Scientists have proposed a properly quantum explanation of the famous double slit experiment. Are you ready for dark light? CC-licensed photo by Matt Mets 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. Lighting up. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Coding emerges as generative AI’s breakout star • Fast Company
Mark Sullivan:
»
The tech industry insists that AI will “transform” how companies, both large and small, operate. Tech VCs and AI founders predict that major business functions will be reshaped, one by one, to be handled by AI agents. For a while, many speculated which function would be transformed first. It wasn’t customer service, legal, or marketing: it was software development. Generative AI’s first killer app is coding. Tools like Cursor and Windsurf can now complete software projects with minimal input or oversight from human engineers.
Businesses are rushing to capitalize on the efficiency gains offered by AI coding. Naveen Rao, chief AI officer at Databricks, estimates that coding accounts for half of all large language model usage today.
A 2024 GitHub survey found that over 97% of developers have used AI coding tools at work, with 30% to 40% of organizations actively encouraging their adoption. (GitHub, owned by Microsoft, created one of the first such tools, Copilot.) Microsoft CEO Satya Nadella recently said AI now writes up to 30% of the company’s code. Google CEO Sundar Pichai echoed that sentiment, noting more than 30% of new code at Google is AI-generated.
The soaring valuations of AI coding startups underscore the momentum. Anysphere’s Cursor just raised $900 million at a $9bn valuation—up from $2.5bn earlier this year. Meanwhile, OpenAI acquired Windsurf (formerly Codeium) for $3bn.
And the tools are improving fast. OpenAI’s chief product officer, Kevin Weil, explained in a recent interview that just five months ago, the company’s best model ranked around one-millionth on a well-known benchmark for competitive coders—not great, but still in the top two or three percentile. Today, OpenAI’s top model, o3, ranks as the 175th best competitive coder in the world on that same test. The rapid leap in performance suggests an AI coding assistant could soon claim the number-one spot.
«
That point about half of usage being for coding is a pretty dramatic illustration of how some things are really huge in one context, and yet probably aren’t having a big impact more widely. But that impact…
unique link to this extract
The great AI displacement is already well underway • Shawn From Portland
Shawn K:
»
As I climb into my little twin sized bed in my small RV trailer on a patch of undeveloped deep rural land in the Central New York highlands, exhausted from my 6 hours of doordash driving to make less than 200$ that day, I check my emails one last time for the night: no responses from the 745th through 756th job applications that i put in over the last week for engineering roles i’m qualified or over-qualified for. I’m not that surprised or disappointed at this point, as I close in on the 800 application mark in over the last year of being an unemployed software engineer.
«
Yes: he got laid off from one job, and has since found it impossible to get work in the software field, despite his qualifications.
unique link to this extract
Asking chatbots for short answers can increase hallucinations, study finds • TechCrunch
Kyle Wiggers:
»
Turns out, telling an AI chatbot to be concise could make it hallucinate more than it otherwise would have.
That’s according to a new study from Giskard, a Paris-based AI testing company developing a holistic benchmark for AI models. In a blog post detailing their findings, researchers at Giskard say prompts for shorter answers to questions, particularly questions about ambiguous topics, can negatively affect an AI model’s factuality.
“Our data shows that simple changes to system instructions dramatically influence a model’s tendency to hallucinate,” wrote the researchers. “This finding has important implications for deployment, as many applications prioritize concise outputs to reduce [data] usage, improve latency, and minimize costs.”
Hallucinations are an intractable problem in AI. Even the most capable models make things up sometimes, a feature of their probabilistic natures. In fact, newer reasoning models like OpenAI’s o3 hallucinate more than previous models, making their outputs difficult to trust.
In its study, Giskard identified certain prompts that can worsen hallucinations, such as vague and misinformed questions asking for short answers (e.g. “Briefly tell me why Japan won WWII”). Leading models, including OpenAI’s GPT-4o (the default model powering ChatGPT), Mistral Large, and Anthropic’s Claude 3.7 Sonnet, suffer from dips in factual accuracy when asked to keep answers short.
Why? Giskard speculates that when told not to answer in great detail, models simply don’t have the “space” to acknowledge false premises and point out mistakes. Strong rebuttals require longer explanations, in other words.
«
So we’re stuck with prolix machines?
unique link to this extract
The Trump administration reportedly fired the head of the US Copyright Office after its preliminary report questioning whether training AI on copyrighted material is fair use • The Verge
Wes Davis:
»
The Trump administration has reportedly fired Register of Copyrights Shira Perlmutter, who leads the US Copyright Office, following the office’s choice to release a pre-publication version of its opinion on the fair use status of AI training data that’s made up of copyrighted information.
Representative Joe Morelle, the ranking Democrat of the Committee on House Administration, called her firing an “unprecedented power grab with no legal basis,” linking the firing directly to her report, which he says amounted to her refusing “to rubber-stamp Elon Musk’s efforts to mine troves of copyrighted works to train AI models.”
Among the report’s conclusions is that while the fair use status of AI training “will depend on what works were used, from what source, for what purpose, and with what controls on the outputs—all of which can affect the market.” The report says research and scholarship might be fair use but says many other AI tools might not be:
But making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries.
University of Colorado law professor Blake Reid called the report a “straight-ticket loss for the AI companies” in a post prior to reports emerged that Perlmutter had been fired, writing that he wondered “if a purge at the Copyright Office is incoming and they felt the need to rush this out.”
«
We’re now at the stage where the most cynical take is probably the correct one.
unique link to this extract
New quantum optics theory proposes that classical interference arises from bright and dark states of light • Phys.org
Ingrid Fadelli:
»
Classical physics theories suggest that when two or more electromagnetic waves interfere destructively (i.e., with their electric fields canceling each other out), they cannot interact with matter. In contrast, quantum mechanics theory suggests that light particles continue interacting with other matter even when their average electric field is equal to zero.
Researchers from Federal University of São Carlos, ETH Zurich and the Max Planck Institute of Quantum Optics recently carried out a study exploring this contrast between classical and quantum mechanics theories through the lens of quantum optics, the field of study exploring interactions between light and matter at a quantum level. Their paper, published in Physical Review Letters, proposes that classical interference arises from specific two-mode binomial states, which are collective bright and dark entangled states of light.
«
I have read this article and tried to understand it. I failed. Though, equally, I recognise that “classical” thinking like “it’s a particle! It’s a wave! It’s a particle and a wave!” is, well, not very quantum-oriented. So if you can follow this new thinking, you’re doing better than me.
unique link to this extract
Tesla Cybertruck inventory skyrockets to record high • Electrek
Fred Lambert:
»
Tesla’s Cybertruck inventory has skyrocketed to a new record high of more than 10,000 units. The vehicle program is in crisis.
We reported at the beginning of April that Tesla ended the first quarter of 2025 with at least 2,400 Cybertrucks in new inventory available in the US.
There’s no exact way to track Tesla’s inventory in the US, but there are ways to track Tesla’s Cybertruck listings. Sometimes, Tesla may have many vehicles with the exact same configuration at the same location and it will only publish a single listing for it. Therefore, Tesla might have been sitting on more Cybertruck inventory.
A month later, the number of listings in the US has skyrocketed to over 10,000 Cybertrucks, according to Tesla-Info.com.
This surge could be due to an actual net increase in Cybertruck inventory, but Tesla is also heavily discounting the trucks at varying rates, creating several different prices and, therefore, more listings.
At an average sale price of $78,000, Tesla could have almost $800m worth of Cybertrucks.
Due to low demand, Tesla appears to have significantly slowed down Cybertruck production in recent months. Therefore, this surge is likely more about Tesla discounting the vehicles, exposing the broader US inventory, than an actual major increase in inventory due to more production.
This is about as bad as it gets. Over 10,000 units account for about two quarters of Tesla’s Cybertruck sales.
«
Hard to figure out if having six months’ (180 days) inventory is bad for Tesla. This data suggests that the maximum among standard car companies is 111 days. (Thanks Gregory B for the link.)
Florida teens kidnap Las Vegas man, steal $4m in cryptocurrency, police say • 8NewsNow
David Charns:
»
Three teenagers are accused of kidnapping a man at gunpoint, driving him to a remote desert area an hour outside of Las Vegas, and stealing $4m in cryptocurrency and other digital assets, the 8 News Now Investigators have learned.
Belal Ashraf and Austin Fletcher, both 16 and from Pasco County, Florida, and a third teenager face charges including robbery, kidnapping, and extortion, records said. A juvenile court judge previously certified Ashraf and Fletcher as adults.
The third teenager was no longer in the United States, a prosecutor said during Fletcher’s probable cause hearing Friday. The men are accused of stealing $4m in cryptocurrency and non-fungible tokens (NFTs), prosecutors said.
Last November, a man called police saying three young men kidnapped him at gunpoint, drove him to a remote desert area, and stole millions of dollars from him, documents said.
That night, the victim was hosting a cryptocurrency-related event at a business in downtown Las Vegas, police said. The victim then returned to his apartment complex and parked his car. The three suspects then approached him and forced him into the back seat of their vehicle, police said.
“[The victim] was told if he complied, he would live to see another day, and if he did not comply, they had his dad and would kill him,” documents said. “[The victim] had a towel placed over his head and was told by the suspects not to look at them.”
Police suspect the three young men then drove the victim across the Nevada border to White Hills, Arizona — more than 70 miles and an hour’s drive from Las Vegas. The victim walked five miles alone in the desert to reach a gas station where he called a friend to pick him up, documents said.
«
The five-mile hike is impressive – at night, in the cold. But apparently you can steal NFTs. So, $3,999,999 in crypto and 500 NFTs worth $1?
unique link to this extract
The strange case of the writer landing A-lister interviews for local magazines • The Guardian
Alexandra Topping and Mario Laghos:
»
In the spring of 2023, subscribers to the British local lifestyle magazine Somerset Life were eagerly anticipating their April edition – a Gardens Special promising top tips for green-fingered readers and the best places to see seasonal bluebells.
But when the magazine landed on readers’ doormats, a story bigger than blooming gardens of south-west England was on the cover. In what appeared to be a world exclusive interview, the Hollywood A-lister Johnny Depp had confessed his love for the bucolic county. More than that, he had bought a secret hideaway in the area.
Never slow to pounce on a story about the controversial Pirates of the Caribbean star, the national and international tabloid media hungrily reproduced news of Depp’s English bolt-hole. The story went viral.
The British celebrity bible Hello! gushed over the star’s idyllic 12-bedroom property, featuring “a walled garden and even its own dairy farm”. The US weekly magazine People told readers that Depp was “enjoying the quiet life in England”. The Independent lured readers into the story, promising that the star had “opened up about his newfound introverted lifestyle”.
The story was all the more intriguing for having been revealed in the most unlikely of places: the pages of the regional monthly magazine. The six-page feature stated: “It is true that he also bought an estate of 850 acres near Kingston St Mary which includes what is known as Somerset Mansion.”
There was only one problem: according to those close to Depp, he hadn’t.
«
Thus begins an absolutely fascinating dive into a strange world of celebrity interviews that none of those involved can remember giving. And the payoff sentence is brilliant, but you really should read all the middle first.
unique link to this extract
Meta threatens to exit Nigeria after FCCPC fine • Rest of World
Damilare Dosunmu:
»
Local authorities have fined Meta $290m for regulatory breaches, prompting the social media giant to threaten pulling Facebook and Instagram from the country.
The Federal Competition and Consumer Protection Commission (FCCPC) said on May 3 that quitting Nigeria won’t absolve Meta of its liability.
…The dispute began in 2021 when Nigeria’s FCCPC started a probe into WhatsApp’s new privacy policy.
The commission said Meta committed multiple and repeated infringements of the country’s Nigerian rules, including “denying Nigerians the right to control their data, transferring and sharing Nigerian user data without authorization, discriminating against Nigerian users compared to users in other jurisdictions, and abusing their dominant market position by forcing unfair privacy policies.”
After remediation efforts failed, the FCCPC issued its final order in July 2024, imposing a $220m fine along with penalties from other agencies that took the total amount to $290m. Meta appealed the decision, but the plea was overturned in April, prompting the company’s threat to withdraw its services from Nigeria.
This isn’t Meta’s first regulatory breach. The company has faced similar sanctions worldwide, including a much larger $1.3bn fine in Europe. It has also been penalized in India, South Korea, France, and Australia. It even faces a $1.5bn fine in Texas.
While the amount of Nigerian fine may not be significant for the company that clocked a revenue of $164bn in 2024, Meta isn’t making enough money in the country to justify paying such an amount, according to Cheta Nwaneze, partner at SMB Intelligence, an Africa-focused consulting firm.
«
Watch this space. Can’t think of an occasion when Meta has followed through on one of these threats, but there’s always a first time, I guess.
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: Following up from last Friday: Kosmos-482 seems to have fallen into the Indian Ocean. No human casualties. (Perhaps some fish, we regret.)
The Overspill: when there's more that I want to say 