Scientists reckon they’ve found the source of a hum that a small percentage of people hear – and it’s a lot closer to home than you might think. CC-licensed photo by SNSF Scientific Image Competition on Flickr.
You can sign up to receive each day’s Start Up post by email. You’ll need to click a confirmation link, so no spam.
A selection of 9 links for you. Listening in. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Meta’s smartglasses mean any child can be covertly filmed. In the age of AI, how do we tackle that risk? • The Guardian
Gaby Hinsliff:
»
It took me a while to realise what the man sitting next to me on the train was doing. He was flicking through pictures of a little girl on his phone; zooming in on some, cropping the images. Hardly unusual, of course, except that on closer inspection the pictures were very clearly of the toddler sitting opposite us, playing with her parents. On a crowded train of people mostly staring obliviously at our own phones, a stranger had been covertly photographing that little girl over and over again. And if it hadn’t been for that very visible phone screen giving him away, he wouldn’t have been caught.
That encounter was a while ago but my memory was jogged last week listening to a clip of Meta’s VP of wearables, Alex Himel, talk fondly to the BBC about using his new smartglasses to film his daughter singing in a talent contest. They let him capture the memory without having to watch her through a screen, he said, freeing him to be “kind of living in the moment, head up and hands free”. So far, so sweet. But what about the risk of creeps potentially using them to film other people’s daughters covertly, heads up and hands free? Meta’s smartglasses have a tiny flashing warning light on the frame to indicate when the wearer is recording, but the technology is still new enough that many people don’t think to look for it. And in the age of AI, stolen images of children matter. Taken in real life or scraped from the internet, they’re the raw material for generating realistic-looking child abuse material of the most extreme and disturbing kind. (If that sounds like a grim but ultimately fairly victimless crime, the Internet Watch Foundation warned in a recent report that AI-generated material risks fuelling sexual interest in children, normalising violence and increasing the risk of real-life offending.)
When Google launched the first smartglasses in 2013, they didn’t catch on: too nerdy-looking but also potentially too creepy, given the possibilities they afforded stalkers, trolls, criminals gathering intelligence and law enforcement in repressive regimes. But big tech hasn’t given up on the dream of connecting humans more seamlessly to the internet, and newer versions learn from past mistakes. Meta’s are cooler, featuring collaborations with Ray-Ban and Kylie Jenner, and are pitched unashamedly at parents under the wholesome slogan: “never miss a moment”.
…Meta has argued that it is for individuals to ensure they don’t “actively exploit” the technology, which makes you wonder if it learned anything at all from watching humans get to grips with its social media platforms.
«
She implies that she said something about the man on the train; one certainly hopes so, though it would be a tricky conversation to open. On the part of Meta, though, there’s a real indifference to the downsides these could have. It’s like letting everyone have guns and trusting nobody will get shot. (Thanks Gregory B for the link.)
unique link to this extract
A device hidden in cars across the US leaves them vulnerable to hacking and paralysis. Patch it now • WIRED
Andy Greenberg:
»
As modern cars have evolved into multi-ton computers on wheels, drivers are beginning to learn they need to install security updates for their vehicles’ code, just as they would for a phone or laptop. Yet not even the most tech-savvy car owners would expect they’d need to install a patch for an insecure third-party component they never installed or requested—and likely aren’t even aware of—that’s been wired into some of the most sensitive systems of their vehicle, leaving it vulnerable to stealthy hacking, tracking, and even roadside paralysis.
That’s the disturbing discovery of a team of security researchers at UC San Diego, who found that a model of aftermarket car alarm known as the KARR Security System, installed in more than two million vehicles across the US by their estimate, can let any hacker within Bluetooth range send radio commands to silently unlock the car at will, turn off its alarm, honk the car’s horn or flash its lights, or even disable its ignition and leave a driver stranded.
The KARR alarm devices are typically installed by car dealers, not manufacturers or owners, and used as a measure to prevent auto theft from dealer lots. Yet when the cars are sold, the alarms typically aren’t removed, even if the buyer declines to pay for it as an additional feature. That means car owners across the US have a hackable device under their hood whose code they’ll need to update to protect their vehicle—but one that, in many cases, they never purchased and have no idea is there.
…The company that sells the KARR Security System, Acrisure Protection Group, on Monday rolled out a firmware update for the vulnerable Bluetooth model of its aftermarket KARR alarm to fix the security issues UCSD uncovered. Car owners who already have the KARR Security smartphone app installed should receive an alert about the firmware update, the UCSD team says. Those who don’t have it installed will need to download the KARR Security System smartphone app (Android or iOS), connect it to their vehicle’s KARR alarm, then tap “customer service” and “firmware update.”
«
Got to love those car dealers.
unique link to this extract
AI just disproved the 87-year-old Jacobian conjecture • The Next Web
Ana Maria Constantin:
»
On Sunday evening, as Spain and Argentina played out the World Cup final, the mathematician Levent Alpöge posted a short message on X. The Jacobian conjecture, he wrote, is false. He thanked a friend for asking about it, and another friend, “fable,” for working during the match.
That second friend was Fable 5, Anthropic’s latest AI model. Together they had produced a counterexample to a problem that had resisted mathematicians for 87 years. It was 216 characters long.
The Jacobian conjecture was set out by Ott-Heinrich Keller in 1939. In rough terms, it says that a certain kind of polynomial map, one whose Jacobian determinant is a non-zero constant, must be reversible with a neat polynomial inverse. It became one of the field’s most stubborn open problems. Stephen Smale put it on his famous 1998 list of challenges for the 21st century. New Scientist calls it the hardest maths problem an AI has yet cracked.
Alpöge and Fable 5 did not prove it. They broke it. They found a single map, from three-dimensional space to itself, that has the required constant determinant yet still cannot be reversed. One valid counterexample is all it takes to sink a conjecture.
…The result is easy to check even if it was hard to find. The counterexample is a plain list of polynomials that any mathematician can verify by hand. Finding it was the needle-in-a-haystack part, and that is where the machine came in.
It was not a one-line prompt, cautioned Bartósz Naskręcki, one of many mathematicians now waiting for Alpöge’s full write-up. Searching for a counterexample like this, he said, takes real insight.
…Not everyone is stirred. Andrew Blumberg, a mathematician at Columbia who sits on a project testing AI on research maths, was pointedly unmoved. “This did not cause me to update my priors,” he told Mashable. This is exactly what he expects AI to be good at.
His point is that a counterexample and a proof are very different prizes. A proof teaches you why something is true. A counterexample just ends the argument. Smale wanted the conjecture solved because a solution might reveal how nature is structured.
«
Where did all the computer-science professors go? To AI companies • The Atlantic
Lila Shroff and Rose Horowitch:
»
AI companies are turning into something like mini-universities in their own right. OpenAI employs top mathematicians and physicists—including one who studies black holes, and another who specializes in string theory. At least three computer-science professors joined Meta’s AI lab in late June. DeepMind, like Anthropic, is home to a crew of philosophers. And Anthropic’s recent job postings indicate an interest in hiring legal scholars and political scientists.
It’s unclear exactly how many current and former professors are working at AI companies. Across these four firms, we found more than 80—the majority of whom are computer scientists. Some have left academia entirely; others are still working part-time at a university. That number is likely a significant undercount, because we don’t have access to internal data; it also doesn’t include the many professors who have started their own companies, those at other AI start-ups, and those working with the industry in a less formal capacity.
Particularly for AI researchers, these companies have a strong gravitational pull. “Much of the important research is being done in industry now,” Humphrey Shi, a computer-science professor at Georgia Tech who joined Nvidia as a vice president last fall, told us. As he sees it, “If you want to do something that really, truly matters, you probably want to join one of those entities.” Tech firms are making offers—including very enticing salaries—that are hard for academics to refuse. In the process, research that previously would have happened in the open is getting locked up behind closed doors.
For decades, universities were the center of AI research. The field itself officially began at a gathering of researchers at Dartmouth in 1956, and federal funding provided much of the field’s early support. In the early 2010s, Silicon Valley executives began to take AI’s commercial potential more seriously, and set out to hire the best researchers. In 2013, Google paid $44m to acquire a start-up run by a trio of AI researchers from the University of Toronto.
Facebook then hired Yann LeCun, an NYU professor, to establish the company’s AI-research lab; Uber poached some 40 Carnegie Mellon researchers to work on driverless cars. But for the most part, even as more work was being done inside of private companies, many of the newly hired academics at tech companies retained their professorships and established a culture of open research. “Researchers will be strongly encouraged to publish their work,” OpenAI wrote in its founding announcement. (Note the organization’s name.) This norm helped lead to the current AI boom
«
Pollution from Musk’s unpermitted xAI power project hits hardest in Black communities • Reuters
Disha Raychaudhuri and Valerie Volcovici:
»
Elon Musk’s artificial intelligence company xAI has installed 59 natural gas turbines for its Colossus 2 data centre project in Tennessee without securing federal clean air permits, according to communications between regulators and xAI representatives.
Potential emissions from the turbines are far beyond the threshold that would require a federal permit, and would be released near predominantly Black communities already estimated to be suffering disproportionately high rates of lung disease, according to a Reuters analysis based on government data and information in the correspondence with regulators.
The findings, which have not been previously reported, reflect how exploding electricity demand from AI data centres is driving companies to build off-grid power plants at a pace outstripping environmental oversight, with potentially big risks to public health.
The number of unpermitted turbines identified by Reuters is about double what xAI has publicly acknowledged. The company previously said it was running 27 unpermitted turbines for Colossus 2 as of January and has argued the permits are not required. At least 57 of the 59 turbines are located in Mississippi, just over the state line from Tennessee where the data centre is located.
The xAI turbines are among scores of off-grid power plants for data centres proposed or under construction around the country. Local authorities often fast-track approvals in just weeks or months, without the years of environmental studies and public hearings typically required for such power generation projects that connect to the grid, Reuters has reported.
Mississippi regulators in March issued a permit for permanent turbines for Colossus 2, allowing construction of 41 gas-fired turbines. The approval came three weeks after the state’s only public hearing on the project.
The xAI cluster of temporary turbines in Mississippi is already among the biggest off-grid data centre power projects, according to Ben King, an analyst with think tank Rhodium Group, who reviewed the Reuters analysis.
“This looks to be an unprecedented level of behind-the-meter gas being installed in one place,” he said, referring to off-grid natural gas plants serving just one customer.«
Very typical of Musk to have a centre that just ignores the federal rules. Those are for the little people.
unique link to this extract
Ways to think about token pricing • Benedict Evans
Benedict Evans is trying to look beyond the present crunch:
»
On the supply side, a trillion dollars or more of data centre capex is coming down the pipe (and plenty more semiconductor capex behind that), inference efficiency continues to improve very quickly, and new models are far more (or far less!) efficient in their token use. On the demand side, although the market has been capacity-constrained since 2022, the crunch in the first half of this year has been driven by sudden product-market fit in really just one use case, software development, and that’s actually a pretty small field (imagine if we had product-market fit for a consumer use case with hundreds of millions of DAUs – today’s infrastructure couldn’t support it at any price). We don’t know what the next use-cases to scale will be, nor when that would be, nor what their token needs would be.
Going up a level, it’s been pretty widely reported that inference today has 40-50% gross margins: this includes deprecation of the associated server costs (or the cost of renting them), but we don’t really know the asset life (five years? Seven years?) and obviously this doesn’t include the cost of training the next model a couple of times a year, which is currently far larger than revenue. In principle, inference is a marginal cost and training is a fixed cost, so with high enough revenue you can reach profitability, but we don’t know how training costs will change. On the other side of the table, it’s unclear how much of the surge in use in the last few months has an ROI (or at least has an ROI that can be quantified to a CFO), let alone any future use cases, and hence what prices people might be prepared to pay for them.
So, all the variables will move all over the place over the next 12 months, and move again over the next three to five years. How could we suggest where this will settle? How and where will supply, demand, price, capacity and capex get back into equilibrium?
«
Absolutely true: it’s quite lucky that multiple sectors didn’t have product-market fit all at once. Token costs would have gone to the Moon.
unique link to this extract
The mysterious hum heard around the world may finally have an explanation • ScienceDaily
Norwegian University of Science and Technology:
»
The cochlea inside the inner ear naturally creates faint sounds at different frequencies, generally between about 500 and 5000 Hertz. These noises do not serve a direct purpose. Instead, they are a byproduct of the ear’s physiological sound amplification process.
“Most of us don’t hear these sounds. However, a few people can actually hear the sounds that the ear itself produces. And these sounds can be measured objectively,” Drexl said.
Known as oto-acoustic emissions, these sounds can be detected by placing a sensitive microphone inside the ear canal. In certain individuals, spontaneous oto-acoustic emissions may be perceived as distressing tinnitus. “One hypothesis was that the participants in our group could hear oto-acoustic emissions at low frequencies. That’s why we tested whether they had them,” says Drexl.
But… the answer was no.
“Then there are people who hear something that cannot be measured objectively. We believe people in this category have a form of low-frequency tinnitus,” Drexl said.
Tinnitus, often described as ringing in the ears, occurs when someone perceives a sound inside the ear or head even though no external source is producing it. Tinnitus may be temporary or persistent. People often initially interpret the sound as something coming from their surroundings. When the noise continues after they change locations, however, they may eventually realize that it is not being generated by anything nearby.
Based on current knowledge of hearing and the results of the participant tests, Drexl believes the most likely explanation has two parts. A small number of people who hear The Hum may genuinely have unusually sensitive low-frequency hearing. For most, however, the experience may be a type of tinnitus in which the sound begins within the auditory system.
“Based on our results, although we haven’t ruled out cases of physical external sound sources, we suggest that subjective tinnitus in the low-frequency range is often the cause of hearing pulsations of low-frequency sound perceptions,” he said.
«
Oh OK find. (Thanks Joe S for the link.)
unique link to this extract
Why American ambulance rides are so expensive • David Oks
David Oks:
»
The heaviest users of ambulances are the elderly, who are covered by Medicare; and since Medicare is the largest purchaser of medical services in the country and is backed by the force and authority of the US government, it enjoys the rare privilege of setting its own prices. And since 2002, Medicare has imposed a national fee schedule on ambulance services—a fixed maximum payment for each category of ride.
And notably, the fees that Medicare sets run far below cost. The average ambulance transport costs $2,673 to provide; Medicare pays only about $329 of that. A typical ambulance ride for a Medicare patient, in other words, loses the ambulance service thousands of dollars. With Medicaid it’s even worse: state programs typically pay a fraction of what Medicare does, and the services lose even more per ride. Normally, when an insurer pays less than the provider charges, the provider bills the patient for the balance, a practice known as “balance billing”; but with Medicare and Medicaid, balance billing is illegal.
On most of their book, then—rides for Medicare and Medicaid beneficiaries—EMS providers lose money on every run. And that’s before accounting for the fixed costs that make up the large majority of their expenses.
That leaves two groups from whom ambulance services can try to recover the cost of their operations: the uninsured and the privately insured.
Let’s start with the uninsured. On paper, they should be the most lucrative patients an ambulance service could ask for: they receive the full, undiscounted charge, with no insurer to negotiate it down. But uninsured patients are disproportionately unlikely to pay, since they’re disproportionately poor; and EMS providers typically end up selling most of that debt to collections firms for pennies on the dollar. So in practice, the uninsured look less like a profit center and more like the Medicare and Medicaid patients: another class of patients that pays below cost.
And that leaves the one group who actually can be made to pay: the privately insured.
«
Which means that the latter get royally screwed if they set foot in an ambulance. The inherent dysfunctionality in the US health system remains astonishing to an outsider. And to many insiders.
unique link to this extract
US judge approves Anthropic’s $1.5bn settlement of copyright lawsuit • Reuters
Blake Brittain:
»
A federal judge in San Francisco on Monday signed off on artificial intelligence company Anthropic’s landmark $1.5bn settlement of a class action lawsuit brought by a group of authors who accused it of misusing their books to train its AI chatbot Claude.
US District Judge Araceli Martinez-Olguin granted final approval of the settlement, the largest known settlement of a US copyright case, rejecting arguments that it was too small.
The case is one of dozens brought by copyright owners including authors and news outlets against tech companies over the training of their large language models, and the first major US case to settle.
Now-retired Judge William Alsup initially approved the deal last September.
“We reached this settlement in 2025, after the court’s landmark ruling that training AI on books is fair use under copyright law — which remains the law today,” Anthropic deputy general counsel Aparna Sridhar said in a statement. “We are pleased that more than 91% of authors and publishers covered by the settlement have claimed their share of the payment, and we’re looking forward to bringing this matter to a close.”
The authors’ lead attorney, Justin Nelson, welcomed what he called a “historic settlement.”
“It is the largest known copyright recovery in history. We look forward to making distributions to the Class as promptly as possible,” Nelson said in a statement, referring to payments to authors covered by the settlement.
The writers sued Anthropic in 2024, arguing that the company, which is backed by Amazon and Alphabet, used pirated versions of their books without permission to teach Claude to respond to human prompts.
Alsup ruled last June that Anthropic made fair use of the authors’ work to train Claude, but found that the company violated their rights by saving more than seven million pirated books to a “central library” that would not necessarily be used for AI training.
«
Note that last point: it’s not the training, it’s the retention that breached copyright.
unique link to this extract
| • Why do social networks drive us a little mad? • Why does angry content seem to dominate what we see? • How much of a role do algorithms play in affecting what we see and do online? • What can we do about it? • Did Facebook have any inkling of what was coming in Myanmar in 2016? Read Social Warming, my latest book, and find answers – and more. |
Errata, corrigenda and ai no corrida: none notified
The Overspill: when there's more that I want to say 