More Dakka On Your Expectations
After hearing my friend talk about his roommate’s brash decision-making from the despair at getting rejected by girls he liked several times, my friend mentioned that his roommate had asked out a total of three people since high school. Only three!
While there are more factors in the story involved, I’ve heard similar enough troubles that it seems worth saying: Three people is not a lot. Certainly not enough rejections to merit the magnitude of self-worth issues people can walk away with that few from.
If you had the expectation that if the first person you ask out didn’t like you then you’re doomed to loneliness, then a (probable enough) failure would be such a damaging experience you might not try again for a long time. If you instead believed that number was two, the first rejection would hurt considerably less. The higher you go, the less it hurts.
Maybe the expectation we implicitly have from culture is low enough to make three rejections somewhat sting. Why should it? Why shouldn’t that threshold be something like sixty? Or a hundred?
If you can only alter your chances by acquiring skills, improving yourself, and looking harder, then each rejection is valuable new data on what to do better next time. None should be felt as a failure towards your goal by any means. Rejections are an indicator of progress.
Given recent discussion of short timelines & take off on LessWrong, AlignmentForm, and broadly, I've been quite worried. I want to be as skeptical as I can, but it’s hard to judge anything: I don’t know what information I’m missing from the timeline estimates I hear, and increasingly strange or concerning things are put out in public regularly. I can’t really say how likely short timelines are, but given what’s happening, it’s absurd to dismiss this as all collective delusion, and it seems something serious is happening.
Because of my information deficit, I don’t have any idea what work will lead to positive outcomes in these conditions, which of those problems I could likely contribute the most to, and how I could start on that work.
So I feel my current task is to gather a theory of positive impact in this situation and work out from there. I think it would be extremely useful to talk with as many people who think about this as I can 1:1 to understand what they think are the relevant factors.
When I first learned about social status as a concept, I somehow got the mistaken impression that any kind of status seeking is amoral. This caused me harm because I didn't want to violate any social boundaries, and trying to avoid violating status seeking behavior hobbles your ability to find and follow up on opportunities.
I think status seeking can be zero sum, and in such cases it should be avoided (like playing school with the intention of becoming valedictorian).
Status seeking can be positive sum while consisting of iterated zero sum games (like playing in a tennis club).
Status seeking behavior in positive sum environments generally consists of good things, like working harder at the gym.
The concept is extremely useful to keep in mind when designing environments. What constitutes status seeking should be legible, enable and encourage prosocial behavior, and allow social norms to be learned in a healthy way. Losing in iterated zero-sum games is often a common factor in environments with this attribute, since losing is then an expected outcome of playing, and the game can altered so that an individual loss is seen as providing a gain in knowledge, and continuing to play becomes the source of reward.
This can be actively implemented into zero-sum social situations by setting up a situation to expose oneself to frequent but non-comprimising losses. Like starting debates to entertain others with the intention of being roasted.
Your timeline was off, but I think your original comment will turn out to have had the right idea. Given the leaps from GPT-3.5 to GPT-4 to Devin to Claude 3.5-Sonnet w/scaffolding, marginal seeming updates to models are turning out to be quite substantial in effective capability. It's hard to create evaluation harnesses for fuzzy, abstract things like the syntax complexity models can handle, and those abilities do transfer to using the models to automate their own post-training tasks, e.g. like what the self-alignment backtranslation paper's scaling charts showed. The better the model, the better they accomplish these tasks with worse task definitions and less help. The piles of prompts necessary for current agents will be less and less necessary, at some point generated on the fly to meek descriptions like "make a code agent to do ${task}" by the models themselves. Whatever human effort will go into the next generation of unprincipled scaffolding will provide yet greater returns to future models. These factors combined, I expect SWE-Bench progress to be discontinuous and rapid, as it has been so far.
A very naive extrapolation using polynomial regression from SWE-Bench scores suggests ≥80% by November 2025. I used model release dates for my x-value. Important to note models may be contaminated too.
What about 3.5 pushes it over the threshold to you that was missing in previous models?
Avoidant behavior is more interesting to think of in reverse: why do people do anything in the first place?
Procrastination (in a serial way, i.e. burnout) is due to a failure to respond to the normal incentives people act on in your situation. It can be solved by finding another motivation for the activity.
Emotional or cognitive pain is something we don't treat usually as a learning signal as we treat other sources of pain. This is troublesome since it underpins many self-destructive behaviors and all of our neurotic thinking. Sometimes, gut-wrenching dysphoria is a signal you've touched a hot stove (realizing you said something hurtful), and other times it's a major reaction to what's substantively a small insult.
After having been debilitated for a few days many times following the latter kind of pain, I think the right approach is to run head on into desensitization based on their different causes. Eventually, the logical assessment of "this isn't a threat" will correspond to the perceived reality.
As an example, I used to get offended by a lot of minor politically-charged statements before I read Paul Graham's essay on heresy, then realized the pain of getting offended makes you stronger – it's even fun to seek – and felt it much less intensely after that. The hinge of that phenomenon was the realization that offense comes from things we fear might be true – but what's actually the case can't hurt you since it's always been that way, knowing can only help you make better decisions. In a similar vein, overblown neuroticism comes from the possibility that an insult might be true. Accepting things we don't want to believe about ourselves as data can only help us improve.
My favorite so far.