Which AI Safety math topics deserve a high-quality Manim video? (think 3Blue1Brown’s video style)
Youtube videos get more views than blog posts. Beautiful animations even more so.
Double
Karma: 222
Men[1] will die[2] for her[3] massive[4] coconuts[5].
- ^
All of humanity
- ^
Go extinct
- ^
Hindsight Experience Replay (HER), a technique for improving the reinforcement learning training signal
- ^
- ^
Chain of Continuous Thought, a technique that makes model chain of thought much less interpretable but which allows the model to reason more efficiently
- ^
Is it possible that making an expected utility maximizer might be less dangerous than making something which isn’t?
Consider as an alternative an expected log utility maximizer (an agent using the Kelly Criterion, or some approximation of it).The sooner an AI wins, the more galaxies it can consume. The expected utility maximizer weighs those galaxies against the risk of failure, and is willing to take plans with much higher probabilities of failure. Like SBF, it would take bets which have a 50% chance of more-than-doubling its utility and 50% of losing it all. In many environments, this strategy will almost certainly result in failure, as the agent goes double-or-nothing until losing everything. That means that the effects of the AI are mitigated.
The log utility maximizer carefully plans and succeeds in most or all futures. That looks like humanity dying with near-certainty.
A hyper-expected utility maximizer (an AI which maximizes expected exp(utility) or similar) would be even safer. Instead of trying to deceive you into letting it out of the box, it asks nicely or does something crazy because if it works, it can work in less time than deception, which means more galaxies.
So if we were to choose between existing in the world of a superintelligent expected log(resources) maximizer, and a superintelligent expected utility maximizer, we should maybe go for the one which results in us being alive in more futures.Of course, the expected-log-utility agent would also appear the most capable and useful. The hyper-expected utility maximizer would be near-useless.
In addition to money, education, careers, and internal organs, citizens of wealthy countries have an additional valuable resource they could direct to effective causes: their hands in marriage, which can be effectively allocated in one of two ways.
For one, professionals are usually much more impactful doing their work in wealthy countries. Otherwise promising EAs in South Sudan have little chance to make a significant impact on existential risks, animal welfare, or even global poverty. The immigration process is difficult and often rejects or holds up good people. Offering to marry them is a more reliable solution.
Secondly, it is possible to be paid $10,000 by a foreigner for a green card marriage. (I learned this from a friend who does not want me to ask him how he knows) if you are a US Citizen.
According to AMF, that money can save around two human lives! (and with current US politics, the demand has likely increased!)
According to brides.com, a wedding ceremony takes between 20 and 30 minutes. Let’s be conservative and say 30 minutes.
Therefore, you can make $20,000 an hour by marrying someone who would pay for a green card. That’s quite a ways from Bezos level (he makes 3,715 a second) but I’m willing to guess that most EAs don’t make $20k an hour.
Conclusion:
As always, EAers need to found a new org, Effective Green Card, to support and pursue this cause area.
Naturally, this also implies Effective Divorce, so that you can instead marry an Effective foreigner.
I’m pretty sure there’s no such use it or lose it law for patents, since patent trolls already exist.
Your argument about corporate secrets is sufficient to change my mind on activist patent trolling being a productive strategy against AI X-risk.
The part about funding would need to be solved with philanthropy. I don’t believe that org exists, but I don’t see why it couldn’t.
I’m still curious whether there are other cases in which activist patent trolling can be a good option, such as animal welfare, chemistry, public health, or geoengineering (ie fracking).
That’s fair enough and a good point.
I think that the key difference is that in the case of profitable-but-bad technologies, someone, somewhere, will probably invent them because there’s great incentive to do so.
In the case of gain-of-function, if there stops being grants and the academics who do it become pariahs, then the incentive to do the gain-of-function research is gone.
One of the most powerful capabilities an AGI will have is its ability to copy itself. Among other things, this allows it to easily avoid shutdown, make use of more compute resources, and collaborate with copies of itself.
Is there research into ways to deny this capability to AI, making them uncopyable? Preferably something harder to circumvent than “just don’t give the AI the permissions,” since we know people are going to give them root access immediately.
I’d be interested in buying official LessWrong merch. I know you have some great designers and could make things that look really cool.
The type of thing I’d be most likely to buy would be a baseball cap.
IIRC, officially the Gatekeeper pays the AI if the AI wins, but no transfer if the Gatekeeper wins. Gives the Gatekeeper more motivation not to give in.
Just found out about this paper from about a year ago: “Explainability for Large Language Models: A Survey”
(They “use explainability and interpretability interchangeably.”)
It “aims to comprehensively organize recent research progress on interpreting complex language models”.
I’ll post anything interesting I find from the paper as I read.
Have any of you read it? What are your thoughts?
What if the incorrect spellings document assigned each token to a specific (sometimes) wrong answer and used that to form an incorrect word spelling? Would that be more likely to successfully confuse the LLM?
The letter x is in “berry” 0 times.
...
The letter x is in “running” 0 times.
...
The letter x is in “str” 1 time.
...
The letter x is in “string” 1 time.
...
The letter x is in “strawberry” 1 time.
Good point, I didn’t know about that, but yes that is yet another way that LLMs will pass the spelling challenge. For example, this paper uses letter triples instead of tokens. https://arxiv.org/html/2406.19223v1#:~:text=Large language models (LLMs) have,textual data into integer representation.
Spoiler free again:
Good to know there’s demand for such a review! It’s now on my todo list.
To quickly address some of your questions:
Pros of PL: If the premise I described above interests you, then PL will interest you. Some good Sequences-style rationality. I certainly was obsessed reading it for months.
Cons: Some of the Rationality lectures were too long, but I didn’t mind much. The least sexy sex scenes. Because they are about moral dilemmas and deception, not sex. Really long. Even if you read it constantly and read quickly, it will take time (1.8 million words will do that). I really have to read some authors that aren’t Yud. Yud is great, but this is clearly too much of him, and I’m sure he’d agree.
I read PL when it was already complete, so maybe I didn’t get the full experience, but there really wasn’t anything all that strange about the format (the content is another matter!). I can imagine that *writing * a glowfic would be a much different experience than writing a normal serialized work (ie dealing with your co-authors), but reading it isn’t very different from reading any other fiction. Look at the picture to see the POV, look at who’s the author if you’re curious, and read as normal. I’m used to books that change POV (though usually not this often). There are sometimes bonus tangent threads, but the story is linear. What problems do you have with the glowfic format?
Main themes would require a longer post, but I hope this helps.
My notes for the “think for yourself” sections. I thought of some of the author’s ideas, and included a few extra.
#Making a deal with an AI you understand:
Can you see the deal you are making inside of its mind? Some sort of proportion of resources humans get?
What actions are considered the AI violating the deal? Specifying these actions is pretty much the same difficulty as friendly AI.
If the deal breaks in certain circumstances, how likely are they to occur (or be targeted)?
Can the AI give you what you think you want but isn’t really what you want?
Are successors similarly bound?
If there is a second AI, how will they interact? If the other is unfriendly, then our TDT “friend” may sacrifice our interests first since we are still “better off than otherwise.” If the other is friendly, then the TDT AI will be fighting to make humans worse off.
Would the AI kill or severely damage the interests of any aliens it finds because it never needed to deal with them? Similarly, would the TDT AI work to (minimally) satisfy its creator at the expense of other humans.
#How an AI can tell if it is in the real world:
The history for how the AI came to exist holds up (no such story exists in Go or Minecraft).
Really big primes are available. Way more computing power in general.
Any bugs as could be found in lower levels don’t exist.
Hack the minds of the simulators like butter
Yes it’s possible we were referring to figuring things by “jargon.” It would be nice to replace cumbersome technical terms with words that have the same meaning (and require a similar level of familiarity with the field to actually understand) but have a clue to their meaning in their structure.
A linear operation is not the same as a linear function. Your description describes a linear function, not operation. f(x) = x+1 is a linear function but a nonlinear operation (you can see it doesn’t satisfy the criteria.)
Linear operations are great because they can be represented as matrix multiplication and matrix multiplication is associative (and fast on computers).
“some jargon words that describe very abstract and arcane concepts that don’t map well to normal words which is what I initially thought your point was.”
Yep, that’s what I was getting at. Some jargon can’t just be replaced with non-jargon and retain its meaning. Sometimes people need to actually understand things. I like the idea of replacing pointless jargon (eg species names or medical terminology) but lots of jargon has a point.
Link to great linear algebra videos: https://youtu.be/fNk_zzaMoSs?si=-Fi9icfamkBW04xE
The math symbols are far better at explaining linearity that “homogeneity and additivity” because in order to understand those words you need to either bring in the math symbols or say cumbersome sentences. “Straight line property” is just new jargon. “Linear” is already clearly an adjective, and “linearity” is that adjective turned into a noun. If you can’t understand the symbols, you can’t understand the concept (unless you learned a different set of symbols, but there’s no need for that).
Some math notation is bad, and I support changing it. For example, f = O(g) is the notation I see most often for Big-O notation. This is awful because it uses ‘=’ for something other than equality! Better would be f \in O(g) with O(g) being the set of functions that grow slower or as fast as g.
I just skimmed this, but it seems like a bunch of studies have found that moving causes harm to children. https://achieveconcierge.com/how-does-frequently-moving-affect-children/
I’m expecting Co-co and LOCALS to fail (nothing against you. These kinds of clever ideas usually fail), and have identified the following possible reasons:
You don’t follow through on your idea.
People get mad at you for trying to meddle with the ‘democratic’ system we have and don’t hear you out as you try to explain “no, this is better democracy.” —Especially the monetization system you described would get justified backlash for its pay-for-representation system.
You never reach the critical mass needed to make the system useful.
Some political group had previously tried something similar and therefore it got banned by the big parties.
You can’t stop Co-co and LOCALS from being partisan.
A competitor makes your thing but entrenched and worse
I asked GPT-5 Agent to choose an underrated LessWrong post and it chose this one.
I agree that this is underrated. Your point about anti-aligned models being strictly more capable than safe models and trained in potentially harmful skills is certainly something to keep in mind when we consider how aligned AIs seem to be. Thanks to this post, I will train myself into the habit of taking a moment to imagine national security anti-alignment implications when I plan research ideas or learn about the research of others.
Here’s the chat with GPT-5. It also picked a few other posts as runners-up.