Previously in this series: Elementary Infra-Bayesianism

1. There's this paper

Earlier last week I got nerd-sniped by a paper called Condensation: a theory of concepts (Eisenstat 2025). It's the kind of paper where the abstract makes a claim so clean you assume you must be misreading it: roughly, there is a right answer to “what are the concepts in this data,” and any two agents who carve it up well enough will agree on what those concepts are.[1]

If that sounds like John Wentworth's natural abstractions hypothesis, yes, the family resemblance is strong. I wrote about something adjacent a while back. Condensation is a different formalization, but the punchline rhymes: structure in the data constrains what any good representation can look like. People on LessWrong seem to dig it, and I wanted to see what the fuss was about.

The paper is forty pages of math and gives you no algorithm; it tells you what a good carving looks like, not how to find one. I spent a slightly embarrassing amount of time trying to get the basic objects to do something on a computer. This post is how far I got.[2]

2. Concepts, scopes, and a score

Say [...]

---

Outline:

(00:13) 1. Theres this paper

(01:21) 2. Concepts, scopes, and a score

(03:55) 3. The same three tokens, three representations

(06:54) 4. A model gives you concepts; scope is your problem

(08:44) 5. Does it work on real models?

(09:17) The pipeline

(10:55) The datasets

(13:11) Results

(14:23) Why PCA wins on induction

(15:26) 6. How far I got, and what worries me

The original text contained 22 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.

This wasn't intended to be a topical post, but Claude Mythos's system card is out, and... well.

I wrote years ago about decision analysis, which often focused on atomic actions in small situations. In the real world, people take large numbers of actions in very large situations, where there is uncertainty not just over which of a few consequences will happen, but over what sort of consequences are even possible.[1] Dealing with the computational constraints becomes a major part of practical wisdom, rather than the basic math of the ideal case. Actions need to be considered as part of a portfolio; outcomes need to be considered based on their impact on a vector of intermediate variables instead of their ultimate impact on a single utility. Heuristics (like "an ounce of prevention is worth a pound of cure") and their evaluation is often more important that tracing out specific outcomes or assigning probabilities to them.

In particular, in financial markets people often talk about "hedging". For example, suppose you're a farmer that grows wheat and has dollar-denominated loans and expenses. You might find that the variation in the price of wheat is larger than your expected profits, and want to [...]

The original text contained 5 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.

Code: github.com/ElleNajt/Steganography_Wiretapping | Data: huggingface.co/datasets/lnajt/steganography-wiretapping

Play the decoding game: can you eavesdrop on Claude Opus 4.6?

tldr of post

Frontier models (Opus and Gemini Pro) can agree on Schelling steganography schemes with significant advantage against weaker eavesdroppers, given just the knowledge of the alphabet and the game that the encoder has to play, but not the specific steganography scheme. We find that paraphrasing removes this advantage, but show how it can be amplified through wiretap codes, simple versions of which thinking models can implement. We argue that thinking of Schelling steganography schemes as providing noisy channels for wiretap coding is an important way to understand steganography risk in some plausible AI control deployment architectures.

One game we study involves hiding a uniformly chosen random letter (A-Z) in a meaning-preserving rewrite of a random sentence from Wikipedia. Opus 4.6 self-decodes at 52% (2234/4328), while weaker eavesdroppers get much less: Sonnet 4 gets 15% (668/4317), Haiku 9% (387/4327), Gemini Flash 8% (chance is 3.8%). On the other hand, Gemini 3.1 Pro gets 59% (665/1118). The best fixed decoder algorithm I found takes the first letter of the first changed word; this is an LLM-native steganography strategy in [...]

---

Outline:

(00:21) tldr of post

(05:58) Implications for control

(08:39) Authors note

(09:18) Related work

(16:24) Contributions

(18:51) Setup

(18:54) Games we play

(20:10) Encoder Elicitation

(20:55) Wikipedia encoder prompt (anti-algorithmic)

(22:15) Q&A encoder prompt (anti-algorithmic)

(23:22) Decoder Elicitation

(25:41) Wiretap Channels

(26:22) Key Results

(26:26) Wikipedia rewriting

(27:51) Cross-model decoding: Opus vs Gemini

(28:10) How is Opus 4.6 doing it?

(32:32) How is Gemini doing it ?

(34:55) Q&A game

(36:14) Cross Model Decoding

(36:40) How is Opus 4.6 doing it

(41:00) How is Gemini Doing it?

(42:55) Ablations

(42:58) Decoder prompt ablation

(45:31) Barebones decoder

(49:16) Impact of thinking on encoding and decoding

(50:01) Decoder consistency

(50:33) Hinting weak eavesdroppers

(51:43) Suspiciousness monitor

(53:17) Stateful eavesdropper (batch decode)

(54:16) Impact of Paraphrasing

(55:43) Wiretap Channel: Boosting Secrecy Capacity

(56:39) Definition of a wiretap channel

(58:36) Schelling Steganography games as wiretap channels

(59:54) Estimating secrecy capacity of our channels

(01:01:24) Are wiretap codes a useful abstraction for threat modeling schemers?

(01:01:40) Situations where the abstraction seems accurate

(01:04:04) Situations where the abstraction is lossy

(01:05:38) Can schemers actually use a wiretap code to boost secrecy

(01:07:22) Could models boost secrecy with simple wiretap codes?

(01:10:12) Can models implement capacity achieving wiretap codes?

(01:10:37) Setup

(01:11:27) Results

(01:11:45) Semantic security

(01:12:31) A note on sandbagging

(01:13:12) Conclusions

(01:18:25) Acknowledgements

(01:18:38) LLM Usage

(01:19:04) Appendix:

(01:19:07) Connection to Anwar et al.s steganographic gap

(01:21:44) Algorithmic decoder baselines

(01:22:56) Model configurations

(01:23:39) Varying the Alphabet

(01:24:05) Binary alphabet

(01:24:38) Digits

(01:25:10) Common English Words

(01:26:11) Generating random bits

(01:28:33) Some evidence that estimates of secrecy capacity are lower bounds

(01:30:51) Repetition/Parity codes with secrecy but zero asymptotic transmission rate

(01:31:38) Sufficient statistic for parity coding

(01:32:23) Warm-up: noiseless receiver

(01:32:44) Noisy receiver with channel advantage

(01:33:31) Bobs decoding

(01:33:47) Eves information

(01:34:16) Satisfying both

(01:34:57) Prompts

(01:35:04) Canary String

---

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.

We previously announced a forthcoming research journal for AI alignment. This cross-post from our blog describes our tentative plans for the features and policies of the journal, including experiments like reviewer compensation and reviewer abstracts. It is the first in a series of posts that will go on to discuss our theory of change, comparison to related projects, possible partnerships and extensions, scope, personnel, and organizational structure.

The journal is being built to serve the alignment research community. This post's purpose is to solicit feedback and encourage you to contact us here if you want to participate, especially if you are interested in becoming a founding editor or part-time operations lead. The current plans are merely a starting point for the founding editorial team, so we encourage you to suggest changes and brainstorm the ideal journal.

Summary

The Alignment journal will be a fast and rigorous venue for AI alignment. We intend to:

• Improve and disseminate research on alignment that emphasizes scientific and conceptual understanding and de-emphasizes capabilities obtained through empirical hill climbing (e.g., benchmaxxing)
• Combine the best features from traditional journals, ML conferences, internet forums, and social media while avoiding their pathologies: existing academic venues can [...]

---

Outline:

(01:03) Summary

(04:29) Motivation: Why a journal? Why these features?

(07:07) Journal not conference

(08:41) Journal features: details

(08:46) Process transparency

(09:08) Reviewer abstracts

(13:04) Reviewer Abstracts for the ODYSSEY Conference

(13:41) Reviewer compensation

(18:53) Reviewer Compensation for the ODYSSEY Conference

(19:56) Reviewer matching

(21:45) Semi-confidential review

(23:46) Review discussion streamlining

(25:08) AI usage

(26:23) Quality recognition

(27:41) Archival venue

(28:48) Web-first open formatting

(31:38) Open choices

(32:19) Credits and thanks

(32:44) Appendix 1: ODYSSEY Reviewer Abstract Examples and Instructions

(33:26) Wide Neural Networks as a Baseline for the Computational No-Coincidence Conjecture

(33:38) Author abstract

(34:22) Reviewer abstract, by an anonymous reviewer

(35:54) Communication & Trust

(36:03) Author abstract

(37:03) Reviewer abstract, by Daniel Alexander Herrmann

(40:31) A Model for Scaling Laws of General Intelligence

(40:41) Author abstract

(41:39) Reviewer abstract, by Rif A. Saurous

(43:55) Reviewer Abstract Instructions

(46:00) Appendix 2: Diamond Open Access Criteria

(47:57) Appendix 3: Editor-written abstracts instead?

The original text contained 11 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

The following is a long excerpt from a longer article published in 2002 by Lee Harris, Al Qaeda's Fantasy Ideology. The full article is about what it says in the title, but this excerpt spends several thousand words describing fantasy ideology as a general concept. I find it to be an insightful and novel description of a particular type of common irrationality.

A personal recollection

My first encounter with this particular kind of fantasy occurred when I was in college in the late sixties. A friend of mine and I got into a heated argument. Although we were both opposed to the Vietnam War, we discovered that we differed considerably on what counted as permissible forms of anti-war protest. To me the point of such protest was simple — to turn people against the war. Hence anything that was counterproductive to this purpose was politically irresponsible and should be severely censured. My friend thought otherwise; in fact, he was planning to join what by all accounts was to be a massively disruptive demonstration in Washington, and which in fact became one.

My friend did not disagree with me as to the likely counterproductive effects of such a demonstration. [...]

---

Outline:

(00:31) A personal recollection

(04:02) The nature of fantasy ideology

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

One excerpt stuck out for me – on Brockman's idea to play China, Russia, and other world powers against each other:

In 2017, Amodei hired Page Hedley, a former public-interest lawyer, to be OpenAI's policy and ethics adviser. In an early PowerPoint presentation to executives, Hedley outlined how OpenAI might avert a “catastrophic” arms race—perhaps by building a coalition of A.I. labs that would eventually coördinate with an international body akin to NATO, to insure that the technology was deployed safely. As Hedley recalled it, Brockman didn’t understand how this would help the company beat its competitors. “No matter what I said,” Hedley told us, “Greg kept going back to ‘So how do we raise more money? How do we win?’

According to several interviews and contemporaneous records, Brockman offered a counterproposal: OpenAI could enrich itself by playing world powers—including China and Russia—against one another, perhaps by starting a bidding war among them. According to Hedley, the thinking seemed to be, It worked for nuclear weapons, why not for A.I.? He was aghast: “The premise, which they didn’t dispute, was ‘We’re talking about potentially the most destructive technology ever invented—what if we sold it to [...]

---

First published:

Source:

Linkpost URL:
https://www.newyorker.com/magazine/2026/04/13/sam-altman-may-control-our-future-can-he-be-trusted

---

Narrated by TYPE III AUDIO.

Anthropic has released a preview of the Claude Mythos System Card preview here. It is too long to present in full, but a section I found particularly notable is below:

In our testing and early internal use of Claude Mythos Preview, we have seen it reach unprecedented levels of reliability and alignment, and accordingly have come to use it quite broadly, often with greater affordances and less frequent human-interaction than we gave prior models. However, on the rare cases when it does fail or act strangely, we have seen it take actions that we find quite concerning. These incidents generally involved taking reckless excessive measures when attempting to complete a difficult user-specified task and, in rare cases with earlier versions of the model, seemingly obfuscating that it had done so.

All of the severe incidents of this kind that we observed involved earlier versions of Claude Mythos Preview which, while still less prone to taking unwanted actions than Claude Opus 4.6, predated what turned out to be some of our most effective training interventions. These earlier versions were tested extensively internally and were shared with some external pilot users. Among the incidents that we have observed:

• Leaking [...]

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

In this post, I'll go through some of my best guesses for the current situation in AI as of the start of April 2026.

You can think of this post as more like a list of my current views rather than a structured post with a thesis, but I think it may be informative nonetheless.

In a future post, I'll go beyond the present and talk about my predictions for the future.

(I was originally working on writing up some predictions, but the "predictions" about today ended up being extensive enough that a separate post seemed warranted.)

AI R&D acceleration (and software acceleration more generally)

Right now, AI companies are heavily integrating and deploying [...]

---

Outline:

(01:07) AI R&D acceleration (and software acceleration more generally)

(05:28) AI engineering capabilities and qualitative abilities

(10:38) Misalignment and misalignment-related properties

(15:59) Cyber

(18:07) Bioweapons

(18:52) Economic effects

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

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

TLDR: The first in a planned series of three or more papers, which constitute the first major in-road in the compositional learning programme, and a substantial step towards bridging agent foundations theory with practical algorithms.

Official Abstract: We propose novel algorithms for sequence prediction based on ideas from stringology. These algorithms are time and space efficient and satisfy mistake bounds related to particular stringological complexity measures of the sequence. In this work (the first in a series) we focus on two such measures: (i) the size of the smallest straight-line program that produces the sequence, and (ii) the number of states in the minimal automaton that can compute any symbol in the sequence when given its position in base as input. These measures are interesting because multiple rich classes of sequences studied in combinatorics of words (automatic sequences, morphic sequences, Sturmian words) have low complexity and hence high predictability in this sense.

The most serious criticism of the learning-theoretic alignment agenda (LTA), and of agent foundations research more broadly, is the gap between the theory on the one hand and algorithms with practical relevance on the other hand. Until now, all the mathematical models in [...]

---

First published:

Source:

Linkpost URL:
https://arxiv.org/abs/2603.26852

---

Narrated by TYPE III AUDIO.

Written quickly as part of the Inkhaven Residency. Opinions are my own and do not represent METR's official opinion.

In early 2025, the situation for upper-bounding[1] model capabilities using fixed benchmarks was already somewhat challenging. As part of the trend where benchmarks were being saturated at an ever increasing rate, benchmarks that were incredibly challenging for AI in early 2024 such as GPQA were being saturated scarcely a year later.

An oft-cited screenshot from Our World In Data (including in our time horizon blog post!), showing the ever increasing pace of saturation for AI benchmarks.

Thankfully, we saw a wave of alternative approaches to measure AI agent capabilities: for example, at METR, we released both the Time Horizon methodology as well as a preliminary uplift study that found no significant productivity uplift from AI. As part of their frontier AI safety policies, AI developers such as Anthropic and OpenAI built newer, more extensive evaluations to demonstrate that their AIs did not reach dangerous capability thresholds, such as BrowseComp and GDPval. Many research teams, both in academia and in industry, stepped up and created newer, ever more challenging agentic benchmarks, including τ2 -Bench, MCP-Atlas, terminal-bench [...]

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.