My website is here.
bilalchughtai
Karma: 1,067
readers of this post may also enjoy the simpler following post: how to run without all the pesky agonizing pain
When evaluating whether to invest time in making things more efficient, I often see people compare the one off cost of make the thing more efficient to the expected future saved time when doing the thing. I think there is often an important third variable to track, that often swings such decisions from not worth the effort to definitely worth the effort. Namely, the expected increase in usage of the thing due to the reduction in friction of utilizing the thing. I in practice often find the final consideration to dominate.
Recent examples from my life:
Reducing the number of button presses needed for common workflows on my laptop means I both can navigate my laptop quicker and end up navigating more instead of procrastinating navigating because it is annoying.
Moving to a more central location in my city has meant I both save time commuting to things and also end up going to more things.
Automating the loading of context from my personal apps into AIs means I spend less time copy pasting context into AIs and also end up asking AIs more questions about my personal context.
thanks for the comment!
i keep hoping ai agents get good enough
i am also keen to think through how ai can make my workflows better! i agree that they are not quite there yet for entirely automating parts of task management.
no, I just want them to go away upon completion
seems good, seems eventually possible
The core is I want the system/ tool to be invisible … Touch tasks as close to once as possible … Why are you making a “today” list?
i agree that that many of the best systems require literally zero cognitive overhead. but i’m skeptical that optimal task management should ever literally be zero effort.
touching each task literally only once (to add it) requires you to perfectly predict the future. the world is a complex system and out of your control. this implies that you cannot perfectly anticipate what tasks you will need to do at some point in future. something new and urgent may come up, that displaces things you initially thought you might want to do on some particular day. something may later become irrelevant reality changes and new information comes to light.
now, it does in principle seem possible to automate away the task of touching the tasks to ai. if you had an ai assistant with an incredible amount of context on your life, and could predict actions you might want to take to maintain your tasks with reasonable accuracy, you could build some system that just serves you tasks. i think there are some problems with this vision. firstly, i expect in such a world almost all of the tasks on your stack could just be done AI too. second, i have local preferences. some days i just “don’t feel” like doing some task, and would rather punt it to the next day and do something else instead. AIs could in principle understand this in the limit, but i expect this to take longer than the first AIs capable of serving you reasonable tasks in a reasonable order. third, i personally would feel a bit like a robot here, if i were just executing actions that some system has told me to complete. i would rather be able to exert agency and choose what i do for myself, even if that contradicts the “most optimal” next task.
You seem to enjoy the planning and to get satisfaction from task completion.
agreed, ymmv
pretty often! i schedule a lot and time box a bit. the main reason i schedule is so that i get a reminder at the right time of day to do the task. sometimes that time is more just a proxy for some event like “when i’m at the office” or “when i get home”. it would be better to specify that precisely but i havn’t seen good support for it anywhere yet.
Jotting down “deal with X” on the todoist mobile app and then later figuring out exactly how to deal with it on my laptop is a pretty common workflow of mine. I find it pretty frictionless. I also use the mobile app for other things (e.g. glancing at my daily to-do list while on the go).
Here are some screenshots of the app:
An opinionated guide to building a good to-do system
two weeks of basically full time effort … $1000
Above some P(you want to do a PhD), I claim this is cheap, given it could effect the next 1-6 years of your life. I think I agree with you that you need to be at something like at least 25% (discussed offline) or so here already to actually commit to doing the applications. But I think your probability can be lower for spending some smaller amount of time “considering” applying.
Also you don’t get much information. You buy a few lottery tickets. The number that come back winners is a weak signal of how good your application was.
Mostly agreed. The information I think you get is whether a PhD is a good option for you. It forces you to think through the prospect and you get to chat to PhD students and professors through the process. I updated positively on PhDs after applying through this process.
Consider applying to PhDs soon!
Last November, I wrote a blog post titled You should consider applying to PhDs (soon!), where I argued it is probably a good use of time for junior AI safety researchers (e.g. people who have recently participated in an upskilling or research program like ARENA or MATS) to apply to PhDs in the current cycle, even if they are on the fence about whether they want to do a PhD.
My core arguments were that academic timelines are very slow (i.e. if you apply this year you would not start until Fall 2026), applications are generally cheap and high information value, and that applying strictly increases your future optionality. I applied to PhDs two years ago, got several offers, did not end up doing a PhD, but still post-hoc endorse spending time on this.
My post last year was very late relative to deadlines; you should probably start thinking about applications soon in order to gather the required application materials (e.g. asking for references) for the December 15th deadlines.
typo: unambitious → unambiguous
karpathy reviews sleep trackers: https://karpathy.bearblog.dev/finding-the-best-sleep-tracker/
Detecting Strategic Deception Using Linear Probes
Paper: Open Problems in Mechanistic Interpretability
As a general rule, I try and minimise my phone screen time and maximise my laptop screen time. I can do every “productive” task faster on a laptop than on my phone.
Here are some things object level things I do that I find helpful that I haven’t yet seen discussed.
Use a very minimalist app launcher on my phone, that makes searching for apps a conscious decision.
Use a greyscale filter on my phone (which is hard to turn off), as this makes doing most things on my phone harder.
Every time I get a notification I didn’t need to get, I instantly disable it. This also generalizes to unsubscribing from emails I don’t need to receive.
What is the error message?
Yep, this sounds interesting! My suggestion for anyone wanting to run this experiment would be to start with SAD-mini, a subset of SAD with the five most intuitive and simple tasks. It should be fairly easy to adapt our codebase to call the Goodfire API. Feel free to reach out to myself or @L Rudolf L if you want assistance or guidance.
How do you know what “ideal behaviour” is after you steer or project out your feature? How would you differentiate a feature with sufficiently high cosine sim to a “true model feature” and a “true model feature”? I agree you can get some signal on whether a feature is causal, but would argue this is not ambitious enough.
Yes, that’s right—see footnote 10. We think that Transcoders and Crosscoders are directionally correct, in the sense that they leverage more of the models functional structure via activations from several sites, but agree that their vanilla versions suffer similar problems to regular SAEs.
Also related to the idea that the best linear SAE encoder is not the transpose of the decoder.
I took up running about a year ago, shortly after starting to wear bearfoot shoes. I only run in barefoot shoes and have had no problems. I suspect it helped that I “learned to run” in barefoot shoes, and that ordinarily one would have to change their stride to fit the shoe better, e.g. by landing mid foot instead of heel striking. I also, as I was a new runner, started off with extremely low volume (~10km a week), which also probably helped build foot strength.
I’m pretty happy overall, though am now tempted to get some proper running shoes for races, as the zero energy return is sad if you care about speed.