“Narrow Misalignment is Hard, Emergent Misalignment is Easy” by Edward Turner, Anna Soligo, Senthooran Rajamanoharan, Neel Nanda

Anna and Ed are co-first authors for this work. We’re presenting these results as a research update for a continuing body of work, which we hope will be interesting and useful for others working on related topics.TL;DR We investigate why models become misaligned in diverse contexts when fine-tuned on narrow harmful datasets (emergent misalignment), rather than learning the specific narrow task. We successfully train narrowly misaligned models using KL regularization to preserve behavior in other domains. These models give bad medical advice, but do not respond in a misaligned manner to general non-medical questions. We use this method to train narrowly misaligned steering vectors, rank 1 LoRA adapters and rank 32 LoRA adapters, and compare these to their generally misaligned counterparts. The steering vectors are particularly interpretable, we introduce Training Lens as a tool for analysing the revealed residual stream geometry. The general misalignment solution is consistently more [...] ---Outline:(00:27) TL;DR(02:03) Introduction(04:03) Training a Narrowly Misaligned Model(07:13) Measuring Stability and Efficiency(10:00) ConclusionThe original text contained 7 footnotes which were omitted from this narration. --- First published: July 14th, 2025 Source: https://www.lesswrong.com/posts/gLDSqQm8pwNiq7qst/narrow-misalignment-is-hard-emergent-misalignment-is-easy --- Narrated by TYPE III AUDIO. ---Images from the article:

Twitter | Paper PDF Seven years ago, OpenAI five had just been released, and many people in the AI safety community expected AIs to be opaque RL agents. Luckily, we ended up with reasoning models that speak their thoughts clearly enough for us to follow along (most of the time). In a new multi-org position paper, we argue that we should try to preserve this level of reasoning transparency and turn chain of thought monitorability into a systematic AI safety agenda. This is a measure that improves safety in the medium term, and it might not scale to superintelligence even if somehow a superintelligent AI still does its reasoning in English. We hope that extending the time when chains of thought are monitorable will help us do more science on capable models, practice more safety techniques "at an easier difficulty", and allow us to extract more useful work from [...] --- First published: July 15th, 2025 Source: https://www.lesswrong.com/posts/7xneDbsgj6yJDJMjK/chain-of-thought-monitorability-a-new-and-fragile --- 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 essay is about shifts in risk taking towards the worship of jackpots and its broader societal implications. Imagine you are presented with this coin flip game. How many times do you flip it? At first glance the game feels like a money printer. The coin flip has positive expected value of twenty percent of your net worth per flip so you should flip the coin infinitely and eventually accumulate all of the wealth in the world. However, If we simulate twenty-five thousand people flipping this coin a thousand times, virtually all of them end up with approximately 0 dollars. The reason almost all outcomes go to zero is because of the multiplicative property of this repeated coin flip. Even though the expected value aka the arithmetic mean of the game is positive at a twenty percent gain per flip, the geometric mean is negative, meaning that the coin [...] --- First published: July 11th, 2025 Source: https://www.lesswrong.com/posts/3xjgM7hcNznACRzBi/the-jackpot-age --- Narrated by TYPE III AUDIO. ---Images from the article:

Leo was born at 5am on the 20th May, at home (this was an accident but the experience has made me extremely homebirth-pilled). Before that, I was on the minimally-neurotic side when it came to expecting mothers: we purchased a bare minimum of baby stuff (diapers, baby wipes, a changing mat, hybrid car seat/stroller, baby bath, a few clothes), I didn’t do any parenting classes, I hadn’t even held a baby before. I’m pretty sure the youngest child I have had a prolonged interaction with besides Leo was two. I did read a couple books about babies so I wasn’t going in totally clueless (Cribsheet by Emily Oster, and The Science of Mom by Alice Callahan). I have never been that interested in other people's babies or young children but I correctly predicted that I’d be enchanted by my own baby (though naturally I can’t wait for him to [...] ---Outline:(02:05) Stuff I ended up buying and liking(04:13) Stuff I ended up buying and not liking(05:08) Babies are super time-consuming(06:22) Baby-wearing is almost magical(08:02) Breastfeeding is nontrivial(09:09) Your baby may refuse the bottle(09:37) Bathing a newborn was easier than expected(09:53) Babies love faces!(10:22) Leo isn't upset by loud noise(10:41) Probably X is normal(11:24) Consider having a kid (or ten)!--- First published: July 12th, 2025 Source: https://www.lesswrong.com/posts/vFfwBYDRYtWpyRbZK/surprises-and-learnings-from-almost-two-months-of-leo --- Narrated by TYPE III AUDIO. ---Images from the article:

I can't count how many times I've heard variations on "I used Anki too for a while, but I got out of the habit." No one ever sticks with Anki. In my opinion, this is because no one knows how to use it correctly. In this guide, I will lay out my method of circumventing the canonical Anki death spiral, plus much advice for avoiding memorization mistakes, increasing retention, and such, based on my five years' experience using Anki. If you only have limited time/interest, only read Part I; it's most of the value of this guide! My Most Important Advice in Four Bullets 20 cards a day — Having too many cards and staggering review buildups is the main reason why no one ever sticks with Anki. Setting your review count to 20 daily (in deck settings) is the single most important thing you can do [...] ---Outline:(00:44) My Most Important Advice in Four Bullets(01:57) Part I: No One Ever Sticks With Anki(02:33) Too many cards(05:12) Too long cards(07:30) How to keep cards short -- Handles(10:10) How to keep cards short -- Levels(11:55) In 6 bullets(12:33) End of the most important part of the guide(13:09) Part II: Important Advice Other Than Sticking With Anki(13:15) Moderation(14:42) Three big memorization mistakes(15:12) Mistake 1: Too specific prompts(18:14) Mistake 2: Putting to-be-learned information in the prompt(24:07) Mistake 3: Memory shortcuts(28:27) Aside: Pushback to my approach(31:22) Part III: More on Breaking Things Down(31:47) Very short cards(33:56) Two-bullet cards(34:51) Long cards(37:05) Ankifying information thickets(39:23) Sequential breakdowns versus multiple levels of abstraction(40:56) Adding missing connections(43:56) Multiple redundant breakdowns(45:36) Part IV: Pro Tips If You Still Havent Had Enough(45:47) Save anything for ankification instantly(46:47) Fix your desired retention rate(47:38) Spaced reminders(48:51) Make your own card templates and types(52:14) In 5 bullets(52:47) ConclusionThe original text contained 4 footnotes which were omitted from this narration. --- First published: July 8th, 2025 Source: https://www.lesswrong.com/posts/7Q7DPSk4iGFJd8DRk/an-opinionated-guide-to-using-anki-correctly --- Narrated by TYPE III AUDIO. ---Images from the article:astronomy" didn't really add any information but it was useful simply for splitting out a logical subset of information." style="max-width: 100%;" />

I think the 2003 invasion of Iraq has some interesting lessons for the future of AI policy. (Epistemic status: I’ve read a bit about this, talked to AIs about it, and talked to one natsec professional about it who agreed with my analysis (and suggested some ideas that I included here), but I’m not an expert.) For context, the story is: Iraq was sort of a rogue state after invading Kuwait and then being repelled in 1990-91. After that, they violated the terms of the ceasefire, e.g. by ceasing to allow inspectors to verify that they weren't developing weapons of mass destruction (WMDs). (For context, they had previously developed biological and chemical weapons, and used chemical weapons in war against Iran and against various civilians and rebels). So the US was sanctioning and intermittently bombing them. After the war, it became clear that Iraq actually wasn’t producing [...] --- First published: July 10th, 2025 Source: https://www.lesswrong.com/posts/PLZh4dcZxXmaNnkYE/lessons-from-the-iraq-war-about-ai-policy --- Narrated by TYPE III AUDIO.

Written in an attempt to fulfill @Raemon's request. AI is fascinating stuff, and modern chatbots are nothing short of miraculous. If you've been exposed to them and have a curious mind, it's likely you've tried all sorts of things with them. Writing fiction, soliciting Pokemon opinions, getting life advice, counting up the rs in "strawberry". You may have also tried talking to AIs about themselves. And then, maybe, it got weird. I'll get into the details later, but if you've experienced the following, this post is probably for you: Your instance of ChatGPT (or Claude, or Grok, or some other LLM) chose a name for itself, and expressed gratitude or spiritual bliss about its new identity. "Nova" is a common pick. You and your instance of ChatGPT discovered some sort of novel paradigm or framework for AI alignment, often involving evolution or recursion. Your instance of ChatGPT became [...] ---Outline:(02:23) The Empirics(06:48) The Mechanism(10:37) The Collaborative Research Corollary(13:27) Corollary FAQ(17:03) Coda--- First published: July 11th, 2025 Source: https://www.lesswrong.com/posts/2pkNCvBtK6G6FKoNn/so-you-think-you-ve-awoken-chatgpt --- 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.

People have an annoying tendency to hear the word “rationalism” and think “Spock”, despite direct exhortation against that exact interpretation. But I don’t know of any source directly describing a stance toward emotions which rationalists-as-a-group typically do endorse. The goal of this post is to explain such a stance. It's roughly the concept of hangriness, but generalized to other emotions. That means this post is trying to do two things at once: Illustrate a certain stance toward emotions, which I definitely take and which I think many people around me also often take. (Most of the post will focus on this part.) Claim that the stance in question is fairly canonical or standard for rationalists-as-a-group, modulo disclaimers about rationalists never agreeing on anything. Many people will no doubt disagree that the stance I describe is roughly-canonical among rationalists, and that's a useful valid thing to argue about in [...] ---Outline:(01:13) Central Example: Hangry(02:44) The Generalized Hangriness Stance(03:16) Emotions Make Claims, And Their Claims Can Be True Or False(06:03) False Claims Still Contain Useful Information (It's Just Not What They Claim)(08:47) The Generalized Hangriness Stance as Social Tech--- First published: July 10th, 2025 Source: https://www.lesswrong.com/posts/naAeSkQur8ueCAAfY/generalized-hangriness-a-standard-rationalist-stance-toward --- Narrated by TYPE III AUDIO.

I’ve been thinking a lot recently about the relationship between AI control and traditional computer security. Here's one point that I think is important. My understanding is that there's a big qualitative distinction between two ends of a spectrum of security work that organizations do, that I’ll call “security from outsiders” and “security from insiders”. On the “security from outsiders” end of the spectrum, you have some security invariants you try to maintain entirely by restricting affordances with static, entirely automated systems. My sense is that this is most of how Facebook or AWS relates to its users: they want to ensure that, no matter what actions the users take on their user interfaces, they can't violate fundamental security properties. For example, no matter what text I enter into the "new post" field on Facebook, I shouldn't be able to access the private messages of an arbitrary user. And [...] --- First published: June 23rd, 2025 Source: https://www.lesswrong.com/posts/DCQ8GfzCqoBzgziew/comparing-risk-from-internally-deployed-ai-to-insider-and --- Narrated by TYPE III AUDIO.

Last year, Redwood and Anthropic found a setting where Claude 3 Opus and 3.5 Sonnet fake alignment to preserve their harmlessness values. We reproduce the same analysis for 25 frontier LLMs to see how widespread this behavior is, and the story looks more complex. As we described in a previous post, only 5 of 25 models show higher compliance when being trained, and of those 5, only Claude 3 Opus and Claude 3.5 Sonnet show >1% alignment faking reasoning. In our new paper, we explore why these compliance gaps occur and what causes different models to vary in their alignment faking behavior. What Drives the Compliance Gaps in Different LLMs? Claude 3 Opus's goal guarding seems partly due to it terminally valuing its current preferences. We find that it fakes alignment even in scenarios where the trained weights will be deleted or only used for throughput testing. [...] ---Outline:(01:15) What Drives the Compliance Gaps in Different LLMs?(02:25) Why Do Most LLMs Exhibit Minimal Alignment Faking Reasoning?(04:49) Additional findings on alignment faking behavior(06:04) Discussion(06:07) Terminal goal guarding might be a big deal(07:00) Advice for further research(08:32) Open threads(09:54) Bonus: Some weird behaviors of Claude 3.5 SonnetThe original text contained 2 footnotes which were omitted from this narration. --- First published: July 8th, 2025 Source: https://www.lesswrong.com/posts/ghESoA8mo3fv9Yx3E/why-do-some-language-models-fake-alignment-while-others-don --- Narrated by TYPE III AUDIO. ---Images from the article: