1,589 karma · joined July 10, 2019
I'm using an old Android phone (Pixel 2 from 2017), a 5$ a month unlimited SMS plan from Tello, and https://github.com/capcom6/android-sms-gateway. For bonus points (I wanted to roll my own security, route messages from different numbers to prod and ppe instances of my backend, and dedup messages) I built a little service in Go that acts as an intermediary between my backend and android-sms-gateway. I deploy this service to my android device using ADB, android-sms-gateway talks to it, and it talks to my backend. I also rooted the android device so I could disable battery management for all apps (don't do this if you want to walk around with the phone of course). It works pretty well!
I plan to open-source this eventually TM, but first I need to decouple my personal deployment infra from the bits useful to everyone else
I bring this up because this article discusses context management mostly in terms of context windows having a maximum size. I think that context management is far more than that. I'm still new to this building agents thing, but my experience suggests that context problems start cropping up well before the context window fills up.
There is more to engineering than understanding the fundamental physics.
I also have to point out that the author's maligning of the now famous Cloudflare experiment is totally misguided.
"There are no controls or alternate experiments" -- there are tons and tons of implementations of the OAuth spec done without AI.
"We also have to take their (Cloudflare’s) word for it that this is actually code of an equal quality to what they’d get by another method." -- we do not. It was developed publicly in Github for a reason.
No, this was not a highly controlled lab experiment. It does not settle the issue once and for all. But it is an excellent case study, and a strong piece of evidence that AI is actually useful, and discarding it based on bad vibes is just dumb. You could discard it for other reasons! Perhaps after a more thorough review, we will discover that the implementation was actually full of bugs. That would be a strong piece of evidence that AI is less useful than we thought. Or maybe you concede that AI was useful in this specific instance, but still think that for development where there isn't a clearly defined spec AI is much less useful. Or maybe AI was only useful because the engineer guiding it was highly skilled, and anything a highly skilled engineer works on is likely to be pretty good. But just throwing the whole thing out because it doesn't meet your personal definition of scientific rigor is not useful.
I do hear where the author is coming from on the psychological dangers of AI, but the author's preferred solution of "simply do not use it" is not what I'm going to do. It would be more useful if instead of fearmongering, the author gave concrete examples of the psychological dangers of AI. A controlled experiment would be best of course, but I'd take a Cloudflare style case study too. And if that evidence can not be provided, then perhaps the psychological danger of AI is overstated?