I think more people should say what they actually believe about AI dangers, loudly and often. Even if you work in AI policy.

I’ve been beating this drum for a few years now. I have a whole spiel about how your conversation-partner will react very differently if you share your concerns while feeling ashamed about them versus if you share your concerns as if they’re obvious and sensible, because humans are very good at picking up on your social cues. If you act as if it's shameful to believe AI will kill us all, people are more prone to treat you that way. If you act as if it's an obvious serious threat, they’re more likely to take it seriously too.

I have another whole spiel about how it's possible to speak on these issues with a voice of authority. Nobel laureates and lab heads and the most cited [...]

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Summary

To quickly transform the world, it's not enough for AI to become super smart (the "intelligence explosion").

AI will also have to turbocharge the physical world (the "industrial explosion"). Think robot factories building more and better robot factories, which build more and better robot factories, and so on.

The dynamics of the industrial explosion has gotten remarkably little attention.

This post lays out how the industrial explosion could play out, and how quickly it might happen.

We think the industrial explosion will unfold in three stages:

1. AI-directed human labour, where AI-directed human labourers drive productivity gains in physical capabilities.
   1. We argue this could increase physical output by 10X within a few years.
2. Fully autonomous robot factories, where AI-directed robots (and other physical actuators) replace human physical labour.
   1. We argue that, with current physical technology and full automation of cognitive labour, this physical infrastructure [...]

---

Outline:

(00:10) Summary

(01:43) Intro

(04:14) The industrial explosion will start after the intelligence explosion, and will proceed more slowly

(06:50) Three stages of industrial explosion

(07:38) AI-directed human labour

(09:20) Fully autonomous robot factories

(12:04) Nanotechnology

(13:06) How fast could an industrial explosion be?

(13:41) Initial speed

(16:21) Acceleration

(17:38) Maximum speed

(20:01) Appendices

(20:05) How fast could robot doubling times be initially?

(27:47) How fast could robot doubling times accelerate?

---

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.

1 Introduction

(Crossposted from my blog--formatting is better there).

Very large numbers of people seem to think that climate change is likely to end the world. Biden and Harris both called it an existential threat. AOC warned a few years ago that “the world is going to end in 12 years if we don’t address climate change.” Thunberg once approvingly cited a supposed “top climate scientist” making the claim that “climate change will wipe out humanity unless we stop using fossil fuels over the next five years.” Around half of Americans think that climate change will destroy the planet (though despite this, most don’t consider it a top political issue, which means that a sizeable portion of the population thinks climate change will destroy the planet, but doesn’t think that means we should prioritize it over taxes and healthcare).

I was taught by my science teacher in ninth [...]

---

Outline:

(00:13) 1 Introduction

(04:36) 2 How much warming will there be?

(07:33) 3 How past warming periods have gone

(11:25) 4 Agriculture

(14:42) 5 Ecosystem collapse and threats to agriculture

(20:57) 6 Heat stress and sea level rise

(22:51) 7 Tipping points

(28:25) 8 Economic costs

(31:06) 9 Migration

(33:42) 10 Conflict

(40:01) 11 Conclusion

---

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.

My post on Neutrality was kind of abstract, and I think it's time to do a follow-up that's more concrete.

Specifically, the place where I think there's something to be done is in information technology, and the ways it shapes the perceived “world”.

One way of looking at this: we had the institutional media era in the 20th century (newspapers and magazines, book publishers, broadcast television and radio), then the “Web 2.0” era of blogs and social media as more and more of the world joined the Internet, and now we may be at the beginning of a new information paradigm that's mediated by LLMs.

Some people are interested in making that new information paradigm “good” in some sense.

As I said earlier:

[most media & internet tech today] just isn’t prepared [...]

---

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.

The Dead Biologist

Suppose you're a biologist working with a bacterium. It's a nasty little bug, so you want to make sure your disinfecting protocol is effective. You have a rather good model of how the bacterium works, and it goes like this:

• The bacterium contains DNA, which is like a master copy of the instructions as to how it works
• This DNA is transcribed to RNA, which are little working copies of parts of the instructions and are disseminated throughout the bacterium
• This RNA is translated into proteins, which actually do the stuff that makes the bacterium function

Yes, there are a few special cases where RNA actually does something, and your model seems to be missing some things---for one, the amount of protein produced from each RNA molecule seems to vary, and you can only account for about 90% of this variance; the other 10% appears [...]

---

Outline:

(00:09) The Dead Biologist

(01:33) The Thing in the Noise

(03:29) Heavy Pressure

(04:30) Almost-All is not All

(05:42) The Most Forbidden Technique

(06:40) Beyond Imperfect Models

(07:41) Coda

The original text contained 3 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Abstract: We propose a new variant of online learning that we call "ambiguous online learning". In this setting, the learner is allowed to produce multiple predicted labels. Such an "ambiguous prediction" is considered correct when at least one of the labels is correct, and none of the labels are "predictably wrong". The definition of "predictably wrong" comes from a hypothesis class in which hypotheses are also multi-valued. Thus, a prediction is "predictably wrong" if it's not allowed by the (unknown) true hypothesis. In particular, this setting is natural in the context of multivalued dynamical systems, recommendation algorithms and lossless compression. It is also strongly related to so-called "apple tasting". We show that in this setting, there is a trichotomy of mistake bounds: up to logarithmic factors, any hypothesis class has an optimal mistake bound of either _Theta(1)_, _Theta(sqrt(N))_ or _N_.

This work is my first [...]

---

First published:

Source:

Linkpost URL:
https://arxiv.org/abs/2506.19810

---

Narrated by TYPE III AUDIO.

Throughout the first half of 2025 I did a blinded experiment to see how low to moderate melatonin intake affects me. In this post I summarize my findings.

Tl;dr: In my blinded self-experiment on melatonin, conducted over n = 60 days and analyzed with hypothesize.io, I found significantly positive effects for time to fall asleep (averaging ~25 instead of ~35 minutes, p ~= 0.001) and feeling awake the following morning (5.74/10 instead of 4.95/10, p = 0.006; but: this effect only persisted for the lower of two dosages, and did not persist throughout the rest of the day; could well be a false positive, as I didn’t correct for multiple hypothesis testing) for a dosage as low as 0.15mg of melatonin, taken on average ~1h before going to bed.

Feel free to jump ahead to the Results section if you don't care much about my methodology.

Experiment Setup

I [...]

---

Outline:

(01:14) Experiment Setup

(04:51) Hypotheses

(05:43) Results

(07:15) Limitations

(09:26) Detailed Results

(09:30) Sleep Quality

(10:41) Fall Asleep Duration

(11:46) Wakefulness after Waking Up

(12:55) Wakefulness at Noon

(13:36) Wakefulness at 4pm

(14:14) Wake-Up Time

(15:27) When Best to Take Melatonin

(18:29) Learnings for Future Self-Experiments

The original text contained 1 footnote which was 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.

This is more speculative and confusing than my typical posts and I also think the content of this post could be substantially improved with more effort. But it's been sitting around in my drafts for a long time and I sometimes want to reference the arguments in it, so I thought I would go ahead and post it.

I often speculate about how much progress you get in the first year after AIs fully automate AI R&D within an AI company (if people try to go as fast as possible). Natural ways of estimating this often involve computing algorithmic research speed-up relative to prior years where research was done by humans. This somewhat naturally gets you progress in units of effective compute — that is, as defined by Epoch researchers here, "the equivalent increase in scale that would be needed to match a given model performance absent innovation". [...]

---

Outline:

(04:09) The standard deviation model

(10:59) Differences between domains and diminishing returns

(13:02) An alternative approach based on extrapolating from earlier progress

(18:14) Takeaways

The original text contained 10 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Crosspost from my blog.

Regime change

Conjecture: when there is regime change, the default outcome is for a faction to take over—whichever faction is best prepared to seize power by force.

One example: The Iranian Revolution of 1978-1979. In the years leading up to the revolution, there was turmoil and broad hostility towards the Shah, across many sectors of the population. These hostilities ultimately combined in an escalation of protest, crack-down, more protest from more sectors (protests, worker strikes). Finally, the popular support for Khomeini as the flag-bearer of the broad-based revolution was enough to get the armed forces to defect, ending the Shah's rule.

From the Britannica article on the aftermath:

On April 1, following overwhelming support in a national referendum, Khomeini declared Iran an Islamic republic. Elements within the clergy promptly moved to exclude their former left-wing, nationalist, and intellectual allies from [...]

---

Outline:

(00:14) Regime change

(03:32) The LLM revolution

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

There was what everyone agrees was a high quality critique of the timelines component of AI 2027, by the LessWrong user and Substack writer Titotal.

It is great to have thoughtful critiques like this. The way you get actual thoughtful critiques like this, of course, is to post the wrong answer (at length) on the internet, and then respond by listening to the feedback and by making your model less wrong.

This is a high-effort, highly detailed, real engagement on this section, including giving the original authors opportunity to critique the critique, and warnings to beware errors, give time to respond, shares the code used to generate the graphs, engages in detail, does a bunch of math work, and so on. That is The Way.

So, Titotal: Thank you.

I note up front that at least Daniel Kokotajlo has indeed adjusted his estimates, and has [...]

---

Outline:

(02:00) The Headline Message Is Not Ideal

(09:19) An Explanation of Where Superexponentiality Is Coming From

(10:24) Three Methods

(10:53) Time Horizon Extension Method

(17:53) The Public Versus Internal Gap

(19:41) The Difficulty Gap

(21:38) Recent Progress

(23:36) Infinite Time Horizons

(24:53) Intermediate Speedups

(28:02) Is There A Flawed Graph Still Up?

(29:07) Some Skepticism About Projection

(29:58) Part 2: Benchmarks and Gaps and Beyond

(30:03) Benchmarks

(35:16) The Time Horizon Part of the Second Model

(37:44) Why The Thresholds?

(39:14) The Gap Model

(39:42) Eli Responds On LessWrong

(45:10) On Eli's Recent Update

(46:09) Conclusion

(50:38) Perhaps The Most Important Disagreement

The original text contained 1 footnote which was 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.