That’s an interesting post, thanks. But I think you can bring value with your creations even if you’re neither pareto optimal nor one-dimension top performer. So if you write scifi stories, you might still write some interesting stuff and deliver inspiring idea even if you don’t have unique mixture of non-usual-sci-fi-writer traits in any tangible sense. You’re not the best skateboard-scifi writer. You’re just an average scifi writer and it’s fine. So my point is in this case you might still need motivation. Maybe to develop your unique style or hit the niche market in the future. But if from the very beginning you compare yourself to superhumans boosted to the right side of the distribution by their unique genome or other irreproducible factors, you might never even start.
Jan Czechowski
[Question] Correcting human error vs doing exactly what you’re told—is there literature on this in context of general system design?
Coping with being average: local group. If you want to produce some content like writing, you’re very likely to take popular writers as a baseline. But in times of the Internet, those are super-high-performers from the far end of the distribution. You and I are just not as good them (I’m 99% sure the reader is not in top 1%).
This might make you lose motivation. You will produce less then your reference point. Your content will feel medicore. You will get orders of magnitude smaller audience.
Idea how to cope better: instead of taking the whole internet, take a random sample of 100 people from your reference group and select top achievers (in the desired area) from this 100. They will probably be only slightly better than you.
A simple mechanism partially emulating this random sampling is your local group.
Comparing yourself to people who are closer to you on performance axis will probably be more productive. They might have struggled with similar problems as you only recently. They might have some tricks that are applicable in your case. You have higher chance of reaching their level 2 years from now.
In other words, if you want to be a better writer, aspire to be like the best writer in your local EA group, not like EY or Scott Alexander.
Hm, I was also thinking of moral value of children in this context. At least in my perception, important part of the moral value is the potential to become a conscious, self-aware being. In what sense does this potential translate to artificially created beings?
Maybe if in neural network parameter space there’s a subspace of minds with moral value, also points close to this subspace would have moral value?
I think the idea with internal activations manipulation is interesting. It might require some refinement—I think activations of encoder-decoder transformer model are a function of inputs, and they change with every token. At first, the input is your prompt, then it’s your prompt + generated tokens. So the protocol / task for GPT3 would be: generate now 5 tokens, so with the last generation this logit is maximized? Also, it depends on hyperparameters of beam search which are controlled by human
I just logged in there, thanks—however they have a couple of disclaimers that this is not a place for beginner / technical support questions. But I suppose it’s more to avoid people asking “how do I install pytorch?” rather than AI safety questions, so maybe it will suit the needs of OP
There’s an AI safety camp slack with #no-stupid-questions channel. I think people stay there even after the camp ends (I’m still there although this year edition ended last week). So you can either apply for next years edition (which I very much recommend!) or maybe contact organizers if they can add you without you being AISC participant/alumni? Just a disclaimer, I’m not sure how active this slack is between camps, and it might be that lot of people leave after the camp ends.
Thought experiment: Can question answering language model without memory, instantiated separately for every session (like GPT-3) be pursuing a goal? Does such setup exclude agency? ----Imagine you put thousand super-geniuses in prison, each in separate cell. You will call them in random order to interrogation room and always ask only one question, listen to the answer and send them back. Each person will be called only once.----- Super-geniuses are allowed to devise a strategy before you lock them. Their goal is to manipulate you to release them. Will they manage? If they are intelligent enough, the fact that each one answers only one question without knowing how many others were interrogated before and what their answers were shouldn’t be a unworkable problem.
Steganography and the CycleGAN—alignment failure case study
Current president of Poland went from being relatively unknown parliament member (I’d bet, less than 5% of general public knew who he was? Also there’s another public figure with the same last name, so maybe even some people who knew the name were confusing him with the other guy) to being the president within one year. So with longer timescale, big support from one of leading political powers, and some political backlog it is possible. There are surely similar stories around the world, of outsiders rising to leaders. Of course what you’re asking is much more extreme, but that provides some staring datapoint.
I’m not really sure if food shakes are the cheapest nutrition possible. They are optimized for time and convenience. Yeah they are probably cheaper than eating out, but I’m not sure how low you can get with cooking yourself from basic products, theoretically scaling up for a larger group of people. I guess thinks like armies and monasteries might have it figured out. Maybe interesting to check what’s the average price of feeding a soldier / monk?
As usual, the deadweight loss is quadratic in the size of the tax.
Can you give some clarifications for this concept? I’m not sure what you mean here.
Thanks for this piece of advice—I actually invested in custom earplugs, it cost me around 60eur and 2h of visits and commute. They are not really perfectly comfortable in the night, but they help in the morning, when my Wife gets up earlier than me, she doesn’t wake me up anymore.
Well, for me ‘logging in to work after 10AM’ seems much worse than ‘starting to get ready to sleep after 10PM’. It’s a very natural and strong Schelling fence. Therefore the idea to play in the mornings rather than nights.
for ripgrep. This is my standard code analysis and config searching tool, I use it multiple times a day. Always feels like papercut when I work in a server with only grep
Free course review — Reliable and Interpretable Artificial Intelligence (ETH Zurich)
Did you consider looking at it rather from “options” than “goals” perspective? Rather than defining goals and looking for the optimal path to get there, you can look at /brainstorm exploitable options that you have available and seem to have high returns. And then prioritize them. I recently spent half a day writing down cool ideas for things to do, then collected them in todoist, and since then, whenever I have time I go through them. And add something new.
Recent spreadsheet situations:
I had a free day and I didn’t have much inspiration on how to spend it. So I decided to sit down and rethink my goals, habits. This made me realize that I have a time-tracking record collected from the last 8 months, so it’s a good moment to put them in a spreadsheet and analyze
I was organizing a birthday party and wanted to invite various groups of friends. I didn’t want the party to be too big, but I wanted to know when I still have space to invite something more. So I created a spreadsheet, grouping people in columns with the probability that they will actually show up (got an invitation but no answer: 30%, confirmed no: 0%, confirmed yes: 90%).
I was organizing a weekend stag do for a friend, with ~10 participants. There were various activities and costs involved, and different people paid for different things. So I asked everyone to put the costs in a spreadsheet and then made a detailed calculation about who put how much money, who should carry costs for what (not everyone used both nights of accommodation for example), calculated a balance for everyone involved.
I notice that with regards to many things I always think of at least one of the following aspects:
Money
Time
Risk
As each of those is quantifiable, it prompts me to actually put some numbers on the given problem.
What can EA learn from Scouting
epistemic status: random notes for writing something more structured in the future
After EA fUnconference in Berlin, I’m thinking more and more about what we can learn from the scouting movement for purpose of community building/cooperation. Some elements characteristic of scouting (note, it’s a rough idea, I don’t think all of them are worth reproducing):
EDIT: to clarify (as I’m afraid now in this chaotic draft this may be misunderstood) this is not a list of things I observed on fUnconference. There there was only a partial implementation of the ‘small groups’ idea, as indicated in the text. Everything else is my memories from scouting times, and I never observed them in any EA group.
Arbitrary splitting larger groups into smaller troops (up to 10 people) with an appointed leader. On fUnconference this was implemented as “Huddle” groups, they were used only for daily checkups (how is everyone doing, does everybody feel welcome and is having fun) and once as (soft) default split for team-based games.
Assigning lots of meaning to color, name, and symbols related to your small group / larger community
Incorporating music both for general singing together but also specific songs for specific occasions (morning, before a meal, evening)
Also, spread and evolution of songs—various smaller communities will sing similar things but will develop slight differences in details in time. This gives you both sense of belonging to the whole movement (because you know the same songs) and your local group (because you sing them in a very specific way)
Rituals/traditions related to gathering around the campfire—leader of the gathering, a special role for the person sustaining the fire, a special role for the person starting the fire, singing together, order of picking songs to sing, order of speaking, gestures indicating you want to speak next