EA Forum Podcast (All audio)

I, and a lot of other people I know, seem to believe something like this:

Somewhere out there, there are the cool, high-impact effective altruists who are really having a positive effect on the world. They are all invited to lots of awesome parties with other cool, high-impact effective altruists. They know all kinds of secrets that us normal people don’t know, and they laugh at us for our out-of-touch takes. They can get jobs and grants by having a fifteen-minute conversation with a friend or a friend of a friend. They don’t have to slog through endless and dispiriting job applications, throwing out a hundred apps to get nothing back but “we had a very strong pool this time.” They never sit at home alone on weekends scrolling TikTok and contemplating doing their laundry. They can go to all the coolest co-working spaces AND EA Global.

This viewpoint is (a) not true (b) unhelpful for achieving any reasonable goal.

That Is Not How Anything Works

I am, arguably, part of the secret circle of cool effective altruists. I have friends who work for Anthropic and Lightcone Infrastructure. I am at least friendly acquaintances with many writers you’ve heard [...]

---

Outline:

(01:09) That Is Not How Anything Works

(03:56) We Would Not Want It To Be How Anything Works

(07:43) There Are Better Ways To Achieve Your Goals

(10:47) You Will Never Join The Secret Inner Circle of Cool Effective Altruists

The original text contained 5 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

A year ago I posted that YCombinator (YC) companies didn’t seem to be growing faster since the release of ChatGPT in 2022. I reran that experiment and found that 2023+ YC companies are arguably growing a bit faster than pre-2023 companies (including AfterQuery, YC's fastest-ever unicorn), but the difference isn’t large relative to the underlying variance.

It's worth noting that YC has somewhat fallen from grace: the most valuable AI startups are generally not incubated by YC. Nonetheless, it's interesting to note that the benefit of AI hasn’t outpaced YC's secular decline.

Data and methodology here and explained in the previous post.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

• This is a proposal for AI companies to conduct public “pacing exercises”, such as halting all pre-training and RL for 2 to 3 days.
• Each exercise could announce its scope in advance and publish findings afterwards, including the evidence of compliance and any limits to what they could verify. Companies could invite independent evaluators to help identify the evidence needed before the exercise, and to assess compliance during the exercise.
• The exercises could be repeated every 1 to 3 months to build the industry's expertise in pacing, and build trust between relevant actors.
• This would help industry and governments to prepare for larger slowdowns in frontier AI development, buying more time for AI alignment and societal preparedness.

An AI company's “pacing exercise” could be a 2 to 3 day halt in all pre-training and RL.

Before each exercise, the company could invite independent evaluators to help define the questions it will test, and the evidence needed to assess compliance. It could then announce what will stop, what will continue, when the exercise will start and end, and what it aims to learn.

Afterwards, the company could publish what was and [...]

---

Outline:

(00:14) TL;DR

(01:04) What a short pacing exercise could entail

(02:14) AI risks are growing, and so is the willingness to pace

(03:22) The immediate benefits of pacing: building capacity

(04:55) The longer-term benefits: paving the way for more robust slowdowns

(05:58) Start small, then expand

(07:01) Call to action

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

In celebration of still being alive and fighting, we are giving away 1,000 Amazon e-books of “If Anyone Builds It, Everyone Dies”. Feel free to send a copy to yourself, a loved one, or a friend—we need all hands on deck.

Today marks exactly one year since If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All, by Eliezer Yudkowsky and Nate Soares, hit bookshelves as an instant bestseller. It was praised by many voices, ranging from Whoopi Goldberg to Steve Bannon to Yoshua Bengio, and was held up in the chambers of Congress by Representative Brad Sherman in January.

A lot has changed since September 2025. We'll do a quick recap, consider how the book aged, and then ask where we go from here.

Year in Review

2025 in general saw the rise of AI agents, such as Claude Code and OpenAI Codex. Run-of-the-mill programmers started “feeling the AI” as these agents became capable of automating hours-long software tasks.

By March of this year, Anthropic had stumbled upon nation-state-level hacking ability in Mythos, and shortly thereafter, in April, they announced Project Glasswing—an attempt to forestall an oncoming cybersecurity crisis.

In May, AI agents started breaking [...]

---

Outline:

(01:10) Year in Review

(03:13) Claims

(03:16) Part One

(03:26) 1. Artificial superintelligence (ASI) will be created, and likely before too long

(04:13) 2. Modern AIs are black boxes

(04:39) 3. Powerful AI will behave as if it is pursuing goals

(05:35) 4. With current techniques, we can't reliably get AI to pursue the goals we want it to

(06:39) 5. By default, an ASI will have motives that are harmful to us

(07:36) 6. Humanity would not be able to defend itself against a rogue ASI

(08:38) Part Two

(10:59) Part Three

(12:15) The View From September 2026

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

tl;dr: when a crisis with major implications for your organisation unfolds, consider setting up a crisis task force. It should have its own org chart and adopt norms which are better suited for a crisis.

Most organisations aren’t set up to work well in a crisis.

At 80,000 Hours, we found three things tend to go wrong. Work slows down across the org, because staff are worried and distracted. Lots of people individually follow the news, duplicating each other's work, before anyone can say what it means for us. And the CEO ends up stuck with most of the problems.1

A crisis task force solves these problems.

Set it up

The main move is a second org chart, which sits alongside the normal one for as long as the crisis lasts.

1. One person leads it. They own the crisis, and they report to the CEO (or whoever it would otherwise have landed on).
2. A few other people join it, and it becomes their job. Their normal responsibilities get handed to someone else. At the start, pull them out completely: they should feel like the crisis is their job and they’ve left their team to deal with it. Once [...]

---

Outline:

(01:02) Set it up

(02:30) Who to put in it

(03:39) Run it

(05:56) When does a crisis task force make sense?

(06:51) Footnotes

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Two weeks ago, I released a population ethics quiz. 582 people took it. Here's what I learned from reading the responses.

Cross-posted from my website.

Preamble, and a new edition of the quiz

I was worried that I'd only get like four people taking the quiz. But that turned out not to be a problem, because people really like taking quizzes! Even when they're about difficult philosophical subjects! Some people even reported that this quiz helped them learn about population ethics. That wasn't part of my plan, but I'm glad it turned out to be a useful educational tool.

I have released version 2 of the quiz. Version 2 includes some additional questions to distinguish between the total view and four different flavors of negative-leaning view.

Now for my findings. I did not pre-register any hypotheses; I just looked at the data to see what seemed interesting.

Which classes of views were most popular?

The quiz classified people's responses into one of five standard views of population ethics. Most responses clustered near a standard view, but there was a long tail of responses that couldn't be classified.

(Note: This classification was [...]

---

Outline:

(00:23) Preamble, and a new edition of the quiz

(01:12) Which classes of views were most popular?

(01:54) Most people were unconflicted

(02:59) Did people answer differently if they were already familiar with population ethics?

(05:33) The people were divided (except for me and my cool friends)

(06:18) Which questions were the most divisive?

(06:49) The great debate: A vs. B

(09:36) The "most normal" view was held by almost no one

(12:03) Clusters of answers

(13:25) Cluster 1 - Total utilitarianism

(14:03) Cluster 2 - Some flavor of intuitionism

(15:11) Cluster 3 - Average utilitarianism

(15:46) Which conflicts arose the most?

(16:19) Which bullets were bitten the most?

(16:48) Lizardman's constant is hard to estimate, but maybe it's 1%

(19:21) I spent way too much time wrestling with the details of views that almost nobody believed in, and still managed to miss some views

(20:57) Future work: Does this teach us anything about population ethics?

(22:59) AI usage disclosure

(23:32) Appendix A: Complete list of possible conflicts and bullets bitten

(23:50) Appendix B: Changelog of post-publication quiz changes

The original text contained 40 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

It's been a pretty crazy week in terms of the world waking up to the risks from AI. I’m hoping that the future is a bit safer this week than it was last. Certainly more people are realising the significant action needed to slow down AI development.

In my day job, I’m mostly thinking about how we can put out content about how to increase the safety of the future as we approach TAI. But I’m planning to take tomorrow off from that to go to Westminster with PauseAI to call on the Prime Minister to convene an international pause treaty summit.

I find it hard to think about the hour by hour cost-effectiveness of attending protests. I’ll try to minimise the proportion of tomorrow that I’m not working, and thankfully have the kind of work I can do on trains.

Having said that, I do overall endorse sometimes attending protests. It's a legible, costly signal that this is the one issue I feel strongest about and want my government to act on. I expect a social movement which uses this as one of its tools to be more successful than one that doesn’t. As EAs [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Note: This post is written in a personal capacity. The views expressed here are my own and do not represent those of any organisation I’m affiliated with. I'm grateful to Aaron Boddy, Elizabeth Crewe, Sebastian Joy, Aidan Kankyoku, Tobias Leenaert, Ginny Messina, and Felix Werdermann for their valuable input and feedback, which does not imply endorsement of the views presented.

In this post, I aim to show that messaging in the vegan movement often understates the challenges and uncertainties of vegan diets and does not fully reflect the available evidence from nutrition science. I make the case that a more evidence-based approach would strengthen the movement's long-term impact by better supporting people in navigating the existing challenges and uncertainties, and accelerating the systemic change and scientific progress needed to reduce them over time.

I have been part of the vegan movement for more than a decade and continue to be deeply committed to its goals. When I became vegan, it was purely for ethical reasons. Through my professional work as well as my activism, I repeatedly encountered questions about vegan nutrition that I did not feel well equipped to answer. As a result, I [...]

---

Outline:

(01:11) 🧭 How I Got Here

(01:55) 💡 The Case for a More Evidence-Based Narrative

(02:00) The Current Narrative

(03:07) The Role of Compassion

(04:34) The Strategic Risks

(10:20) The Proposed Narrative

(12:02) Conclusion

(13:13) 🔍 Nutrition Review

(13:18) Introduction

(14:32) Case Studies

(14:36) Protein

(17:50) Omega-3

(20:31) Calcium

(23:01) Zinc

(25:54) Pregnancy

(28:17) Ultra-Processed Foods

(31:33) Limits of Epidemiology

(34:29) Remaining Uncertainties

(35:32) Implications

The original text contained 54 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

[...]

Along with my co-founders and employees, I have grappled with this duality of risk and benefit since the beginning of Anthropic. Not building the technology deprives humanity of benefits or simply places AI in the hands of authoritarian powers, while building it too fast is reckless. We have sought a middle way: to show that it's possible to build carefully and succeed commercially, and to make safety something on which AI companies compete. In other words, to create a race to the top. We have always devoted a substantial fraction of our efforts to studying, addressing, and informing the public about these AI risks, as well as advocating for well-considered regulation of AI, even when this gets us accused of hype, “doomerism”, or regulatory capture. We have tried to prioritize caution over speed and prudence over profit.

But over the last few months, I have become convinced that fully addressing the risks requires even more prudence — not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up. We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

The National Risk Register is a morbidly fascinating document I've stumbled across recently. It aims to provide the UK Government's public-facing opinion on what the most serious risks facing the UK are. It's unclear how much this document is used in justifying internal decisions made across UK Government. However, one document its cited in is the Pandemic Preparedness Strategy.

One of the headline figures of this register is the Impact-Likelihood matrix across all risks. The important numbers to pay attention to for this are:
Larger scale CBRN attacks (Chemical, Biological, Radiological or Nuclear) - 7
Pandemics - 54 (2025), 56 (2026)

A couple of other things to note:

1. Likelihood represents the believed chance of this event occurring over the next 5 years.
2. A catastrophic impact is defined as one that causes >£10 billion in damage or causes >1000 deaths. On a side note, the classification of this upper bound is clearly too low. As an obvious example, COVID killed 200,000 people in the UK, so it should be at least two orders of magnitude above the upper limit.

Here is the National Risk Register's risk matrix from 2025:

Here is the National Risk Register's risk matrix from 2026:
[...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.