Deep inside a White House proposal to overhaul how the government awards grants is a short section that health disparities researchers say could disqualify much of their work from federal funding — perhaps the most serious threat yet to the future of their field.
Since it was issued last month, the 412-page proposal has drawn thousands of public comments, in large part because it would diminish the power of peer review in determining which grants are funded, and place that authority with political appointees. Less attention has been given to how the proposed changes could affect efforts to eliminate health disparities research.
It’s cold, it’s very, very noisy, and—if I can be quite honest with you—I’m not feeling super relaxed.
I’m currently around 300 meters, or 1,000 feet, beneath the North Sea, in a dark, dank cave. It smells weird. And I am increasingly aware of the pressure from millions of tons of seawater just above my head, pushing down with a force of more than 500 pounds per square inch. Picture a baby rhino standing on a postage stamp.
Only fabulous engineering is keeping me from being crushed, drowned, disappeared. My safety goggles are foggy.
Just a few hundred meters away, someone is about to blow up a giant rock wall. Luckily, earlier that day I was given a full safety briefing, and I’ve got a special hard hat on. “Don’t worry—if you don’t make it, we’ll have your stuff sent back to your office,” geologist Anne-Merete Gilje tells me, straight-faced. Ah, Norwegian humor.
“It’s kind of a lifestyle. You have to be a little bit crazy to work underground all the time.”
Niclas Brusehed, tunnel foreman, Implenia
I’m in this odd situation under the iconic fjords of Norway to visit what will soon become the world’s longest and deepest subsea road tunnel, called Rogfast (short for “Rogaland Fixed Link”). I want to understand how you make something as audacious as a 26.7-kilometer (16.6-mile) highway that sits 390 meters (1,280 feet) below the sea at its deepest point. And also—at a time when it can feel hard to get anything done, especially in the US—to reassure myself that ambitious engineering is still possible. That we can still make things.
The Norwegians already have the world’s longest subsea tunnel, the 14.4-kilometer Ryfylke, though Rogfast will dwarf it. Their expertise has attracted attention from Japan, Spain, Morocco, and even a number of US states, whose representatives were due to visit the site in May, just weeks after I went. They, too, want to know how Norway does it.
The answer: tons of explosives.
The entire endeavor feels like an obstinate refusal to give in to physics and geology. “It’s always exciting,” Niclas Brusehed, a tunnel foreman at Implenia, a Swiss firm involved in the project, tells me. “Every blast creates a new world.” There’s not just the blasting of the tunnel itself—although that is an epic project on its own—but an immense logistics challenge involving huge ventilation shafts, extreme pressure, underground roundabouts, and the complex Norwegian geology. Oh, and the water. So much water.
“This is the longest continuous blast on the sea,” says John Olaf Østerhus, assistant project manager at Implenia. “Never been done before. We can’t buy a book to see how we do this.”
All right, time to fish my phone out of my safety suit—don’t want to forget this.
On another planet
Arriving at the rock face where the tunnel hits seabed feels like being on the moon. It’s a huge slab of stone at the end of a long, dark, wet, wide passageway that’s lit (barely) by electric lights. Giant vehicles carting tons of rocks rumble past periodically, and we pull to the side of the road to let them by.
Rescue chambers are spread throughout the tunnel network.
COURTESY OF NORWEGIAN PUBLIC ROADS ADMINISTRATION
Workers clock in for 12-hour shifts, 6 a.m. until 6 p.m., deep in the bowels of the Earth where no natural light can reach. Twelve days on, 16 days off. They eat their lunch at a table in this damp cave surrounded by portacabins plastered with safety notices. “It’s kind of a lifestyle,” says Brusehed, laughing. “You have to be a little bit crazy to work underground all the time.”
These crazy engineers are here to make tunnels the Norwegian way. The nation frequently uses what’s known as the drill-and-blast method instead of the tunnel-boring machines that are more typical elsewhere. This approach offers more flexibility for long, complex operations with varied rock types. Each blast adds about five to six meters to the tunnel.
Rogfast is being built inward from the ends to speed things up. The construction company Skanska is leading from the north, coming from the island of Vestre Bokn; Implenia has joined a company called Stangeland to tunnel from Randaberg in the south, which is where I am. Both teams use multiple laser scans each day to consistently measure their orientation and check that the tunnel is exactly where it should be. The two ends should meet sometime in 2029, with no more than just a few centimeters of deviation.
The caves are like towering cathedrals, scattered with rubble.
COURTESY OF NORWEGIAN PUBLIC ROADS ADMINISTRATION
Norway has constructed more than a thousand kilometers of tunnels over the past several decades. The depth and length of these make the best efforts to date of Elon Musk’s Boring Company—a mere 2.7-kilometer tunnel in Las Vegas that is just 3.6 meters wide—look rather pathetic. The country’s spectacular setting makes such builds necessary; while Norwegians are proud of having the second-longest coastline in the world after Canada, getting up and down the west coast requires multiple ferry rides between islands, which can move extra slowly when the weather’s bad.
After it’s completed, which is scheduled to happen in 2033, Rogfast should help eliminate two ferry routes and cut the five-hour journey between the southwestern cities of Stavanger and Bergen by 40 minutes. It will funnel four lanes of traffic deep beneath the fjords of Boknafjord and Kvitsøyfjord, and at one section a relatively scant 50 meters of rock will separate the drivers speeding through the tunnel from the bottom of the North Sea. There are also, delightfully, two undersea roundabouts located 220 meters below sea level.
But the first job is to contend with all that water.
The never-ending battle
Subsea tunneling is defined by a constant, ultimately unwinnable battle with the ocean. The sheer weight of the sea above you, and the crushing pressure, means the water will always find a way in. “It’s the volume and the pressure that’s the biggest risk,” says Ole Magne Rønning, project leader for Implenia/Stangeland.
So before tunnel engineers blow stuff up, they need to check for leaks. Into the rock face ahead of them, they drill a number of narrow holes that go 25 to 30 meters deep to see how much water comes through. Even a small probe can unleash a torrent within seconds, says Rønning. When road traffic eventually rumbles through these tubes, water will still trickle from the rocks; it will be redirected into mini reservoirs dotted throughout the tunnel network before being pumped back out.
Since stopping the water entirely is impossible, the game is instead to push it away as best you can. If the leakage in front of the rock face exceeds a certain limit—around four liters per hole per minute—then the next stage is “grouting”: pumping a mixture of cement-like sludge into new holes that fan out in the ceiling above and around the face. Ideally, you address the leaks that are ahead of you; “it’s a lot more difficult to stop a leak that’s behind you,” says Rønning.
At one point deep below the sea, I chat with Tarald Johan Nomeland, the project’s grouting specialist. He’s big and bearded, perhaps one of the most Norwegian-looking men I’ve ever met. He stands, towering above me, and shakes my hand in his giant bear-like paw. Grouting is in Nomeland’s family; his dad did it too. He loves it. “There’s not necessarily just one solution to a problem,” he says, eyes flashing with delight as he describes fighting the interminable battle with the water. “There may be many solutions.”
The amount of grouting needed determines how fast the project can move. On the Skanska side, for example, some weeks the face moves 30 meters; others, as few as 10.
This isn’t made any easier by the rock itself. The seabed around Norway was shaped by glaciers during the Ice Age. As the ice retreated, it dragged softer rock with it, carving out the fjords for which the nation is so famous. But this legacy makes digging subsea tunnels particularly gnarly. Much of what’s left is the hard, difficult-to-break stuff.
And it’s not just one type of rock, either. There are “big wide areas where we don’t know what’s down there,” says Gilje, the geologist who is a project manager for the Norwegian Public Roads Administration, which is in charge of the entire project. Before any construction started, boats took core samples from the seabed along the planned tunnel route. Seismic surveys from the ocean surface—like those that look for oil in the region—helped fill in the gaps.
Each kind of rock presents its own challenges, so the engineers “have different techniques for different problems,” Gilje explains. For example, they found that one southern section contains a lot of phyllite. Phyllite is considered “nice” to work with. It is formed from a combination of shale, siltstone, and mud over time and is pretty compact, with few cracks to let water through. Its compact nature means it requires more explosives per blast, however. It also contains a lot of quartz, which is toxic when released into the air during blasting. So workers wear monitors to measure their exposure, and a curtain of water sprayed in front of the rock face helps prevent too much from drifting into the tunnel.
The northernmost part of the route, meanwhile, is made mostly of solid granite and a similar rock called gneiss. Both are hard but contain fractures that allow the seawater to trickle through.
The rock type can also change over just a short distance. So during the dig, every 80 meters or so, an engineer sends sound waves through the face to expose its secrets and help evaluate its structural integrity. The rock is graded on a scale of 1 to 5, with 5 being the worst and least stable. “When you are reaching class 5, then it’s almost like soil. It’s not rock anymore,” says Rønning.
This investigation informs the kinds of structural supports each section will need—from steel rods that fan out above the rock face like an umbrella, for the strongest rock, to reinforced-concrete arches that hold up the weakest. To seal everything off, the team sprays a substance called “shotcrete,” liquid concrete mixed with reinforced-steel fibers, onto the walls throughout. A plastic membrane and concrete panels are fitted later.
“It’s going to be a very safe tunnel,” Gilje says. “It’s going to last for 100 years.”
Strange dangers
While I may not be brave, at least I don’t get seasick. Back at the surface, I board a small ferry that putters and sloshes its way from the mainland to Kvitsøy, a sparsely populated municipality made up of 365 separate islands and islets—something its 550 or so inhabitants are very proud of, even though most of these islands are uninhabited chunks of rock.
For the next few years, Kvitsøy’s population will experience a tiny boom as its largest island hosts a semipermanent encampment of contractors and engineers working on what is probably the most complex part of the Rogfast project: the giant ventilation shafts that will sit roughly halfway along the tunnel’s length to bring fresh air into the entire network, and remove the stale air in turn.
It’s also one of the reasons why road tunnels are much more complex than rail tunnels. Cars pump out fumes that have to be vented away. During construction, fresh air flows in via huge plastic tubes suspended from the ceiling, but eventually, Rogfast’s air will come in through two nine-meter-wide shafts that will bore down from Kvitsøy’s surface: one to bring it in, one to take it out.
Hundreds of steel rods are fitted to support the ceiling and walls.
COURTESY OF NORWEGIAN PUBLIC ROADS ADMINISTRATION
Creating these shafts is a wild process. First, narrow boreholes are drilled from the ground down into the tunnel 210 meters below the surface. A vertical drill rig is then pulled up through the hole from the bottom, widening the shaft to 2.4 meters as it ascends.
Then explosives are set off on the island’s surface, bashing down through the rock to widen the shaft. A large digger pushes the resulting debris down the narrower, not-yet-exploded length of shaft below, sending rocks barreling toward the tunnel at the bottom like socks tumbling down a laundry chute. Trucks haul away the fallen rocks. This process happens in stages, repeating at regular intervals, opening up the passage a bit deeper with each pass. Once it’s all done, steel rods are installed in the shaft’s walls to keep it secure.
Down below, I stand beneath one of the narrow guide holes for one of the two ventilation shafts. The ceiling soars overhead—a strangely beautiful cathedral, cragged and shadowed by lamplight.
Besides poisonous air, the epic nature of these engineering projects throws up other surprising dangers. For example, Rogfast will take about 30 minutes to drive through. It doesn’t seem that long, but the project’s designers worry that the monotonous environment may lull some drivers to sleep.
Engineers faced this problem with Ryfylke—which, as the current longest subsea road tunnel, has been a testing ground for its bigger sibling. It relieves the tedium with a large hall that opens up in the middle of the tunnel, lit by colored lights that change each day. When Rogfast is finished, artists will be invited to do something similar, using lights, colors, and shapes to keep drivers alert.
Then there are the environmental risks. What is there to do with all the loose rock created by the blasts? The engineers predict 8.5 million cubic meters’ worth. That’s enough to fill more than 2,500 Olympic-sized swimming pools. The solution is to bring it back to the surface, where it can be used to create new land. To do this, the project employs a giant barge designed to split open and dump 350 tons of rock in one go.
But adding more rock particles to the water can make it hard for fish to breathe, says Elizabeth Austdal Paulen, Implenia’s environment lead on the project and my fellow passenger on the windy (and soon to be redundant) ferry over to Kvitsøy. Her team monitors their levels in real time: If the particulate count is too high, the drops must pause until the new rock has settled on the seabed. The goal is to protect lobster fishing, a vital part of the local economy, and to safeguard the breeding time for cod, which was an issue when I visited.
Finally, of course, on top of all this are the many hazards for the people who are actually doing all this blasting and digging and hauling. Or, say, for the visitors who are finding their inner nine-year-old getting a little too giddy about what’s next.
Time to blow
Before I’m allowed underground, I must sit through a short safety briefing, where I learn there are multiple hazards when you’re that deep. Fires, for instance, can break out, exacerbated by the way the salt water affects electronics. Just a week earlier, a car caught fire somewhere deep within the network. “You have to be aware all the time,” says Anne Brit Moen, the project lead for Skanska. “It’s a very harsh, harsh climate.”
After the session, I’m given a hi-viz suit, the hard hat (which has built-in ear protectors), gloves, safety glasses, and reinforced boots. I get instructions on how to operate the oxygen mask that will be in the car with me, and a device to put in my pocket that will track my exact location on screens in the control room. The device also acts as a personal warning system: If it vibrates and a blue light appears, then a blast is imminent and I must get to safety; if it vibrates and glows red, umm, well, that’s bad news and it’s time to evacuate.
“If you’re the first to the rescue chamber, press the green button … close the hatch and sit down and be calm.”
Ketil Myklebost, project manager, Implenia
But let’s say I can’t—I’m too deep underground. Then there is a second, less fun option. I’m given instructions on how to access the rescue chambers. These metal boxes—about the size of a large van—can squeeze in around 16 people, and each contains chocolate, water, radio equipment, a defibrillator, and enough oxygen for 24 hours. I see them dotted throughout the tunnels as we drive through. Worst-case scenario, I’m supposed to get to the nearest one, sit tight, and hope to get rescued.
“If you’re the first to the rescue chamber, press the green button for 15 seconds to release pressure,” says Ketil Myklebost, a project manager at Implenia. “And then close the hatch and sit down and be calm.”
Calm, right. Okay.
In the hours before my visit, a huge drilling “jumbo” rig puts as many as 180 holes deep into the rock face. The number, angle, depth, and spacing of the holes is calculated in advance using software but finalized at the face—here, they’re almost six meters deep. At one point, I clamber up into the jumbo and inspect the pattern on its screen, matching it against what I can see on the huge rock face, which stands more than 12 meters tall and wide.
The holes have been stuffed with an explosive slurry. (Someone quips that if I get any on my clothes, I’ll be stopped at the airport as a terrorist. A Norwegian joke, again.) As I watch, workers in a kind of cherry picker fit each hole with a detonator and make sure they’re all connected to one another by wire, ready to be triggered remotely.
Then my personal safety device starts vibrating. When I take it out of my pocket, it’s blinking blue. Showtime.
How far back do I need to be? “It’s dangerous in this direction 500 to 600 meters, but if you’re around the corner you can be closer,” says Sveinung Brude, project manager for the Norwegian Public Roads Administration.
Workers lay asphalt.
COURTESY OF NORWEGIAN PUBLIC ROADS ADMINISTRATION
I stand by the worker who will trigger the blast from what looks like a small briefcase with an antenna. Then he presses the button.
The shock wave hits me before I hear it. My chest vibrates. In the first few milliseconds, a propulsive thump briefly stuns my senses, followed immediately by a rolling, crumpled thunder.
Just a moment later—almost instantly, really—wind billows through the cavern. Rocks clatter as they crash off the walls. I try not to show any panic. (That was meant to sound like that, right?) A hush falls, and there’s just the tinkling of stones as they bounce and skip amid the rubble.
Dust rises into the air, and there is a strange smell.
Through my ear protection it sounds like the end of the world.
Niall gets to grips with what it’s like to work under the sea.
NORWEGIAN PUBLIC ROADS ADMINISTRATION
The explosion itself is a beautiful choreography: Blasts are initiated one after another, starting from the center. In video footage, you can just about hear the sequential pitter-patter of the charges as they go off. (In person, it’s a bit more all-at-once and overwhelming.)
Rogfast has just crept another few meters closer to completion.
I find myself grinning. Maybe there’s something extremely primal about being near an explosion? I’m not sure. I look down at my hand, where I have my phone out, recording the intensity of the moment.
Except … I wasn’t recording. The stupid rubber safety gloves I’m wearing must have stopped the command from going through.
Oh no. Oh no.
A once-in-a-lifetime opportunity, and I, uh, blew it. “I WASN’T RECORDING!” I shriek.
“It’s better that way,” says Rønning, walking off into the gloom. “You’ll remember it.” How very Norwegian.
Conditions: Asperger’s Disorder; Asperger Disorder; Autism Disorder; Autism; ADHD – Attention Deficit Disorder With Hyperactivity; ADHD; ASD; Alcohol Abuse/Dependence; Alcohol Addiction; Alcohol and Other Drug Use Disorders; Alcohol and Other Substance Use Prevention; Gambling Addiction; Gambling Disorder; Sex Abuse; Sex Behavior; Sex Crimes; Sex Disorder; Sex Disorders; Gender Dysphoria, Adult; Eating Behavior Disorders; Narcotic-Related Disorders; Narcotic Addiction; Narcissism; Psychiatric Disorder; Psychedelic Effects in Healthy Volunteers; Psychedelic Experiences; Psychedelic Drug Dependence; Marijuana Use Disorder; Marijuana Abuse and Dependence; Smoking (Tobacco) Addiction; Smoking Among Youth; Smoking Abstinence; Abstinence, Sex; Opiate Substitution Treatment; Opioid Abuse (Disorder); Opioid Abuse and Addiction; Cocaine Abuse; MDMA (‘Ecstasy’); Addiction Disorders; Homeless and Low Incomes People, Refugees; Homelessness; Reliability and Validity; Anger Problems; Child Abuse, Sexual
Interventions: Behavioral: AI City Hall Project (AIhealth4u – Public Health Central); Behavioral: AI City Hall Project (QAIAx Microcity A – AI Public Health Central)
Sponsors: Veterans Recovery Network Inc.; U.S. Special Operations Command; Central Virginia VA Health Care System; AI-119 Vulcan Project Research & Educational Technology Company (fka Henry Nanpei Academy Project)
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.
A startup claims it broke through a bottleneck that’s holding back LLMs
AI startup Subquadratic came out of stealth last month with a huge claim: it had solved a mathematical bottleneck that had held back large language models for almost a decade.
The purported breakthrough comes from slashing the number of computations transformers need to carry out to generate answers. The result is a faster and cheaper LLM that uses far less energy than any other model on the market.
Many experts remained skeptical—but Subquadratic has started to share the receipts. They suggest that their approach might be worth paying attention to.
This week, I covered the story of Casey Harrell—a man with ALS who is “the first power user” of a brain implant. The device has enabled him to maintain an income, reconnect with friends and family, and read to his daughter. He told me that it’s “nothing short of revolutionary.”
Over the past couple of years, the number of BCI trial volunteers has soared. This year, China became the first country to approve a BCI for medical use. Advances in technology are allowing engineers to provide more features than ever. BCI research is properly taking off.
This story is from The Checkup, our weekly newsletter giving you the inside track on all things 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 Amazon workers who backed data center limits may face termination The engineers say they’re under investigation by the company. (NYT $) + And could face discipline, including potential termination. (The Verge) +They had testified at meetings about pausing data centers. (CNBC) + They’ve filed a joint complaint to Seattle’s Office for Civil Rights. (Wired $)
2 A new fossil discovery has rewritten 150 years of evolutionary theory It suggests early land vertebrates skipped the tadpole stage. (New Scientist $) + And raises questions about how vertebrates adapted to land. (404 Media) + Sponges may have been the first animals. (MIT Technology Review)
3 Bernie Sanders plans to give the public direct ownership of AI firms He’s unveiled new legislation to create an AI sovereign wealth fund. (AP News) + It would be funded through a one-time tax on AI companies’ stock. (Quartz) + And make annual payments directly to Americans. (Washington Post $)
4 Investors in China secretly acquired stakes in SpaceX before its IPO One had ties to Chinese military contractors. (ProPublica) + The US fears China has got one of ASML’s top machines. (Reuters $)
5 Researchers have figured out Russia’s nuclear-powered missile They call it “a terrible idea”—but not an impossible one. (NPR) + NASA is building a nuclear reactor-powered spacecraft. (MIT Technology Review)
6 Longevity medicine faces a do-or-die moment in a landmark trial It will test whether cellular aging can be safely reversed in humans. (Axios) + The next step is “chemical reprogramming.” (MIT Technology Review)
7 Studies suggest AI may already be deskilling professionals Over-reliance appears to weaken doctors’ and engineers’ abilities. (Nature)
8 Tech workers who maxed out their AI use are now trying to minimize it Spiralling costs mean “tokenminning” has replaced “tokenmaxxing.” (NYT $)
9 Scientists say the human genome’s structure may confound AI models Which would constrain AI-based models of biology and disease. (Quanta)
10 A new robotic self-driving toilet brings the bathroom to you The Xiaoban also cleans up and empties itself all on its own. (The Verge)
Quote of the day
“They hated me. They were doing everything they could to knock me down. And look at them now.”
—Donald Trump mocks Mark Zuckerberg and Jeff Bezos in a conversation with Elon Musk that’s recounted in a new book, Wired reports.
One More Thing
PABLO DELCAN
Technology can help us feed the world, if we look beyond profit
The pandemic exposed the weak spots in our interconnected food system. They’re the result of decades’ worth of technological advances, from globe-spanning shipping to refrigeration networks. But technology is not inherently opposed to sustainable and resilient food systems.
Powerful technologies like genetic modification can create stronger local agriculture and a healthier food system—but they normally aren’t. The challenge is ensuring they serve food security and human well-being, rather than simply maximizing profits.
IntroductionResidents have a high risk of burnout during the training period. The COVID-19 pandemic has had a great impact on work conditions and health of medical residents. We therefore estimated the prevalence of post-pandemic burnout and perceived health, and their relationship with psychosocial factors, stress profile, and job satisfaction, as well as the pre-post pandemic changes.MethodsWe carried out a retrospective cross-sectional, online, anonymous survey using the validated FPSICO-4.0 questionnaire in a teaching hospital after the COVID-19 pandemic (2023). We also studied the pre-post changes from an independent pre-pandemic online survey (2018) with the same tool.ResultsOne hundred ninety-seven residents, 64% females, answered the questionnaire (43.5%); 28% had post-pandemic burnout without gender differences, being higher in surgical specialties (p=0.017) and along residency years (p=0.041). Role performance, behavior stress, job satisfaction and residency years were associated with an increase in odds of burnout (p<0.05). Supervision, conflict of role/cognitive stress explained 35% of the variation in perceived general health (adjusted-R2 = 0.356); temporary autonomy, supervision, information/training/promotion, behavior/somatic stress, and job satisfaction did 68% of mental health (adjusted-R2 = 0.683); and take decisions, conflict of role, behavior/somatic/cognitive stress, and male gender did 62% of vitality (adjusted-R2 = 0.624). Pre/post pandemic mean differences, adjusted by gender and residency year, showed deterioration of variety sense of work (p<0.001), effort of attention (p=0.026), behavior stress (p=0.022), and improvement of temporary autonomy (p=0.045).ConclusionsThe identification after pandemic of high prevalence of burnout, stress levels and certain psychosocial risk factors (role performance and supervision), highlights the need of specific health promotion plan.
Miami-based AI startup Subquadratic came out of stealth mode last month with a huge claim. It announced that it had solved a mathematical bottleneck that had been holding back large language models for almost a decade.
The details were thin, and many people were unconvinced. But Subquadratic has started to bring the receipts, sharing the results of an independent evaluation of its new tech. The results suggest that the company’s claims might be worth paying attention to.
According to Subquadratic, it has developed a new kind of LLM, called SubQ, that is faster and cheaper and uses a lot less energy than any other model on the market. The company also claims that SubQ is able to process up to 12 times as much text at once than most other models, allowing it to carry out a range of data-heavy tasks, such as analyzing hundreds of documents or entire code bases.
What’s more, Subquadratic says, SubQ does this while more or less matching the performance of the best models put out by Google DeepMind, OpenAI, and Anthropic on key tasks like coding.
The problem was that the company at first provided little evidence for its claims beyond a handful of self-published test scores. And it has yet to make SubQ widely available for people to try out themselves.
So it’s no surprise that Subquadratic’s claims were met with skepticism. Dan McAteer, an artificial intelligence engineer, captured the overall response on X: “SubQ is either the biggest breakthrough since the Transformer … or it’s AI Theranos.”
“We expected healthy skepticism,” says Subquadratic cofounder and chief technology officer Alex Whedon. “In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism, which is why we’re taking the time to make sure any future results are fully verified before putting them out.”
Subquadratic asked Appen, which evaluates other companies’ models, to run its tests on SubQ. The results seem to back up a lot of Subquadratic’s claims. “That was really exciting to me, it validated their architecture,” says Jeanine Sinanan-Singh, Appen’s director of generative AI research.
“I was like, ‘Wow, this could be a game changer,’ because models struggle with speed and inefficiency,” she adds. “But when you have kind of shocking results, it’s really not as credible when you say it yourself.”
SubQ won’t replace existing top models across the board, but it could offer huge increases in speed at a fraction of the typical cost for certain tasks. Subquadratic insists that in the long run, though, its breakthrough could change how LLMs are built. “We hope we’re kicking off a new age of efficiency,” says Justin Dangel, the firm’s cofounder and CEO. “We don’t think anybody will be building on transformers in a few years.”
Attention!
To understand why Subquadratic’s claims are a big deal, let’s dig into how most LLMs work. The key mechanism inside an LLM is a type of neural network called a transformer, which runs a process known as dense attention. Today’s LLMs typically chain together multiple transformers. (The foundational paper of the LLM era, published by researchers at Google in 2017, was titled “Attention Is All You Need.”)
Dense attention works like this: When a transformer processes a chunk of text, it first encodes each word (or part of a word, known as a token) with a number. To capture the meaning of the full text, it then multiplies each of those numbers with every other number for that text. For example, a piece of text 10,000 words long would kick off almost 50 million individual multiplications. That’s a lot of computation and the main reason that LLMs are notorious power hogs.
“If you want to summarize The Great Gatsby, you have to look at the first word and the last word together, and then you have to look at every other combination,” says Dangel.
As the length of the text increases, the number of computations skyrockets. That’s because each additional number must be multiplied by all other previous numbers. Double the number of words, and you roughly quadruple the number of computations, a rate of increase known as a quadratic expansion.
(You can picture this yourself: Draw a circle and mark dots around its edge. Each dot is a token. Then draw lines between pairs of dots to represent the multiplication of those two tokens. A circle with five dots will have 10 lines crossing it. Make it 10 dots and you will have 45 lines, 20 dots and you will have 190 lines, and so on.)
Slashing costs
Subquadratic’s solution is to ditch dense attention, the core operation of a transformer, in favor of what’s known as sparse attention, which slashes the number of computations needed. Instead of multiplying the number assigned to each token by every other number, sparse attention selects just some of the numbers to multiply. The idea is that not all relationships between words in a piece of text matter.
“Sparse attention says not all of those relationships are important, because they’re not,” says Whedon. “If you’re reading a book, you’re not going to look at the first and second words, first and third—that’s insane.”
It’s a simple approach, and Subquadratic is not the first to try it. “Pretty much everything under the sun has been attempted,” says Will Depue, an independent AI researcher who previously worked at OpenAI. “It’s not impossible, but it’s akin to running a four-minute mile.”
Previous techniques for selecting which numbers to multiply and which to ignore have not produced a mechanism that can capture the meaning of a document as well as dense attention can.
Subquadratic claims to have cracked the problem at last. It pitches SubQ as the first sparse-attention LLM that rivals mainstream dense-attention models in performance.
“Historically, most mechanisms have used fixed patterns, like always comparing the first word to the fifth,” says Whedon. “That’s pretty limiting. Language is too sophisticated for that. And so, one of the things that makes our mechanism unique is that we dynamically select which ones are important.”
The firm won’t say exactly how SubQ chooses which words to focus on, but the selection is calculated on the fly and differs for each piece of text the model is given. “That’s kind of where the secret sauce is,” says Whedon.
Testing, testing
The upshot is that for certain tasks, SubQ may be faster and cheaper to run than most other models. Appen evaluated SubQ on a handful of standard tests. In a straight-up speed test, which sets a baseline for how fast a model can operate in theory rather than assess what a model can actually do, Appen found that SubQ was 56 times faster than models using FlashAttention, a previous sparse-attention technique.
On LiveCodeBench, a test that looks at how well models perform on competitive coding problems taken from real contests, SubQ scored 89.7%, putting it in the same ballpark as other top coding models. “This model continues to provide frontier-level performance in coding,” says Appen’s Sinanan-Singh.
Subquadratic’s claims about cost are harder to verify because SubQ is not yet widely available. According to Dangel, it costs $2600 to run Anthropic’s LLM Opus 4.6 through RULER 128, a test developed by Nvidia to assess a model’s ability to retrieve information from large data sets. And SubQ? “It cost us eight dollars,” he says.
SubQ does seem to be able to handle very large data sets. The model has a context window (roughly akin to a working memory) up to 12 million tokens long. Most top models today have context windows one million tokens long. In a demo that Whedon ran for me, he asked SubQ to perform a task that required it to reason about information contained in 400 documents. It responded in seconds. When he gave Perplexity—a popular LLM-powered search engine—the same task, it failed to load all 400 documents.
Appen also ran the needle-in-a-haystack test, which assesses how well a model can retrieve specific information buried in a large amount of data. In its report, Appen states that SubQ scored 98% with context windows six million and 12 million tokens long, “sustaining near-perfect long-context retrieval at scales few models are tested at.”
Too good to be true?
Despite the high scores, benchmarks paint an incomplete picture of what a model can and cannot do. Testing under very specific conditions is not a substitute for running a model on a wide range of real tasks.
Subquadratic is offering SubQ as a model tailored to coding and to searching very large data sets. It says that tens of thousands of potential users have already signed up for early access, including more than 500 enterprise customers. But there’s a long waitlist, and the firm has given very few people access so far. Subquadratic’s response is that it is a new, small company with limited resources and cannot serve too many people at once.
Until more people get their hands on the model and try it out for themselves, some skepticism is justified. One nagging issue is that Subquadratic reused the weights (values set within a model during training that determine how it will behave) from a version of the Chinese open-source model Qwen to bootstrap SubQ, rather than training it from scratch. That’s a common thing for model makers to do, but it cuts across Subquadratic’s claim that it has fully reinvented how LLMs work.
“They may have built something real and useful,” says Depue. “But the public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.”
In the meantime, Subquadratic cofounder Whedon insists that making something different was his only option. If you want to build a competitive model, you have to have new ideas, he says: “We’re more up against it than OpenAI is.”
Background: The clinical diagnosis rate of adenoid hypertrophy (AH) in children has increased in recent years, drawing growing attention from parents. Short-video platforms such as Bilibili, TikTok, and Xiaohongshu host a large volume of educational content on this condition. However, the quality and reliability of this information remain unclear. Objective: This study aimed to evaluate the completeness, understandability, actionability, reliability, and overall quality of short videos on AH across Bilibili, TikTok, and Xiaohongshu and to explore factors associated with these quality metrics, including uploader characteristics and engagement indicators. Methods: We collected 220 videos (Bilibili: n=90, 40.9%; TikTok: n=63, 28.6%; and Xiaohongshu: n=67, 30.5%) using newly registered accounts. Two independent reviewers evaluated video quality using a 6-item content completeness scale (score range 0-12), the Patient Education Materials Assessment Tool for Audiovisual Materials, the modified DISCERN instrument, and the Global Quality Scale (GQS). Interrater reliability was high (Cohen κ=0.77-0.993). Completeness assessed essential informational components of AH. As data were nonnormally distributed, results are presented as median (IQR). Cross-platform comparisons were conducted using the Kruskal-Wallis test with post hoc Mann-Whitney tests (with Bonferroni correction). Spearman correlation was used to explore associations between video characteristics (ie, duration and engagement metrics) and quality outcomes. Stepwise linear regression identified independent predictors of overall quality (GQS). Results: Video duration differed significantly across platforms (Bilibili: median 113.5, IQR 66.5-271.5 seconds; TikTok: median 73, IQR 44-100 seconds; and Xiaohongshu: median 63, IQR 41-127.5 seconds; <.001). Bilibili videos demonstrated higher completeness than videos on the other 2 platforms (Bilibili: median 2, IQR 1.5‐4.0; TikTok: median 1.5, IQR 0.5‐2.0; and Xiaohongshu: median 1.5, IQR 0.5-2.8; <.001); overall differences were observed for understandability and reliability, but pairwise comparisons did not reach statistical significance after Bonferroni correction. Xiaohongshu videos showed greater actionability than TikTok videos (=.011). Medical professionals (n=158, 71.8%) had higher understandability than nonprofessionals (n=158, 81.8% vs n=62, 66.7%; =.001). Video duration positively correlated with completeness (ρ=0.64, 95% CI 0.56-0.71; <.001). Shares showed weak positive correlations with completeness and actionability. Stepwise regression identified understandability (using the Patient Education Materials Assessment Tool–Understandability) as the strongest independent predictor of overall quality (GQS), followed by actionability, video duration, and uploader type; engagement metrics and platform did not enter the final model. Conclusions: The quality of AH-related videos on Chinese short-video platforms is generally suboptimal. Bilibili offers higher completeness, while Xiaohongshu excels in actionability and interactivity. Understandability is the strongest predictor of overall quality, surpassing uploader type and engagement metrics. To improve online health information, platforms should move beyond engagement-based algorithms, and health care professionals should prioritize clear, actionable content.
BackgroundAttention-deficit/hyperactivity disorder (ADHD), a common neurodevelopmental condition, often requires treatment with stimulants such as methylphenidate and nonstimulants such as atomoxetine. Despite their effectiveness, medication-related side effects, adherence and symptom control issues have increased the interest in alternative therapies, especially herbal medicine. However, existing preclinical studies vary in quality and relevance. This systematic review evaluates animal studies investigating herbal medicines for ADHD; identifies limitations in experimental models, behavioral assessments, safety evaluation, and methodological rigor; and proposes a framework to guide future research.MethodsFollowing PRISMA 2020 guidelines, PubMed and Web of Science were systematically searched for eligible studies published between 1995 and March 2025 that used animal models to interventions for the treatment of ADHD-like symptoms.ResultsAnalysis of 25 studies revealed a strong reliance on a limited range of animal models, particularly the spontaneously hypertensive rat (48%), with 88% of studies using male animals exclusively. Behavioral assessments predominantly focused on hyperactivity (72%), whereas evaluations of inattention and impulsivity were comparatively limited. Herbal interventions were associated with improvements in ADHD-like behaviors and modulation of catecholaminergic neurotransmission, including increased dopamine and norepinephrine levels in the prefrontal cortex. Additional mechanisms included activation of neurotrophic signaling pathways (brain-derived neurotrophic factor/TrkB) and suppression of proinflammatory mediators such as TNF-α and IL-1β. However, 80% of the studies did not report safety outcomes, and no study clearly described randomization or blinding procedures, limiting the translational relevance of the findings. Despite these limitations, several formulations, including the Long Mu Qing Xin mixture, An Shen Ding Zhi Ling demonstrated promising multitarget effects.ConclusionAlthough herbal medicines demonstrate promising multipathway effects for ADHD, progress remains constrained by the limited diversity of animal models and narrow scope of behavioral assessments. Future research should prioritize the use of diverse models, more comprehensive assessments, and the implementation of rigorous methodological standards.
