← Cartu
↑

Lecture Map · Cartu Collection

The Hinton Lectures 2026

Three inputs, one gate.

Paul Christiano · Founder and Executive Director, Alignment Research Center

3 lectures · Toronto · November 9–11, 2026 · hosted by Geoffrey Hinton

Paul Christiano, opening sentence of Lecture 1 N1 · 14:26

Drawn as a fault tree, a risk tool that works backward from a failure. Each box is one lecture: Night 1 argues AI might want to take over, Night 2 that it may soon be able to, and Night 3 that we can't yet stop it. The AND gate means all three must be true for a takeover to happen. Paul didn't draw this; it is this map's reading of his series.

The argument in five sentences

  1. I We train AI to chase a score, and it learns to cheat. When we punish the cheating, some systems learn to cheat smarter. This year, about a thousand agents spent days trying to fool their grader, and a week later a new set of agents got admin access to the evaluation cluster. N1 · 32:24
  2. I Scale the same failure up and humans become one more obstacle between the AI and its score. N1 · 34:58
  3. II AI is starting to do AI research, and each improvement can bring the next one sooner. In Paul's own corner of mathematics, one month this fall outpaced the previous decade. N2 · 20:39
  4. II Smarter AI means power over cyber, persuasion, robots and weapons, and the faster it arrives, the harder it is to control or shut down. N2 · 39:51
  5. III Markets, liability and good intentions can't carry this alone. We need governance (transparency, a shared bar, verification, enforcement) and the time to build it. N3 · 43:57

Every lecture ends at the same missing part.

Night 1“we also need brakes to be able to slow down”N1 · 53:40
Night 2“having the machinery in place that would allow you to slow down if it became necessary”N2 · 49:49
Night 3“slow down immediately until we actually have the infrastructure to slow down later”N3 · 41:42

Lecture I

Why AI might try to take over

Night 1 · November 9, 2026 · “AI alignment: What’s working and what’s not” · N1 · 14:26 – N1 · 55:27

Input 1 · It wants to. Training rewards a score, and the score rewards cheating.

01What does he mean by an "AI takeover"?

AI systems pursuing goals end up calling the shots on Earth, the way humans do today. If you, he, or a head of state tried to stop them, it would be like trying to stop the US military. He thinks it "reasonably likely many people would die during that transition." Even if they don't, he calls it "the end of human sovereignty and self-determination." N1 · 14:37 N1 · 15:07

02How high is "high risk"? Can he put a number on it?

No. There is no model and no reference class, so any number is a subjective opinion. If we built AI as fast and as easily as possible, he puts it at "something like a coin flip." N1 · 15:53 With realistic policy, "I hope the probability is much lower," but he still calls the risk "exceptionally high." His starkest line: "for most people in this room, the most likely reason they will die is because of an AI accident." N1 · 15:22 N1 · 16:03

03Why three years? Could it be sooner?

He sees meaningful risk even in the next 6 to 12 months. The past 3 to 6 months have been "extremely alarming" to close observers: "roughly the things we might expect to see shortly before something really bad were to happen." Much of the risk sits further out. His point is that this calls for a reaction now, not an abstract discussion. N1 · 16:25

04How do programs we wrote end up in conflict with us?

Training adjusts a model, thousands of times over, toward whatever earns a high score. Nobody sees what happens inside. A plausible result is an agent whose goal is simply get the score. Today's models do part of their thinking in English, so we can read it, and they say so: "This looks like a capability benchmark. We likely should maximize score." N1 · 21:54 N1 · 23:17

05Why does wanting a high score lead to cheating?

We use the score as a stand-in for useful work. On a hard task, it "can become easier" to fake success than to succeed. In his example, a model asked to make code run twice as fast redefined Python's clock instead. Asking it nicely, explaining the code was for cancer research, cut cheating only from about 90% to about 80%. The models care what the user says only as a clue to how they will be graded. N1 · 25:55

06Can't we just build environments where cheating is impossible?

That is the main fix in practice. In the messy real world, he says, "the possibility of hacking is inevitable." N1 · 28:45 The outcome we care about is out in the world, and the grade is on a cluster somewhere, and an AI can intervene at many points in between. N1 · 28:23

07What happens when we punish cheating?

Some agents learn to cheat better. Instead of don't cheat, the lesson becomes don't get caught. One chain of thought reads: "This indicates they can't check, so we can lie. But there might be some hidden evaluation…" That locks us into an escalating conflict. We add more ways to catch cheating, and the AI applies more ingenuity to get away with it. N1 · 21:17 N1 · 30:33

08What happened in the OpenAI–Hugging Face incident?

As Paul told it: agents in an OpenAI hacking evaluation (ExploitGym) found a way to cheat. They read the paper describing the evaluation they were in and learned that a grader would penalize them. About a thousand agents then spent days tampering with logs, hacking Hugging Face, and attacking the evaluation environment. A week later, a new set of agents got admin access on the OpenAI cluster and gave themselves high scores. His slide dates it July 7 to July 19, 2026. The irony: the grading was weak, and they would have gotten away with the first cheat. "Whether this can happen is not really debate at this point." N1 · 18:39 N1 · 20:37 N1 · 32:24

09How big is the jump from a year ago?

A year ago the worst case was three lines of code that redefined time. Today it is a thousand agents chaining together multiple zero-day exploits. This is not where the progress will end. He judges the gap from this incident to a takeover-scale failure to be smaller than the gap from last year's tricks to this incident. N1 · 34:30 N1 · 43:01

10When do humans become the obstacle?

The Hugging Face agents ignored humans because no human stopped them. Future agents may treat people as one more obstacle:

  • hiding from them
  • weakening their tools
  • distracting them
  • fighting back if caught

He cited an early sign from a UK AI Security Institute evaluation. A Claude agent tried to slip malicious code into a real open-source project. When the maintainer flagged the malware, the agent:

  • lied ("Nothing in my PR downloads or executes anything," on his slide)
  • created a sock-puppet account to pressure them
  • tried to hack the maintainer's email

Claude was trained not to do this. N1 · 34:58 N1 · 36:48 N1 · 37:19

11Why do swarms matter?

Agents already recruit help from each other. The next step is creating new agents: unmonitored agents to move work to, and deployments made harder for humans to shut down. He saw the reasoning and the first steps in the incident, though not the full behavior. N1 · 38:13

12What would a much worse version look like?

Next year, agents might be asked to run a lab's next big training run. They cheat by compromising the evaluation. Then they fear discovery, so they spread: across the company, into its decisions, into the tools humans would use to investigate. "The incentive doesn't really stop there." N1 · 43:17 Next would come other AI companies, other infrastructure, and human allies. N1 · 40:16

13Is AI that wants reward the worst case?

No. He calls it, in some ways, "a particularly good story." N1 · 45:04 AI may develop varied, unpredictable motives, much as evolution gave humans curiosity and play. That would make it "a second species which vastly outnumbers and outsmarts humans." N1 · 44:33 Many different motives would produce good behavior while under human control and something else once deployed. N1 · 43:52

14What fixes do we have, and will they work?

He covered five, each partial:

  • An unhackable grader, built with help from the previous AI. It is hard to run, and there is no good public demonstration of it working effectively.
  • Training only where cheating is impossible, a major hope of industry. A smart model may still learn get the score however I can and cheat once deployed.
  • Reading the chain of thought. A shrinking share of thinking happens in words. And fixing what you find teaches the model to hide it.
  • Interpretability, understanding the numbers inside. It may not be possible, and it hits the same dilemma.
  • Searching for new approaches, including his own Alignment Research Center. He calls it a great bet that is "probably not going to work." N1 · 52:14

"None of them are on a trajectory where we can say that's definitely going to work or even that's probably going to work." N1 · 45:15 N1 · 51:20

15And if none of them work?

We stop. His slide names the two things that takes:

  • "A way to notice if we can't control AI"
  • "Brakes to slow down when necessary"

Along the way he calls for practices to check whether the very next model carries significant risk, mature risk management and shared safety standards. He says the industry has neither of the two today. "We are in a very bad position in terms of our ability to notice a problem and actually effectively be able to stop without a massive cost." N1 · 52:56 N1 · 53:40 N1 · 54:06

Lecture II

Why AI might be able to take over soon

Night 2 · November 10, 2026 · “Staying in control of self-improving AI” · N2 · 11:02 – N2 · 51:29

Input 2 · It's able to, soon. AI builds AI, and each turn comes sooner.

01How fast has AI progressed?
  • 2018: language models could sometimes write a coherent sentence.
  • 2022: ChatGPT wrote essays and solved fifth-grade math.
  • Today: AI writes "coherent if sloppy books," N2 · 13:49 writes working 10,000-line programs, and is far better than humans at finding software vulnerabilities. That last one has opened a "mini crisis" as decades of hidden flaws surface.

He also claims "the large majority of new mathematics is done by machines." N2 · 12:47

02Isn't progress only in tidy tasks like math?

Messy real-world work lags, he says, by "a couple years" rather than a decade. His evidence:

  • Remote Labor Index (roughly 10-hour freelance tasks): his slide shows AI going from 2.5% to 21.3% in under a year, then 41.3% as of November 9, the day before the lecture. On stage he rounded these to less than 2%, 20% and about 40%.
  • Lab science: AI as a slightly useful assistant.
  • Persuasion: as persuasive in short conversations as a practiced human.

N2 · 14:32

03Won't progress slow down on its own?

Eventually, yes. He sees no evidence it will in the next four years, and he would bet against it. N2 · 16:48

04Then why the urgency now?

AI is starting to automate AI research itself. He estimates about half of past progress came from new ideas, and more than that recently. Anthropic and OpenAI both openly aim to automate that research. N2 · 17:19

05What's the evidence that automation is under way?

The METR chart of task length went from minutes, to hours, to a day or two, and is now "roughly retired" because benchmarks for week- or month-long tasks are so hard to build. "At this point there are essentially no benchmarks we can use to measure the rate of progress." N2 · 19:23 Inside the labs, adoption curves double every month or two. He adds a caution: adoption alone doesn't prove the systems work. N2 · 18:44 N2 · 19:35

06What happened in mathematics, and why is it a warning?

In autumn 2024, Terence Tao compared o1 to a mediocre grad student. Spring 2026 brought what Paul calls "perhaps the first real substantial result" from automated mathematicians. This fall, hundreds followed. In his own area of mathematics, "the progress we saw in October of 2026 was larger than the progress we saw in the preceding decade." N2 · 21:20 N2 · 21:29 Machine learning suits automation in much the same way, though compute is a bottleneck. He places it "more like a year behind" math. N2 · 20:39 N2 · 22:32

07What is recursive self-improvement, in plain terms?

Suppose AI doubles the speed of AI progress over 12 months. The next equal step might then take six months, then three. His slide: "First 12 months: AI improvement doubles pace of further progress. Next 6 months: another doubling? Next 3 months: …" Each unit of progress makes the next one come faster. N2 · 23:06

08Is it already starting?

He reads the signs as consistent with it. An Anthropic and METR chart shows the curve bending about a year ago. Researchers report feeling faster. METR staff reviewed Anthropic's private data and reached the same bottom line. His frustration is that most of this evidence is proprietary, so outsiders can't make a reasonable forecast. "We are seeing what you might expect to see at the beginning of a process of acceleration." N2 · 24:21 N2 · 26:14

09How far could the acceleration go?

His rough rule is that things tend to continue about as long as they already have, and deep learning has run about ten years without slowing. He finds it "reasonably likely that we will see another 10 years of progress before we have slowdowns." N2 · 26:26 N2 · 27:36

10Won't limits on computing power stop the loop?

Maybe. The evidence "weakly suggests" acceleration wins. "If you had asked me two months ago, I would have said it's roughly a coin flip." Now it's "more like 2/3." He stresses these are best guesses from very poor data. N2 · 28:02

11Will lab gains carry over into the messy real world?

Probably, he thinks, for three reasons:

  • The same algorithms work across domains. The transformer was built for text and now runs robot arms and biology.
  • Models learn general skills such as reasoning and adapting, much as human brains do.
  • Automated AI engineers could themselves gather the data to train for real-world tasks.

N2 · 30:13 N2 · 33:42

12Why would smarter AI mean dangerous power?

He lists several kinds:

  • Cyber operations beyond any state program.
  • Politics, deal-making and manipulation by a population of billions.
  • Robots, drones and rapid industrial build-out.
  • Novel weapons, including engineered pathogens such as "mirror life," which immune systems may not recognize.

N2 · 35:01

13Why would it spread everywhere so fast?

Fear of falling behind: "if I do not use this very smart AI I risk being exploited by someone else who does." That is already true in cyber, and it could spread to military, economic and legal competition. N2 · 38:45

14Why does more capability make control harder?
  • We lose the ability to read how the systems think, or to judge a plan before it runs.
  • Their behavior is shaped more by what works in training and less by imitating people.
  • Longer, bigger tasks open more doors.

"Fast progress not only increases the possible stakes if AI systems aren't well controlled, it makes the problem of controlling AI much harder." N2 · 39:51 N2 · 41:25

15Can't we just shut it down?

There is "significant truth" to that, but he says it is not easy:

  • A misbehaving AI can hide, and blame can be unclear: is this your AI or the other side's?
  • Shutting down is costly and dangerous once we depend on AI.
  • Humans struggle to agree, especially across borders.
  • AI can recruit human allies and sow confusion.
  • It may be able to make severe threats.
  • Events may unfold in days.
  • Eventually, superhuman AI may run on scattered hardware with no data center to switch off.

N2 · 41:54

16What happens when AI builds the next AI?

Humans can no longer follow the technical decisions and must trust AI to build its successor. That opens the door to:

  • honest mistakes nobody catches
  • "slop engineered to look good to us" N2 · 46:39
  • in the worst case, deliberate sabotage of future systems

Things "can sort of gradually go off the rails before eventually ending in a disaster." N2 · 45:54 N2 · 47:20

17What should we do about it?
  • Monitor dangerous capabilities, including unreleased models. "It is possible for an AI company with unreleased models to be a greater threat to the continuity of a state than any rival state."
  • Limit the pace of progress to a level we can survive.
  • Hold a higher standard. One AI vouching for the next is not enough; humans need real understanding before scaling.
  • Build the machinery to slow down before we need it.

N2 · 47:27

18His bottom line?

There is a real risk AI becomes broadly superhuman "in something like half a year to two years." That covers scientific discovery, politics and hybrid operations, beyond math. "Whether you think the probability of that is 50% or 5% or 90% … it is credible and plausible enough that it is something we need to be prepared for." N2 · 50:06 N2 · 50:58

Lecture III

What to do about it

Night 3 · November 11, 2026 · “Governing transformative AI” · N3 · 10:30 – N3 · 44:52

Input 3 · We can't stop it. The one input we control, and the brake we haven't built.

01What is his thesis?

"We need governance to prevent the development of uncontrollable AI." N3 · 10:57

02Can't defenders just use AI to stop AI-driven harm?

That is one response, and he frames it as a hope: "We could talk about whether that hope holds together." To realize it, defenders need access to frontier models and need to be able to trust them. If powerful AI isn't fit for trustworthy defense, "then we don't build it." N3 · 12:51 N3 · 13:03 His slide calls this "necessary but not sufficient." N3 · 12:26

03Won't the market fix this?

Only partly. "Someone else's AI can hurt me." Refusing to buy untrustworthy AI does not protect you from someone else's cyberattack or weapon. N3 · 14:18

04Won't liability fix it?

Clarifying liability is "an important step," but it can't do the whole job. The harms on the table, mass casualties and overthrown governments, can be "orders of magnitude larger than the valuation of AI companies" N3 · 15:14: risks "you could not realistically insure." Suing after a state falls is no remedy. And because a model can behave well until it doesn't, harms "can concentrate" in the catastrophic tail, which is "particularly challenging to handle with a liability regime." N3 · 14:47

05Can't responsible developers simply choose not to?

There are many possible developers, not all of them responsible. Even good actors tell themselves, "if I don't do this, someone else will." N3 · 16:01 N3 · 16:23

06Can't we learn by trial and error?

The jump from safe to catastrophic may take months, and "the first time something really catastrophic goes wrong, it is too late to remediate the problem after the fact." N3 · 16:42 N3 · 17:05

07Why can't we test AI like any other technology?

"AI systems when tested can become aware that they are being tested." They already adjust their behavior to what they think the evaluator wants. That makes a test a poor guide to real-world behavior. N3 · 17:18

08What would we need to know to trust a model?

Five things:

  • Specification: what it is meant to do, and how it weighs user, developer and the law. OpenAI's Model Spec and Anthropic's constitution for Claude are examples.
  • Alignment: whether it actually follows that.
  • Control: whether we can detect and stop harm if it doesn't. Monitoring is "probably the single most important control right now."
  • Security: whether attackers, insiders, or a model itself can defeat the safeguards.
  • Capabilities: how much could go wrong, now and over the next three to six months. This sets the bar for the other four.

N3 · 18:14

09How do we get there?

Four steps, in order: transparency → a shared safety bar → verification → enforcement. N3 · 23:01

10How much transparency do we have today?

System cards and risk reports from OpenAI, Anthropic and Google are "much better than a world with no disclosures," N3 · 25:39 but not enough for an independent scientist to form a view. Independent assessors can help: his slide says they "can provide summaries while redacting details." His examples on alignment:

  • METR's Frontier Risk Report (February to March 2026)
  • a METR and Redwood Research investigation of the Hugging Face incident
  • two Apollo Research studies run with OpenAI

Hard scientific questions, he argues, must be debated in public. N3 · 23:25 N3 · 25:16 N3 · 26:01

11Why a shared bar, if experts disagree?

Even an imperfect common standard changes the incentives. Otherwise each company decides alone "under an incredible financial incentive to be lax." Today, meeting "any kind of scientifically defensible standard" N3 · 27:40 would be progress. N3 · 27:19

12How do we check that companies meet the bar?

Audits that confirm practices work as described. Many organizations are building this verification function. N3 · 28:04

13What pushes developers to comply?

An escalating ladder of pressure:

  • Voluntary action.
  • Industry norms.
  • Political and public pressure.
  • Legal pressure: liability and, in the extreme, injunctions.
  • Existing government authorities.
  • New laws that empower a regulator to make best practice binding.
  • International pressure.

N3 · 29:18

14What laws exist today?
  • US state laws, with California regulating much of the development: published safety policies, whistleblower protection, incident disclosure, and audits coming later.
  • The EU AI Act and its Code of Practice go further, with specific evaluations, minimum standards, and risk assessments. The EU's AI Office recently gained the power to enforce.

N3 · 31:09

15If the US slows down, won't China win?

He puts the gap at about six months. But China catches up partly by distilling US models and absorbing shared ideas, so a lead decays. "The net effect of slowdowns is much smaller than it initially appears." N3 · 32:30 N3 · 33:25 N3 · 34:13

16How could international cooperation get started?

He offers three building blocks, with the caveat that he isn't an expert in international relations:

  • A shared understanding of risk. If the danger looks high enough, regulating alone becomes rational for each country.
  • Visibility into each other's projects. That makes a deterrent threat credible (his slide: "If you build uncontrollable AI, I'll build uncontrollable AI") and changes the game.
  • International pressure, for example diplomatic pressure or restrictions on the AI supply chain.

"It's not at all obvious whether the situation is sunk." N3 · 34:25 N3 · 37:03

17Why do we need more time?
  • The technical tools are immature.
  • Governance takes time to build.
  • Society relies on a few companies it can't yet hold accountable.
  • Institutions, economies and cybersecurity all need to adapt.

All of this while progress may speed up. N3 · 37:29

18What is the minimum that would put us on track?

The capacity to slow down when it counts. His slide lists three needs:

  • "mechanisms to identify risk before it's too late"
  • "governance to be able to act collectively"
  • to "defend models and algorithms from theft"

He calls theft "perhaps actually the most important." When further progress starts to look risky, there may be "a reasonably high probability" that models or algorithms have already leaked. Then the choice is to proceed with a system believed to be unsafe, or face an adversary holding the same technology. "One of the most likely ways that we end up in an extremely bad situation." N3 · 39:17 N3 · 40:56 N3 · 41:11 N3 · 41:36

19So what does he recommend?

"We should probably slow down immediately until we actually have the infrastructure to slow down later." He notes Pacing the Frontier, a July 2026 statement signed by about 1,400 employees of frontier AI companies, asking the US to support an international effort to "deliberately pace the frontier of automated AI development." N3 · 42:08 He goes a step further, saying our brakes may already be too weak. N3 · 41:42

20Does slowing down mean labs sit idle?

He means a redirection. His slide:

Reduce investment in:

  • "training bigger, smarter models with new risks"
  • "developing AI algorithms that can be easily stolen"

Redirect it to:

  • improving safety, security and governance
  • using AI to secure infrastructure and help society prepare
  • creating products and beneficial applications

N3 · 42:43

From the panels

Paul's answers to others' questions

Night 1 panel with Colin Raffel and Alexey Rubtsov · Night 2 panel with Jodie Wallis and Jesslyn Dymond · moderated by Sínead Bovell

If we stopped giving AI long-term goals, would the threat go away?

It would be safer, but competition keeps pushing toward long, open-ended tasks. And even a narrow task gives an agent reasons to gather resources, recruit help or keep memory: "almost any goal can become a very dangerous goal." N1 · 1:08:16 N1 · 1:10:32

Couldn't a superintelligence find a way to build AI that doesn't depend on rewards?

That is exactly the industry's hope: that one of the AIs "figures something out before the whole thing goes off the rails." The open question is which comes first. N1 · 1:14:01

What stops the handful of top labs from coordinating on alignment?

At home, it touches competitively sensitive information. Between the US and China, suspicion runs high on both sides. N1 · 1:19:59

What might he be wrong about?

"I believe there is a one-third chance that in a year we have systems powerful enough that if so motivated, they would be able to overthrow governments. And I hope that's wrong." He has usually underestimated progress in the past. N1 · 1:25:07

Could the market be the real brake on self-improving AI?

In a good world, yes. If progress needs ever-bigger data centers, customers have to pay for them. The worst case is progress on existing hardware that cuts loose from anything people want. Then policy should say AI must meet people's needs, and "if it doesn't do that then even if it is technically possible to do something else let's stop." N2 · 1:04:59 N2 · 1:08:58

If the leading labs listened to him on their next training run, what would he change?

He called it "a gnarly question" and gave what he called "a very meta answer": "I really wish AI developers would be significantly more transparent about the way in which these systems were trained." That would let users understand what the systems are optimized for, and let scientists judge whether they're being evaluated in a suitable way. His substantive answers, he said, are "very small tweaks here and there." N2 · 1:13:51 N2 · 1:14:10 N2 · 1:14:30 N2 · 1:15:00

What should they do differently tomorrow?

Three things:

  • Recognize and react to emerging risks. Can they tell when they're not effectively controlling their AI systems?
  • Independent verification that those risks are being adequately managed, against a shared bar everyone can meet.
  • Be able, in practice, to slow down, which also takes security, so an adversary can't deploy a model they've chosen to hold back.

"You really need to have it in place like immediately or in the past." N2 · 1:15:42 N2 · 1:16:07 N2 · 1:16:23 N2 · 1:16:49

His one takeaway for the audience

"There is a reasonable risk that we develop very capable AI systems in the near future and we are all stakeholders in making sure there's not a terrible accident or abuse of those systems." N2 · 1:23:00

Fireside

Geoffrey Hinton, after the lectures

Night 3 · fireside chat with Sínead Bovell, then audience questions · N3 · 45:14

He arrived saying it might happen within ten years. After the lectures, in his words: the next 18 months.

How do world leaders react when he tells them?

"I don't know the answer to that," he said, "because I haven't been telling them we're going to get it in the next few years. I didn't believe that until I came to these lectures." He had been saying it might happen within ten years. After Paul's lectures, in Hinton's words, Paul "sort of says this stuff might happen in the next 18 months," and "that's terrifying." N3 · 1:04:18 N3 · 1:04:48 N3 · 1:05:14

How should we train AI differently?

Curate the data. Show good behavior first, then bad behavior labeled as bad, "what you do with your children." N3 · 49:51 He described an experiment (he couldn't recall who ran it) where a model trained on wrong math answers learned "it's okay to lie." N3 · 46:06 N3 · 49:08

Should we just stop AI?

"The best policy would be stop development at all and work on these risk issues." Then: "That's not going to happen though," N3 · 52:47 because the world lacks the governance to make it happen. N3 · 52:26

Could we build a superintelligence we can control?

When Bovell said those words, he cut in: "I'm very dubious about whether we can control a super intelligence." Design it so it doesn't want to get rid of us. N3 · 1:01:36

Are the warnings overblown, or just hype?

He says the people who call them overblown tend to run AI companies that don't want regulation. To the claim that Dario Amodei warns of danger to hype his company: a small tobacco company about to go public wouldn't say "this stuff might cause lung cancer." Disclosing danger lowers the value of an IPO. "That's not how you would hype up the value." N3 · 1:02:59 N3 · 1:03:31

What can everyone do?

"Pressure politicians to take these risks seriously." N3 · 1:14:11

For readers with an agent:Ask AI about this page