easily distracted by bikes // mts @adaption_ai // @neo @middlebury
/sf
Joined July 2018
- Tweets2.2K
- Following441
- Followers1.2K
- Likes11.5K
found some code to scrape prices from traderjoes, got claude to rip with this and get a meal list in my notes and hooked it up to my todolist and calendar
another example of the Feynman “computer disease”
This is a very apt thing
I have noticed a certain malaise i am seeing. it honestly reminds me a lot of 2021 neovim config andies.
You can build anything, just some time and some prompts and you got your fully custom piece of tech for exactly what you want!!
but wait, here is another thing to optimize! wait... there is another thing..
oh no! your first thing doesn't quite do it the way you want it to, better fix that...
...
pretty soon you are having 8 agents building all sorts of stuff and you consistently feel like you are going no where and you have nothing to show for it, yet your brain is moving at 100x the speed, your sleep is suffering, your attention with your kids and wife are dwindling... but bro, just one more prompt, just one more prompt and it will fix this issue, i swear
The burden of being able to build anything is the burden of having to maintain everything and in the day and age of vibing, contracts are even less held tight and change is even faster.
I recently read Feynman's lecture "Los Alamos From Below". There are some great tales of him picking locks and avoiding censorship, but my favorite is this short bit about the _computer disease_.
It's the disease you get when you realize everything that is possible without thinking about if it's necessary. We have few remedies for this disease and its symptoms can be seen today in heavy drop-shadow kanban boards, story points, and second brain systems managed by teams of ai agents.
Has the Pragmatic Interp Agenda killed sparse learning methods?
I spent some time circuit tracing gpt2 today as a learning exercise.
The llm biology work is quite satisfying, but I’m too pragmatic pilled to see how this helps safety. I could see a circuit tracer like interface that attempts to describe the computation alongside an agent trace, but the moment my model generates more than one token, this graph becomes untenable.
