Might some of the books on these shelves literally poison you if you read them? Scottish scientists can save you. CC-licensed photo by Cornell University Library 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’s another post coming this week at the Social Warming Substack on Friday at 0845 UK time. Free signup.

---

A selection of 10 links for you. Untouchable. 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 is creating a new A.I. lab to pursue ‘superintelligence’ • The New York Times

Cade Metz and Mike Isaac:

»

Meta is preparing to unveil a new artificial intelligence research lab dedicated to pursuing “superintelligence,” a hypothetical A.I. system that exceeds the powers of the human brain, as the tech giant jockeys to stay competitive in the technology race, according to four people with knowledge of the company’s plans.

Meta has tapped Alexandr Wang, 28, the founder and chief executive of the A.I. start-up Scale AI, to join the new lab, the people said, and has been in talks to invest billions of dollars in his company as part of a deal that would also bring other Scale AI employees to the company. Meta has offered seven- to nine-figure compensation packages to dozens of researchers from leading A.I. companies such as OpenAI and Google, with some agreeing to join, according to the people.

The new lab is part of a larger reorganization of Meta’s A.I. efforts, the people said. The company, which owns Facebook, Instagram and WhatsApp, has recently grappled with internal management struggles over the technology, as well as employee churn and several product releases that fell flat, two of the people said.

Mark Zuckerberg, Meta’s chief executive, has invested billions of dollars into turning his company into an A.I. powerhouse. Since OpenAI released the ChatGPT chatbot in 2022, the tech industry has raced to build increasingly powerful A.I. Mr. Zuckerberg has pushed his company to incorporate A.I. across its products, including in its smart glasses and a recently released app, Meta AI.

Staying in the race is crucial for Meta, Google, Amazon and Microsoft, with the technology likely to be the future for the industry. The giants have pumped money into start-ups and their own A.I. labs. Microsoft has invested more than $13bn in OpenAI, while Amazon has plowed $8 billion into the A.I. start-up Anthropic.

«

Well this will be joy-filled news for all the people in the metaverse division, won’t it?

Quite apart from the hubris of chasing superintelligence, which sounds like Zeno’s Paradox in machine form.
unique link to this extract

---

Web-scraping AI bots cause disruption for scientific databases and journals • Nature

Diana Kwon:

»

Internet bots have been around for decades, and some have been useful. For example, Google and other search engines have bots that scan millions of web pages to identify and retrieve content. But the rise of generative AI has led to a deluge of bots, including many ‘bad’ ones that scrape without permission.

This year, the BMJ, a publisher of medical journals based in London, has seen bot traffic to its websites surpass that of real users. The aggressive behaviour of these bots overloaded the publisher’s servers and led to interruptions in services for legitimate customers, says Ian Mulvany, BMJ’s chief technology officer.

Other publishers report similar issues. “We’ve seen a huge increase in what we call ‘bad bot’ traffic,” says Jes Kainth, a service delivery director based in Brighton, UK, at Highwire Press, an Internet hosting service that specializes in scholarly publications. “It’s a big problem.”

The Confederation of Open Access Repositories (COAR) reported in April that more than 90% of 66 members it surveyed had experienced AI bots scraping content from their sites — of which roughly two-thirds had experienced service disruptions as a result. “Repositories are open access, so in a sense, we welcome the reuse of the contents,” says Kathleen Shearer, COAR’s executive director. “But some of these bots are super aggressive, and it’s leading to service outages and significant operational problems.”

One factor driving the rise in AI bots was a revelation that came with the release of DeepSeek, a Chinese-built large language model (LLM). Before that, most LLMs required a huge amount of computational power to create, explains Rohit Prajapati, a development and operations manager at Highwire Press. But the developers behind DeepSeek showed that an LLM that rivals popular generative-AI tools could be made with many fewer resources, kick-starting an explosion of bots seeking to scrape the data needed to train this type of model.

«

unique link to this extract

---

He’s a master of outrage on X. The pay isn’t great • The New York Times

Stuart Thompson:

»

Mr. McGee started swiping and poking on his iPhone, rapidly downloading the video, adding his watermark, exporting the video and tapping out some words in just a few minutes. Then he pressed send.

“WNBA star Brittney Griner is going viral on the internet after fans discovered what she sounds like for the first time after releasing this video,” the post read.

This was his “jump starter,” as he called it: a post sent early in the morning aimed at juicing the algorithm in his favor. He believes that if a post performs well, the algorithm will reward his subsequent posts with more reach. Since X’s algorithm is largely a black box, creators rely on these kinds of intuitions and speculations to guide their decisions.

The post also had what he called an “alley-oop” — a phrase that encourages users to post a comment. He made sure to phrase the post so he could call Ms. Griner “she,” baiting commenters to call her the wrong gender.

“I know for a fact that Brittney Griner will go viral, and it’s a great post to start the day,” he said. His plan worked. The views ticked higher: a few thousand at first, then tens of thousands. Comments filled with harassment flooded in, as he’d expected. It didn’t matter to Mr. McGee what they wrote, as long as they wrote something. Each time a premium user — who pays at least $8 per month for special features — interacted with his posts, he earned a little bit of money under X’s revised revenue program.

“Honestly, Brittney Griner should be proud, because of the fact that Brittney Griner can get 500 comments in an hour,” Mr. McGee said.

…In total, X has paid Mr. McGee about $157,000 since the program started in 2023, according to payment records reviewed by The Times. He made $67,000 in his first year and just $12,000 last year after he was kicked from the advertising program. After X overhauled its payment program last October, Mr. McGee complained to Mr. Musk that he still wasn’t being paid. Mr. Musk replied: “Will fix.” Mr. McGee has collected about $16,000 since.

He has also earned about $62,000 directly from his X subscribers since 2023, who pay $10 monthly for insider content. He has separately earned a little money from YouTube and TikTok, though he said it hadn’t amounted to much. His posts to Facebook and Instagram earn minimal views, and those platforms don’t pay most creators.

Unlike YouTube and TikTok, X provides no tools for creators to see how much money they earn from each post.

«

Interestingly, the NYT never says what his account is. (It’s dom_lucre.) Life is tough even with 1.5 million followers, it seems.
unique link to this extract

---

News sites are getting crushed by Google’s new AI tools • WSJ via MSN

Isabella Simonetti and Katherine Blunt:

»

The AI armageddon is here for online news publishers.

Chatbots are replacing Google searches, eliminating the need to click on blue links and tanking referrals to news sites. As a result, traffic that publishers relied on for years is plummeting.

Traffic from organic search to HuffPost’s desktop and mobile websites fell by just over half in the past three years, and by nearly that much at the Washington Post, according to digital market data firm Similarweb.

Business Insider cut about 21% of its staff last month, a move CEO Barbara Peng said was aimed at helping the publication “endure extreme traffic drops outside of our control.” Organic search traffic to its websites declined by 55% between April 2022 and April 2025, according to data from Similarweb.

At a companywide meeting earlier this year, Nicholas Thompson, chief executive of the Atlantic, said the publication should assume traffic from Google would drop toward zero and the company needed to evolve its business model.

Google’s introduction last year of AI Overviews, which summarize search results at the top of the page, dented traffic to features like vacation guides and health tips, as well as to product review sites. Its U.S. rollout last month of AI Mode, an effort to compete directly with the likes of ChatGPT, is expected to deliver a stronger blow. AI Mode responds to user queries in a chatbot-style conversation, with far fewer links.

“Google is shifting from being a search engine to an answer engine,” Thompson said in an interview with The Wall Street Journal. “We have to develop new strategies.”

«

That zero figure for The Atlantic is surely echoed at other sites. And that’s pretty worrying.
unique link to this extract

---

How Expedia’s CMO is preparing for a future dominated by AI • Business Insider

Jochen Koedijk, speaking to Lara O’Reilly:

»

The way consumers are using the internet to make their decisions is fundamentally changing.

I have two kids, 10 and seven years old, and I think it’s very unlikely they will ever put a query into a search box, as I’ve grown up doing. The way people are using voice and having real conversations is going to be a long-term change in how people are using the internet.

What we’re seeing today is that a lot of the inspirational travel searches are evolving rapidly with things like Copilot, ChatGPT, and even Instagram Reels.

There are a lot of moments where my wife sends me Reels, and she’s like, “Where is this?” And I have no idea. So then, we were like, what if we can develop something where you send those Reels to Expedia, and then we will tell you: “It’s this destination, or this hotel. Here are a couple of things to do. The best time to visit is between May and July, and here are a couple of booking options so you can get started.”

Right now, it’s still early access, which means it’s only with Reels, but we’re looking at other forms of content as well.

It’s a very different journey versus going to Google.com and typing in “best hotels in Miami with pool, minus spa.”

We’re also focusing on the visibility of our brands in agentic search engines. It’s really evolving.

We’ve launched with Operator for OpenAI, which I still see as a precursor to an agentic interface because you’re looking at the cursor moving on your screen. Of course, true agentic, where it’s going, will be more behind-the-scenes. But it’s very important to be early so that we can experiment and iterate.

«

Some years ago Benedict Evans said that his child would never see a pixel (after Apple introduced “retina” screens on all its products). The “never type a query into a search box” point feels similar.
unique link to this extract

---

This is what really happened with Siri and Apple Intelligence, according to Apple • TechRadar

Lance Ulanoff:

»

As Apple was working on a V1 of the Siri architecture, it was also working on what Federighi called V2, “a deeper end-to-end architecture that we knew was ultimately what we wanted to create, to get to a full set of capabilities that we wanted for Siri.”

What everyone saw during WWDC 2024 were videos of that V1 architecture, and that was the foundation for work that began in earnest after the WWDC 2024 reveal, in preparation for the full Apple Intelligence Siri launch.

“We set about for months, making it work better and better across more app intents, better and better for doing search,” Federighi added. “But fundamentally, we found that the limitations of the V1 architecture weren’t getting us to the quality level that we knew our customers needed and expected. We realized that V1 architecture, you know, we could push and push and push and put in more time, but if we tried to push that out in the state it was going to be in, it would not meet our customer expectations or Apple standards, and that we had to move to the V2 architecture.

“As soon as we realized that, and that was during the spring, we let the world know that we weren’t going to be able to put that out, and we were going to keep working on really shifting to the new architecture and releasing something.”

That switch, though, and what Apple learned along the way, meant that Apple would not make the same mistake again, and promise a new Siri for a date that it could not guarantee to hit. Instead. Apple won’t “precommunicate a date,” explained Federighi, “until we have in-house, the V2 architecture delivering not just in a form that we can demonstrate for you all…”

He then joked that, while, actually, he “could” demonstrate a working V2 model, he was not going to do it. Then he added, more seriously, “We have, you know, the V2 architecture, of course, working in-house, but we’re not yet to the point where it’s delivering at the quality level that I think makes it a great Apple feature, and so we’re not announcing the date for when that’s happening. We will announce the date when we’re ready to seed it, and you’re all ready to be able to experience it.”

I asked Federighi if, by V2 architecture, he was talking about a wholesale rebuilding of Siri, but Federighi disabused me of that notion.

«

Apple seems to be going round in circles on this. It’s coming! It’s not coming! It’s going to be better! It might be better! It’s coming sometime!
unique link to this extract

---

WWDC25: macOS Tahoe breaks decades of Finder history • 512 Pixels

Stephen Hackett:

»

Something jumped out at me in the macOS Tahoe segment of the WWDC keynote today: the Finder icon is reversed.

You can see that in the image below. On the left is macOS Sequoia, and on the right is macOS Tahoe:

I know I am going to sound old and fussy, but Apple needs to roll this back.

«

Agree absolutely. (“Tahoe” is the forthcoming release of the desktop OS, and all these details can still be changed before it’s released in September or October.) As Hackett points out, it’s been this way (in colour) for 30 years, and it is just jarring to have it swapped. It’s unnecessary – pointless change for the sake of pointless change. There are people inside Apple’s design team who don’t know when not to move our cheese.
unique link to this extract

---

Tool to identify poisonous books developed by University of St Andrews • The Guardian

Ella Creamer:

»

A new tool to quickly identify books that are poisonous to humans has been developed by the University of St Andrews.

Historically, publishers used arsenic mixed with copper to achieve a vivid emerald green colour for book covers. While the risk to the public is “low”, handling arsenic-containing books regularly can lead to health issues including irritation of the eyes, nose and throat along with more serious side-effects. The toxic pigment in the book bindings can flake off, meaning small pieces can easily be inhaled.

In recent years, many libraries have prevented access to all suspect green books as a precaution, as testing has until now been costly and time-consuming. For example, the University of Bielefeld, along with several other German universities, isolated 60,000 books as a precautionary measure last year.

The new device can quickly and cheaply detect the presence of toxic pigment. “A device used in the School of Earth Sciences to detect minerals in rocks was the starting point,” said Pilar Gil, who led the research. “The Eureka moment was discovering the unique reflectance pattern from emerald green pigment in the visible spectrum. The idea was then to apply this discovery to an instrument which we could use and share with the sector.”

Two scientists from the university’s astronomy and physics school, Graham Bruce and Morgan Facchin, developed a portable tool.

«

You did not know there was such a thing as a poisonous book. Fabulous possibility for a murder mystery, surely.
unique link to this extract

---

About Us: Deepware – Scan & Detect Deepfake Videos With a Simple tool

»

Deepfakes are going to be the first real punch from AI to humanity. The cybersecurity industry has a very short time to get ahead of deepfakes before they undermine the public’s trust in reality.

We first recognized the danger while our parent company Zemana researched methods to develop an AI-based antivirus engine. Later, in mid-2018, we started our research on deepfake detection and generation as the deepware AI team.

«

Perhaps a useful tool. Certainly one to start using on the videos that flit by on social media.
unique link to this extract

---

Major labels in talks to license AI use of music, report says • Investopedia

Bill McColl:

»

Warner Music Group (WMG) shares fell on word the big music distributor was one of three firms negotiating with artificial intelligence (AI) startups to monetize AI use of its music catalog.

The Wall Street Journal and Bloomberg reported that Warner Music, along with Universal Music Group and Sony Group Corp unit Sony Music Group, are discussing licensing deals with Suno and Udio to receive compensation when music by artists they represent is used to train generative AI models and produces new music.

The Journal said that the music companies want the AI firms to develop fingerprinting and attribution technology that will track when and how a song is used. Plus, they want to be able to actively participate in the products Suno and Udio release, which includes which songs are developed and how they work.

The Journal noted both Suno and Udio, which are being sued for copyright infringement, have argued that they aren’t infringing on the music companies’ business. However, it added because of “a more uncertain regulatory environment and investor pressure to develop commercial frameworks for the use of music in generative AI products,” both firms are eager to come to an agreement.

«

It’s going to be hard to monetise exactly from AI training. Will it be done on a per-song-scanned basis? That’s going to be pretty thin pickings.
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