Some things I learned while researching air purifiers for my house, to reduce COVID risk during jam nights.
Eight beliefs I have about technical alignment research
Written up quickly; I might publish this as a frontpage post with a bit more effort.
Agency/consequentialism is not a single property.
It bothers me that people still ask the simplistic question "will AGI be agentic and consequentialist by default, or will it be a collection of shallow heuristics?". A consequentialist utility maximizer is just a mind with a bunch of properties that tend to make it capable, incorrigible, and dangerous. These properties can exist independently, and the first AGI probably won't have all of them, so we should be precise about what we mean by "agency". Off the top of my head, here are just some of the qualities included in agency:
See Yudko...
I'm a little skeptical of your contention that all these properties are more-or-less independent. Rather there is a strong feeling that all/most of these properties are downstream of a core of agentic behaviour that is inherent to the notion of true general intelligence. I view the fact that LLMs are not agentic as further evidence that it's a conceptual error to classify them as true general intelligences, not as evidence that ai risk is low. It's a bit like if in the 1800s somebody says flying machines will be dominant weapons of war in the future and get rebutted by 'hot gas balloons are only used for reconnaissance in war, they aren't very lethal. Flying machines won't be a decisive military technology '
I don't know Nate's views exactly but I would imagine he would hold a similar view (do correct me if I'm wrong ). In any case, I imagine you are quite familiar with the my position here.
I'd be curious to hear more about where you're coming from.
I think the framing "alignment research is preparadigmatic" might be heavily misunderstood. The term "preparadigmatic" of course comes from Thomas Kuhn's The Structure of Scientific Revolutions. My reading of this says that a paradigm is basically an approach to solving problems which has been proven to work, and that the correct goal of preparadigmatic research should be to do research generally recognized as impressive.
For example, Kuhn says in chapter 2 that "Paradigms gain their status because they are more successful than their competitors in solving a few problems that the group of practitioners has come to recognize as acute." That is, lots of researchers have different ontologies/approaches, and paradigms are the approaches that solve problems that everyone, including people with different approaches, agrees to be important. This suggests that to the extent alignment is still preparadigmatic, we should try to solve problems recognized as important by, say, people in each of the five clusters of alignment researchers (e.g. Nate Soares, Dan Hendrycks, Paul Christiano, Jan Leike, David Bau).
I think this gets twisted in some popular writings on LessWrong. John Wentworth w...
Possible post on suspicious multidimensional pessimism:
I think MIRI people (specifically Soares and Yudkowsky but probably others too) are more pessimistic than the alignment community average on several different dimensions, both technical and non-technical: morality, civilizational response, takeoff speeds, probability of easy alignment schemes working, and our ability to usefully expand the field of alignment. Some of this is implied by technical models, and MIRI is not more pessimistic in every possible dimension, but it's still awfully suspicious.
I strongly suspect that one of the following is true:
I'm only going to actually write this up if there is demand; the full post will have citations which are kind of annoying to find.
After working at MIRI (loosely advised by Nate Soares) for a while, I now have more nuanced views and also takes on Nate's research taste. It seems kind of annoying to write up so I probably won't do it unless prompted.
Edit: this is now up
Maybe this is too tired a point, but AI safety really needs exercises-- tasks that are interesting, self-contained (not depending on 50 hours of readings), take about 2 hours, have clean solutions, and give people the feel of alignment research.
I found some of the SERI MATS application questions better than Richard Ngo's exercises for this purpose, but there still seems to be significant room for improvement. There is currently nothing smaller than ELK (which takes closer to 50 hours to develop a proposal for and properly think about it) that I can point technically minded people to and feel confident that they'll both be engaged and learn something.
The cost of goods has the same units as the cost of shipping: $/kg. Referencing between them lets you understand how the economy works, e.g. why construction material sourcing and drink bottling has to be local, but oil tankers exist.
It's really impressive that for the price of a winter strawberry, we can ship a strawberry-sized lump of...
I'm worried that "pause all AI development" is like the "defund the police" of the alignment community. I'm not convinced it's net bad because I haven't been following governance-- my current guess is neutral-- but I do see these similarities:
The obvious dis-analogy is that if the police had no funding and largely ceased to exist, a string of horrendous things would quickly occur. Murders and thefts and kidnappings and rapes and more would occur throughout every country in which it was occurring, people would revert to tight-knit groups who had weapons to defend themselves, a lot of basic infrastructure would probably break down (e.g. would Amazon be able to pivot to get their drivers armed guards?) and much more chaos would ensue.
And if AI research paused, society would continue to basically function as it has been doing so far.
One of them seems to me like a goal that directly causes catastrophes and a breakdown of society and the other doesn't.
There are less costly, more effective steps to reduce the underlying problem, like making the field of alignment 10x larger or passing regulation to require evals
IMO making the field of alignment 10x larger or evals do not solve a big part of the problem, while indefinitely pausing AI development would. I agree it's much harder, but I think it's good to at least try, as long as it doesn't terribly hurt less ambitious efforts (which I think it doesn't).
Say I need to publish an anonymous essay. If it's long enough, people could plausibly deduce my authorship based on the writing style; this is called stylometry. The only stylometry-defeating tool I can find is Anonymouth; it hasn't been updated in 7 years and it's unclear if it can defeat modern AI. Is there something better?
Tech tree for worst-case/HRAD alignment
Here's a diagram of what it would take to solve alignment in the hardest worlds, where something like MIRI's HRAD agenda is needed. I made this months ago with Thomas Larsen and never got around to posting it (mostly because under my worldview it's pretty unlikely that we need to do this), and it probably won't become a longform at this point. I have not thought about this enough to be highly confident in anything.
I was going to write an April Fool's Day post in the style of "On the Impossibility of Supersized Machines", perhaps titled "On the Impossibility of Operating Supersized Machines", to poke fun at bad arguments that alignment is difficult. I didn't do this partly because I thought it would get downvotes. Maybe this reflects poorly on LW?
A toy model of intelligence implies that there's an intelligence threshold above which minds don't get stuck when they try to solve arbitrarily long/difficult problems, and below which they do get stuck. I might not write this up otherwise due to limited relevance, so here it is as a shortform, without the proofs, limitations, and discussion.
A task of difficulty n is composed of independent and serial subtasks. For each subtask, a mind of cognitive power knows different “approaches” to choose from. The time taken by each approach is at least 1 but drawn from a power law, for , and the mind always chooses the fastest approach it knows. So the time taken on a subtask is the minimum of samples from the power law, and the overall time for a task is the total for the n subtasks.
Main question: For a mind of strength ,
I'm looking for AI safety projects with people with some amount of experience. I have 3/4 of a CS degree from Caltech, one year at MIRI, and have finished the WMLB and ARENA bootcamps. I'm most excited about making activation engineering more rigorous, but willing to do anything that builds research and engineering skill.
If you've published 2 papers in top ML conferences or have a PhD in something CS related, and are interested in working with me, send me a DM.
I had a long-ish conversation with John Wentworth and want to make it known that I could probably write up any of the following distillations if I invested lots of time into them (about a day (edit: 3 days seems more likely) of my time and an hour of John's). Reply if you're really interested in one of them.
The LessWrong Review's short review period is a fatal flaw.
I would spend WAY more effort on the LW review if the review period were much longer. It has happened about 10 times in the last year that I was really inspired to write a review for some post, but it wasn’t review season. This happens when I have just thought about a post a lot for work or some other reason, and the review quality is much higher because I can directly observe how the post has shaped my thoughts. Now I’m busy with MATS and just don’t have a lot of time, and don’t even remember what posts I wanted to review.
I could have just saved my work somewhere and paste it in when review season rolls around, but there really should not be that much friction in the process. The 2022 review period should be at least 6 months, including the entire second half of 2023, and posts from the first half of 2022 should maybe even be reviewable in the first half of 2023.
Below is a list of powerful optimizers ranked on properties, as part of a brainstorm on whether there's a simple core of consequentialism that excludes corrigibility. I think that AlphaZero is a moderately strong argument that there is a simple core of consequentialism which includes inner search.
Properties
Has anyone made an alignment tech tree where they sketch out many current research directions, what concrete achievements could result from them, and what combinations of these are necessary to solve various alignment subproblems? Evan Hubinger made this, but that's just for interpretability and therefore excludes various engineering achievements and basic science in other areas, like control, value learning, agent foundations, Stuart Armstrong's work, etc.
Suppose that humans invent nanobots that can only eat feldspars (41% of the earth's continental crust). The nanobots:
Does this cause human extinction? If so, by what mechanism?
Antifreeze proteins prevent water inside organisms from freezing, allowing them to survive at temperatures below 0 °C. They do this by actually binding to tiny ice crystals and preventing them from growing further, basically keeping the water in a supercooled state. I think this is fascinating.
Is it possible for there to be nanomachine enzymes (not made of proteins, because they would denature) that bind to tiny gas bubbles in solution and prevent water from boiling above 100 °C?
Is there a well-defined impact measure to use that's in between counterfactual value and Shapley value, to use when others' actions are partially correlated with yours?
I'm planning to write a post called "Heavy-tailed error implies hackable proxy". The idea is that when you care about and are optimizing for a proxy , Goodhart's Law sometimes implies that optimizing hard enough for causes to stop increasing.
A large part of the post would be proofs about what the distributions of and must be for , where X and V are independent random variables with mean zero. It's clear that
The most efficient form of practice is generally to address one's weaknesses. Why, then, don't chess/Go players train by playing against engines optimized for this? I can imagine three types of engines:
The first tool would simply be an opponent when humans are inconvenient or not available. The second and third tools wo
...I looked at Tetlock's Existential Risk Persuasion Tournament results, and noticed some oddities. The headline result is of course "median superforecaster gave a 0.38% risk of extinction due to AI by 2100, while the median AI domain expert gave a 3.9% risk of extinction." But all the forecasters seem to have huge disagreements from my worldview on a few questions:
Question for @AnnaSalamon and maybe others. What's the folk ethics analysis behind the sinking of the SF Hydro, which killed 14 civilians but destroyed heavy water to be used in the Nazi nuclear weapons program? Eliezer used this as a classic example of ethical injunctions once.
People say it's important to demonstrate alignment problems like goal misgeneralization. But now, OpenAI, Deepmind, and Anthropic have all had leaders sign the CAIS statement on extinction risk and are doing substantial alignment research. The gap between the 90th percentile alignment concerned people at labs and the MIRI worldview is now more about security mindset. Security mindset is present in cybersecurity because it is useful in the everyday, practical environment researchers work in. So perhaps a large part of the future hinges on whether security m...
The author of "Where Is My Flying Car" says that the Feynman Program (teching up to nanotechnology by machining miniaturized parts, which are assembled into the tools for micro-scale machining, which are assembled into tools for yet smaller machining, etc) might be technically feasible and the only reason we don't have it is that no one's tried it yet. But this seems a bit crazy for the following reasons:
In nanotech? True enough, because I am not convinced that there is any domain expertise in the sort of nanotech Storrs Hall writes about. It seems like a field that consists mostly of advertising. (There is genuine science and genuine engineering in nano-stuff; for instance, MEMS really is a thing. But the sort of "let's build teeny-tiny mechanical devices, designed and built at the molecular level, which will be able to do amazing things previously-existing tech can't" that Storrs Hall has advocated seems not to have panned out.)
But more generally, that isn't so at all. What I'm looking for by way of domain expertise in a technological field is a history of demonstrated technological achievements. Storrs Hall has one such achievement that I can see, and even that is doubtful. (He founded and was "chief scientist" of a company that made software for simulating molecular dynamics. I am not in a position to tell either how well the software actually worked or how much of it was JSH's doing.) More generally, I want to see a history of demonstrated difficult accomplishments in the field, as opposed to merely writing about the field.
Selecting some random books from my shelves (literally...
Is it possible to make an hourglass that measures different amounts of time in one direction than the other? Say, 25 minutes right-side up, and 5 minutes upside down, for pomodoros. Moving parts are okay (flaps that close by gravity or something) but it should not take additional effort to flip.
I don't see why this wouldn't be possible? It seems pretty straightforward to me; the only hard part would be the thing that seems hard about making any hourglass, which is getting it to take the right amount of time, but that's a problem hourglass manufacturers have already solved. It's just a valve that doesn't close all the way:
Unless you meant, "how can I make such an hourglass myself, out of things I have at home?" in which case, idk bro.
Given that social science research often doesn't replicate, is there a good way to search a social science finding or paper and see if it's valid?
Ideally, one would be able to type in e.g. "growth mindset" or a link to Dweck's original research, and see:
An idea for removing knowledge from models
Suppose we have a model with parameters , and we want to destroy a capability-- doing well on loss function -- so completely that fine-tuning can't recover it. Fine-tuning would use gradients , so what if we fine-tune the model and do gradient descent on the norm of the gradients during fine-tuning, or its directional derivative where ? Then maybe if we add the accumulated parameter vector, the new copy of the model wo...
We might want to keep our AI from learning a certain fact about the world, like particular cognitive biases humans have that could be used for manipulation. But a sufficiently intelligent agent might discover this fact despite our best efforts. Is it possible to find out when it does this through monitoring, and trigger some circuit breaker?
Evals can measure the agent's propensity for catastrophic behavior, and mechanistic anomaly detection hopes to do better by looking at the agent's internals without assuming interpretability, but if we can measure the a...
Somewhat related to this post and this post:
Coherence implies mutual information between actions. That is, to be coherent, your actions can't be independent. This is true under several different definitions of coherence, and can be seen in the following circumstances:
Many people think that AI alignment is intractable (<50% chance of success) and also believe that a universe optimized towards elephant CEV, or the CEV of aliens that had a similar evolutionary environment to humans, would be at least 50% as good as a universe optimized towards human CEV. Doesn't this mean we should be spending significant effort (say, at least 1% of the effort spent on alignment) finding tractable plans to create a successor species in case alignment fails?
Are there ring species where the first and last populations actually can interbreed? What evolutionary process could feasibly create one?
2.5 million jobs were created in May 2020, according to the jobs report. Metaculus was something like [99.5% or 99.7% confident](https://www.metaculus.com/questions/4184/what-will-the-may-2020-us-nonfarm-payrolls-figure-be/) that the number would be smaller, with the median at -11.0 and 99th percentile at -2.8. This seems like an obvious sign Metaculus is miscalibrated, but we have to consider both tails, making this merely a 1 in 100 or 1 in 150 event, which doesn't seem too bad.
Current posts in the pipeline:
I'm reading a series of blogposts called Extropia's Children about the Extropians mailing list and its effects on the rationalist and EA communities. It seems quite good although a bit negative at times.
Eliezer Yudkowsky wrote in 2016:
At an early singularity summit, Jürgen Schmidhuber, who did some of the pioneering work on self-modifying agents that preserve their own utility functions with his Gödel machine, also solved the friendly AI problem. Yes, he came up with the one true utility function that is all you need to program into AGIs!
(For God’s sake, don’t try doing this yourselves. Everyone does it. They all come up with different utility functions. It’s always horrible.)
...His one true utility function was “increasing the compression of environ
What was the equation for research progress referenced in Ars Longa, Vita Brevis?
...“Then we will talk this over, though rightfully it should be an equation. The first term is the speed at which a student can absorb already-discovered architectural knowledge. The second term is the speed at which a master can discover new knowledge. The third term represents the degree to which one must already be on the frontier of knowledge to make new discoveries; at zero, everyone discovers equally regardless of what they already know; at one, one must have mastered every
Showerthought: what's the simplest way to tell that the human body is less than 50% efficient at converting chemical energy to mechanical work via running? I think it's that running uphill makes you warmer than running downhill at the same speed.
When running up a hill at mechanical power p and efficiency f, you have to exert p/f total power and so p(1/f - 1) is dissipated as heat. When running down the hill you convert p to heat. p(1/f - 1) > p implies that f > 0.5.
Maybe this story is wrong somehow. I'm pretty sure your body has no way of recovering your potential energy on the way down; I'd expect most of the waste heat to go in your joints and muscles but maybe some of it goes into your shoes.
Are there approximate versions of the selection theorems? I haven't seen anyone talk about them, but they might be easy to prove.
Approximate version of Kelly criteron: any agent that follows a strategy different by at least epsilon from Kelly betting will almost surely lose money compared to a Kelly-betting agent at a rate f(epsilon)
Approximate version of VNM: Any agent that satisfies some weakened version of the VNM axioms will have high likelihood under Boltzmann rationality (or some other metric of approximate utility maximization). The closest thing I'...
Is there somewhere I can find a graph of the number of AI alignment researchers vs AI capabilities researchers over time, from say 2005 to the present day?
Is there software that would let me automatically switch between microphones on my computer when I put on my headset?
I imagine this might work as a piece of software that integrates all microphones connected to my computer into a single input device, then transmits the audio stream from the best-quality source.
A partial solution would be something that automatically switches to the headset microphone when I switch to the headset speakers.