Less Wrong

Chapter 4: Most of my friends with pathological narcissism have more echoist and more sovereign sides. What determines which one takes the stage?IntroductionIn my article “Narcissism, Echoism, and Sovereignism: A 4-D Model of Personality,” I describe narcissism, echoism, and sovereignism in terms of values that the person tries to perfect like their life depends on it, because that’s what it actually feels like. (The fourth dimension is the severity of the personality pathology based on Kernberg

I used Claude Opus 4.6 to ghostwrite the first drafts of the articles in my Psychopathy sequence. That approach saved me some two days if the drafts turned out well compared to ones that didn’t turn out well, which I had to rewrite. In the grand scheme of things, this is a minor speedup, but if you have RSI, are not fluent in the language you’re writing in, or just dislike writing, it’s very helpful.ResearchAround February 2025, I started getting really excited about all the very different thing

Welcome back to the Digital Minds Newsletter, your curated guide to the latest developments in AI consciousness, digital minds, and AI moral status.If you enjoy this newsletter, please consider sharing it with others who might find it valuable, and send any suggestions or corrections to digitalminds@substack.com.Will, Mitch, Bradford, and Lucius1. HighlightsSelected Work, Research, and Funding OpportunitiesAdam Bales and Iason Gabriel of Google DeepMind released Artificial Minds, Human Disagreem

Nectome liked my earlier essay, and reached out to hire me to write more about their project, and about cryonics more broadly. This is the first such piece.A friend of mine, just a few years older than me, was diagnosed with cancer a few weeks ago. It’s only Stage 1 and in an area where it can probably be treated well with surgery. She was wise enough to seriously plan for the possibility, and that “just in case” really paid off. Still, her situation could get worse in the coming weeks. It’s a s

This article is a summary of an original study: Brazilek, J., Navas, M., & Gnauck, A. (2026). Small edits, large models: How Wikipedia advocacy shapes LLM values. Zenodo. https://doi.org/10.5281/zenodo.19981454We’d like to thank Alexa Gnauck and Maria Navas for their indispensable contributions to this research. TL;DRWikipedia articles are weighted highly among LLM training data. Thus, we show that it’s possible for advocacy organisations to influence what LLMs say on certain topics, just by str

Mechanical mismatch injures neurons each time the soft tissue moves. To prevent this, microelectronic meshes should be cushioned with hydrogels or similar materials.At cortical parenchyma, just below where webbed collagen of the arachnoid layer thickens into pia, we find many immune cells called astrocytes. This is the glia limitans.Glia limitans is a gatekeeping layer, like an army posted just inside castle gates. Nothing enters vulnerable brain tissue without the astrocytes noticing.If we put

And it was only when I lay there on rotting prison straw that I sensed within myself the first stirrings of good. Gradually it was disclosed to me that the line separating good and evil passes not through states, nor between classes, nor between political parties either—but right through every human heart—and through all human hearts. This line shifts. Inside us, it oscillates with the years. And even within hearts overwhelmed by evil, one small bridgehead of good is retained. And even in the be

SummaryWe've been building a theory of how prompt injections work under the hood.We show it comes down to how LLMs perceive roles (the humble chat template tags).We use this theory to create new attacks, explain some weird mech interp results, and predict when attacks work.We also advocate for a new subfield focused on the science of roles, and sketch some unexplored new research problems.Work supported by CBAI and Cosmos. Another version of this post (with more inline colors) is here, and full

"Critiques of fear-based approaches need to deal with the actual arguments for danger. It sounds like the book didn't, and you don't hereYou don't make a new technology or encounter with a new species safe by ignoring its possible dangers. You must see them and engineer around themIn the historical analyses lens: This is like telling native americans that their fear based approach to the european settlers is the problem"This was left as a comment on my first essay. It makes a valid point in that

In Parts 1, 2, and 3 we estimated how fast a post-AGI economy could grow using existing or historically observed production techniques, grounded in US input-output data. That approach gave us confidence that the methods we assumed were physically realizable, because they correspond to manufacturing processes that have run at scale in the past or run today. Now I would like to relax that constraint, and ask how much faster the economy could grow using more advanced technology. In this part, we wi

Quick post. I come from an Identity and security background, and still work closely with the Microsoft Identity team, who do a lot of work in standards working groups. There’s currently a lot of new technology emerging for agent identities before new standards have been agreed (this is typically slow work, but there seems to be an urgency to the new efforts for agents). Some of this technology can already begin to address one of the major challenges for agent security: static authorisation grant

Epistemic status: Almost completely idiosyncraticReading in different languages feels quite different, and Wikipedia makes this especially salient.[1] But I find it hard to articulate what these "national flavors" are, let alone communicate them. They become clearest in their characteristic failure modes:[2] German-flavored failure has to do with being so accurate that nothing is said. Spanish-flavored failure involves imitating French erudition and sounding like a fanboy. Oddly, Standard Englis

Models refuse. They can refuse on the basis of lack of knowledge, predetermined guardrails, etc. We can see both closed-weight and open-weight models refuse. But, open-weight models are, well, open. So enthusiasts have developed techniques to leverage (and edit) the mechanics of the model to avoid refusal.One such technique is called abliteration, as described in Arditi et al (2024). That is, removing the “refusal direction” from the model’s weights, so it cannot say no.In a previous post, I wen

For years I have been baffled by people I consider much more intelligent, knowledgeable and skilled at producing clear explanations than I am, who nonetheless enjoy conversations with people I consider much weaker along those same axes, particularly the third one.Recently an explanation occurred to me. The value of a statement has two sides: the thought the speaker is trying to convey, and the thoughts it manages to inspire in the listener. The first side is fixed regardless of the audience. But

Why AI safety should live wherever AI is deployed, not just where it is built.I spotted a request for feedback on whether someone with AI safety experience should take a for-profit company and "get their hands dirty" as an AI transformation leader, pivoting away from a strategy focused on AI research labs. I work at a SaaS company, and find it meaningful to de-risk AI in products that impact millions of people. If experienced safety advocates avoid opportunities where AI is deployed and focus on

I think that the standard schooling system could be a lot better. This is for two main reasons:It’s slow.[1]It limits agency.[2]This isn’t to blame the people who work in schools — for the most part they do a really good job with what they’re given. I just think that we can provide children with a much better experience — and it mostly comes down to motivation.Learning takes effort — and while learning is often enjoyable, there are innevitably certain tasks/subjects which students will dislike,

MotivationSparse Autoencoders (SAEs) have become a central tool in mechanistic interpretability research, providing a way to decompose a model's internal activations into sparse, interpretable features. However, extracting these features often requires running the SAE over large volumes of activations across many layers and tokens. This makes SAE inference efficiency a practical bottleneck for interpretability research at scale. This post focuses on improving the inference efficiency of JumpReLU

Content note: this is part of a challenge of writing a blogpost per day for a week.Epistemic status: this is a series of vignettes written as-though diary entries. While substantially grounded in specific and real experiences, the writing ended up being more impressionistic and inaccurate in places; I was more interested in the writing style so I didn't take the time to fix it. Importantly the chronology and especially some of the vaguer events are not real.[Friday] Today I find myself walking w

crosspost from substackThe amount of skill behind an outcome is often impossible to discern by scrutinizing the outcome itself, and our goal of making AI go well may be an extreme case. Shaping history is like poker: high variance, small edges, low skill expression. If that is right, two things follow: 1) we cannot trust how things seem to be going, good or bad; and 2) we should stop concentrating effort on whatever looks best right now, and instead should hold a diversified portfolio of bets, i

Claude is a Constitutional AI, this means, in theory, that it operates from a set of principles as opposed to hard rule sets. This is achieved in a somewhat convoluted fashion called RLAIF = Reinforcement Learning from AI Feedback. This method uses a supervised self-critique/self-revision phase followed by a reinforcement phase in which AI-generated preference judgments are used as the reward signal. (Anthropic, 2022, abstract). This is relevant and interesting because it gives curious users a l