Audio note: this article contains 31 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

Lewis Smith*, Sen Rajamanoharan*, Arthur Conmy, Callum McDougall, Janos Kramar, Tom Lieberum, Rohin Shah, Neel Nanda

* = equal contribution

The following piece is a list of snippets about research from the GDM mechanistic interpretability team, which we didn’t consider a good fit for turning into a paper, but which we thought the community might benefit from seeing in this less formal form. These are largely things that we found in the process of a project investigating whether sparse autoencoders were useful for downstream tasks, notably out-of-distribution probing.

• To validate whether SAEs were a worthwhile technique, we explored whether they were useful on the downstream task of OOD generalisation when detecting harmful intent in user prompts
• [...]

---

Outline:

(01:08) TL;DR

(02:38) Introduction

(02:41) Motivation

(06:09) Our Task

(08:35) Conclusions and Strategic Updates

(13:59) Comparing different ways to train Chat SAEs

(18:30) Using SAEs for OOD Probing

(20:21) Technical Setup

(20:24) Datasets

(24:16) Probing

(26:48) Results

(30:36) Related Work and Discussion

(34:01) Is it surprising that SAEs didn't work?

(39:54) Dataset debugging with SAEs

(42:02) Autointerp and high frequency latents

(44:16) Removing High Frequency Latents from JumpReLU SAEs

(45:04) Method

(45:07) Motivation

(47:29) Modifying the sparsity penalty

(48:48) How we evaluated interpretability

(50:36) Results

(51:18) Reconstruction loss at fixed sparsity

(52:10) Frequency histograms

(52:52) Latent interpretability

(54:23) Conclusions

(56:43) Appendix

The original text contained 7 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.

Epistemic status: Reasonably confident in the basic mechanism.

Have you noticed that you keep encountering the same ideas over and over? You read another post, and someone helpfully points out it's just old Paul's idea again. Or Eliezer's idea. Not much progress here, move along.

Or perhaps you've been on the other side: excitedly telling a friend about some fascinating new insight, only to hear back, "Ah, that's just another version of X." And something feels not quite right about that response, but you can't quite put your finger on it.

I want to propose that while ideas are sometimes genuinely that repetitive, there's often a sneakier mechanism at play. I call it Conceptual Rounding Errors – when our mind's necessary compression goes a bit too far .

Too much compression

A Conceptual Rounding Error occurs when we encounter a new mental model or idea that's partially—but not fully—overlapping [...]

---

Outline:

(01:00) Too much compression

(01:24) No, This Isnt The Old Demons Story Again

(02:52) The Compression Trade-off

(03:37) More of this

(04:15) What Can We Do?

(05:28) When It Matters

---

First published:

Source:

Narrated by TYPE III AUDIO.

I’ve spent the past 7 years living in the DC area. I moved out there from the Pacific Northwest to go to grad school – I got my masters in Biodefense from George Mason University, and then I stuck around, trying to move into the political/governance sphere. That sort of happened. But I will now be sort of doing that from rural California rather than DC, and I’ll be looking for something else – maybe something more unusual – do to next.

A friend asked if this means I’m leaving biosecurity behind. No, I’m not, but only to the degree that I was ever actually in biosecurity in the first place. For the past few years I’ve been doing a variety of contracting and research and writing jobs, many of which were biosecurity related, many of which were not. Many of these projects, to be clear, were incredibly cool [...]

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 127 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

(Work done at Convergence Analysis. The ideas are due to Justin. Mateusz wrote the post. Thanks to Olga Babeeva for feedback on this post.)

In this post, we introduce the typology of structure, function, and randomness that builds on the framework introduced in the post Goodhart's Law Causal Diagrams. We aim to present a comprehensive categorization of the causes of Goodhart's problems.

But first, why do we care about this?

Goodhart's Law recap

The standard definition of Goodhart's Law is: "when a proxy for some value becomes the target of optimization pressure, the proxy will cease to be a good proxy.".

More specifically: we see a meaningful statistical relationship between the values of two random variables _I_ [...]

---

Outline:

(00:56) Goodharts Law recap

(01:47) Some motivation

(02:48) Introduction

(07:36) Ontology

(07:39) Causal diagrams (re-)introduced

(11:40) Intervention, target, and measure

(15:05) Goodhart failure

(16:23) Types of Goodhart failures

(16:27) Structural errors

(20:47) Functional errors

(25:48) Calibration errors

(28:02) Potential extensions and further directions

(28:54) Appendices

(28:57) Order-theoretic details

(30:11) Relationship to Scott Garrabrants Goodhart Taxonomy

The original text contained 10 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.

Download the latest PDF with links to court dockets here.

---

First published:

Source:

Linkpost URL:
https://chatgptiseatingtheworld.com/2025/03/24/latest-map-of-all-40-copyright-suits-v-ai-in-u-s/

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.

In this post, I'll list all the areas of control research (and implementation) that seem promising to me.

This references framings and abstractions discussed in Prioritizing threats for AI control.

First, here is a division of different areas, though note that these areas have some overlaps:

This division is imperfect and somewhat [...]

---

Outline:

(01:43) Developing and using settings for control evaluations

(13:51) Better understanding control-relevant capabilities and model properties

(28:21) Developing specific countermeasures/techniques in isolation

(36:01) Doing control experiments on actual AI usage

(40:43) Prototyping or implementing software infrastructure for control and security measures that are particularly relevant to control

(41:58) Prototyping or implementing human processes for control

(46:50) Conceptual research and other ways of researching control other than empirical experiments

(49:37) Nearer-term applications of control to practice and accelerate later usage of control

---

First published:

Source:

Narrated by TYPE III AUDIO.

We recently released Subversion Strategy Eval: Can language models statelessly strategize to subvert control protocols?, a major update to our previous paper/blogpost, evaluating a broader range of models (e.g. helpful-only Claude 3.5 Sonnet) in more diverse and realistic settings (e.g. untrusted monitoring).

Abstract

An AI control protocol is a plan for usefully deploying AI systems that prevents an AI from intentionally causing some unacceptable outcome. Previous work evaluated protocols by subverting them using an AI following a human-written strategy. This paper investigates how well AI systems can generate and act on their own strategies for subverting control protocols whilst operating statelessly (i.e., without shared memory between contexts). To do this, an AI system may need to reliably generate effective strategies in each context, take actions with well-calibrated probabilities, and coordinate plans with other instances of itself without communicating.

We develop Subversion Strategy Eval, a suite of eight environments [...]

---

Outline:

(00:38) Abstract

(02:13) Paper overview

(03:55) Results summary

(05:31) Qualitative analysis

(08:50) How do these results inform risk?

(11:08) How should this eval be used with more capable models?

(11:13) Tuning conservatism

(14:19) Difficulties with attaining a conservative measurement

(17:41) Reducing noise

The original text contained 3 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.

Last week I covered Anthropic's relatively strong submission, and OpenAI's toxic submission. This week I cover several other submissions, and do some follow-up on OpenAI's entry.

Google Also Has Suggestions

The most prominent remaining lab is Google. Google focuses on AI's upside. The vibes aren’t great, but they’re not toxic. The key asks for their ‘pro-innovation’ approach are:

---

Outline:

(00:21) Google Also Has Suggestions

(05:05) Another Note on OpenAI's Suggestions

(08:33) On Not Doing Bad Things

(12:45) Corporations Are Multiple People

(14:37) Hollywood Offers Google and OpenAI Some Suggestions

(16:02) Institute for Progress Offers Suggestions

(20:36) Suggestion Boxed In

---

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.

A Note on GPT-2

---

Outline:

(00:57) A Note on GPT-2

(02:02) What Even is AGI

(05:24) Seeking Deeply Irresponsibly

(07:07) Others Don't Feel the AGI

(08:07) Epoch Feels the AGI

(13:06) True Objections to Widespread Rapid Growth

(19:50) Tying It Back

---

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 seems the company has gone bankrupt and wants to be bought and you can probably get their data if you buy it. I'm not sure how thorough their transcription is, but they have records for at least 10 million people (who voluntarily took a DNA test and even paid for it!). Maybe this could be useful for the embryo selection/editing people.

---

First published:

Source:

Narrated by TYPE III AUDIO.