costco

Mainly Claude and Muse Spark. It seemed happy to help if it was convinced that I controlled the website with the test captchas on it. I’ve also found the more Claude saves memories of working on the project the less conflicted it gets.

Author here. I submitted this a few days ago but seems like it was added to the second chance pool so now it’s on the front page. Anyways, I have received a report from a user that this solver did not work for them with European proxies but did with US proxies… so they may be serving different versions for different regions. I will fix that soon. I’m not at my computer right now but it was working when I tested it a few hours ago (Cloudflare constantly changes things).

My understanding is that there is some sort of rate limiting around PATs per Apple ID (that’s why they require you to be signed in). Also if you want to scale to millions per day you don’t want to have to use headless browsers.

For the last few builds released by Cloudflare, I basically just told an agent “Hi they changed something. Please fix.” And waited for a few hours (mostly spent in Rust compilation) until it came back fixed. Keeping the codebase clean while doing that long term is honestly the harder part.

Yes, that was my goal. I will update this for the next couple of weeks, but probably not forever. However I think this is a pretty good foundation for future efforts.

You know this is easily verifiable right…. I just put 10 random pages into Pangram and they all came back 100% human. And now I feel bad for making an even greater proportion of comments on this thread not about graph theory.

This is a nice spin on the Aaronson oracle:

> In a class I taught at Berkeley, I did an experiment where I wrote a simple little program that would let people type either “f” or “d” and would predict which key they were going to push next. It’s actually very easy to write a program that will make the right prediction about 70% of the time. Most people don’t really know how to type randomly. They’ll have too many alternations and so on. There will be all sorts of patterns, so you just have to build some sort of probabilistic model. Even a very crude one will do well. I couldn’t even beat my own program, knowing exactly how it worked. I challenged people to try this and the program was getting between 70% and 80% prediction rates. Then, we found one student that the program predicted exactly 50% of the time. We asked him what his secret was and he responded that he “just used his free will.”

Wow that was quick :) Meh, I feel like comment points have become somewhat untethered to comment quality (at least assuming threads are roughly sorted by comment score). It’s a good heuristic usually but it’s wrong often enough for it to be difficult to rely on.

I want to improve the classifier scores too. There’s some parameters in the loss function I could play around with more when training which basically optimize for more false negatives or false positives. Problems with no ground truth are very frustrating lol (though I suppose dang/tomhow’s manually written warnings could be treated as labels instead of using labels from GPT, but I suspect there’s not enough of them).