Hide table of contents

Thanks to Siao Si for helpfwl debate on several aspects of this post. Update: seems like the idea already exists on LessWrong: iterated trust kickstarters.

In a chat with Seth Baum in gather town during EAGxVirtual we were talking about the risk of nuclear escalation in the Russian invasion of Ukraine, and the question came up,

"What can we personally do to contribute?"

In this post, I want to give an introspective demonstration of my particular approach to building an inside view as a complete beginner to a topic. I do think my approach is somewhat unusual, and it may strike people as hubristic, but I claim that this is closer to how I ought to reason if my end goal is to fix something in the real world as opposed to fitting into people's expectations. Research in the real world is messy, and an optimised methodology shouldn't fit neatly into a genre that's evolved for looking professional.

Feel free to give yourself ~5 minutes to come up with your own idea. Try to pay attention to how you approached it, and compare it to what I did. My main project this year is about figuring out how to figure out things, so I'd be very grateful if you tried to describe your approach. Even if you think your process "looks dumb" or somesuch, I think having the courage to reveal your inner workings and talk openly about it reflects some extremely important virtues--especially if you think it's dumb. This is a collaboration, not a contest.


Conceptual research methodology

The Inventor's Paradox is the observation that when you're trying to solve a specific problem, it's often easier to solve a bigger problem that includes the specific problem as a special case.

Why? I don't know yet, but one aspect of it is that when you go up an abstraction level to solve a more general problem, you have fewer objects you need to reason over and therefore a much smaller search tree. E.g. if you're solving a specific problem involving all integers between 1 and 99, it takes more information to specify that set of particular numbers than it takes to just specify "all integers". You're compressing the detail into smaller objects whose essential features now pop out to your mind's eye.[1]

In problem-solving, the hardest and most productive step is often figuring out which parts you can ignore. How can you turn a problem with seemingly many moving parts, into a problem with fewer moving parts?

Anyway, solving the nuclear conflict between Russia and other countries seems like a really hard problem. So many details! Thus, the level of analysis I start out with is not "how can I personally contribute?" nor is it even "how can we resolve the Ukraine-Russia conflict". Instead, I find what I call the "shower-thought level", and ask myself,

"What's a general mechanism for ending all nuclear war"?

When I reframe the problem like this, it forces me to think about the most general features, and allows me to go into problem-solving mode instead of "I have no clue, I need to read more". Zooming in and dealing with the details becomes necessary at some point, but there's an appropriate time for that which is not yet. First, I need to have a high-level inside-view model that lets me have an idea of what specific details to even start looking for. In other words, I'm looking for something I can think productively about while in the shower.

For real-world thinking with enormous search trees, collecting as many ideas that seem like they can make metaphors to other stuff is a good general strategy. When the search space is so large, it's nearly impossible to build a new functional model from scratch. It becomes necessary to look for existing models from other fields (or nature), and iteratively testing them against the new use-case and tweaking them accordingly.[2]

Furthermore, I think the popular advice of "hold off on proposing solutions" is often overapplied. As long as I'm confident in my ability to avoid imprinting on the first idea I find, I often propose several dumb solutions as a way to learn about the problem in the first place. If I don't believe in myself, I might be anxiously attached to the first thing I produce because I doubt I'll be able to find anything else.


"Nuke swapping" to de-escalate nuclear stockpiles

When two competitors are in direct conflict but neither of them wants to waste resources on a fight, I have two metaphors that spring to mind: assurance contracts and vote swapping. I've long thought that assurance contracts are the best thing ever, but I couldn't see a way to make them work in this case (~2 minutes). So I move on to explore the other pattern.

Vote swapping is when two voters for opposing sides in a single-winner election agree to vote for a third candidate instead. If the election is overwhelmingly bipolar, then the two votes would by default directly cancel each other out. But they might both value a third candidate above zero, so vote swapping lets them net a positive sum instead of a zero sum.

This seems sorta kinda like a nuclear standoff, but I can't immediately see how. So I spend a few minutes trying to force the metaphor until something clicks. What I end up with doesn't really fit with vote swapping in the election case. Instead, I find new pattern that I could only arrive at by trying to force the metaphor and deliberately trusting myself in that I can find something that is valuable. (~4 minutes)

In a situation where two sides depend on mutually assured destruction for their safety, neither of them would want to disarm unilaterally. Even if both sides strongly prefer the world were neither side had nuclear weapons, disarming asymmetrically would upset the balance of power, which could in the worst case have the consequence of making a nuclear attack more likely. So what's a mechanism for ensuring symmetric disarmament that doesn't upset the balance?

(On reflection, I think I was mistaken here about this being one of the main bottlenecks for this to work. But while I'm trying to generate something, I endorse not spending too much time trying to critique what I'm coming up with at every point. I want to have something coherent that snaps into place, and that complete pattern may let me see further metaphors that I can draw on to understand the problem. I can more rigorously vet each step after I've already made the model.)

Here the concept of an escrow comes to mind because it's used it variants of assurance contracts, which is a metaphor I previously activated. One side of the conflict places an insignificant quantity of weapons (a 'unit') into the hands of a mutually trusted third party who's responsible for verifying and destroying it. If the unit is small enough, it won't shift the balance of power enough for it to matter, so the cost to the initial 'bid' is insignificant.

But once an insignificant bid has been made, there is pressure on the other party to match it, because it'll be politically unreasonable to refuse such a marginal sacrifice for the global good. This will then hopefwly trigger the opposite of an escalation of conflict--an iterative game of "nuke swapping". Just like how arms races build themselves up because there's momentum in the direction of escalation, the marginal nuke swapping will hopefwly spark momentum for de-escalation.

Now I just need to tell Putin!


After the aha moment

I recognise how important it is to be innocently excited about the ideas I produce, otherwise my brain won't be very motivated to comply next time I ask it to produce something for me. So I feel pretty excited about this idea. Humility has its uses when the stakes are high, but I'm just exploring. And as long as I don't fool myself into thinking I have a finished product, I don't want to constrain my creativity by worrying about whether it's socially appropriate.

Although I think the idea is wrong, I think it's productively wrong in the sense that there are many parts that can be tweaked to look for less wrong ideas nearby. 

I doubt in the applicability of my idea, because if I don't have a good understanding of the problem I'm fitting my solution to, it's unlikely to fit. But I'm still able to feel excited about this because I think it's productively wrong. Models that are wrong for a particular use-case may still still be part of the arsenal I can use against other problems, or I can tweak them until they do fit. And moreover, I can feel excited about having gone through a process which I endorse and would like to see universally adopted.


Where I think it goes wrong

After generation, I move on to analysing and critiquing. Again, I use metaphorical thinking, because now that I have a new pattern, I may be able to see similarities to different models that do slightly different things. Metaphorical thinking lets me ask questions like "why does that model do it like that when my model does it like this?" and it hints me about where I might've gone wrong.

I can quickly see at least two possible mechanistic reasons, but I'm sure there are plenty that I'm not immediately seeing here.

  1. The side that's weaker by conventional weapons does not want to agree to start the chain, because they can get trampled if they cooperate. MAD benefits the weaker side because it ensures stalemate when the weaker side would otherwise overwhelmingly lose a conventional war.
  2. The analogy to elections breaks down because the incentive to swap marginal weapons goes down the more weapons you've already swapped. It becomes less and less like a bipolar conflict because the fewer nukes you have, the more you have to worry about other non-nuclear actors now becoming a threat.  (H/T Ruth)
  1. ^

    Note, when you compress the problem into general abstract categories, you have fewer details to talk about, so you don't have as many opportunities to reveal your expertise by talking about the many details you know. Writing about the finer details will often look more impressive and professional, although imo that doesn't necessarily translate to being more usefwl, and I think it can often bias people to start at the wrong level of abstraction.

  2. ^

    I have a flashcard collection in RemNote of >100 ideas/patterns/tools/perspectives that I feel will be the most generally applicable to other problems. Sometimes when I'm stuck, I flip through this list, and it's been surprisingly effective for me. I call them "germs of generality", taking inspiration from David Hilbert.

Comments2


Sorted by Click to highlight new comments since:

This is great! I'm really glad you wrote this, and I'd be interested in seeing more work like this.

owo thanks!

Curated and popular this week
 ·  · 10m read
 · 
I wrote this to try to explain the key thing going on with AI right now to a broader audience. Feedback welcome. Most people think of AI as a pattern-matching chatbot – good at writing emails, terrible at real thinking. They've missed something huge. In 2024, while many declared AI was reaching a plateau, it was actually entering a new paradigm: learning to reason using reinforcement learning. This approach isn’t limited by data, so could deliver beyond-human capabilities in coding and scientific reasoning within two years. Here's a simple introduction to how it works, and why it's the most important development that most people have missed. The new paradigm: reinforcement learning People sometimes say “chatGPT is just next token prediction on the internet”. But that’s never been quite true. Raw next token prediction produces outputs that are regularly crazy. GPT only became useful with the addition of what’s called “reinforcement learning from human feedback” (RLHF): 1. The model produces outputs 2. Humans rate those outputs for helpfulness 3. The model is adjusted in a way expected to get a higher rating A model that’s under RLHF hasn’t been trained only to predict next tokens, it’s been trained to produce whatever output is most helpful to human raters. Think of the initial large language model (LLM) as containing a foundation of knowledge and concepts. Reinforcement learning is what enables that structure to be turned to a specific end. Now AI companies are using reinforcement learning in a powerful new way – training models to reason step-by-step: 1. Show the model a problem like a math puzzle. 2. Ask it to produce a chain of reasoning to solve the problem (“chain of thought”).[1] 3. If the answer is correct, adjust the model to be more like that (“reinforcement”).[2] 4. Repeat thousands of times. Before 2023 this didn’t seem to work. If each step of reasoning is too unreliable, then the chains quickly go wrong. Without getting close to co
JamesÖz
 ·  · 3m read
 · 
Why it’s important to fill out this consultation The UK Government is currently consulting on allowing insects to be fed to chickens and pigs. This is worrying as the government explicitly says changes would “enable investment in the insect protein sector”. Given the likely sentience of insects (see this summary of recent research), and that median predictions estimate that 3.9 trillion insects will be killed annually by 2030, we think it’s crucial to try to limit this huge source of animal suffering.  Overview * Link to complete the consultation: HERE. You can see the context of the consultation here. * How long it takes to fill it out: 5-10 minutes (5 questions total with only 1 of them requiring a written answer) * Deadline to respond: April 1st 2025 * What else you can do: Share the consultation document far and wide!  * You can use the UK Voters for Animals GPT to help draft your responses. * If you want to hear about other high-impact ways to use your political voice to help animals, sign up for the UK Voters for Animals newsletter. There is an option to be contacted only for very time-sensitive opportunities like this one, which we expect will happen less than 6 times a year. See guidance on submitting in a Google Doc Questions and suggested responses: It is helpful to have a lot of variation between responses. As such, please feel free to add your own reasoning for your responses or, in addition to animal welfare reasons for opposing insects as feed, include non-animal welfare reasons e.g., health implications, concerns about farming intensification, or the climate implications of using insects for feed.    Question 7 on the consultation: Do you agree with allowing poultry processed animal protein in porcine feed?  Suggested response: No (up to you if you want to elaborate further).  We think it’s useful to say no to all questions in the consultation, particularly as changing these rules means that meat producers can make more profit from sel
michel
 ·  · 4m read
 · 
I'm writing this in my personal capacity as someone who recently began working at the Tarbell Center for AI Journalism—my colleagues might see things differently and haven’t reviewed this post.  The rapid development of artificial intelligence could prove to be one of the most consequential technological transitions in human history. As we approach what may be a critical period in AI development, we face urgent questions about governance, safety standards, and the concentration of power in AI development. Yet many in the general public are not aware of the speed of AI development, nor the implications powerful AI models could have for them or society at large. Society’s capacity to carefully examine and publicly debate AI issues lags far behind their importance.  This post makes the case for why journalism on AI is an important yet neglected path to remedy this situation. AI journalism has a lot of potential I see a variety of ways that AI journalism can helpfully steer AI development.  Journalists can scrutinize proposed regulations and safety standards, helping improve policies by informing relevant actors of potential issues or oversights. * Example article: Euractiv’s detailed coverage of the EU AI Act negotiations helped policymakers understand the importance of not excluding general-purpose models from the act, which may have played a key role in shaping the final text.[1] * Example future opportunity: As individual US states draft AI regulations in 2025, careful analysis of proposed bills could help prevent harmful loopholes or unintended consequences before they become law. Journalists can surface (or amplify) current and future risks, providing activation energy for policymakers and other actors to address these risks.  * Example article: 404 Media's investigation revealing an a16z-funded AI platform's generation of images that “could be categorized as child pornography” led a cloud computing provider to terminate its relationship with the platf