Gave me back ~3hr per day when nothing else worked[2].
Wikipedia can be downloaded via kiwix (~90GB for English WP with images), programming documentation with zeal & devdocs.io. Logseq as a replacement for obsidian/roam, yt-dlp for downloading YouTube videos (and video/audio from many many other sources) to watch/listen to later. wget for downloading whole websites+assets to read at some later point.
No great solution for LLMs (yet…), all the ones I can run on my laptop are not good enough—maybe I should bite the bullet and get a GPU/digits that can run of the big LLaMas/DeepSeek V3 locally. I have internet at my workplace, and a local library with internet 10 minutes away by walk that closes at 7PM.
(No mobile internet either, that’d defeat the purpose. Haven’t yet quite figured out to get people to call me when they want stuff…)
For >3 years now. The benefits reduce after a while as homeostasis kicks in (e.g. moving sleeping times back by ~4 hrs got halved to ~2 hrs), but it’s still net positive: I used to lose ≥4½ hrs to random aimless websurfing, now it’s only about one. Not all time gained is spent productively, I still randomly click through articles of the local Wikipedia copy, but that feels much less unproductive than watching YouTube videos.
Website blockers like browser extensions are too easy to turn off (especially since I have complete control over my OS). Accountability didn’t work well either. Behavioral interventions (like exercising/meditation/whatever) did ~nil.
I think especially if you have a competent coding LLM in the air-gapped network, then probably yes, if you mean software engineering.
The biggest bottlenecks to me look like having nice searchable documentation for libraries the LLM doesn’t know well—the documentation for most projects can’t easily be downloaded, and the ones for which it can be downloaded easily aren’t in a nicely searchable format. Gitbook isn’t universal (yet. Growth mindset).
(Similar with the kiwix search—you need to basically know what you’re looking for, or you won’t find it. E.g. I was trying to think of the Nazi Bureaucrat who in the Sino-Japanese war in the 1930s had rescued many Chinese from getting killed in warcrimes, but couldn’t think of it—until LLaMa-2-13b (chat) told me the name was John Rabe—but I’ve despaired over slightly more obscure questions.)
A well-resourced actor could try to clone the Stackoverflow content and their search, or create embeddings for a ton of documentation pages of different software packages. That’d make it much nicer.
Also, a lot of software straight up doesn’t work without an internet connection—see e.g. the incident where people couldn’t do arithmetic in Elm without an internet connection. Thankfully it’s the exception rather than the norm.
Solution: No internet at home[1].
Gave me back ~3hr per day when nothing else worked[2].
Wikipedia can be downloaded via kiwix (~90GB for English WP with images), programming documentation with zeal & devdocs.io. Logseq as a replacement for obsidian/roam, yt-dlp for downloading YouTube videos (and video/audio from many many other sources) to watch/listen to later. wget for downloading whole websites+assets to read at some later point.
No great solution for LLMs (yet…), all the ones I can run on my laptop are not good enough—maybe I should bite the bullet and get a GPU/digits that can run of the big LLaMas/DeepSeek V3 locally. I have internet at my workplace, and a local library with internet 10 minutes away by walk that closes at 7PM.
(No mobile internet either, that’d defeat the purpose. Haven’t yet quite figured out to get people to call me when they want stuff…)
For >3 years now. The benefits reduce after a while as homeostasis kicks in (e.g. moving sleeping times back by ~4 hrs got halved to ~2 hrs), but it’s still net positive: I used to lose ≥4½ hrs to random aimless websurfing, now it’s only about one. Not all time gained is spent productively, I still randomly click through articles of the local Wikipedia copy, but that feels much less unproductive than watching YouTube videos.
Website blockers like browser extensions are too easy to turn off (especially since I have complete control over my OS). Accountability didn’t work well either. Behavioral interventions (like exercising/meditation/whatever) did ~nil.
Do you think this hints at “doing engineering in an air gapped network can be made somewhat reasonable”?
(I’m asking in the context of securing AI labs’ development environments. Random twist, I know)
I think especially if you have a competent coding LLM in the air-gapped network, then probably yes, if you mean software engineering.
The biggest bottlenecks to me look like having nice searchable documentation for libraries the LLM doesn’t know well—the documentation for most projects can’t easily be downloaded, and the ones for which it can be downloaded easily aren’t in a nicely searchable format. Gitbook isn’t universal (yet. Growth mindset).
(Similar with the kiwix search—you need to basically know what you’re looking for, or you won’t find it. E.g. I was trying to think of the Nazi Bureaucrat who in the Sino-Japanese war in the 1930s had rescued many Chinese from getting killed in warcrimes, but couldn’t think of it—until LLaMa-2-13b (chat) told me the name was John Rabe—but I’ve despaired over slightly more obscure questions.)
A well-resourced actor could try to clone the Stackoverflow content and their search, or create embeddings for a ton of documentation pages of different software packages. That’d make it much nicer.
Also, a lot of software straight up doesn’t work without an internet connection—see e.g. the incident where people couldn’t do arithmetic in Elm without an internet connection. Thankfully it’s the exception rather than the norm.
I’ll add: No internet on my phone.
My friend recommended I delete the browser on my phone. It’s saved me time from going down curiosity rabbit holes.