I have also tried the maze from a photo test a few times and never seen a one-shot success. But yesterday I was determined to succeed so I allowed Gemini 3 to write a python gui app that takes in photos of physical mazes (I have a bunch of 3d printed ones) and find the path. This does work.
Gemini 3 then one-shot ported the whole thing (which uses CV py libraries) to a single page html+js version which works just as well.
I gave that to Claude to assess and assign a FAANG hiring level to, and it was amazed and said Gemini 3 codes like an L6.
Since I work for Google and used my phone in the office to do this, I think I can't share the source or file.
I've read all of Ted's work and I hope you do try going back to Exhalation and just skipping TLOSO to get to more good stuff. It is far too long and I thought there was some value in it, but it was a slog, and in the bottom 1/4 of the stories IMO, so just skip it and get more great value.
I was at Meta when it was forced by the FTC to start adding this compliance stuff. It SUCKED to retrofit everything.
Now I'm at google, and onboarded on to the version of the infra that already went through that, and I can take it all at face value. It is a PAIN still, but this is the reality of a system that interfaces with O(10^8) users, O(10^2) governments.
I wish I had an answer for you. I spend at least half of the past year trying to make that decision. The internal LLM that can read all the docs and code, you'd think, could get the context to know what the optimal state is, but it easily gets confused by out of date documentation and recommends paths that are going to be marked as "why didn't you use the new thing?" at review time, OR it builds out a solution using "oh, this isn't ready for use yet" parts.
IT IS THAT GOOD.
reply