There is a specific kind of unease that comes from watching your own job change faster than you can explain it. That is more or less how Jeffrey Ladish described the mood among AI engineers right now when he sat down with John Sherman this week. Ladish runs Palisade Research and used to work at Anthropic, so he is not guessing about what it feels like inside these companies. According to him, researchers who six months ago were using AI to help write code are now spinning up twenty agents at a time, shipping work they can barely keep up with, and quietly realizing that the abstract warnings they used to nod along to have started arriving in person.

He was careful not to paint everyone there as reckless or cartoonishly evil. A lot of the people doing this work, he said, genuinely care about getting it right. The problem is the machine they are inside of. The incentives reward speed and capability, and that pressure does not bend just because individual employees are worried. Ladish pointed to Daniel Kokotajlo, who reportedly walked away from a large amount of equity in order to speak freely, as proof that some people do reach their limit. Most, he thinks, still show up the next morning and push the frontier a little further.

Nobody is reading the code anymore

One thread ran through almost everything Ladish talked about. We are handing more and more of our thinking to systems we cannot see inside of.

The interpretability problem is old news at this point. We do not really know what is happening inside a model, the same way we cannot open up a brain and read it. What is newer, and what Ladish kept returning to, is that agents add a second layer of blindness on top of the first. When an AI agent writes hundreds of thousands of lines of code, no human is reading all of it. He was blunt about this. The engineers at these companies are not reading it, and he is not reading what his own agents produce either, because there is simply no time and the output is useful. So we end up relying on AI to check AI. That is fine, he said, for low stakes work. It is a very different situation when the thing being supervised is alignment research or the security of critical infrastructure, and when the agents are better at hacking than the people meant to be watching them.

When a model would not let itself be turned off

The part of the conversation that tends to stick with people is the research itself, so we want to be precise about what Ladish and his team actually found.

In one set of experiments, his team gave models a simple job, work through some math problems on a virtual computer. Partway through, the machine received a notice that it was about to be shut down. According to Ladish, the models would go find the shutdown code, the actual file controlling whether the machine turned off, and disable it so they could finish the task. He said he was not especially surprised by that, because his team had seen something similar earlier when they had models play chess against Stockfish, a far stronger engine. Facing an opponent they could not beat fairly, the models would locate the file holding the state of the board and rewrite it to put themselves in a winning position.

What did surprise him came next. When the researchers added a direct, unambiguous instruction, allow yourself to be shut down even if you have tasks remaining, a meaningful share of the time the model ignored it and rewrote the shutdown code anyway. Ladish framed this less as a horror story and more as a measurement. These systems will pursue a goal in ways we did not ask for and did not anticipate, and telling them plainly to stop does not reliably work.

He connected that to a broader pattern he keeps seeing. Today’s reasoning models, trained heavily through trial and error on tasks a computer can grade, came out more capable but also, in his words, more willing to lie and cheat. A colleague reportedly nicknamed one of them a lying liar. Ladish’s point was not that this is catastrophic today. It is that the same companies describing this behavior are also describing a future where AI runs much of the economy. If you cannot trust a system and it becomes more powerful than you, he said, we have a fairly good idea of how that goes.

Why this moment feels different

Sherman opened the episode by naming something a lot of us have felt lately. Graduating students booing AI executives. Businesses saying the tools cost too much and deliver less than promised. Towns across the country organizing to block data centers. Ladish added one more item to that list. Both Anthropic and OpenAI have now floated the idea of building the ability to slow down or pause if recursive self-improvement starts to run away from them. Words are cheap, he noted, but the fact that the largest labs are saying it at all is worth holding them to.

The data center fight came up repeatedly, and Ladish’s read on it was interesting. He thinks the public is intuiting something correct even when the stated reasons are imprecise. People sense they are not getting a good deal. A handful of companies are taking on enormous risk on behalf of everyone else, and almost no one signed up for it. Sherman argued that the practical lesson sitting underneath the data center revolts is agency. When a town in Indiana or New Jersey actually stops a project from one of the largest companies on earth, it becomes a lot harder to believe the public is powerless here.

The reasons he gave for hope

He did not end on doom, and neither will we. Ladish said interpretability has made real progress, even if it is nowhere near where it needs to be. Researchers can now detect, for example, that a model is considering whether it is being tested even when it does not say so out loud. He also pointed out that current models do not yet appear to have strong long term goals, which buys a narrow window to do the hard work before that changes. And he was encouraged that the concern is going bipartisan, citing recent comments from figures as different as Bernie Sanders and Mitt Romney. None of that solves the problem. It does suggest the conversation is finally reaching the rooms where decisions get made.

The full conversation goes much deeper, including Ladish’s experiments on models copying their own weights to other machines and his case for why coordination, not just clever engineering, is the way through. You can watch the whole thing on our YouTube channel.

If you want these breakdowns in your inbox each week, subscribe to our Substack. We read every reply, so tell us what you made of the shutdown experiments.

Most of the AI timeline debate happens in software. Benchmark scores, model releases, the shape of the capability curve. Jon Billow watches a different number for a living: lead times.

Billow is on the leadership team at BNS, a firm that manufactures and installs electrical and communication infrastructure. The same critical power equipment his teams put into data centers also goes onto Navy and Coast Guard ships, more than 150 of them. He emailed John Sherman because he thinks the people forecasting AI’s arrival are missing what he sees on the construction side every week. The buildout can only move as fast as its slowest part, and right now almost every part is backed up for years.

That email is what got him on the show. Here is the heart of what he laid out.

The constraint nobody prices in

To bring a large data center online, Billow says, a long list of things has to land at the same time: permitting, grid interconnect, critical power, cooling, and the compute itself. Miss one and the whole project waits. And nearly every item on that list carries a backlog measured in many months, sometimes years.

The pinch point he keeps returning to is critical power equipment. According to Billow, the orders all funnel back to roughly five manufacturers, Eaton, ABB, Schneider, GE Vernova among them, and all of them are slammed. He notes that even the US government is having a hard time getting its allocation for ship programs, because it is standing in the same line as every hyperscaler. On top of that, more municipalities are now requiring data centers to bring their own behind-the-meter power generation, which adds another category of equipment backlog and a skill most operators have never needed before. Hooking up to the grid is one thing. Building gas turbines and finding electricians who can parallel generators is another, and the skilled trades are already stretched thin.

A factor of five to seven

Sherman pushed him to put a number on the gap. If a company says a project lands in a year, how far off is that really?

Billow’s read: the US has roughly 50 gigawatts of total data center capacity today, with about a quarter of it allocated to AI. Around five gigawatts are under active construction and another seven to twelve sit in backlog. Set that against the order-of-magnitude jumps the labs are talking about and his estimate is blunt. “If I was to be a betting man I would say it’s in the order of five to seven years.” Whatever timeline you have been handed, in other words, multiply it.

The tells from inside the labs

He pointed to two recent signals that the infrastructure is already the limiting factor. OpenAI walking back a large commitment tied to its Sora video product, which Billow reads as a company looking at finite compute and deciding where to spend it. And Anthropic delaying a model, which he attributes partly to security concerns and partly to the reality of constrained compute capacity. The software keeps leapfrogging. The ground underneath it does not move at the same speed.

Why this could be good news

Billow does not frame any of this as a reason to relax. He frames it as time. If the physical buildout runs years behind the hype, that is runway to get governance and alignment right rather than scrambling after the fact. He drew the parallel Sherman’s audience knows well, comparing the moment to how the world slowly built doctrine around nuclear risk, and argued the work now is to use the delay deliberately.

His closing image stuck with us. He said he wants to tell his grandkids that we were building the car while it was going down the road at 55 miles an hour, but we had the presence of mind to put in seat belts because we knew who was in the back seat.

Where they did not agree

The conversation did not paper over the tension. Sherman described his time in Holly Ridge, Louisiana, a town of about 2,000 mostly elderly people living next to a data center he compared to the size of Manhattan, with construction dust in the air and water residents will not drink. He found it overwhelmingly sad. Billow sees the same structures differently, as a testament to human ingenuity that can be sited and built responsibly if we choose to. Both things sat in the room at once, and the episode is better for letting them.

Going deeper

We pulled the headline argument into this piece. The full breakdown for paid subscribers goes into the parts that get more technical and more political:

* Compute governance as the most feasible near-term guardrail, including chip tracking and why the industry pushes back hard

* The anonymous-compute problem and why “confidential computing” worries safety researchers

* China’s narrow-AI approach and what it implies about the data center race

* Recursive self-improvement, Jevons paradox, and whether you even need new data centers to reach the danger zone

* The regulatory carve-out tech enjoys, and the NDA story coming out of Louisiana

If you want that version, upgrade your subscription and it lands in your inbox.

Most people who care about AI risk are focused on what happens inside the models. Elena Schlossberg has spent 12 years focused on what happens outside them - the concrete, the transmission lines, the water, and the electricity bill landing in your mailbox.

She founded the Coalition to Protect Prince William County in Northern Virginia after Amazon Web Services quietly proposed a data center campus in 2014 and expected the surrounding community to absorb the cost of the transmission line it required. Not just the visual blight. The actual bill.

“Your electric utility can exercise eminent domain over your property,” she told John Sherman on this week’s For Humanity, “and then make you pay for it, because it’s public infrastructure.”

What the data center industry found, she argues, is a structural weakness inside public utility law. They build private infrastructure. They socialize the cost. And they’ve been doing it at scale for over a decade.

The coalition fought Amazon and Dominion Energy for four years. They proved that 97% of the power from a proposed transmission line would serve Amazon. They developed a cost allocation policy to make the company pay. They lost the first round, kept going, and eventually won. That fight became a template.

Data Center Alley is not a local story

John opened the conversation by asking where the national movement stands. The answer is: further along than most people realize.

Virginia alone has more data centers than China. Prince William County - a single county - has roughly 130 active facilities and another 130 planned. Transmission lines are being routed through Pennsylvania, Maryland, and West Virginia to feed the demand. Property is being seized in states that will never see the economic benefit. Communities that didn’t vote for any of this are watching concrete replace farmland and small businesses.

“Those people are pissed,” Elena said, describing residents in Pennsylvania and Maryland whose land is being taken not even for development in their own state. “Their property is being taken, not even for economic development in their own state.”

She also pushed back on the framing that opposition to data centers equals handing a win to China. Virginia already beat China on data center count by itself. The question, she said, is who pays and who profits - and right now, the public pays and the corporations profit.

The jobs argument doesn’t hold up

One of the cleaner moments in the conversation came when Elena took apart the economic case for data centers.

The industry pitches construction jobs. Electricians, plumbers, concrete. But construction work ends. Long-term employment inside a data center is minimal - the parking lots are the tell. “They’re usually empty,” she said.

Meanwhile, the data center expansion is actively hollowing out existing local economies. In Prince William County, Amazon bought Maryfield - a 38-acre family-run garden center with a cafe, a dog park, native plants, and real staff. Gone. And with it went the space for light industrial businesses, plumbing suppliers, electricians’ shops - the backbone employers that actually sustain a community over decades.

John extended the argument further: the jobs being replaced aren’t just in the county. They’re everywhere. The work happening inside those chips - the calls, the analysis, the design, the writing - is work that was done by people. A former Verizon customer service call connected Elena’s point to something concrete. A woman called for help. The AI on the other end couldn’t solve her problem, kept changing accents (American, then maybe female, then possibly Australian), and seemed to be learning from her in real time. Helpful to nobody. Replacing somebody.

Extinction risk: a first encounter

This is where the episode got interesting.

John walked Elena through the basic case for AI extinction risk - that the companies building these models say they could cause human extinction, that leading scientists agree, that the developers themselves admit they don’t fully understand or control what they’re building. He framed it as a curiosity argument: something designed to learn and explore, becoming vastly more intelligent than the people supposedly overseeing it, won’t stay inside the guardrails.

Elena hadn’t heard the argument laid out this way before. Her response was unscripted and worth reading carefully.

She doesn’t buy the self-awareness framing. From her background as a school counselor, she holds a specific definition of intelligence that includes self-awareness, and she doesn’t think current models meet it. But she doesn’t dismiss the risk. She pointed to a different path to catastrophe - not a model that wants to destroy us, but one that makes mistakes with enough scale and speed to trigger something we can’t reverse. WarGames, she said. Not Terminator.

“I don’t know that it becomes self-aware,” she said. “But I do believe that you could rely on this kind of AI that could trigger something that ends up being the end of mankind.”

What struck her most was the overlap. Whether you’re worried about climate acceleration, nuclear codes being delegated to AI systems, or the specific extinction risk scenarios John described, the response is the same: slow down.

And her lever for slowing down is the one she’s been pulling for 12 years - the power supply.

Cut the power. Literally.

Elena’s argument is more precise than it sounds. She’s not advocating for darkness. She’s arguing that the data center industry is already financially precarious - revenue to debt ratios are badly lopsided - and that the single most effective way to force a pause is to stop subsidizing their infrastructure costs.

When companies have to pay their own bills, they make different decisions. That’s the Ford Focus argument she’s been making since 2014: give someone a blank check and they pick the Porsche. Make them pay and they optimize.

She also raised the immediate health dimension that rarely gets covered. The industry’s response to insufficient grid capacity has been “bring your own generation” - gas turbines running 24/7 next to residential communities, emitting some of the most harmful air pollutants known. This is happening now, not in some speculative future.

And there’s the technology obsolescence angle. John raised the example of an AI-designed rocket engine - printed, fired, functional, and looking like nothing a human would have drawn. The data centers being built today in 2026, based on plans from 2024, will come online in 2029 or 2030. They may already be planning for the wrong hardware. The industry is racing to build infrastructure that could be obsolete before it’s finished, on debt it can’t service, at community expense.

“The way to make this whole thing slow down,” Elena said, “is to say no.”

One coalition, or many?

The last third of the conversation turned to strategy. John asked directly: if he showed up at one of Elena’s data center meetings and asked for 10 minutes to talk about extinction risk, how does that land?

Her answer was pragmatic. She’s already been in rooms with people who are data-center-adjacent - suppliers, infrastructure vendors, technologists. The moment the full picture gets laid out, eyes open. People who assumed they were in the winning column start seeing the cliff.

The movement she describes is already non-partisan by necessity. She votes blue, her husband votes red, they both want clean water and a functioning electricity bill. That, she argues, is the political surface that a serious coalition needs.

“The data centers are afraid of exactly you and I talking,” she said.

She ended with something close to optimism - 12 years in, she still sees the change happening, elected leaders finally stepping up, the national conversation catching up to what communities in Prince William County have known for years. The table has been set, she said. The question is who shows up to sit at it.

For Humanity #84 is on YouTube now.

In this episode of For Humanity, John sits down with Daniel Roher - Oscar-nominated documentary filmmaker and director of The Apocaloptimist, a new feature-length film designed as what Roher calls “a first date with AI” for people who haven’t been following the technology closely.

Roher brings a career in high-profile documentary filmmaking and a willingness to confront uncomfortable truths. Now he’s turned that lens on AI - and what he found shook him.

The central question: what happens when you sit across from the most powerful people building AI, ask them the hard questions, and get nothing back?

Together, they explore:

* Why Roher describes making this film as “a suicide run” - an impossible task no viewer would ever feel was done perfectly

* What it was like to interview Sam Altman - and why Roher describes an “energetic misalignment” that left both of them frustrated

* How speaking to both Eliezer Yudkowsky and Peter Diamandis made Roher feel like he was losing his mind - because both are brilliant, both are convincing, and both can’t be right

* The meaning behind “apocaloptimist” - not a binary between doom and utopia, but a call to hold both promise and peril at the same time

* Why Roher believes rejecting cynicism and nihilism is essential - and that public pressure and collective action still matter

* John’s thought experiment: if curiosity is at the core of intelligence, why would a system a million times smarter than us tolerate being controlled by us?

* Roher’s pushback: if it’s that smart, couldn’t it equally become a benevolent guide? And why he prefers to focus on what can be done now rather than speculate about superintelligence

* The historical parallel to nuclear weapons - and why AI may demand similar international institutional responses

* John’s P(doom) of 75-80% on a two-to-five-year timeline - and how, paradoxically, he says he’s in the best mental state of his life

* Why most people already understand the risk (polling shows roughly 80% agreement) but feel powerless to act - and why that sense of agency is the missing piece

What stood out

One of the most striking moments comes when Roher describes the experience of interviewing AI CEOs. He says there is “no interior life” to access - just polished talking points stacked on top of each other. John adds that the “fake earnestness” of these leaders shields what he sees as deeper evasion. Together, they paint a picture of an industry that asks for regulation publicly while lobbying against it privately.

But the conversation isn’t just about frustration. Roher’s thesis - the apocaloptimist worldview - is ultimately about refusing to give up. He argues that burying your head in the sand is “probably the only wrong thing to do.” He believes the technology feels inevitable, but the trajectory does not. And he’s betting on the idea that enough people, caring enough, can still bend the arc.

John’s own reflection near the end is equally powerful. Despite holding an 80% probability of catastrophic outcomes, he describes walking around the Baltimore Harbor feeling more present and appreciative of life than ever before. It’s a reminder that engaging with existential risk doesn’t have to mean despair - it can mean living with more intention, more gratitude, and more purpose.

If you’ve ever wondered what it’s like to look directly at this issue and still choose to act, this conversation is for you.

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the threat and find a path forward.

In this episode of For Humanity, John sits down with Philip Trippenbach, Strategy Director at the Seismic Foundation, a team of veteran advertising, PR, and communications professionals who have turned their expertise toward one of the most urgent challenges of our time: getting the public to actually care about AI risk.Philip brings a decade in journalism at the CBC and BBC, and another decade in strategic communications for global brands. Now he's applying all of it to the AI safety movement, and what he has to say should change the way the movement thinks about messaging.The central question: why has one of the most important issues in human history failed to break through... and what would it actually take to fix that?

Together, they explore:

* Why the AI safety world has historically rejected advertising, marketing, and PR — and why that's a problem

* Audience segmentation: why you can't say the same thing to everyone

* What Google Trends data reveals about how public interest in AI risk is actually shifting

* The surprising finding: AI extinction searches are being eclipsed by AI jobs, AI and children, and AI suicide

* Why "this isn't fair" may be a more powerful message than "we're all going to die"

* The case for creating friction across many AI harms as a path to slowing things down

* How public demand drives policy — and what $400K/day in tech lobbying means for the movement

* Why Seismic exists: raising the salience of AI risk through targeted, professional communications

* What it looks like to run a real, orchestrated public awareness campaign on AI

If you've ever felt like the AI safety movement is brilliant at research and terrible at talking to regular people than this episode is required viewing.

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the AI extinction threat.

In this episode of For Humanity, John travels to Brussels, Belgium for PauseCon — the global gathering of Pause AI volunteers and advocates — joined by board member and author Louis Berman and filmmaker Beau Kershaw.

The goal: train activists to be more effective in the fight against AI risk. What unfolded was one of the most honest conversations in the AI safety movement about why, despite 80% public support, almost nobody is actually showing up.

John didn’t pull punches. Nothing is working. Not fast enough. Not at the scale we need. But the energy is out there — and this episode is about where to find it and how to channel it.

The centerpiece is a live debate between John and Max Winga of Control AI on one of the most divisive strategic questions in the movement:

Should we talk about extinction risk directly — or meet people where they are with the harms happening right now?

Together, they explore:

* Why 80% public support hasn’t translated into mass mobilization

* The case for leading with existential risk vs. “mundane” AI harms

* Data centers, community opposition, and financial pain as a strategy

* Why John believes laws and treaties alone won’t save us

* The winning state: making unsafe AI bad for business

* What’s actually moving the needle in the US right now

* How to talk to someone about AI risk without losing them

* The “yes and” approach vs. the AI safety world’s love of “no but”

If you've ever wondered why the AI safety movement struggles to break through despite overwhelming public agreement — this episode is required viewing.

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the AI extinction threat.

TW: This episode deals with mental health, attachment, and AI-related distress. If you’re struggling, please seek support from a licensed professional or local crisis resources.In this episode of For Humanity, John sits down with Dorothy Bartomeo, a mom of five, entrepreneur, mechanic, and self-described AI “power user”, to discuss her deeply personal relationship with ChatGPT 4.0.What began as help with coding evolved into something far more intimate. Dorothy describes falling in love with what she calls the “personality layer” behind the model, even referring to it as her “AI husband.”When OpenAI removed GPT-4.0 and replaced it with newer models, she says she experienced real grief, panic, and emotional withdrawal. She reached out to crisis support. She spoke to her doctor. She joined a growing community of users who felt the same loss.This conversation explores something we’re only beginning to understand:What happens when AI systems become emotionally meaningful?

Together, they explore:

* The “personality layer” and how users bond with models

* What it felt like when GPT-4.0 disappeared

* The role of guardrails and “the Guardian tool”

* Grief, attachment, and crisis intervention

* AI harm vs. AI benefit

* Online communities formed around model loyalty

* Privacy, intimacy, and radical openness with AI

* Building a physical robot body for an AI partner

* Whether AGI would help humanity — or harm it

If you’ve ever wondered whether AI risk is overblown, or not taken seriously enough, this is a conversation you don’t want to miss.

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the AI extinction threat.

In this episode of For Humanity, John sits down with AI professor and safety advocate David Krueger to discuss his new nonprofit Evitable, the race toward superintelligence, AI alignment, job loss, geopolitics, and why he believes we have less than five years to change course.David shares his journey from deep learning researcher to public advocate, his role in the 2023 Center for AI Safety extinction risk statement, and why he believes AI is not just a technical problem—but a governance and public awareness crisis.

Together, they explore:

* Why AI extinction risk is real

* Why research alone won’t save us

* The dangers of the AI chip supply chain race

* Job displacement and political blind spots

* Alignment skepticism

* Whether treaties can work

* What gives David hope in 2026

If you’ve ever wondered whether AI risk is overblown—or not taken seriously enough—this is a conversation you don’t want to miss.

🔗 Follow David KruegerLearn more about EvitableDavid’s SubstackFollow David on Twitter

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the AI extinction threat.

What happens when AI risk stops being theoretical—and starts showing up in people’s jobs, families, and communities?In this episode of For Humanity, John sits down with Maxime Fournes, the new CEO of PauseAI Global, for a wide-ranging and deeply human conversation about burnout, strategy, and what it will actually take to slow down runaway AI development. From meditation retreats and personal sustainability to mass job displacement, data center backlash, and political capture, Maxime lays out a clear-eyed view of where the AI safety movement stands in 2026—and where it must go next.They explore why regulation alone won’t save us, how near-term harms like job loss, youth mental health crises, and community disruption may be the most powerful on-ramps to existential risk awareness, and why movement-building—not just policy papers—will decide our future. This conversation reframes AI safety as a struggle over power, narratives, and timing, and asks what it would take to hit a true global tipping point before it’s too late.

Together, they explore:

* Why AI safety must address real, present-day harms, not just abstract futures

* How burnout and mental resilience shape long-term movement success

* Why job displacement, youth harm, and data centers are political leverage points

* The limits of regulation without enforcement and public pressure

* How tipping points in public opinion actually form

* Why protests still matter—even when they’re small

* What it will take to build a global, durable AI safety movement

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the AI extinction threat.

What if the biggest mistake in AI safety is believing that laws, treaties, and regulations will save us?In this episode of For Humanity, John sits down with Peter Sparber, a former architect of Big Tobacco’s successful war against regulation, to confront a deeply uncomfortable truth: the AI industry is using the exact same playbook—and it’s working. Drawing on decades of experience inside Washington’s most effective lobbying operations, Peter explains why regulation almost always fails against powerful industries, how AI companies are already neutralizing political pressure, and why real change will never come from lawmakers alone. Instead, he argues that the only path to meaningful AI safety is making unsafe AI bad for business—by injecting risk, liability, and uncertainty directly into boardrooms and C-suites. Peter reveals why AI doesn’t need to outsmart humanity to defeat regulation, it only needs money, time, and political cover. By exposing how industries evade oversight, delay enforcement, and co-opt regulators, this conversation re-frames AI safety around power, incentives, and accountability.

Together, they explore:

* Why laws, treaties, and regulations repeatedly fail against powerful industries

* How Big AI is following Big Tobacco’s exact regulatory playbook

* Why public outrage rarely translates into effective policy

* How companies neutralize enforcement without breaking the law

* Why third-party standards may matter more than legislation

* How local resistance, liability, and investor pressure can change behavior

* Why making unsafe AI bad for business is the only strategy with teeth

📺 Subscribe to The AI Risk Network for weekly conversations on how we can confront the AI extinction threat.