Made a quick neural network (reaching about 70% accuracy), and checked all available scores.
Its favorite result was: +2 Cha, +8 Wis. It would have like +10 Wis if it were possible.
For at least the top few results, it wanted to (a) apportion as much to Wis as possible, then (b) as much to Cha, then (c) as much to Con. So we have for (Wis, Cha, Con): 1. (8, 2, 0) 2. (8, 1, 1) 3. (8, 0, 2) 4. (7, 3, 0) 5. (7, 2, 1) …
Any favored resources on metta?
If we solve the problem normally thought of as “misalignment”, it seems like this scenario would now go well. If we solve the problem normally thought of as “misuse”, it seems like this scenario would now go well. This argues for continuing to use these categories, as fixing the problems they name is still sufficient to solve problems that do not cleanly fit in one bucket or another
Sure!
I see people on twitter, for example, doing things like having GPT-3 provide autocomplete or suggestions while they’re writing, or doing grunt work of producing web apps. Plausibly, figuring out how to get the most value out of future AI developments for improving productivity is important.
There’s an issue that it’s not very obvious exactly how to prepare for various AI tools in the future. One piece of work could be thinking more about how to flexibly prepare for AI tools with unknown capabilities, or predicting what the capabilities will be.
Other things that come to mind are:
Practice getting up to speed in new tool setups. If you are very bound to a setup that you like, you might have a hard time leveraging these advances as they come along. Alternatively, try and be sure you can extend your current workflow
Increase the attention you pay to new (AI) tools. Get used to trying them out, both for the reasons above and because it may be important to act fast in picking up very helpful new tools
To be clear, it’s not super clear to me how much value there is in this direction. It is pretty plausible to me that AI tooling will be essential for competitive future productivity, but maybe there’s not much of an opportunity to bet on that
Now, it’s still possible that accumulation of slow-turnover senescent cells could cause the increased production rate of fast-turnover senescent cells.
Reminds me of this paper, in which they replaced the blood of old rats with a neutral solution (not the blood of young rats), and found large rejuvinative effects. IIRC, they attributed it to knocking the old rats out of some sort of “senescent equilibrium”
If timelines are short, where does the remaining value live? Some fairly Babble-ish ideas:
Alignment-by-default
Both outer alignment and inner by default
With full alignment by default, there’s nothing to do, I think! One could be an accelerationist, but the reduction in suffering and lives lost now doesn’t seem large enough for the cost in probability of aligned AI
Possibly value could be lost if values aren’t sufficiently cosmopolitan? One could try and promote cosmopolitan values
Inner alignment by default
Focus on tools for getting good estimates of human values, or an intent-aligned AI
Ought’s work is a good example
Possibly trying to experiment with governance / elicitation structures, like quadratic voting
Also thinking about how to get good governance structures actually used
Acausal trade
In particular, expand the ideas in this post. (I understand Paul to be claiming he argues for tractability somewhere in that post, but I couldn’t find it)
Work through the details of UDT games, and how we could effect proper acausal trade. Figure out how to get the relevant decision makers on board
Strong, fairly late institutional responses
Work on making, for example, states strong enough to (coordinately) restrict or stop AI development
Other things that seem useful:
Learn the current hot topics in ML. If timelines are short, it’s probably the case that AGI will use extensions of the current frontier
Invest in leveraging AI tools for direct work / getting those things that money cannot buy. This maybe a little early, but if the takeoff is at all soft, maybe there are still >10 years left of 2020-level intellectual work before 2030 if you’re using the right tools
Even more so, I would love to see your unjustifiable stab-in-the-dark intuitions as to where the center of all this is
Curious why this in particular (not trying to take umbrage with wanting this info; I agree that there’s a lot of useful data here. Would be a thing I’d also want to ask for, but wouldn’t have prioritised)
Seems like you’re missing an end to the paragraph that starts “Related argument”
I liked your example of being uncertain of your probabilities. I note that if you are trying to make an even money bet with a friend (as this is a simple Schelling point), you should never Kelly bet if you have discounted rate of 2⁄3 or less of your naïve probabilities.
The maximum bet for is when is 1, which is which crosses below 0 at
In the pie chart in the Teams section, you can see “CooperateBot [Larks]” and “CooperateBot [Insub]”
Yeah, that’s what my parenthetical was supposed to address
(particularly because there is large interpersonal variation in the strength of hedging a given qualifier is supposed to convey)
Perhaps you are able to get more reliable information out of such statements than I am.
I like qualifiers that give information on the person’s epistemic state or, even better, process. For example:
“Technological progress happens exponentially (70%)”
“My rough intuition is that technological progress happens exponentially”
“Having spent a year looking into the history of technological progress, I concluded it happens exponentially: [here are specific data sources, or links, or some sort of pointer]”
Given that I don’t start thinking that anyone can report directly the state of the world (rather than their beliefs and understanding of it), “From what I understand, technological progress happens exponentially” does not provide much information (particularly because there is large interpersonal variation in the strength of hedging a given qualifier is supposed to convey).
Sometimes I feel forced to add qualifiers because it makes it more likely that someone will keep engaging with me. That is, I am confused about something, and they are helping refine my model. By adding qualifiers, I signal that I’m reasonable / understand I should be uncertain of my conclusions / am open to updating.
You can steer a bit away from catastrophe today. Tomorrow you will be able to do less. After years and decades go by, you will have to be miraculously lucky or good to do something that helps. At some point, it’s not the kind of “miraculous” you hope for, it’s the kind you don’t bother to model.
Today you are blind, and are trying to shape outcomes you can’t see. Tomorrow you will know more, and be able to do more. After years and decades, you might know enough about the task you are trying to accomplish to really help. Hopefully the task you find yourself faced with is the kind you can solve in time.
I think you’re right. I think inline comments are a good personal workflow when engaging with a post (when there’s a post I want to properly understand, I copy it to a google doc and comment it), but not for communicating the engagement
My understanding is the first thing is what you get with UDASSA and the second thing would be what you get is if you think the Solomonoff prior is useful for predicting your universe for some other reason (ie not because you think the likelihood of finding yourself in some situation covaries with the Solomonoff prior’s weight on that situation)
It is at least the case that OpenAI has sponsored H1Bs before: https://www.myvisajobs.com/Visa-Sponsor/Openai/1304955.htm
He treated it like a game, even though he was given the ability to destroy a non-trivial communal resource.
I want to resist this a bit. I actually got more value out of the “blown up” frontpage then I would have from a normal frontpage that day. A bit of “cool to see what the LW devs prepared for this case”, a bit of “cool, something changed!” and some excitement about learning something.
That’s a very broad definition of ‘long haul’ on duration and on severity, and I’m guessing this is a large underestimate of the number of actual cases in the United Kingdom
If the definition is broad, shouldn’t it be an overestimate?
Should I be reading all the openings of transparent envelopes as actual openings, or are they sometimes looking at the sealed envelope and seeing what it contains (the burnings incline me towards the second interpretation, but I’m not sure)?
EDIT: Oh, I think I understand better now