Why did you use the weak AGI question? Feels like a motte-and-Bailey to say “x time until AGI” but then link to the weak AGI question.
Logan Riggs
Karma: 2,281
For those with math backgrounds not already familiar with InfraBayes (maybe people share the post with their math-background friends), can there be specifics for context? Like:
If you have experience with topology, functional analysis, measure theory, and convex analysis then...
Or
You can get a good sense of InfraBayes from [this post] or [this one]
Or
A list of InfraBayes posts can be found here.
These arguments don’t apply to the base models which are only trained on next word prediction (ie the simulators post), since their predictions never affected future inputs. This is the type of model Janus most interacted with.
Two of the proposals in this post do involve optimizing over human feedback, like:
Creating custom models trained on not only general alignment datasets but personal data (including interaction data), and building tools and modifying workflows to facilitate better data collection with less overhead
, which they may apply to.
I don’t think the “Only Game in Town” argument works when EY in the OP says
I have a few stupid ideas I could try to investigate in ML,
As well as approving redwood research.
I really like this feature, and will probably switch to reading on greaterwrong because of it. The only implementation choice I disagree with is the confirmation notification if you want to toggle it back on.
A reason to keep the confirmation would be to create a trivial inconvenience for people who are tempted to be biased by the Who and the karma value.
A reason to remove the confirmation is when I want to verify who is replying back to whom. Not the specific person’s name, just if B comments on A’s comment, and then someone comments on B, I’d like to toggle quickly to see if that someone is A or someone else.
Remember in high school when you wanted to rent a bounce house for your birthday? I went around to everyone trying to raise money for the cause, but I could only get $20 (all from myself, haha).
$20 isn’t enough to rent a bounce house, but you found some employees testing bounce houses, paid them $20 to bounce on one for 20 minutes and left like a king.
I know, I know, I’m a day late, but still,
happy birthday man
My understanding is: Bob’s genome didn’t have access to Bob’s developed world model (WM) when he was born (because his WM wasn’t developed yet). Bob’s genome can’t directly specify “care about your specific family” because it can’t hardcode Bob’s specific family’s visual or auditory features.
This direct-specification wouldn’t work anyways because people change looks, Bob could be adopted, or Bob could be born blind & deaf.
[Check, does the Bob example make sense?]
But, the genome does do something indirectly that consistently leads to people valuing their families (say ~80% of people). The bulleted list (e.g. reaction to being scammed, etc) are other extremely common human values & biases that seems improbable for the genome to directly specify, so the alternative hypothesis is the genome set the initial conditions (along with the environment) such that these are generally convergently learned.
The hope is that this is true, the mechanisms of which can be understood, and these mechanism can be applied to AGI convergently learning desired values.
I agree. You can even get career advice here at https://www.aisafetysupport.org/resources/career-coaching
Or feel free to message me for a short call. I bet you could get paid to do alignment work, so it’s worth looking into at least.
I don’t understand why shminux’s comment was down to −6 (as of 11⁄17). I think this comment is good for thinking clearly. How reality is perceived to you is based off how you collect data, update, and interpret events. You can get really different results by changing any of those( biased data collection, only updating on positive results, redefining labels in a motte and bailey, etc.)
Going from a “one truth” to a “multiple frames” model helps communicating with others. I find it easier to tell someone
from a semantics viewpoint, ‘purpose’ is a word created by people to describe goals in normal circumstances. From this standpoint, to ask “What’s my purpose in life?” doesn’t make sense since a goal doesn’t make sense applied to a whole life [Note: if you believe in a purposeful god, then yes you can ask that question]
than stating more objectively (ie without the “from a semantics viewpoint”).
This is also good for clarifying metrics because different frames are better at different metrics, which should be pointed out (for clear communication’s sake).Instead of denying whole viewpoints, this allows zeroing in on what exactly is being valued and why. For example, Bob is wishing people loving-kindness and imagining them actually being happy as a result of his thoughts. I can say this is bad on a predictive metric, but good on a “Bob’s subjective well-being” metric.
Happy belated birthday, brother. My grandfather got married yesterday, so I was away from my laptop. You’ve missed a lot. I think it would’ve been real fun to discuss Gamestop stocks with you; I think you would’ve invested!
I’ve almost finished grad school as well and intend to study math for a bit on my own. You got to experience life after college for a bit; I wonder how your business and overall life would’ve gone since now. I think we would’ve argued vaccines, lockdowns, and mask when we saw each other. I remember arguing with you about politics and religion, but I don’t remember us ever getting mad at each other (except freshman year, but that was over a girl, haha).
I don’t even know where your grave is, but I’ll continue to talk to you here.
~Take care
I’m excited about sensory substitution (https://eagleman.com/science/sensory-substitution/), where people translate auditory or visual information into tactile sensations (usually for people who don’t usually process that info).
I remember Quintin Pope wanting to translate the latent space of language models [reading a paper] translated to visual or tactile info. I’d see this as both a way to read papers faster, brainstorm ideas, etc and gain a better understanding of latent space during development of this.
No, “why” is correct. See the rest of the sentence:
Write out all the counter-arguments you can think of, and repeat
It’s saying assume it’s correct, then assume it’s wrong, and repeat. Clever arguers don’t usually devil advocate themselves.
High-detachment is great!...for certain situation for certain times. I really enjoy Rob Burbea’s “Seeing That Frees” meta-framework regarding meditation techniques/ viewpoints: they are tools to be picked up and put down. If viewing the world in complete acceptance helps your suffering in that moment, then great! But you wouldn’t want to do this all the time; eating and basic hygiene are actions of non-acceptance at a certain conceptual level. Same with impermanence and no-self. Your math friend may be open to that book recommendation.
TurnTrout argues that Tsuyoku Naritai is not it, and maybe he is right. I do not know what the correct emotion feels like, but I think maybe DF knew.
I’ve had a similar experience with feeling Tsuyoku Naritai, but it being a temporary state (a few hours or days at a time maybe). I’m currently extremely interested in putting on different mindsets/perspectives for different purposes. An example is purposely being low-energy for sleeping and high-energy for waking up (as in purposely cultivating a “low-energy” mindset and having a TAP to tune into that mindset when you’re trying to sleep). In this case, Tsuyoku Naritai may be good for lighting the fire for a day or two to get good habits started. But I think people may use unnecessary muscles when tuning into this mindset, causing it to be tense, head-ache inducing, and aversive.
This is speculation though, but I am, again, very interested in this topic and discussing it more. Feel free to dm me as well if you’d like to chat or call.
Looking back, this all seems mostly correct, but missing a couple, assumed steps.
I’ve talked to one person since about their mild anxiety talking to certain types of people; I found two additional steps that helped them.
Actually trying to become better
Understanding that their reaction is appropriate for some situations (like the original trauma), but it’s overgeneralized to actually safe situations.
These steps are assumed in this post because, in my case, it’s obvious I’m overreacting (there’s no drone) and I understand PTSD is common and treatable. Step 2 is very much Kaj Sotala’s internal family system’s post while this post is mainly about accessing lower-level sensory information about the trauma-reaction.
Thanks renato!
Regarding your first set of questions:
Reading: originally 3 hr. This changed to 0-3 hours depending on when I woke up in the mornings, which meant going to bed around 9-10, which meant making a habit of trying to fall asleep when I get in bed. I did try to read in the evenings as well, but my eyes would glaze over after working for the day.
Tensor flow: originally 2 hr. I dropped this after the first week due to work taking up 2.2x as much time as I initially predicted. I also felt like I wasn’t actually learning anything while going through Google’s tutorials, and TurnTrout convinced me to just learn the pre-reqs and theory of ML first.
AI-Safety reading: originally 1 hr. This became 1-3 hours depending on how interested I was in. I dropped this though after I moved 40 days ago due to habit changes and then simply forgetting! I really enjoyed it and it was a low-spoons activity for me.
Meditate: originally 1 hr. This became 0-1 hours depending on the day. I experimented with doing it at different times, but during my lunch break is probably the best for me at the moment.
Weekends are a different animal. If I had a free one & wasn’t experiencing emotional problems, I would tear into a book, meditate a lot, read AI papers, and just get a lot of reading on LW and SlateStarCodex done which was great! I really wish I didn’t have to work so that could be every day.
So to answer your question, I changed what I was doing after giving it a solid try and adjusting from there if I needed to. After doing that for a couple of months, I have a much better idea of how to do these types of things now.
Regarding “How much efficiency during reading sessions?”
I interpret that to mean “how many pages per day”/”How many chapters per week”/”How many books per month”. If that’s correct, then I would say I could (right now) learn a subject/book a month. Like I could read Linear Algebra Done Right in less than a month and Tao’s Analysis I & II in less than two months, while doing all of the exercises.
If I didn’t have to work, I would predict that I could half that time and finish one in less than 2 weeks.
If that wasn’t what you asked, please let me know!
[Note: this one, steelman, and feedback on proposals all have very similar input spaces. I think I would ideally mix them as one in an actual product, but I’m keeping them separate for now]
Task: Obvious EA/Alignment Advice
Context: There are several common mental motions that the EA community does which are usefully applied to alignment. Ex. “Differential Impact”, “Counterfactual Impact”, “Can you clearly tell a story on how this reduces x-risk?”, and “Truly Optimizing for X”. A general “obvious advice” is useful for general capabilities as well, but this is intended to have a strong, EA tilt to it.
Input Type: A plan for reducing x-risk, either a research direction, upskilling, or field-building
Output Type: Useful exploratory questions
Instance 1
Input:
I’m thinking about taking up a job at Google to upskill in machine learning before I apply to Anthropic or Redwood Research.
Output:
Isn’t there a more direct path to upskilling in machine learning? Like taking a course or doing a bunch of dumb, fun projects? Additionally, is programming/machine learning your comparative advantage?
Instance 2
Input:
I’m working on robustness to distributional shift, specifically identifying when an input is out-of-distribution so it can default to a no-op or asking for human feedback on that input.
Output:
How neglected is this field? If you didn’t work in this field, would someone else have done the work you’d do, but a year or two later?
Instance 3
Input:
I’m building an interpretability tool that can tell you what the model was thinking when it was making a classification, for instance, what part of the input it was using to make that classification.
Output:
Let’s say you succeed in making this tool; can you tell a story on how this tool can reduces AI x-risk?
I think I’ve convinced my girlfriend that it’s okay for me to be sad because of what happened to you. She used to try to cheer me up, but I would tell her that it’s okay for me to be sad. It sucks and it’s okay if I act like it sucks.
I had honestly thought the day was July 15th. Then I saw my calendar and saw that it was today. As soon as I noticed, I started watching youtube, I guess to distract myself. When I stopped, it all just weighed on me again.
It’s hard to accept your death. You had your goals and friends and all of your expectations, and then it just ended.
Your sister still posts for you on Facebook. Your friends still think of you.
I remember you. I miss you man
Happy Birthday Will,
I remember in 9th grade you started dating my ex right after consoling me. I was so mad! Haha. I never told you this, but me and the others on our forensics team saw y’all just sitting, holding hands, and having a good time, and Jennifer suggested that me and her hold hands and sit next to y’all giggling.
I said no, though it would’ve made a better story if I went through with it, haha. I think we started getting along again after she moved, although I can’t remember saying anything mean to you because of it.
I’m not sure she knows what happened to ya. I know y’all kept in touch when she moved, and maybe she checks Facebook more than I do.
Anyways, a lot of us are back home cause of the Coronavirus, and I would love to be able to give you a call and see how your life’s progressed these past few years.
Love you Will,
Logan
I’m submitting my sequence Arguing Well. As far as the sequence is planned, most of the posts will be exercise based, and it’s planned to be finished by Friday.
At a pond, my niece was in a child floaty, reached too far and flipped over into the water. I slammed my half-eaten sandwich on my brother’s chest, hoping he would grab it and ran into the water and saved her.
She was fine and I got to finish my sandwich.