schapi1991.bsky.social

There are basically two types of AI users: (1) "I delegated all my thinking to the AI, I assume that's for the best." (2) "After 350 hours of coding I finally have a decent self-evolving harness loop that forces the model to actually follow the instructions in my prompt. It works 53% of the time."

Screw it, it’s worth saying this again

Man Struggling To Pierce Orange Peel With Fingernail Under Impression He Could Kill If He Had To https://theonion.com/man-struggling-to-pierce-orange-peel-with-fingernail-un-1819579648/

“Companies don’t appear to be replacing workers with AI on a significant scale,” suggesting instead that companies may be using the technology as a cover for routine headcount reductions," per Oxford Economics.

GPT-5 regularly asks you why you are not citing its work. GPT-5 always believes your estimator is biased. GPT-5 has more of a comment than a question. Sometimes GPT-5 tells you stories about grad school and expresses regret about getting a PhD. GPT-5's mom wonders when it will start making money.

Once again: LLMs are word-association algorithms. They are incapable of grokking that the words they parse correspond to actual objects & phenomena outside of their own algorithms. To an LLM, its statements depend not on facts but on how often words occur together in the material the're trained on.

Internet Archive Responds to Appellate Opinion in Hachette v. Internet Archive 🔗 blog.archive.org/2024/09/04/i...