Essays and notes on working with AI: why agent-written code needs layered review, what agent work really costs, and how my 2019 predictions held up.
Connected topics, ranked by shared posts.
Per token, GPT didn't get cheaper in 14 months and caching is close to its limit. Per solved task, the newest GPT got 9× cheaper since April. So my $240 of subscriptions have an end date: ChatGPT Pro in about half a year, Claude in about two years. Also, einsager is out: dictation that runs on my Codex login.
My recovered AI history contains 76.85 billion tokens. Most are cached input, which changes what the models cost and how much capability I get for the money.
Every check misses stuff. A bug only reaches production if all of them miss it, which is why I think AI code review needs different models with different prompts.
What I expected from AI in my June 2019 school paper, what happened since, and how I think about it in 2026.
'move fast and break things' is basically the default now, and AI-coding will crank that up even more. the real fix is boring: code review + managing agents well, so you don't ship the next outage or security mess.
What everyone with a single braincell thought would happen, has happened.
Study Helper file management systems, and upcoming technical challenges. Plus some cool finds in tech and anime.
Supercharge your study sessions with an AI assistant! Generate flashcards, summaries.
