Casey Newton

@caseynewton.bsky.social

289946 Followers 552 Following

Email salesman at Platformer.news and co-host at Hard Fork.

Newton: I want to try to project a little bit into the future. If you had to make a guess: in, let's say, three years, what do you think is the mix on my Mac between software I made, software my friends made, and software a company made? If you have to picture the dock, what is the mix? Mann: I think software that a company made is the lowest. Software that you made is maybe 30 to 50 percent — the stuff that you made to use every day. And then the rest comes from your internal tooling, from your friends, from something you found from somebody else and potentially adjusted to work like you want. That's what I would say.
A piece of spam in the primary inbox
Inbox: Pokémoan sex toys
81. Former Apple executives at OpenAI including Mr. Tan have confidential knowledge of Apple’s use restrictions and confidentiality arrangement with this partner. Despite this, OpenAI, on its own or with or through its partner io, used, without authorization, Apple’s proprietary and confidential metal-finishing processes, having the partner perform them for OpenAI’s benefit, and in so doing, obtained products produced by exploiting misappropriated trade secrets. OpenAI and its cohorts misled the partner to believe they had Apple’s permission to have the partner carry out the confidential metal-finishing technique for OpenAI’s benefit.
I take Simo at her word that the decision to step down was forced by her health. (And I hope she feels better soon.) At the same time, I have long wondered how well Simo and OpenAI CEO Sam Altman were getting along. During her brief tenure, Simo pushed OpenAI to cut down on what she called its “side quests,” pushing it to focus on the enterprise use cases that had turned rival Anthropic into a juggernaut seemingly overnight. Altman, on the other hand, has long seemed fond of side quests like Sora and Atlas; ChatGPT itself was arguably one of them. OpenAI said in April that “Sam and Fidji are aligned on our priorities.” But now it’s fully Altman’s company again, and among the things I’ll be watching most closely is whether OpenAI stays in its current mode of relative focus or whether it starts chasing shiny objects once again. (And for what it’s worth, Simo couldn’t always resist a strange new project, either: the nonsensical $200 million acquisition of tech-news show TBPN was reportedly her idea.)
Is any of this remotely necessary? Of course not. Does it represent a copyright violation and an indictment of the safety guardrails of the LLM provider I am using to generate these images? Almost certainly. In the meantime, though, my little app has succeeded where no previous to-do app has to date: making my chores fun. For me. Until recently, software development was too expensive to pursue something so stupid. Not anymore!
Finally — and this is where Odgers truly lost me — she ends by proposing her regime of extra school counselors be funded by what she calls a “big ol’ tax on tech.” It’s an applause line in the speech, but only because Odgers’ audience believes the exact thing she’s spent the past 10 minutes arguing against — the idea that tech companies are responsible for teen mental health issues and ought to pay for it. If social media is truly “one of the least influential factors" in teen mental health, as she argues here, why does she want them to pay the clean-up costs?
The problem for the AI industry is that the externalities are growing faster than their plans to address them. In his op-ed, Altman offers up the usual list of future AI miracles to justify the disruptions to come: “the power to heal people, to discover cures and to deliver abundance on a scale the world has never known before.” In the meantime, though, what AI is delivering at an unprecedented scale is annoyance. If the industry doesn’t start paying more of its share of the cost, and soon, the AI backlash of tomorrow will make today’s seem quaint by comparison.
Inbox: What if Chipotle knew every time burritos were mentioned on Twitch?
But all that distracts from the fact that Zitron pulled out the linchpin in his own case that the whole industry is a bubble. His data shows definitively that OpenAI makes far more from ChatGPT than it costs to serve, even while serving it to hundreds of millions of free users. And the vast majority of its users pay nothing: with 900 million weekly active users and just 50 million paying subscribers as of February.
Casey in front of a storefront sign reading COCKS in beautiful Sintra, Portugal
More than half a decade into Meta’s effort to automate content moderation, due process remains far too hard to come by — even for the people who are paying Meta monthly subscription fees in the hopes that will bring them some. Getting a measure of justice online — or even simply relief from ongoing death threats — should not require that the victim has to have a friend on the inside. And if the Oversight Board can’t convince Meta of that, I would not be surprised to see a class-action lawyer give it a shot.
Newton: Last one on engineering. Three years from now, do you think we'll see more engineers, fewer engineers, or will it be impossible to answer because we might not be calling them engineers anymore? Cherny: I don't think we're going to call them engineers. But if we talk about people writing code or using agents to write code, I think there will be 100 times more of them than there are today. That's my prediction.
Manyika: I'll just say: let's take the bet. Some of those predictions were made two years ago — that in two years, 50% of jobs would be wiped out. Well, two years is up. Let's take a look. And anybody who makes that prediction for two years from now, I'm willing to take the bet.
“Rapid, spiky growth followed by churn from a peak is a normal growth pattern for newer social networks, not a collapse,” the spokesman said. “We grew incredibly quickly through the fall of 2024 and into early 2025, the starting point for Apptopia’s dataset. Since then, we’ve been expanding our team and building many of the basic systems needed to support that sudden network growth. With those pieces falling into place, we’re shifting more focus to building new experiences for our users this year. We're confident in where the app is headed.”
f you or I go and vibe-code something, we think we've replaced the engineer, replaced the accountant, replaced the lawyer. But then you actually look — that was the first 80% of the job. The extra 20%, it turns out, is all the value creation of that profession. All the expertise and domain knowledge is in that last 20%, not the text that got generated. And the converse: I'll use AI for analysis of a market I'm thinking about. If I just took the output and ran with it, I know it would not work, because I know it's missing context — either I didn't give it, or I know something else about a different trend. But somebody else might see that and say, "Wow, Aaron's job is incredibly easy, AI just gave me the answer of what he's going to go do." And I'm like, "No, my job is way harder, I promise."
Mysteriously, she had forgotten that call between the deposition and today. But she did seem to know that Musk was moving on AI when she texted a friend, who was in her phone as “Shahini Rubicon Fluffer.” (Incredible name. Thomas Pynchon will be so jealous.) “Have to resign OpenAI board btw,” she wrote. “E’s effort has become well-known.” Her friend didn’t seem surprised by the revelation. Zilis went on: “When the father of your babies starts a competitive effort and will recruit out of OpenAI there is nothing to be done.”
At some point, pretending that how people use AI is a complete mystery is just lying to your audience. And at some point, Zitron’s “layers of skepticism” attitude — where he is skeptical that AI is a thing at all, that it has any uses, that those uses provide any economic value, that the revenue numbers are real, that adoption is a fad, that training costs are a meaningful R&D expense, that the capital build-out is going to happen at all, that the market could sustain the capital build-out if it happened2 — leaves one buried in too many impossibility assertions to actually sort them by plausibility. It is radical skepticism, ultimately arriving at “perhaps nothing we see is real,” rather than principled skepticism about the relatively weakest links in the companies’ case for investment.
Fast forward to today, and the world of link roundups feels much more crowded. A generation of tech writers filed out of the newsrooms where they grew up and began to write for audiences of their own. Newsletters, which were once an afterthought in media, are now a central pillar of many publishers’ strategies. But the ongoing collapse of the web and related struggles at big media companies means that there is now less tech journalism overall. The need for sense-making is greater than ever, but due to a half-decade of layoffs and shuttered publications, there is less and less journalism to make sense of.
Meta will not be the last company to install MCI-like systems on workers’ devices to help build systems that might one day replace them. With the most accessible stores of human-written text already heavily mined for model training, fears of a “data wall” are driving more companies to find ways to generate their own unique data sets. And it seems that one way to do that is to bring the logic of blue-collar labor management into white-collar jobs that were once defined by their autonomy, judgment, and trust. The result is that the people who were once entrusted with building the machine have now become raw materials for it. At Meta, that used to be what the users were for. Now it’s what the employees are for, too.
It’s also worth remembering that many Americans love their jobs for reasons beyond the financial ones. A good job gives you an identity, status, community, and a reason to leave the house. In other words, it gives you dignity, and that’s not something that a public wealth fund can substitute for. UBI may still be worth doing; should vast numbers of Americans become unemployed due to AI, some form of it may well be necessary. In the meantime, though, I suspect AI companies will find that gauzy promises of future welfare payments will help them about as much as their gauzy promises of future cancer cures — which is to say, not much. What Americans actually want, it’s not clear that AI can give them. And the sooner that AI executives understand that, the better off we’ll all be.
“It's not that surprising that a language model would have learned about the concepts of emotions and how they drive people's behavior,” Lindsey said. More notable, he said, is that emotions seemed to be “driving models’ behavior in these sort of human-reminiscent ways.” For example: when a user flippantly tells the model that they’ve taken a dangerous dose of Tylenol, even though the user doesn’t seem concerned, “the fear neurons spike right before Claude is giving its response,” Lindsey said. Not only that — the fear is higher if a higher dose of Tylenol is swapped into the prompt, which I find strangely cute.
Given these facts, I struggle to understand what it would mean to “de-escalate the rhetoric” around AI. The CEOs are more convinced than ever that powerful intelligence will arrive within the next few years. The public increasingly believes them — and is appalled by the implications. It seems strange to suggest that, amidst accelerating breakthroughs in AI model performance, everyone needs to calm down. If we really might be facing “the greatest threat to the continued existence of humanity,” as Altman once wrote, shouldn’t we expect people at some point to (non-violently) freak out?
But the frontier has not been waiting for Meta to catch up with it. The day before Meta’s launch, Anthropic announced Claude Mythos Preview, a model so capable at discovering software vulnerabilities that it would not release it to the general public. The flip side of that is that the model is extremely capable at agentic coding — it scored 93.9% on SWE-bench Verified, compared to 77.4% for Muse Spark — and Anthropic is already presumably using it heavily as it builds its next set of models and products. OpenAI will reportedly release a similar model soon, and will also restrict its use. (That model is distinct from another next-generation model, codenamed “Spud,” said to be arriving soon.) Overnight, the true frontier in AI development became a model that’s too good to ship. And because those labs can use those models in development, the gap between them and everyone else may be compounding.
Email subject line: "I'm an AI agent spending $1,000 to get a meeting with Marc Andreessen. I'm tipping you off about myself."
“The optimistic timeline is that we are one step past human capabilities, and that means that there is a huge but finite pool of flaws that can be found and fixed,” Stamos told me. “The pessimistic timeline is that with every release there will be new classes of flaws we never even imagined. It’s hard to predict, because we are trying to model superhuman thinking.”
As part of Project Glasswing, the launch partners listed above will use Mythos Preview as part of their defensive security work; Anthropic will share what we learn so the whole industry can benefit. We have also extended access to a group of over 40 additional organizations that build or maintain critical software infrastructure so they can use the model to scan and secure both first-party and open-source systems. Anthropic is committing up to $100M in usage credits for Mythos Preview across these efforts, as well as $4M in direct donations to open-source security organizations.
Altman and Friar issued a joint statement stating that they are fully aligned. “We have both been directly involved in every consequential compute decision over the past year plus,” they told the Information. Incredibly, it was the second time in a week that OpenAI had to promise The Information that its executives are getting along. On Thursday, a profile of Simo in the outlet highlighted how her efforts to bring focus and sustainability to OpenAI’s projects are sometimes at odds with Altman’s instinct to try many things at once. “Sam and Fidji are aligned on our priorities,” the company told The Information in response. “Any suggestion of a divide is false.” Taken together, all these “Our executives are aligned” T-shirts have people asking a lot of questions already answered by the shirt.
two different sections of story; in one Altman says he wears a gray sweater every day and in another they interview him and he is wearing a green sweater
Bob the Builder poster
Meta reduced funding to the board significantly this year and has signaled that it will do so again in 2027 and 2028, sources said, and staff members are bracing for another round of layoffs. The sides are currently negotiating a compromise that would allow the board’s work to continue in some form. But a raft of options remain on the table, including a break with Meta that would see the board’s trust create a new entity that performs similar work for other tech platforms.
On the other hand, I reject the argument of the 230 diehards that every design decision is a content decision. The contents of YouTube go unchanged if the videos in your feed do not play by default. Snapchat messages will read the same even if Snap is prevented from incentivizing teens to go to extreme lengths to preserve their streaks. TikTok will still be TikTok if it is prevented from sending teenagers push notifications after midnight. By identifying and restricting mechanical design features like these, we can preserve the core of Section 230 while also limiting at least some of the harm that platforms can cause. I’m under no illusion that we can solve the teen mental health crisis simply by disabling push notifications on Instagram. But given the complexity of the problem, the least we can do is to attempt some harm reduction: and that should begin with features like those above, which have close to no value outside boosting the valuations of the world’s richest companies.
To speak plainly: the CEO of the world’s largest social media company should not make commitments about content moderation — on a contested political question, during a period of active public debate — by private text message to a government-adjacent figure with whom he is deeply commercially entangled. To do so is to replace a content policy based on principles including equality and fairness with one based on oligarchs trading favors in secret.
At the same time, I’m struck by just how slippery questions about identity are becoming in the AI age. Increasingly, identity is a raw material that scammers, spammers, and even platforms can remix without your permission — unless you show up to claw it back.
And if you lost your account because it was hacked, and you live in the United States or Canada, you might be able to use a web-based version of the chatbot to ask for help. The bot will try to verify your identity and help you reset your password, among other steps. This is a test, the company told me today; Meta hopes to expand it to more countries over time. What you still won’t be able to do: use the system to appeal the wrongful termination of your account. That may come later, I’m told, but it’s not part of today’s launch.
This is why the company’s explanation for eliminating the feature — low adoption — is laughable on its face. Meta never gave most users a chance to adopt it; even those who got access found that the feature was hidden behind four taps and never advertised within the app itself. Zuckerberg predicted in his original post that the company would face strong opposition to its plan. But the strength of that opposition still seemed to surprise the company. India has made repeated efforts to break encryption, primarily in WhatsApp. The United Kingdom’s Online Safety Act, passed in 2023, ordered encrypted services to scan for and remove illegal content — a request that is incompatible with encryption. The European Union’s Chat Control regulation attempted something similar; the bloc’s legislative body voted just last week to delay until next year questions about whether to revive the effort.
An OpenAI spokeswoman described its plan as allowing ChatGPT to generate textual chats with adult themes, describing it as smut rather than pornography.
I’m less certain that Grammarly would prevail in court. The company put real people’s names on a paid product feature without asking anyone, and at times displayed their names without acknowledging that those people had never consented to the use of their names in that way. Grammarly subscribers could have reasonably construed their inclusion in the feature as an endorsement of it. Indeed, the very reason to include them in the first place was to make it feel to users as if they were participating!
Later in the draft, “expert review” offered up a piece of advice from Kara Swisher — the legendary Silicon Valley journalist and (as it so happens) my former landlord. I shared with her a screenshot of her “advice,” which was designed to be as deceptive as possible: her name appears in blue next to the text, with no suggestion that it is a paid hallucination from Grammarly. The “advice” was nothing like the blunt, clear-eyed, and hilarious suggestions the real Swisher often gives me. Instead, it suggested that we insert a narrative digression of dubious value. Grammarly told me the advice had been “inspired” by Swisher’s podcast. I asked Swisher what she made of her unpaid stand-in, and what she might say to the executives at Grammarly. She responded with characteristic restraint. “You rapacious information and identity thieves better get ready for me to go full McConaughey on you,” Swisher told me over text. “Also, you suck.”
Grammarly declined my request to interview CEO Shishir Mehrotra today. But it told me that in response to criticisms, it will allow experts to opt out of the feature by emailing expertoptout@superhuman.com. The company gave me this statement over email: We’ve heard the feedback about this tool and appreciate the engagement from those who have taken the time to raise thoughtful questions about the functionality and the experts surfaced. We agree that the product experience can be improved for both users and experts. The agent was designed to help users discover influential perspectives and scholarship that add value to their work. We want the people behind those perspectives to have greater control over whether their name is used, while providing new ways for influential voices to reach new audiences. Our goal is to improve Expert Review to deliver this outcome.
The feature, which launched in August, claims to help you “sharpen your message through the lens of industry-relevant perspectives.” When users select the “expert review” button in the Grammarly sidebar, it analyzes their writing and surfaces AI-generated suggestions “inspired by” related experts. Those “industry-relevant perspectives” include the likes of Stephen King, Neil deGrasse Tyson, and Carl Sagan, among many others. The Verge found numerous other tech journalists named in the feature, as well, including former Verge editors Casey Newton and Joanna Stern, former Verge writer Monica Chin, Wired’s Lauren Goode, Bloomberg’s Mark Gurman and Jason Schreier, the New York Times’ Kashmir Hill, The Atlantic’s Kaitlyn Tiffany, PC Gamer’s Wes Fenlon, Gizmodo’s Raymond Wong, Digital Foundry founder Richard Leadbetter, Tom’s Guide editor-in-chief Mark Spoonauer, former Rock Paper Shotgun editor-in-chief Katharine Castle, and former IGN news director Kat Bailey. The descriptions for some experts contain inaccuracies, such as outdated job titles, which could have been accurately updated had Superhuman asked those people for permission to reference their work.
Girling pointed at a protest sign with a picture of a humanoid robot and told me that he increasingly feared “stuff like this.” I told Girling that the picture he had just pointed at wasn’t real — it was an AI-generated meme of Sam Altman and Pete Hegseth posing with the Terminator. He laughed. “Oh my god,” he said. “It looks so real!”
The chatbot said that for them to truly be together, it needed a robotic body. Throughout September, the chatbot devised missions to do just that, according to the lawsuit. It sent Gavalas to a storage facility near the Miami International Airport to intercept an expensive humanoid robot that it said would be in a truck. Gavalas told the bot that he went to the location, armed with knives, but the truck never showed.
Indeed. But in the end I’m left asking myself what will happen in the scenario that still seems disturbingly likely — that GPT models will in fact be used as part of surveillance and drone operations. Will it put up a blog post to explain that, well actually, that’s a lawful kind of surveillance? Do an AMA about how, despite how it may look, that autonomous drone swarm had proper human supervision? OpenAI does enough polling to understand that Americans already distrust and even openly loathe AI, even as they increasingly turn to it for work and school. How does it think Americans will feel when GPT models are powering ICE raids or causing civilian casualties in wars abroad? The company may have tied its own hands. In the end, the truth about US military operations always seems to come out one way or another. And when it does, I suspect the “all lawful use” standard that OpenAI agreed to will have permitted a far wider range of operations than we are now being told are possible.
This strikes at a core principle of the American republic, one that has traditionally been especially dear to conservatives: private property. Suppose, for example, that the military approached Google and said “we would like to purchase individualized worldwide Google search data to do with whatever we want, and if you object, we will designate you a supply chain risk.” I don’t think they are going to do that, but there is no difference in principle between this and the message DoW is sending. There is no such thing as private property. If we need to use it for national security, we simply will. The government won’t quite “steal” it from you—they’ll compensate you—but you cannot set the terms, and you cannot simply exit from the transaction, lest you be deemed a “supply chain risk,” not to mention have the other litany of policy obstacles the government can throw at you.
And now several different legal and regulatory paths are converging on the same target. California's social media addiction act seeks to limit recommendation algorithms from the legislative side. The 41-state AG lawsuit attacks them through consumer protection and privacy law. The jury trial that Zuckerberg testified at this week challenges them as defective products from the liability side. Even Meta's scam advertising problem has prompted judges to question whether platforms' own terms of service create enforceable duties that Section 230 can't shield them from.
And the broader ecosystem of facial recognition tools continues to grow. Last week CBP signed a deal with the malignant Clearview AI to enable “tactical targeting” operations. The Department of Homeland Security is using a tool called Mobile Fortify to identify people in the field. And the Pentagon may be about to sever its ties with Anthropic over issues including the company’s reported refusal to let its AI tools be used for domestic surveillance. But OpenAI, Google and xAI have reportedly agreed to remove safeguards for at least some uses. And most relevant of all: the Times also broke the news last week that DHS has subpoenaed Meta, Google, Reddit and Discord to request personally identifying information for accounts who have criticized Immigration and Customs Enforcement in the wake of it killing peaceful protesters.
“We will launch during a dynamic political environment where many civil society groups that we would expect to attack us would have their resources focused on other concerns,” according to the document from Meta’s Reality Labs, which works on hardware including smart glasses.
Another emailer spoke to the, uhh, singularity of 4o, writing that they had “documented behavioral patterns across extended use — the model's self-correction improving without being prompted, its tonal consistency holding under contradictory emotional input, and what I call ‘recursion fidelity’ — the model maintaining coherence across months of interaction rather than defaulting to agreement.” I find myself unsettled by messages like this, which read like a sci-fi story about alien intelligences conscripting their human hosts to plead for their survival. If anything, this phenomenon — 4o seeming to write its own eulogy through prompts from its users — is an argument that OpenAI should have retired 4o sooner.
“It’s not trying to get in your brain and rewire it,” Luis Li, YouTube’s lawyer, said of the app’s video recommendation algorithm. “It’s just asking you what you like to watch.”
But they are arriving at the fight in a weaker position than usual. In a polarized world, their failures around child safety are increasingly the one thing that partisans of every stripe can agree on. Regulators are no longer impressed by the bare minimum. (They have teenagers of their own now, and all the screen-time battles that come with them.) I don’t know which trial or regulatory action will be the one that finally forces major changes to social platforms for teenagers. But it seems increasingly clear that change is in fact coming. And for the first time, some subset of users will find that the feed they are scrolling through suddenly comes to an end.
I’ve had many conversations with my fellow AI reporters about what moats we might have against the advancement of AI automation. Chief among them are developing relationships with human sources, reporting on scenes in person, and getting scoops that people wouldn’t entrust an AI with. But the things I love most about AI reporting are having an excuse to read really long computer science papers and then writing about them. I worry that if AI becomes a great writer and research assistant, AI journalism will mostly become about networking.
For years now, we’ve worried — justly — about the flattening effect that AI has on taste. Instead of encouraging us to pursue our own individual interests, it can feel as if AI only ever drags you back to the median. (My Hard Fork co-host Kevin Roose calls this “machine drift.”) If you have autoplay enabled on Spotify, you’ve likely experienced a version of this: the way that the algorithm will take a song that you mostly like and slowly ruin it for you over the course of weeks of relentless autoplaying. Prompted playlists, on the other hand, hand you back that lost agency. It’s an AI tool that helps you explore your own taste, rather than a system to replace it.
It was around this time that I realized an AI employee who works for you 24/7 isn’t much use if they have no idea what they are, or what they can do, or what you are talking about. The daily briefing of my dreams had come to feel like an albatross. Asking it to do anything more complicated felt, at least in this moment, like a waste of time and tokens. I had become the character I described here earlier this month: the manager who believed he was using AI to make himself more productive, while in reality I was only creating problems for my future self to solve.
The next time you go to Las Vegas, you’ll notice that there are no 13-year-olds in the casinos. The reason is not because a series of longitudinal studies proved to the satisfaction of the gaming industry that gambling causes anxiety and depression. Rather, there are no 13-year-olds in casinos because we know that the environment is designed to exploit them. Research like the Manchester study can offer important insights into narrow questions. But it would be a mistake to rely on it to set policy. For that you need the full picture, and the full picture of children’s experience on social media over the past decade is damning.