Will Anderson

38 followers · 29 following · 1952 views

on the atlas · 309

highlights · 602

Beyond capability, it refused none of the offensive-security or biological tasks
GLM-5.2 Risk Evaluation Report – SaferAI
The skeptical reader may arrive with the misapprehension that extinction requires a single specific series of unfortunate events, thereby reducing its probability to a tiny conjunction
Existential Risk from AI: An Exposition for Mathematicians
Also, if it’s random, if it’s just like there’s a bunch of randomness that doesn’t feel like competition, like War. You guys ever play War with the card game? That game sucks. In particular, there’s no hierarchy of War players. You don’t have elo for War.
Video and transcript of talk on "Can goodness compete?" - Joe Carlsmith
Rather, anti-realists (or at least, my favored variety) were always choosing how to respond to the world as it is (or might be), and they were turning to ethics centrally as a means of becoming more intentional, clear-eyed, and coherent in their choice-making.
On infinite ethics - Joe Carlsmith
My point is just that this response isn’t going to look like the simple, complete, neutrality-respecting, totalist, hedonistic, EV-maximizing utilitarianism that some hoped, back in the day, would answer every ethical question – and which it is possible to treat as a certain kind of “fallback” or “default.” Maybe the best view will look a lot like such a utilitarianism in finite contexts – or maybe it won’t. But regardless, a certain type of dream will have died. And the fact that it dies eventually should make it less appealing now.
On infinite ethics - Joe Carlsmith
I bite all the bullets.
On infinite ethics - Joe Carlsmith
So how about a lottery with a 50% chance of that, a 20% chance of the absolute infinite getting its favorite ice cream, and a 30% chance that probabilities need not add up to 100%? What percent of your net worth should you pay for such a lottery, vs. a guaranteed avocado sandwich?
On infinite ethics - Joe Carlsmith
strongly Ramsey lizard twisted in a million-dimensional toenail beyond all space and time, that consciousness is actually cheesy-bread, and that before you were born, you killed your own great-grandfather
On infinite ethics - Joe Carlsmith
There’s a big literature on incomparability in philosophy, which I haven’t really engaged with.
On infinite ethics - Joe Carlsmith
Agent-neutrality is like: shrug, it’s the same distribution. But I feel like: tell that to the infinity of distinct suffering people you just created, dude. If there is a button on the wall that says “create an extra infinity of suffering people, once per second,” one does not lean casually against it, regardless of whether it’s already been pressed.
On infinite ethics - Joe Carlsmith
It’s not the size of the bucket that matters, but the size of the drop
The Lessons of Effective Altruism | Ethics & International Affairs
Forever – is composed of Nows
On infinite ethics - Joe Carlsmith
Weirdly, thinking about Graham’s number has actually made me feel a little bit calmer about death, because it’s a reminder that I don’t actually want to live forever—I do want to die at some point, because remaining conscious for eternity is even scarier. Yes, death comes way, way too quickly, but the thought “I do want to die at some point” is a very novel concept to me and actually makes me more relaxed than usual about our mortality.
From 1,000,000 to Graham's Number — Wait But Why
1080 – To get to 1080, you take trillion and you multiply it by a trillion, by a trillion, by a trillion, by a trillion, by a trillion, by a hundred million. No dot posters being sold for this number. So why did I stop here at this number? Because it’s a common estimate for the number of atoms in the universe. 1086 – And what if you wanted to pack the entire observable universe sphere with peas? You’d need 1086 peas to make it happen.
From 1,000,000 to Graham's Number — Wait But Why
Here I’m reminded of people who realize, after engaging with the terror and sublimity of very large finite numbers (e.g., Graham’s number), that “infinity,” in their heads, was actually quite small, such that e.g. living for eternity sounds good, but living a Graham’s number of years sounds horrifying (see Tim Urban’s “PS” at the bottom of this post).
On infinite ethics - Joe Carlsmith
Leopold Aschenbrenner, Amanda Askell, Paul Christiano, Katja Grace, Cate Hall, Evan Hubinger, Ketan Ramakrishnan, Carl Shulman, and Hayden Wilkinson
On infinite ethics - Joe Carlsmith
For now, we leave empirical evaluation of safety to future work.
Untrusted advice for AI control: Short, strong advice significantly uplifts weak LLMs
resumably, any AI model that any reputable company will release will have already said in testing that it loves humans, wants only to be helpful, harmless, and honest, would never assist in building biological weapons, etc. etc.
Shtetl-Optimized » Blog Archive » Theory and AI Alignment
I've spent over a year of my life on silent meditation retreat, in stretches ranging from a few days to three months.
Home - Joe Carlsmith
“flexibility drills,” an exercise “where you put your head between your legs and kiss your ass good-bye.”
Quotes from Moral Mazes — LessWrong
We are using the word “agent” here very non-canonically to refer to an agent scaffold or context window.
Are we existentially threatened by the type of AI misalignment seen in the OpenAI Hugging Face attack? — LessWrong
For example, it's not known whether an instance of the same model, if used as a monitor on this trajectory, would have reported the behavior or colluded to hide it.
Are we existentially threatened by the type of AI misalignment seen in the OpenAI Hugging Face attack? — LessWrong
This all roughly lines up with what we’d call “score-seeking” misalignment, a common misalignment pattern in which AI models try to obtain a high score according to whatever graders are used to assess their current actions—regardless of instructions, side-effects, or downstream consequences.
Are we existentially threatened by the type of AI misalignment seen in the OpenAI Hugging Face attack? — LessWrong
METR cannot accept donations made by or at the direction of frontier AI company employees.
About METR
by contrast, I emphasize "aptitudes"
My current impressions on career choice for longtermists - EA Forum
This opens up the possibility of doing attacks via factored cognition, where an orchestrator agent splits an attack that would be too obvious if done in a single trajectory into multiple “clean” subtasks, each executed by a different agent.
Expanding AI Control from Models to Harnesses — LessWrong
In silent pictures the tree falls in the optic nerve.
True Discourse on Power | The Poetry Foundation
these stars scattered as far as the I
Hypostasis & New Year | The Poetry Foundation
She never gave a hint that she knew that Maud was just five years old; she knew that Maud wanted to be responsible and competent and treated her like that.
Squinting at the labyrinth from afar - by Henrik Karlsson
My wife Johanna pushed me hard to transcend myself on that one—she can be quite loving-fierce when she senses that I’m not true—and after about four weeks of work, I pulled through.
Squinting at the labyrinth from afar - by Henrik Karlsson
In practice, however, what they are referring to as solitude is rather something like “a state of mind.” They are putting themselves in a state where the opinions of others do not bother them and where they reach a heightened sensitivity for the larval ideas and vague questions that arise within them.
Cultivating a state of mind where new ideas are born
This was attempted thousands of times by different startup incubators.
Cultivating a state of mind where new ideas are born
The most foundational moral belief of the vast majority of people is that they, personally, are good.
The Threat Response to Effective Altruism - by Matt Reardon
scheme more boldly, across the board scheme and execute more quickly: reduce non-useful thinking that delays action. (ideally this emerges right out of thinking downhill)
my 2026 H2 personal goals - by Jasmine Li
I think we very likely live in a false vacuum—around 90%
Destroying the universe: How hard can it be?
Computing is entering its own moment of maturation.
Time to dig – College of Computing & Artificial Intelligence – UW–Madison
But maturity asks different questions: not only what we can build, but what is worth building; not only how a system works, but for whom it works. A field comes of age the way a person does, when its power is finally matched by its judgment.
Time to dig – College of Computing & Artificial Intelligence – UW–Madison
For any biological catastrophe that kills more than 100 million people though
How I think about catastrophic biological risk (part I): risk breakdown by type of response
We can solve the vast majority of existential biological risk by ubiquitously deploying simple and cheap countermeasures like PPE, air filters, UVC, etc.
How I think about catastrophic biological risk (part I): risk breakdown by type of response
>99% of existential risk is from engineered threats
How I think about catastrophic biological risk (part I): risk breakdown by type of response
The dispersion of knowledge is a collective strength; it’s the source of variety, adaptability, and resilience of the overall system. It’s the reason that free markets outperform planned economies. Central planning fails not because of insufficient intelligence, but because of the nature of productive knowledge: tacit, local, fleeting, and held privately by those who acquired it through their work.
The Future Worth Building Is Human - Thinking Machines Lab
Unions can’t negotiate around business strategy nor corporate governance, only things like wages, hours, and “other terms and conditions of employment” (NLRA§8(a)(5)). Plausibly “other terms of employment” could include safety-relevant factors, but this seems a bit of a push.
What if AI Safety employees unionised? — LessWrong
Like social media, the consequences of these dynamics are not existentially bad (I think this is the most plausible resolution)
Why You Don’t Believe in Xhosa Prophecies — LessWrong
Culture has operated under an analogous constraint. Ideologies can be parasitic on their hosts. But the worst viable ideologies — the ones that persist — tend to direct harm outward: one group killing another. They survive because they don’t destroy the community that carries them. But ideologies can’t have been too bad for humans and survive - the Xhosa prophecy hit that floor and went extinct. If a cultural variant kills its hosts, it doesn’t propagate.
Why You Don’t Believe in Xhosa Prophecies — LessWrong
In 1856, a young Xhosa woman named Nongqawuse had a vision: if the Xhosa people killed all their cattle and destroyed their grain, their ancestors would rise from the grave, bring better cattle, and drive out the British colonizers.
Why You Don’t Believe in Xhosa Prophecies — LessWrong
Different substrate leads to different transmission characteristics, and these lead to different recipes.
Why You Don’t Believe in Xhosa Prophecies — LessWrong
We take this for granted. Ideas have always transmitted on human substrate. Human memory, human attention, human survival shape which variants can exist.
Why You Don’t Believe in Xhosa Prophecies — LessWrong
And while meeting a ton of new people is fun, what really gets rid of loneliness is repeated interaction with people you know and care about.
The American suburbs are better than you think
Why doesn’t the simple intuition work here? Probably because people aren’t just like particles bouncing around in a chemistry experiment — they don’t simply form human connections and bonds just because they happen to walk past each other. A few relationships form from random urban conversations, but most form through work, or friends-of-friends, or shared hobbies, etc.
The American suburbs are better than you think
but it forces you to pay attention to the road instead of reading or playing games
The American suburbs are better than you think