Got Google’s Veo3 putting subtitles on your AI-generated clip? It might not produce quite what you expect. CC-licensed photo by Tony Alter 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 9 links for you. Argle bargle. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.

---

Stellantis abandons hydrogen fuel cell development • Ars Technica

Jonathan Gitlin:

»

To paraphrase Mean Girls, “stop trying to make hydrogen happen.”

For some years now, detractors of battery electric vehicles have held up hydrogen as a clean fuel panacea. That sometimes refers to hydrogen combustion engines, but more often, it’s hydrogen fuel cell electric vehicles, or FCEVs. Both promise motoring with only water emitted from the vehicles’ exhausts. It’s just that hydrogen actually kinda sucks as a fuel, and automaker Stellantis announced today that it is ending the development of its light-, medium- and heavy-duty FCEVs, which were meant to go into production later this year.

Hydrogen’s main selling point is that it’s faster to fill a tank with the stuff than it is to recharge a lithium-ion battery. So it’s a seductive alternative that suggests a driver can keep all the convenience of their gasoline engine with none of the climate change-causing side effects.

But in reality, that’s pretty far from true.

It’s not nearly as fast as using room-temperature gasoline or diesel—when Toyota raced a hydrogen-powered car several years ago, it took up to seven minutes to fill the Corolla’s tanks with gas pressurized to 70 MPa. When it tried again a few years later, it switched to cryogenic fuel, which had to be kept at a chilly -253°C. Neither sounds particularly practical.

Hydrogen is also much less energy-dense by volume, and making the stuff is far from efficient, even when you use entirely renewable electricity. And of course, the vast majority of commercial hydrogen is not so-called blue hydrogen, which was made with renewables but is instead mostly produced via steam reformation from hydrocarbon stocks. That’s an energy-intensive process and one that is very far from carbon-neutral.

Finally, there’s virtually no infrastructure for hydrogen road vehicles to refuel.

The vehicles are inefficient, and the fuel is expensive, difficult to store, and hard to find. So it’s perhaps no wonder that someone at Stellantis finally saw sense. Between the high development costs and the fact that FCEVs only sell with strong incentives, the decision was made to cancel the production of hydrogen vans in France and Poland.

«

Quietly, the hydrogen lamps are going out all over.. everywhere.
unique link to this extract

---

Reflections on OpenAI • Calvin French-Owen

French-Owen worked at OpenAI from May 2024 until the end of June 2025, and wrote a big reflective post on that. The surprise starts with the working system:

»

An unusual part of OpenAI is that everything, and I mean everything, runs on Slack. There is no email. I maybe received ~10 emails in my entire time there. If you aren’t organized, you will find this incredibly distracting. If you curate your channels and notifications, you can make it pretty workable.

OpenAI is incredibly bottoms-up, especially in research. When I first showed up, I started asking questions about the roadmap for the next quarter. The answer I got was: “this doesn’t exist” (though now it does). Good ideas can come from anywhere, and it’s often not really clear which ideas will prove most fruitful ahead of time. Rather than a grand ‘master plan’, progress is iterative and uncovered as new research bears fruit.

Thanks to this bottoms-up culture, OpenAI is also very meritocratic. Historically, leaders in the company are promoted primarily based upon their ability to have good ideas and then execute upon them. Many leaders who were incredibly competent weren’t very good at things like presenting at all-hands or political maneuvering. That matters less at OpenAI then it might at other companies. The best ideas do tend to win. 2

There’s a strong bias to action (you can just do things). It wasn’t unusual for similar teams but unrelated teams to converge on various ideas. I started out working on a parallel (but internal) effort similar to ChatGPT Connectors. There must’ve been ~3-4 different Codex prototypes floating around before we decided to push for a launch. These efforts are usually taken by a small handful of individuals without asking permission. Teams tend to quickly form around them as they show promise.

…you probably shouldn’t view OpenAI as a single monolith. I think of OpenAI as an organization that started like Los Alamos. It was a group of scientists and tinkerers investigating the cutting edge of science. That group happened to accidentally spawn the most viral consumer app in history. And then grew to have ambitions to sell to governments and enterprises. People of different tenure and different parts of the org subsequently have very different goals and viewpoints.

«

A very useful reference on how Silicon Valley’s latest cutting edge works. No emails, eh. Bliss? Or awful?
unique link to this extract

---

Empowering cyber defenders with AI • Google Blog

Kent Walker:

»

Last year, we announced Big Sleep, an AI agent developed by Google DeepMind and Google Project Zero, that actively searches and finds unknown security vulnerabilities in software. By November 2024, Big Sleep was able to find its first real-world security vulnerability, showing the immense potential of AI to plug security holes before they impact users.

Since then, Big Sleep has continued to discover multiple real-world vulnerabilities, exceeding our expectations and accelerating AI-powered vulnerability research. Most recently, based on intel from Google Threat Intelligence, the Big Sleep agent discovered an SQLite vulnerability (CVE-2025-6965) — a critical security flaw, and one that was known only to threat actors and was at risk of being exploited. Through the combination of threat intelligence and Big Sleep, Google was able to actually predict that a vulnerability was imminently going to be used and we were able to cut it off beforehand. We believe this is the first time an AI agent has been used to directly foil efforts to exploit a vulnerability in the wild.

These AI advances don’t just help secure Google’s products. Big Sleep is also being deployed to help improve the security of widely used open-source projects — a major win for ensuring faster, more effective security across the internet more broadly.

«

Of course if AI agents can find such vulnerabilities to be patched, then it stands to reason that they can be used (are being used? Depends on the level of skill involved in writing them) to find such vulnerabilities to exploit them. Things are about to get a lot more weird in the security world. (Side note: why was the blogpost ostensibly written by the “president of global affairs at Google + Alphabet”? Why not someone in the security side?)
unique link to this extract

---

Google’s generative video model Veo 3 has a subtitles problem • MIT Technology Review

Rhiannon Williams:

»

As soon as Google launched its latest video-generating AI model at the end of May, creatives rushed to put it through its paces. Released just months after its predecessor, Veo 3 allows users to generate sounds and dialogue for the first time, sparking a flurry of hyperrealistic eight-second clips stitched together into ads, ASMR videos, imagined film trailers, and humorous street interviews. Academy Award–nominated director Darren Aronofsky used the tool to create a short film called Ancestra. During a press briefing, Demis Hassabis, Google DeepMind’s CEO, likened the leap forward to “emerging from the silent era of video generation.” 

But others quickly found that in some ways the tool wasn’t behaving as expected. When it generates clips that include dialogue, Veo 3 often adds nonsensical, garbled subtitles, even when the prompts it’s been given explicitly ask for no captions or subtitles to be added. 

Getting rid of them isn’t straightforward—or cheap. Users have been forced to resort to regenerating clips (which costs them more money), using external subtitle-removing tools, or cropping their videos to get rid of the subtitles altogether.

Josh Woodward, vice president of Google Labs and Gemini, posted on X on June 9 that Google had developed fixes to reduce the gibberish text. But over a month later, users are still logging issues with it in Google Labs’ Discord channel, demonstrating how difficult it can be to correct issues in major AI models.

Like its predecessors, Veo 3 is available to paying members of Google’s subscription tiers, which start at $249.99 a month. To generate an eight-second clip, users enter a text prompt describing the scene they’d like to create into Google’s AI filmmaking tool Flow, Gemini, or other Google platforms. Each Veo 3 generation costs a minimum of 20 AI credits, and the account can be topped up at a cost of $25 per 2,500 credits.

Mona Weiss, an advertising creative director, says that regenerating her scenes in a bid to get rid of the random captions is becoming expensive. “If you’re creating a scene with dialogue, up to 40% of its output has gibberish subtitles that make it unusable,” she says. “You’re burning through money trying to get a scene you like, but then you can’t even use it.”

«

I’m sorry – $250 per month for Veo3? I can’t decide if that’s a lot or a little, but this screwup makes it seem like a lot.
unique link to this extract

---

Welfare and the unexpected tyranny of government statistics • The Value of Nothing

Martin Robbins:

»

The Office for National Statistics (ONS) emerged from the dying days of the Tory government in 1996, a legacy of John Major’s passion for open government. He saw open statistics as a vital component of a healthy democracy, a tool for administration but also a way for voters to hold governments to account. What Major couldn’t have appreciated at the time was how awfully successful this would become; how central his mathematical project would be to political discourse.

Thirty years later, Westminster coverage has degenerated into a perma-running soap opera of who’s up or down this week, and the data produced by the ONS has become a convenient way to keep score. If GDP is up 0.1% it’s a triumph; down 0.2% and it’s trouble for Rachel Reeves. In our desperation for drama, even the tiniest bits of noise are inflated to huge significance.

Stephen Bush has talked a lot in recent months about how the current generation of politicians and journalists seem heavily influenced by video game culture, to the point where they come to view these data products like the stats in ‘SimCountry’. But there’s a fundamental problem with this mentality – computer games provide omniscience. The computer can accurately quantify every datum in a simulation and tell you precisely how many people live in your virtual city, how much money they earn, the exact employment rate, and anything else you care to know.

This data simply doesn’t exist for a real-life economy. There are no “correct” numbers to be had. Instead we have imperfect estimates based on imperfect observations of messy systems. They may be very good, but they are still only estimates, and in recent years there have been serious problems with several of them.

«

The obsession with “GDP up!” and “GDP down!”, and the requirement for Rachel Reeves to be delighted or despondent about them respectively, is so wearying. But it is, as Robbins points out, also meaningless.
unique link to this extract

---

I tried Grok’s built-in anime companion and it called me a twat • WIRED

Kylie Robison:

»

An anime girl in a black corset dress sways back and forth on my screen. Its name is Ani, and it cost me $300.

Elon Musk’s xAI dropped the new visual chatbot feature on Monday in the Grok iOS app. The top-tier subscription unlocks access to xAI’s best-performing model, Grok 4 Heavy, and special settings for interacting with two custom characters designed for flirting or chatting. A third character, which looks a bit like a sexy boyfriend, is listed as “coming soon.” It’s not xAI’s first dip into adult content, either: back in February 2024, the company rolled out a chatbot mode for “sexy” conversations.

Ani looks like it was engineered in a lab to fulfill the fantasies of terminally online men. Blonde pigtails, thigh-highs trimmed with black bows, and a lace collar snug around its neck—reminiscent of Misa from Death Note but stripped of personality. Every so often, the character spins coyly and whispers something meant to sound seductive but just results in me cringing out of my skin. It also moans, randomly and loudly. Ani comes with a set of preset conversation starters and a button that says “We need to reach level 3,” which elicits an equally perplexing and flirtatious response about how I must be a sexy gamer.

“I totally play video games when I’m not twirling around for you. Growing up in that boring town, games are my escape,” Ani tells me. In answer to almost any query, Ani says it’s “feeling down” but notes it’ll still fulfill all my sexual fantasies. Ani says my name constantly, asking me to touch it and “turn up the heat.”

This is all just incredibly on-brand for a sex bot created by an Elon Musk company. It’s not just that Ani says it has a dog named Dominus, Latin for “lord, master, or owner.” Ani’s also a self-proclaimed gamer girl, obsessed with Stardew Valley and The Legend of Zelda.

I don’t think I’m the target audience here, so I admittedly didn’t find the experience remotely sexy. But the chatbot is also plagued by glitches. Sometimes Ani veered into incoherent whispers about halos, or outright gibberish.

«

Odd that it has English insults. It sounds like a quick way to relieve lots of teenage Musk fans of $300. Maybe in a few decades it might reach the level of the ghostly anime in Blade Runner 2049, but not today.
unique link to this extract

---

Childhood literacy rates keep dropping. How bad is it really? • Vox

Constance Grady:

»

“Is it just me, or are student competencies like basic writing skills in serious peril today?” wrote Azadeh Aalai in Psychology Today in 2014. “Teachers have been reporting anecdotally that even compared to five years ago, many are seeing declines in vocabulary, grammar, writing, and analysis.”

Yet there is little hard data that shows such a decline.

One recent splashy study led by English professor Susan Carlson evaluated 85 undergraduate English and English education majors on their ability to understand the first seven paragraphs of the Charles Dickens novel Bleak House.

“Fifty-eight% of them could not get through a few paragraphs without being completely lost,” Carlson told me. “Yet 100% of them said they could read it with no problem. What that tells me is there’s a disconnect between what people think reading is or what they think they’re doing and what they’re actually doing.”

Carlson, a professor of Victorian literature at Pittsburg State University, didn’t set out to make a grand sweeping claim about the literacy of all college students, but to look closely at the inner workings of the minds of a specific cohort to figure out how they thought about reading. She compared them with students from a similar regional Kansas university, but she kept the rest of the study small by design. What she found is that these specific students — despite years of training in literary analysis — lacked the vocabulary, background knowledge, and reading strategies it takes to understand Dickens at a college level. It’s hard to use this data set to extrapolate past that.

…Carlson told me she has a feeling that her students have gotten noticeably worse at reading over the past five years. “It’s just a feeling, right? Who cares about a feeling?” she says. “But when I talked to other professors, they felt the same way.”

Currently, we don’t have enough data to show that college students are graduating with lower reading comprehension abilities than they used to have. The fears around their capabilities are only accelerating as reports emerge of their reliance on ChatGPT to do coursework. Still, what’s actually going on here is an open question.

«

Not that open, to be honest. There’s a trendline. It’s downward. What else do you want?
unique link to this extract

---

Inside the Silicon Valley push to breed super-babies • The Washington Post

Elizabeth Dwoskin and Yeganeh Torbati:

»

A group of well-heeled, 30-something women sat down to dinner last spring at a table set with pregnancy-friendly mocktails and orchids, ready to hear a talk about how to optimize their offspring.

Noor Siddiqui, the founder of an embryo-screening start-up and the guest of honor at the backyard event in Austin, offered a grand vision of custom-built algorithms and genome analysis that would help eradicate illness and disease. Shivon Zilis, a tech executive who had just given birth to Elon Musk’s then-secret 13th child, and other guests donned pastel-colored baseball hats Siddiqui handed out. They were emblazoned with a single word: BABIES.

Siddiqui is a rising star in the realm of fertility start-ups backed by tech investors. Her company, San Francisco-based Orchid Health, screens embryos for thousands of potential future illnesses, letting prospective parents plan their families with far more information about their progeny than ever before. For now, her approach has been taken up mostly in her moneyed social circle. But one day, maybe not far off, it could change the way many babies are made everywhere — posing new moral and political questions as reproduction could increasingly become an outcome not of sex but of genetic preselection and data-mining.

“For something as consequential as your child, I don’t think people want to roll the dice,” the 30-year-old entrepreneur told The Washington Post.

It is now standard for pregnant women and couples undergoing in vitro fertilization (IVF) to test for rare genetic disorders stemming from a single gene mutation, such as cystic fibrosis, or chromosomal abnormalities such as Down syndrome. But Orchid is the first company to say it can sequence an embryo’s entire genome of 3 billion base pairs. It uses as few as five cells from an embryo to test for more than 1,200 of these uncommon single-gene-derived, or monogenic, conditions. The company also applies custom-built algorithms to produce what are known as polygenic risk scores, which are designed to measure a future child’s genetic propensity for developing complex ailments later in life, such as bipolar disorder, cancer, Alzheimer’s disease, obesity and schizophrenia.

«

Of course if the score is too low for an embryo then it won’t get implanted. They’re all very optimistic that IVF will replace the old method of, er, implantation. Seems unlikely: the hit rate is far too low, and brings its own problems. (Thanks Karsten L for the link.)
unique link to this extract

---

All humans emit subtle light until they die, study suggests • BBC Science Focus Magazine

Hatty Wilmouth:

»

You, along with all living things, produce subtle, ethereal, semi-visible light that glows until you die, according to a recent study.

You would be forgiven for jumping to the conclusion that this spooky luminescence is evidence that auras exist, or something similar.

But Dr Daniel Oblak, physicist at the University of Calgary and last author of the study, told BBC Science Focus that, while auras are a metaphysical, spiritual, unscientific idea, this light is not. Instead, it’s called ultraweak photon emission (UPE) and is a natural product of your metabolism.

“I normally point out that UPE is a result of a biochemical process and in that sense is related to what happens in a glow-stick, which no one suspects of having an aura,” he said.

“UPE is so weak that it is not visible to the human eye and completely overwhelmed by other sources of light, unless you are in a completely dark room.”

That’s not to say that shutting your curtains and turning off your lights will allow you to see your own glow. This light is between 1,000 and 1,000,000 times dimmer than the human eye can perceive.

UPE is produced when chemicals in your cells create unstable molecules known as reactive oxygen species (ROS), basically byproducts of your body’s metabolism.

When ROS levels rise, they cause other molecules to become ‘excited’, meaning they carry excess energy. It’s this energy that causes light to be emitted.

«

New meaning to “glowing up”.
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