researcher at the Florida Space Institute. “I’m happy to see them finally doing this.” One version of this technology is known as nuclear thermal propulsion, or NTP. You start with a nuclear reactor, one that’s cooking at around 5,000°F. Then “you’ve

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Want to understand the current state of AI? Check out these charts.  If you’re following AI news, you’re probably getting whiplash. AI is a gold rush. AI is a bubble. AI is taking your job. AI can’t even read a clock. Stanford’s 2026 AI Index—the field’s annual report card—cuts through the noise.   The data reveals a technology evolving faster than we can manage. From the China-US rivalry and model breakthroughs to public sentiment and the impact on jobs, here are the index’s key findings on the state of AI today.  —Michelle Kim  Why opinion on AI is so divided  Stanford’s 2026 AI Index is full of striking stats. It also reveals a field riddled with inconsistencies, most notably in the gap between experts and non-experts.   On jobs, 73% of US experts view AI’s impact positively, compared to just 23% of the public. Similar divides emerged on the economy and healthcare. What’s driving this disconnect?  Part of the answer may lie in their diverging experiences. Those using AI for coding and technical work see it at its best, while everyone else gets a more mixed bag. The result is two very different realities. Read the full story on what they are—and why they matter.  This story is from The Algorithm, our weekly newsletter on AI. Sign up to receive it in your inbox every Monday.  —Will Douglas Heaven  Job titles of the future: Wildlife first responder  Grizzly bears have made such a comeback across eastern Montana that in 2017, the state hired its first-ever prairie-based grizzly manager: wildlife biologist Wesley Sarmento.   For seven years, Sarmento worked to keep both bears and humans out of trouble. He acted like a first responder, trying to defuse potentially dangerous situations. He even got caught in some himself, which led him to a new wildlife safety tool: drones. Find out the results of his experiments in digital ecology.   —Emily Senkosky  This article is from the next issue of our print magazine, which is all about nature. Subscribe now to read it when it lands on Wednesday, April 22.   The must-reads  I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.  1 Human scientists still trounce the top AI agents at complex tasks  The best agents perform only half as well as experts with PhDs. (Nature) + Can AI really help us discover new materials? (MIT Technology Review)  2 OpenAI is escalating its fight with Anthropic while pulling away from Microsoft A leaked memo exposes plans to attack Anthropic. (Axios) + And says Microsoft “limited our ability” to reach clients. (The Information $) + While touting a budding alliance with Amazon. (CNBC)  3 Carbon removal technology is stalling—and that may be good news Better solutions could now emerge. (New Scientist) + Here are three that are set to break through. (MIT Technology Review)  4 AI is finding bugs faster than we can fix them—and hackers will benefit Welcome to the bug armageddon. (WSJ $)  + AI may soon be capable of fully automated attacks. (MIT Technology Review)  5 A Texas man has been charged with the attempted murder of Sam Altman He allegedly threw a Molotov cocktail at the OpenAI CEO’s home last Friday. (NPR) + The suspect reportedly had a list of other AI leaders. (NYT $)  6 AI is beginning to transform mathematics It’s proving new results at a rapid pace. (Quanta) + One AI startup plans to unearth new mathematical patterns. (MIT Technology Review)  7 Students are turning away from computer science It’s had a massive drop in enrollments. (WP $) + AI coding tools have diminished the degree’s value. (NYT $)   8 India’s bid to become a data center hub is sparking a fierce backlash Farmers are protesting Delhi’s courtship of hyperscalers. (Rest of World)  9 Meta is set to overtake Google in advertising revenue this year And become the world’s largest digital ad platform for the first time. (WSJ)  10 AI influencers are taking over Coachella  Synthetic content creators are “everywhere” at the festival. (The Verge)  Quote of the day  “These people are almost nothing like you. They are most likely sociopathic/psychopathic and, in the case of Altman, consistently reported to be a pathological liar.”  —The alleged firebomber of Sam Altman’s home shares his distrust of AI leaders in a blog post.  One More Thing  We’ve never understood how hunger works. That might be about to change.  A few years ago, Brad Lowell, a Harvard University neuro­scientist, figured out how to crank the food drive to the maximum. He did it by stimulating neurons in mice. Now, he’s following known parts of the neural hunger circuits into uncharted parts of the brain.  The work could have important implications for public health. More than 1.9 billion adults worldwide are overweight, and more than 650 million are obese. Understanding the circuits involved could shed new light on why these numbers are skyrocketing.  Read the full story.  —Adam Piore  We can still have nice things  A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line.)  Top image credit: Stephanie Arnett/MIT Technology Review | Getty Images  + Someone built a mechanical version of Tony Hawk’s Pro Skater from Lego. + Enjoy this wholesome clip of toddlers discovering the existence of hugs. + This interactive body map shows exactly which exercises you need. + Jon McCormack’s photos of nature’s patterns are breathtaking. 

Language models (LMs), at their core, are text-in and text-out systems.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. You have no choice in reading this article—maybe How do humans make decisions? The question has been on Uri Maoz’s mind since he read an article in his early twenties suggesting that… maybe they didn’t.   Had he even had a choice about whether to read that article in the first place? How would he ever know if he was truly responsible for making any decisions? “After that, there was no turning back,” says Maoz, now a professor of computational neuroscience at Chapman University.  Today, Maoz is a central figure in efforts to understand how desires and beliefs turn into actions. He’s also uncovered new wrinkles in the debate. Read the full story on his discoveries. —Sarah Scoles This article is from the next issue of our print magazine, packed with stories all about nature. Subscribe now to read the full thing when it lands on Wednesday, April 22. What’s in a name? Moderna’s “vaccine” vs. “therapy” dilemma  Moderna, the covid-19 shot maker, is using its mRNA technology to destroy tumors through a very, very promising technique known as a cancer vacc—  “It’s not a vaccine,” a spokesperson for Merck said before the V-word could be uttered. “It’s an individualized neoantigen therapy.”  Oh, but it is a vaccine, and it looks like a possible breakthrough. But it’s been rebranded to avoid vaccine fearmongering—and not everyone is happy about the word game. Read the full story.  —Antonio Regalado This article is from The Checkup, our weekly newsletter covering the latest in biotech. Sign up to receive it in your inbox every Thursday.  The must reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 Sam Altman’s home has been attacked twice in two days A driver reportedly fired a gun at his property on Sunday. (SF Standard) + A Molotov cocktail was thrown at his home on Friday. (NBC News) + The suspect wrote essays warning AI would end humanity. (SF Chronicle) + The attacks expose growing divides in opinion on AI. (Axios)  2 AI weapons are ushering in a new kind of arms race Countries are racing to deploy AI in military systems. (NYT $) + The Pentagon wants AI firms to train on classified data. (MIT Technology Review) + Where OpenAI’s technology could show up in Iran. (MIT Technology Review)  3 Artemis II was a success Astronauts did an array of experiments that will be crucial to the future of both the program itself and deep-space missions. (Guardian) + But next steps for the Artemis missions are uncertain. (Ars Technica)  4 OpenAI and Elon Musk are heading toward a massive courtroom clashThe company has accused Musk of a “legal ambush.” (Engadget) + He’s lost a streak of cases ahead of the showdown. (FT $)  5 AI job fears in China are fueling a viral “ability harvester” project It claims to turn human skills into AI tools. (SCMP) + Hustlers are cashing in on China’s OpenClaw AI craze. (MIT Technology Review)  6 Governments are hiding information about the Iran war online Through restrictions on internet access and satellite imagery. (NPR)   7 Apple is testing four smart glasses that could rival Meta Ray-Bans They’re part of a broader wearables strategy. (Bloomberg $)  8 Meta is building an AI version of Mark Zuckerberg to interact with staffIt’s being trained on his mannerisms, voice, and statements. (FT $)  9 Anthropic is asking Christian leaders for guidance It’s seeing advice on building moral machines. (WP $) + AI agents have spread their own religions. (MIT Technology Review)  10 A dancer with MND is performing again through an avatar Her brainwaves powered the digital dancer. (BBC)  Quote of the day “Earth was this lifeboat hanging in the universe.” —Artemis II astronaut Christina Koch describes her view of Earth from space, the Guardian reports. One more thing RAVEN JIANG How AI and Wikipedia have sent vulnerable languages into a doom spiral When Kenneth Wehr started managing the Greenlandic-language version of Wikipedia, he discovered that almost every article had been written by people who didn’t speak the language.   A growing number of them had been copy-pasted into Wikipedia from machine translators—and were riddled with elementary mistakes. This is beginning to cause a wicked problem.  AI systems, from Google Translate to ChatGPT, learn new languages by scraping text from Wikipedia. This could push the most vulnerable languages on Earth toward the precipice.  Read the full story on what happens when AI gets trained on junk pages.  —Jacob Judah  We can still have nice things A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Hungary’s next health minister can throw some serious shapes.  + Here’s a welcome route to an AI-free Google search. + Movievia eschews endless scrolling to find the right film for your needs+ A photography trick has turned a giant glacier into a tiny, living diorama.

If you’re following AI news, you’re probably getting whiplash. AI is a gold rush. AI is a bubble. AI is taking your job. AI can’t even read a clock. The 2026 AI Index from Stanford University’s Institute for Human-Centered Artificial Intelligence, AI’s annual report card, comes out today and cuts through some of that noise.  Despite predictions that AI development may hit a wall, the report says that the top models just keep getting better. People are adopting AI faster than they picked up the personal computer or the internet. AI companies are generating revenue faster than companies in any previous technology boom, but they’re also spending hundreds of billions of dollars on data centers and chips. The benchmarks designed to measure AI, the policies meant to govern it, and the job market are struggling to keep up. AI is sprinting, and the rest of us are trying to find our shoes. All that speed comes at a cost. AI data centers around the world can now draw 29.6 gigawatts of power, enough to run the entire state of New York at peak demand. Annual water use from running OpenAI’s GPT-4o alone may exceed the drinking water needs of 12 million people. At the same time, the supply chain for chips is alarmingly fragile. The US hosts most of the world’s AI data centers, and one company in Taiwan, TSMC, fabricates almost every leading AI chip.  The data reveals a technology evolving faster than we can manage. Here’s a look at some of the key points from this year’s report.  The US and China are nearly tied In a long, heated race with immense geopolitical stakes, the US and China are almost neck and neck on AI model performance, according to Arena, a community-driven ranking platform that allows users to compare the outputs of large language models on identical prompts. In early 2023, OpenAI had a lead with ChatGPT, but this gap narrowed in 2024 as Google and Anthropic released their own models. In February 2025, R1, an AI model built by the Chinese lab DeepSeek, briefly matched the top US model, ChatGPT. As of March 2026, Anthropic leads, trailed closely by xAI, Google, and OpenAI. Chinese models like DeepSeek and Alibaba lag only modestly. With the best AI models separated in the rankings by razor-thin margins, they’re now competing on cost, reliability, and real-world usefulness.  The index notes that the US and China have different AI advantages. While the US has more powerful AI models, more capital, and an estimated 5,427 data centers (more than 10 times as many as any other country), China leads in AI research publications, patents, and robotics.  As competition intensifies, companies like OpenAI, Anthropic, and Google no longer disclose their training code, parameter counts, or data-set sizes. “We don’t know a lot of things about predicting model behaviors,” says Yolanda Gil, a computer scientist at the University of Southern California who coauthored the report. This lack of transparency makes it difficult for independent researchers to study how to make AI models safer, she says. AI models are advancing super fast Despite predictions that development will plateau, AI models keep getting better and better. By some measures, they now meet or exceed the performance of human experts on tests that aim to measure PhD-level science, math, and language understanding. SWE-bench Verified, a software engineering benchmark for AI models, saw top scores jump from around 60% in 2024 to almost 100% in 2025. In 2025, an AI system produced a weather forecast on its own.   “I am stunned that this technology continues to improve, and it’s just not plateauing in any way,” says Gil. However, AI still struggles in plenty of other areas. Because the models learn by processing enormous amounts of text and images rather than by experiencing the physical world, AI exhibits “jagged intelligence.” Robots are still in their early days and succeed in only 12% of household tasks. Self-driving cars are farther along: Waymos are now roaming across five US cities, and Baidu’s Apollo Go vehicles are shuttling riders around in China. AI is also expanding into professional domains like law and finance, but no model dominates the field yet.  But the way we test AI is broken These reports of progress should be taken with a grain of salt. The benchmarks designed to track AI progress are struggling to keep up as models quickly blow past their ceilings, the Stanford report says. Some are poorly constructed—a popular benchmark that tests a model’s math abilities has a 42% error rate. Others can be gamed: when models are trained on benchmark test data, for example, they can learn to score well without getting smarter.  Because AI is rarely used the same way it’s tested, strong benchmark performance doesn’t always translate to real-world performance. And for complex, interactive technologies such as AI agents and robots, benchmarks barely exist yet.  AI companies are also sharing less about how their models are trained, and independent testing sometimes tells a different story from what they report. “A lot of companies are not releasing how their models do in certain benchmarks, particularly the responsible-AI benchmarks,” says Gil. “The absence of how your model is doing on a benchmark maybe says something.”  AI is starting to affect jobs Within three years of going mainstream, AI is now used by more than half of people around the world, a rate of adoption faster than the personal computer or the internet. An estimated 88% of organizations now use AI, and four in five university students use it.  It’s early days for deployment, and AI’s impact on jobs is hard to measure. Still, some studies suggest AI is beginning to affect young workers in certain professions. According to a 2025 study by economists at Stanford, employment for software developers aged 22 to 25 has fallen nearly 20% since 2022. The decline might not be pinned on AI alone, as broader macroeconomic conditions could be to blame, but AI appears to be playing