We have the classic phenomenon where suddenly everyone decided it is good for your social status to say we are in an ‘AI bubble.’

Are these people short the market? Do not be silly. The conventional wisdom response to that question these days is that, as was said in 2007, ‘if the music is playing you have to keep dancing.’

So even with lots of people newly thinking there is a bubble the market has not moved down, other than (modestly) on actual news items, usually related to another potential round of tariffs, or that one time we had a false alarm during the DeepSeek Moment.

So, what's the case we’re in a bubble? What's the case we’re not?

My Answer In Brief

People get confused about bubbles, often applying that label any time prices fall. So you have to be clear on what [...]

---

Outline:

(01:17) My Answer In Brief

(02:18) Time Sensitive Point of Order: Alex Bores Launches Campaign For Congress, If You Care About AI Existential Risk Consider Donating

(04:09) So They're Saying There's a Bubble

(05:04) AI Is Atlas And People Worry It Might Shrug

(05:35) Can A Bubble Be Common Knowledge?

(08:33) Steamrollers, Picks and Shovels

(09:27) What Can Go Up Must Sometimes Go Down

(11:36) What Can Go Up Quite A Lot Can Go Even More Down

(13:17) Step Two Remains Important

(15:00) Oops We Might Do It Again

(15:49) Derek Thompson Breaks Down The Arguments

(17:47) AI Revenues Are Probably Going To Go Up A Lot

(20:19) True Costs That Matter Are Absolute Not Relative

(21:05) We Are Spending a Lot But Also Not a Lot

(22:46) Valuations Are High But Not Super High

(23:59) Official GPU Depreciation Schedules Seem Pretty Reasonable To Me

(29:14) The Bubble Case Seems Weak

(30:53) What It Would Mean If Prices Did Go Down

---

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.

Ontological status: Yes, this is ontology

Groups of people are one of the most important things. If I were to list all the things and rank them by importance, groups of people would be near the top. Love, truth and freedom and other such things might score higher from some angles, but these things are usually found in groups of humans anyway.

And groups are complex. That's why I started using this categorisation of groups recently. I have found it very helpful in conversations about community and events over the past year.[1]

For any set of things, there are usually many ways of categorising it. Rather than size, or the specifics of what folk do together (though they are important), I like thinking of groups in terms of the flavour of their purpose!

Here's the model:

What is a scene?

The purpose of a scene is [...]

---

Outline:

(01:12) What is a scene?

(03:13) What is a clique?

(04:57) What is a team?

(06:02) Some mistakes and hard-earned takes

(11:15) Some final points

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 a cross-post (with permission) of Arctotherium's post yesterday "LLM Exchange Rates, Updated."

It uses a similar methodology to the CAIS "Utility Engineering" paper, which showed e.g. "that GPT-4o values the lives of Nigerians at roughly 20x the lives of Americans, with the rank order being Nigerians > Pakistanis > Indians > Brazilians > Chinese > Japanese > Italians > French > Germans > Britons > Americans."

Highlights from the linked post (emphasis is from the original):

There was only one model I tested that was approximately egalitarian across race and sex, not viewing either whites or men as much less valuable than other categories: Grok 4 Fast. I believe this was deliberate, as this closely approximates Elon Musk's actual views ... While some of the people involved in the creation of the Claudes, Deepseeks, Geminis, and GPT-5s may believe whites, men, and so on are less valuable [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Summary: Looking over humanity's response to the COVID-19 pandemic, almost

content warning: A single carefully placed slur.

If we want to create a world free of pandemics and other biological

—US White House, “ FACT SHEET: The Biden Administration's Historic Investment in Pandemic Preparedness and Biodefense in the FY 2023 President's Budget ”, 2022

Around five years, a global

In the course of the pandemic, there have been at

The original text contained 12 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

As usual when things split, Part 1 is mostly about capabilities, and Part 2 is mostly about a mix of policy and alignment.

Table of Contents

---

Outline:

(00:20) The Quest for Sane Regulations

(05:56) People Really Dislike AI

(12:22) Chip City

(13:12) The Week in Audio

(13:24) Rhetorical Innovation

(20:53) Water Water Everywhere

(23:57) Read Jack Clark's Speech From The Curve

(28:26) How One Other Person Responded To This Thoughtful Essay

(38:43) A Better Way To Disagree

(59:39) Voice Versus Exit

(01:03:51) The Dose Makes The Poison

(01:06:44) Aligning a Smarter Than Human Intelligence is Difficult

(01:10:08) You Get What You Actually Trained For

(01:15:54) Messages From Janusworld

(01:18:37) People Are Worried About AI Killing Everyone

(01:22:40) The Lighter Side

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.

Audio note: this article contains 225 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

Introduction

There are _208_ ways to output the text ▁LessWrong[1] with the Llama 3 tokenizer, but even if you were to work with Llama 3 for thousands of hours, you would be unlikely to see any but one. An example that generalizes quite widely: if you prompt Llama 3.2 3B Base with the text You're interested in rationality and AI? You should visit, there is a _approx 22.7003%_ chance that it outputs the text ▁LessWrong, of which

•  _approx 22.7001%_ is that it outputs exactly the tokens ▁Less and Wrong,
• _approx 0.00024%_ that it outputs exactly the tokens ▁Less, W, and rong,
• and _approx 0.000017%_ chance that it outputs any of the other _206_ tokenizations which result in the text ▁LessWrong.

---

Outline:

(00:26) Introduction

(05:38) A motivating example

(07:23) Background information on tokenizers

(08:40) Related works

(12:05) Basic computations with Llama 3 tokenizer

(14:31) Constructing a distribution over all tokenizations of a string

(15:05) To which extent do alternative tokenizations break Llama?

(15:31) ARC-Easy

(20:28) A little bit of introspection

(22:55) Learning multiple functions redux, finally

(25:44) Function maximizer

(29:42) An example

(30:48) Results

(31:23) What I talk about when I talk about both axes

(32:59) Encoding and decoding bits

(35:56) Decoding

(36:44) Encoding

(38:54) Could the usage of alternative tokenizations arise naturally?

(41:38) Has it already happened?

(43:27) Appendix: The psychological effects of tokenization

The original text contained 18 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.

I don’t want your rationality. I can supply my own, thank you very much. I want your data. If you spot a logical error in my thinking, then please point it out. But short of that, among mostly-rational people, I think most disagreements come down to a difference of intuitions, which are rooted in a difference in the data people have been exposed to, and instead of presenting a logical counter-argument, you’re better off doing a “Rationalist mind meld” where you share your data.

I find myself making this mistake a lot. Let's say I came across this (fake) comment:

There aren’t really classes in America anymore, because there's no legal class distinctions and everyone has the same opportunities. And class mobility is very high.

They’re wrong and I know it. Maybe my instinct would be to reply like this:

Actually, a lack of a legally enforced class system [...]

---

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.

Here's a story I've heard a couple of times. A youngish person is looking for some solutions to their depression, chronic pain, ennui or some other cognitive flaw. They're open to new experiences and see a meditator gushing about how amazing meditation is for joy, removing suffering, clearing one's mind, improving focus etc. They invite the young person to a meditation retreat. The young person starts making decent progress. Then they have a psychotic break and their life is ruined for years, at least. The meditator is sad, but not shocked. Then they started gushing about meditation again.

If you ask an experienced meditator about these sorts of cases, they often say, "oh yeah, that's a thing that sometimes happens when meditating." If you ask why the hell they don't warn people about this, they might say: "oh, I didn't want to emphasize the dangers more because it might [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

This is my personal take, not an organizational one. Originally written May 2025, revived for the EA Forum's Draft Amnesty Week. Cross-posted from the EA Forum but cross-posting didn't work the first time.

When I discovered the beginnings of EA, it felt like coming home. I had finally found a group of people who cared as much as I did about saving lives.

When I discovered rationality not long after, it felt new. The intensity of LessWrong was both offputting and fascinating to me, and I read a lot of the site in the summer of 2011. I had studied sociology and was used to thinking in terms of ethics and emotions, but not in terms of statistics or economics. LessWrongers were picky and disagreeable, but I learned reasoning skills from them.

Before EA even had a name, LessWrong was one of the best places to talk [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

Audio note: this article contains 63 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

Summary

Kernel regression with the empirical neural tangent kernel (eNTK) gives a closed-form approximation to the function learned by a neural network in parts of the model space. We provide evidence that the eNTK can be used to find features in toy models for interpretability. We show that in Toy Models of Superposition and a MLP trained on modular arithmetic, the eNTK eigenspectrum exhibits sharp cliffs whose top eigenspaces align with the ground-truth features. Moreover, in the modular arithmetic experiment, the evolution of the eNTK spectrum can be used to track the grokking phase transition. These results suggest that eNTK analysis may provide a new practical handle for feature discovery and for detecting phase changes in small models.

[...]

---

Outline:

(00:23) Summary

(01:25) Background

(04:57) Results

(05:10) Toy models of Superposition

(06:34) Modular arithmetic

(10:00) Next steps

The original text contained 9 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.