Dawn Drescher

• Dawn Drescher 19 Aug 2026 11:42 UTC

  1 point

  0

  on: Shut­ting Down the Light­cone Offices

  I just wrote a note on Substack that reminded me of this post. I’ll reproduce it below.

  I’m looking for a new role in the AI safety space (or possibly cofounders) to work as a field strategist or active grantmaker. As such it’s particularly important to me to avoid getting locked into any perverse incentive structures. I’m someone who supports people at what they want to do anyway, not someone who goes into an organization to fundamentally change it from within, so I need to create the right environment for my work to bear (non-poisonous) fruit.

  This post stands out to me (even three years later) as a testament to good judgment. So I wonder who the people and organizations are who have continued to maintain a focus on reducing x- and s-risks and not gotten distracted by ML prestige, Goodharting, or funding? Which ones would you warn me against? Any other pointers?

  Since 2019, I’ve been increasingly concerned about the direction that the AI safety ecosystem has taken.

  Early on it was more of a prospective concern that we’re headed in a risky directions; later it turned into a concern about the status quo (one that surely was also felt by Oliver Habryka and Ben Pace when they closed the Lightcone offices); and now it has been infused with a new bitterness as years have passed and the situation has only gotten more dire.

  Or has it? I wish I had better data to tell. My impression is that a real concern for safety against x-risks and s-risks has ceded to a myopic toil on tasks that some Powers That Be have perfumed to smell like safety research.

  The key questions on my mind are, Do the researchers know what they are working on? or Did the Powers That Be do a good job perfuming the right tasks?

  Several friends of mine argue that the researchers largely don’t know what they are working on. They know their narrow fields, but they don’t know the grander political and game-theoretical context in which it is likely situated. In a sense, they are blindly but capabily executing orders like soldiers in a war.

  I find this picture disconcerting. I’d much rather cultivate an ecosystem where there is a high level of strategic overview at every level so that (1) we don’t run into the problems of the Waterfall engineering model where the designers have the least detailed models but if they make any mistake, everything collapses, and (2) when people gain seniority and the field grows, the knowledge of the original seniors is gradually lost.

  It reminds me not only of the Waterfall model but also of widespread parenting methods where you expect your kid to be perfectly subservient and have no preferences and voice no disagreements until age 18 or 21 and then suddenly start their own family and switch to leadership mode from one year to the next.

  Another friend of mine argues that what ought to happen is something in between: When people enter the space, they can still be naive, armed only with the knowledge of their narrow domain acquired in a few months of bootcamps, reading, and chatting with AIs. But these disciples should follow the research agendas set by the elders who have the grand overview. Then, as they come of age as researchers, the wisdom of the elders percolates down, and one day they take the place of the elders, but not before they’ve learned everything that the elders knew when they assumed the position.

  Since my preferred system of widely distributed knowledge is very likely not working, the question is whether this system of timely percolation is working instead?

  It can break down in at least two ways: (1) The percolation is too slow compared to the influx of new disciples, and (2) the elders already lack the wisdom or the disciples don’t know how to select the elders who possess the wisdom.

  A fairly quick way for me to make progress on this is to read some research agendas and assess whether the research, if successful, is robustly positive/​neutral in impact or reckless or outright harmful. If it’s the first, I’ll assume that the elders, at least those who wrote it, are wise. Otherwise, not so much.

• Dawn Drescher 29 Jun 2026 22:45 UTC

  1 point

  0

  in reply to: averyzlim’s comment on: Patholog­i­cal Nar­cis­sism: The Pen­du­lum Swing be­tween Echo­ism and Sovereignism

  Interesting, thanks! How would you define moralizers, stabilizers, and challengers?

  I feel like “economic and social structures reward pathological versions of echoism, sovereignism, and narcissism” is a minor factor. Most of life sucks as a result of those adaptations, and when they come in handy in some situations (that probably resemble typical childhood situations), it doesn’t offset how much they suck in all other situations. But there’s the combined problem that people have had these adaptations from early on and so don’t know anything better, don’t know what they’re missing out on, and that the steps that would lead in the direction of healing feel like suicide.

  Maybe we need more content that contrasts secure, non-disordered ways to engage with the world against disordered ways (but in an authentic, nuanced, relatable way) to drive home the effectiveness of the first?

• Dawn Drescher 29 Jun 2026 9:11 UTC

  1 point

  0

  in reply to: averyzlim’s comment on: Patholog­i­cal Nar­cis­sism: The Pen­du­lum Swing be­tween Echo­ism and Sovereignism

  Wanna have a call some time to chat about this! Really exciting to see others think about these things! :-D

  It seems likely to me that one of the greatest challenges in encouraging change is that it is difficult for people to have the kinds of life experiences that shift their perceived values regarding cooperation, emotional safety, idealism, and attachment.

  Yeah, on both the avoidant and the preoccupied side, there’s a lack of trust in others. Either along the lines of, “No one must know how weak/​flawed/​bad I really am, or they’ll banish me as punishment,” or “All these assholes are out to get me, so I need to stay one step ahead of them and never reveal any of my weaknesses.” In the preoccupied case, everyone else is judge, jury, and executioner; in the avoidant case everyone is an enemy.

  But additionally, especially in the first case, you’re your own judge, jury, and executioner too, so even getting to the point where one can think the above thoughts is hard.

  My hope has been that just by demonstrating that I can think and express things that are similar (hopefully) to what they’re experiencing and make it sound as normal as possible, I can create the kind of safety where they can at least allow themselves to think them.

  I think that can work, but if they’re very disgusted with these parts of themselves, it can also backfire and cause them to become disgusted with me for displaying or accepting these parts. A friend of mine who had this problem had a therapist who did basically the opposite of what I would do. The therapist remained perfectly neutral and probably did the blank slate thing that therapists used to do more often, so on the one hand the experience was more scary for my friend as a result, but she didn’t run into the disgust issue. But her recovery took some 8 years, which is pretty slow, so overall, a less conservative approach might be better.

  I imagine it’s all quite similar to parenting. Amanda Diekman (Low-Demand Parenting) also writes that she prioritizes a trusting relationship between her and her children over almost everything else. Once a parent manages to win the trust of their child, many other things can flow from that because the child is no longer locked inside only their own mind. I’ve heard the term “epistemic trust” in this context.

  Contemporary society makes both pathways difficult to access, and many of these traits also involve behaviors that steer people away from those opportunities in the first place.

  What do you mean by this?

  Riffing off how this article explores different expressions of sovereignism, narcissism, and echoism across different attachment styles and contexts, I wonder whether some people also express these traits differently depending on whether they are evaluating themselves, other people, or social norms. I’ve noticed differences between what people feel comfortable expressing themselves, what they value or expect from other people, and what they perceive as baseline expectations or social norms.

  And this – do you have examples?

Patholog­i­cal Nar­cis­sism: The Pen­du­lum Swing be­tween Echo­ism and Sovereignism

Speedup from AI Ghostwriting

Tac­ti­cal and Oper­a­tional Ex­plo­ra­tory Model­ing for AI Governance

On Re­spon­si­bil­ity and Death: Can We See Real­ity for What It Is or Will It Break Us

• Dawn Drescher 17 May 2026 18:25 UTC

  0 points

  0

  on: Threat-Re­sis­tant Bar­gain­ing Me­ga­post: In­tro­duc­ing the ROSE Value

  I used NotebookLM to generate a summary of the sequence with the prompt:

  Let’s say you’re talking to a bunch of psychopaths who are familiar with basic game theory and the three bargaining solutions and have no notion of human fairness in the first place. So you can skip over all that fair vs. chaa and get right to the chase: we’re bargaining, we endorse different bargaining solutions, and the outcomes we would accept don’t coincide but are wildly different. How do we successfully bargain anyway? Thanks!

  Is the video accurate or do you all notice any issues with it? I want to make this critically important topic more accessible to people.

Psy­chopa­thy: The Choice

Psy­chopa­thy: The Types

Psy­chopa­thy: The Mechanics

• Dawn Drescher 5 May 2026 20:21 UTC

  1 point

  0

  in reply to: Ninety-Three’s comment on: Psy­chopa­thy: The Problem

  Lol, then knock yourself out with this one, because that’s virtually all hand-written! (Inb4 it’s also unaesthetic, just in a different way. 🙈)

• Dawn Drescher 5 May 2026 17:57 UTC

  −1 points

  0

  in reply to: Ninety-Three’s comment on: Psy­chopa­thy: The Problem

  1. The sequence clarifies the term “psychopathy,” preventing unhelpful or outright misdiagnoses

  2. The sequence clarifies the internal experience and differences in the internal experience of people with psychopathy, which is important for self insight and treatment

  3. The sequence explains the adaptive mechanisms reducing stigma, which makes treatment more accessible

  4. The sequence clarifies the term “recovery” and reframes it as a menu that patients can choose from, which breaks black and white thinking around whether one is perfect or broken

  5. And much more

  So you can happily downvote because you don’t have anyone in your life you has any of these conditions so it’s not useful for you personally, but when you downvote it because you don’t like something about my writing style, then I wouldn’t call it downvoting on its merits but more on linguistic-aesthetic taste.

Psy­chopa­thy: The Self

Psy­chopa­thy: The Shaping

• Dawn Drescher 3 May 2026 15:59 UTC

  3 points

  0

  in reply to: Seth Herd’s comment on: Psy­chopa­thy: The Problem

  Thanks, I can expand my LLM note a bit more! I just remembered that I have a backup of my full conversation (up to the point where I took the backup, but almost all of it) with Claude, including the first drafts.

  Having thought and read about psychopathy for so long, I felt very confused about how to structure my mental model, so my input to Claude were countless fairly unorganized thoughts about models, contradictions, advantages and disadvantages of framings, etc., and Claude’s first big contribution was to suggest this tag structure where tags (made up of a letter and a descriptor) get combined to form a personality profile. That was a format that hadn’t occurred to me and that I loved for its power and flexibility. But then it was me again who fleshed out that model – introduced the layers of genetics, neurology, psychodynamics, behavior, etc.

• Dawn Drescher 3 May 2026 15:27 UTC

  1 point

  0

  in reply to: Seth Herd’s comment on: Psy­chopa­thy: The Problem

  Thanks! <3

  Yeah, I didn’t keep track of what words were written by who. It’s quite plausible that I’ve touched every single sentenced and wrote almost half of them from scratch, but it’s also plausible that I touched maybe 80% of sentences and wrote 20% from scratch. The Choice article is mostly hand-written because the AI didn’t have a lot of ideas, but this one is more mixed. It’s hard for me to reconstruct at this point. In the future I can try to commit all changes to Git with correct attribution and then share the whole edit history for transparency. (But really, when two people coauthor an article, it’s also often not clear who wrote which sentences, contributed which ideas, did which edits during the proofreading, etc.)
  
  I’ll move the collapsible box to the top if it lets me! (The editor gets a bit weird when I use these blocks.)

• Dawn Drescher 3 May 2026 13:38 UTC

  6 points

  −1

  in reply to: Jiro’s comment on: Psy­chopa­thy: The Problem

  I throw into my articles whatever I want

Psy­chopa­thy: The Substrate

• Dawn Drescher 2 May 2026 22:46 UTC

  2 points

  0

  in reply to: FlorianH’s comment on: Psy­chopa­thy: The Problem

  Ty! <3