Clojurians Log v2

I think he got most of them, but he's missing a key identifying shibboleth of AI maximalists: the way they talk/write about AI and respond to any criticism. It's some weird combination of standard paid influencer / grifter breathless and bombastic clickbait methods along with some form of bizarre reverence. • "ChatGPT o5.6-mini-thinking-final-pro just shattered these artificial benchmarks that are probably included in the training set, here's what that means:" to start the 50 tweet thread of mostly generated slop • "CEO of company that makes money from AI says AI will cure cancer and all other problems! Wow!" surely an unbiased opinion! If you say that AGI isn't here or maybe even can't ever be made (in the sense that https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html) they respond within the range of • "you just need to prompt better" an ad hominem if there ever was one • "I bet you program only in assembly too" comparing non deterministic stochastic parrots to compilers is nonsense. • "You can't stop it" probably true, if it is indeed possible, but orthogonal to the discussion • "You're going to get left behind" these things have been widely available for about two years. If I'm wrong and they become required I can catch up. • "humans learn by reading so it's ok if a company steals data from the entire internet illegally" when you point out that all models rely on copyright infringement

Note that many, many of the same influencers peddling vibe coding and so on were mere years ago selling JPEGs as currency

Certainly huge overpromising just like in 80’s. Will lead to AI Winter #3 probably. But certainly the tech has neat abilities just like Expert Systems of the 80’s. The cycle of ideas, research, consumer and hype, valley, (repeat) will continue with computer tech.

Yeah I don't deny the tech is interesting. They're extremely good at faking a human interaction

But the more I've tried them the more I run into situations where the limitations show

In a chat with Claude about something and it ended up saying it would pray for me. Obviously because it had been trained to reproduce human conversations where that would be normal to say. Another where I asked one of the models a relatively trivial but hard-to-reason through Clojure question. It was one of the thinking models. It spent about 5 minutes on it (running locally) and ended up thinking itself into responding (incorrectly) in pseudo Python

I also recall the first time I tried GitHub copilot on my own dot files repo and it suggested my own code back to me from a different repo. Verbatim character for character. It was on an extremely niche function I have in my Zsh config. So, copyright laundering. Since although technically I had a license to use that code, and so does the whole world, it does require attribution.