Posts
Comments
For next time: https://www.lesswrong.com/posts/LfrNFfJFcqnG9WuFf/activated-charcoal-for-hangover-prevention-way-more-than-you ;)
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.
I used to get depressed about genetic determinism. It's a two-sentence thought that eliminates your perceived capacity for change.
However, while some predictive models can be built, and many things do revert to the mean – those are tendencies. You only get a pattern from a behavior that repeats. Some things don't. If you're looking for an overarching cause of life-outcomes, you necessarily cancel out individual variation.
No study on demographics that includes a section called "the one-off thing that happened to one guy once in defiance of what usually happens".
Those models don't know what you want. They can't account for your decision to find it.
Most importantly, they can't account for a systematic effort to find it.
Hide and cover clocks to stop procrastinating. There's no five minutes or five years "from now" that's not just "now" – but quantified time creates the illusion you perceive the future. This creates an emotional relationship to approaching deadlines. If you can't see time in your environment (at least when you want to work), the pain of experiencing "the future" immediately subsides but so does the idea you can put things off.
Useful links: Overcoming Bias, Dr. K, J Krishnamurti, Jeffery Kaplan
Writing polite but short emails that have a single intention is hard. The fewer words you use, the more can be wrongly inferred about the tone you hoped to convey. You want to save your recipient's time and energy and to do that consistently, but it's difficult to know if people will read something you didn't intend to say.
While Claude and GPT-4 often understand exactly what I mean when I feed them poorly written word salad that is both long-worded and not acceptable to send, they don't yet do a good job of removing what I want because of what seems to owe to how instruction-tuning and RLHF weight verbosity. Maybe this is easy to fix with the correct prompt, but no prompt I've tried has been universal.
I want software that can totally remove my personality and the extra context of what else I might be thinking about from a piece of writing, while also double checking every interpretation of my tone . I want to be the Bruce Lee of emails.
Unwanted thoughts are continuous and amorphous processes, and you can lessen their severity by focusing on when the focal point of your attention fluctuates, leaving a sort of "negative space" of the thought's form in your mind's periphery.
Another point worth mentioning: Isaac Newton allegedly had the ability to focus on his work for entire consecutive days at a time. This is highly unusual. The only non-chemical intervention I've ever heard of that can take a normal human mind to that ability is meditation. Though most people aren't willing to take their meditation practice to the level of intensity that does that, with proper instruction, it may have a more than marginal effect on prolonged concentration.
I'm coming at this from an absolutely insane angle, but I think I've figured out the important thing that those questions miss - or at least another way to put what's already been said. "Consciousness" cannot be described using positive definitions. This is due to an indexicality error. Your "experience" is everything there is - not in a solipsistic sense, but in the much more important sense that the notion of anything outside of experience is itself happening in experience. As this applies to the future and the past, every perception occurs in a totally "stateless" environment, in which all indication is a sleight of hand. You can't think about the totality of your attention, only shift the focal point of the lens. One of the things we can shift the lens to is a symbol of the whole thing, but that's strictly a component of it. In the only relevant sense, you are always focused on everything there is, and any attempt to look at consciousness or answer the question "why is it like anything at all?" is like a snake trying to eat its own mouth. It's not "like" anything because "likeness" is a comparative term used here to describe one thing relative to an incoherent notion.
So far it's just been my family members and their friends. I'm going to continue interviewing that pool until I exhaust it (which should take a long time), though I'm not sure what I'll do after that.