The Bahamas turn out to be the perfect location for a homeschooling girl to become a maths genius who solves a longstanding puzzle. CC-licensed photo by Christine Warner-Morin 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.
There may be another post coming this week at the Social Warming Substack on Friday at 0845 UK time. Free signup.
A selection of 9 links for you. Wave theory. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Whistleblower alleges Meta artificially boosted Shops ads performance • AdWeek
Trishla Ostwal:
»
Meta wanted advertisers to believe its ecommerce ad product, Shops ads, was outperforming the competition, per a whistleblower complaint filed in a UK court.
The former employee alleges the social media giant artificially inflated return on ad spend (ROAS) by counting shipping fees as revenue, subsidizing bids in ad auctions, and applying undisclosed discounts.
The complaint, viewed by ADWEEK, was filed with the London Central Employment Tribunal on Wednesday August 20 by Samujjal Purkayastha, a former product manager on Meta’s Shops ads team. The document claims Meta artificially inflated performance metrics to push brands toward its fledgling ecommerce ad product.
The company’s motivation, the complaint says, was in part to combat Apple’s 2021 privacy changes that cut the troves of iOS tracking information that had long powered Meta’s ad machine.
Meta’s former chief financial officer (CFO), David Wehner, said the changes would cost “on the order of $10 billion” in losses during the company’s Q4 2021 earnings call. User purchases on Facebook or Instagram Shops pages would provide more first-party data, however.
Purkayastha, who joined Meta (then Facebook) in 2020 as a product manager on the Facebook Artificial Intelligence Applied Research team, was reassigned to the Shops Ads team in March 2022 and remained at the company until Feb. 19, 2025, when he was terminated.
He alleged that during internal reviews in early 2024, Meta data scientists found the return on ad spend (ROAS) from Shops ads had been inflated between 17% and 19%. This discrepancy stemmed from Meta counting shipping fees and taxes as part of a sale, even though that money never went to merchants, he alleged.
«
Employment tribunals are about dismissals; a judge said the case can proceed, but the hearing won’t be until next year. There might be more documents before then.
unique link to this extract
The neural network hierarchy: how foundation models actually rank your content for AI search results • GEO Platform
“AI Content Team”:
»
If you’ve spent a decade tuning pages for Google’s link graph and user signals, the rise of AI-powered “search” feels like a new continent with its own topography. Foundation models — the large, general-purpose neural networks that power chat assistants and AI overviews — aren’t just another ranking algorithm. They evaluate and surface content according to an internal hierarchy that looks nothing like classic web search. And that’s important: recent, large-scale analyses and industry reports show that the signals that make a site authoritative on the web are largely orthogonal to the signals these models use to decide what to cite and surface in AI answers.
In April 2025, an analysis of 41 million AI search results across platforms like ChatGPT, Google AI Overviews, Perplexity, and Copilot showed a near-total disconnect between traditional SEO metrics and AI citation behavior — 95% of citation variance couldn’t be explained by traffic metrics (r² = 0.05), and 97.2% couldn’t be explained by backlink profiles (r² = 0.038).
In plain English: domain authority, backlink counts, and even raw traffic aren’t the primary features these models use to decide what to quote.
[…Instead the process used is:]
• Representation: Text is turned into dense vectors by the model’s encoder. These embeddings capture semantic meaning and content features the model was trained to value (factuality, explicitness, structure)
• Retrieval: Vector search (often approximate nearest neighbor) surfaces candidate documents or passages based on embedding similarity rather than link-based authority
• Reranking / Fusion: Scored candidates are re-evaluated by the model itself in context; the model weighs freshness, explicitness of answers, and internal confidence when deciding what to cite or include
• Generation / Citation: The model constructs a response and decides whether to cite an external source, often based on whether the retrieved passage can be integrated without hallucination.«
This is a detailed post and makes it clear that AI “search” – or what AI systems rely on – is very, very different from what we’ve grown used to in nearly 30-odd years of Google backlinks.
unique link to this extract
At 17, Hannah Cairo solved a major math mystery • Quanta Magazine
Kevin Hartnett:
»
It’s not that anyone ever said sophisticated math problems can’t be solved by teenagers who haven’t finished high school. But the odds of such a result would have seemed long.
Yet a paper posted on February 10 left the math world by turns stunned, delighted and ready to welcome a bold new talent into its midst. Its author was Hannah Cairo, just 17 at the time. She had solved a 40-year-old mystery about how functions behave, called the Mizohata-Takeuchi conjecture.
“We were all shocked, absolutely. I don’t remember ever seeing anything like that,” said Itamar Oliveira of the University of Birmingham, who has spent the past two years trying to prove that the conjecture was true. In her paper, Cairo showed that it’s false. The result defies mathematicians’ usual intuitions about what functions can and cannot do.
So does Cairo herself, who found her way to a proof after years of homeschooling in isolation and an unorthodox path through the math world.
Cairo grew up in Nassau, the Bahamas, where her parents had moved so that her dad could take a job as a software developer. She and her two brothers — one three years older, the other eight years younger — were all homeschooled. Cairo started learning math using Khan Academy’s online lessons, and she quickly advanced through its standard curriculum. By the time she was 11 years old, she’d finished calculus.
Soon she had consumed everything that was readily available online. Her parents found a couple of math professors to tutor her remotely — first Martin Magid of Wellesley College, then Amir Aazami from Clark University. But much of her education was self-directed, as she read and absorbed, on her own, the graduate-level math textbooks that her tutors recommended. “Eventually,” Cairo recalled, Aazami “said something like, he feels uncomfortable being paid, because he feels like he’s not really teaching me. Because mostly I would read the book and try to prove the theorems.”
«
All of this is amazing, except the youth of the person making the discovery: mathematicians who make amazing discoveries are often young. But the Khan Academy as a route to the frontiers of knowledge? That’s novel, and in its way, encouraging. Let’s hope this is the trend that continues, rather than chatbots.
The article does a pretty good job of explaining the Mizohata-Takeuchi conjecture, which sounds like something out of Star Trek. Maybe that’s what will give us faster-than-light travel in 200 years.
unique link to this extract
The AI takeover of education is just getting started • The Atlantic
Lila Shroff:
»
Gone already are the days when using AI to write an essay meant copying and pasting its response verbatim. To evade plagiarism detectors, kids now stitch together output from multiple AI models, or ask chatbots to introduce typos to make the writing appear more human. The original ChatGPT allowed only text prompts. Now students can upload images (“Please do these physics problems for me”) and entire documents (“How should I improve my essay based on this rubric?”).
Not all of it is cheating. Kids are using AI for exam prep, generating personalized study guides and practice tests, and to get feedback before submitting assignments. Still, if you are a parent of a high schooler who thinks your child isn’t using a chatbot for homework assistance—be it sanctioned or illicit—think again.
The AI takeover of the classroom is just getting started. Plenty of educators are using AI in their own job, even if they may not love that chatbots give students new ways to cheat. On top of the time they spend on actual instruction, teachers are stuck with a lot of administrative work: They design assignments to align with curricular standards, grade worksheets against preset rubrics, and fill out paperwork to support students with extra needs.
Nearly a third of K–12 teachers say they used the technology at least weekly last school year. Sally Hubbard, a sixth-grade math-and-science teacher in Sacramento, California, told me that AI saves her an average of five to 10 hours each week by helping her create assignments and supplement curricula. “If I spend all of that time creating, grading, researching,” she said, “then I don’t have as much energy to show up in person and make connections with kids.”
«
Quite a contrast with Hannah Cairo, isn’t it?
unique link to this extract
Sony is raising PS5 prices on Thursday • The Verge
Jay Peters:
»
Sony is raising the price of all PlayStation 5 models by $50 in the US. In a blog post announcing the change, Sony cited the “challenging economic environment,” which includes the tariffs President Trump has placed on imported products.
The changes will go into effect on Thursday, and the new prices are as follows:
• PlayStation 5 – $549.99
• PlayStation 5 Digital Edition – $499.99
• PlayStation 5 Pro – $749.99Sony says that the retail prices for PS5 accessories “remain unchanged.”
In April, Sony raised the price of PS5 hardware in the UK, Europe, Australia, and New Zealand, and in May, the company said it was considering price hikes to cover the Trump administration’s tariffs.
«
The April price rises were 10-15%; these changes are about 1%. Not exactly world-ending, is it, when the devices apparently face 145% tariffs (though who knows if the tariffs are operative?).
unique link to this extract
T-Mobile claimed selling location data without consent is legal—judges disagree • Ars Technica
Jon Brodkin:
»
A federal appeals court rejected T-Mobile’s attempt to overturn $92m in fines for selling customer location information to third-party firms.
The Federal Communications Commission last year fined T-Mobile, AT&T, and Verizon, saying the carriers illegally shared access to customers’ location information without consent and did not take reasonable measures to protect that sensitive data against unauthorized disclosure. The fines relate to sharing of real-time location data that was revealed in 2018, but it took years for the FCC to finalize the penalties.
The three carriers appealed the rulings in three different courts, and the first major decision was handed down Friday. A three-judge panel at the US Court of Appeals for the District of Columbia Circuit ruled unanimously against T-Mobile and its subsidiary Sprint.
…The carriers also argued that the device-location information, which is “passively generated when a mobile device pings cell towers to support both voice and data services,” does not qualify as Customer Proprietary Network Information (CPNI) under the law. The carriers said the law “covers information relating to the ‘location… of use’ of a telecommunications service,” and claimed that only call location information fits that description.
Judges faulted T-Mobile and Sprint for relying on “strained interpretations” of the statute. “We begin with the text. The Communications Act refers to the ‘location… of a telecommunications service, not the location of a voice call… Recall that cell phones connect periodically to cell towers, and that is what enables the devices to send and receive calls at any moment,” the ruling said.
«
Have to wonder if anyone, at any point, raised a hand inside T-Mobile and said “are we sure that’s exactly.. ethical? Legal?” Probably not. Some will probably have thought it, though, and be silently happy at this outcome.
unique link to this extract
Publisher traffic sources: Google steady but social and direct referrals are down • Press Gazette
Charlotte Tobitt:
»
New data from Chartbeat suggests that “search” as a source of total traffic to major news publishers has remained stable over the last year.
This appears to chime with a Google statement earlier this month downplaying the impact of AI Overviews and AI Mode on publisher referrals.
However, this includes Google Discover – which has replaced search as the main source of Google traffic.
Social media has however sharply declined as a source of publisher traffic in recent years, as has direct traffic.
This comes despite a growing theme in the past two years of publishers setting out to grow the audience that comes to them directly to help future-proof in the face of AI search, unpredictable social media algorithms and changing audience habits.
Direct traffic is defined by audience data analytics tool Chartbeat as visitors that arrive directly on the website via typing in the URL or through a bookmark.
Across 565 US and UK news websites that are Chartbeat customers and have opted in to sharing their anonymised traffic data for aggregate research purposes, 16.09% of traffic came directly to homepages and other landing pages in January 2019.
This fell to an initial low of 12.45% in April 2020 before seeing growth over the next two years to 16.26%. However the proportion of direct traffic has largely fallen again since to 11.46% in July.
«
So now we wait to see what effect AI Overview has in the longer term. The signs aren’t encouraging.
unique link to this extract
The new American inequality: the Cooled vs. the Cooked • The New York Times
Jeff Goodell:
»
In the hottest regions of the country, such as Texas, where I live, the climate crisis is not only changing our world; it is also dividing it. When the heat spikes during the summer, we morph into a two-party state: the cooled and the cooked. On one side, there is water, shade and air-conditioning. On the other, there is sweat, suffering and even, in the worst cases, death. And it means that no matter where we live, we have to update our conception of heat as a disruptive and punishing force.
The cooled are people like me, who work mostly indoors, bathed in the soothing breeze of manufactured air. We live hidden from the brutality of summer, except when we run out to the mailbox or the grocery store. There we hit a wall of heat that feels like an alien force field and burn our hands on the car’s steering wheel.
We live vampire lives, out early for a walk or to run errands, retreating indoors to our comfy caves during the afternoon, then out again after sundown to hang out with friends and complain about the heat and plot a getaway to the beach or the mountains. For the cooled, heat is an inconvenience, an intrusion into our lifestyles and a reason to finally pull the trigger on a loan to build a backyard swimming pool.
The cooked are people like Matthew Sanchez, the pit manager at Terry Black’s BBQ in Austin. On a busy Saturday, he and his co-workers might grill about 2,000 pounds of brisket in five long steel wood-fired BBQ pits. In the summer, the pit gets so hot it breaks thermometers that hang on the wall. “Sometimes it feels like we are rendering ourselves,” Mr. Sanchez told me.
I also met a delivery driver in Austin who had been hospitalized with heat exhaustion. Though he’s recovered, on hot days the muscles in his back tingle and his kidneys hurt. I met a former emergency medical technician who described the disturbing number of calls she responded to from workers at an Amazon warehouse in Texas, many of them related to heat stress.
«
How much do electric car batteries degrade? • Sustainability By Numbers
Hannah Ritchie:
»
Before we quantify how big this effect is, it’s interesting to look at how these processes work over the life of a battery. In the chart below, you can see battery retention measured across a large cohort of Teslas up to 200,000 miles (that’s already telling us something about how big the effect is).
But what’s interesting is that degradation tends to happen quickest in the first 20,000 miles or so. This is because initial lithium salts react with other materials and start building that SEI layer we discussed earlier. After this initial drop, degradation is fairly slow and linear.
Of course, this fact might be one of the explanations why even fairly low-mileage electric cars quickly lose a lot of value once they’ve been driven. As soon as you get on the road, you’re entering the steepest part of the decline.
What’s missing, though, is the context that the overall drop in capacity is still small — probably around 3% to 5% within 25,000 miles — and degradation won’t continue at this rate. So if you buy a second-hand electric car that’s done 20,000 miles, it’s not going to degrade at the same pace that it was.
We’ve now had enough electric cars on the road – and for long enough – to have a good idea of how the battery holds up over time.
Here we’ll focus on a metric used to capture the battery’s “State of Health” (SoH). It’s what percentage of a battery’s initial capacity is still usable after a given number of miles or years.
Let’s start with the results of the huge Tesla cohort that we looked at above. In its 2023 Impact Report, Tesla reported that after 200,000 miles of use, the batteries in a Model 3 and Model Y had lost just 15% of their capacity, on average. For the Model S and X, it was just 12%.
«
(Thanks Quentin SF for the link.)
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 
It would seem unwise to take any Tesla self-reported figures at face value.