30.8 = 69.1 and 52.8 > 69.1 ??? wat ??? ig llms still can't do arithmetic

The chart on the left should be an absolute jail-able crime.

gg everybody

But amp here works well and wondering if there was any trick to knowing when enough is enough, when it’s achieved its task or needs to do more, if it’s sidetracked away from previous request, etc

The longer you look at this, the worse it gets

It seems like AI really HATES the new syntax in Go that lets you do: for i := range 25 { ... } Every time I give it code with this syntax it just OBSESSES over how wrong it is. And then mansplains Go syntax to me. And then doesn't answer my original question.

Me: "commits.txt contains 150k git hashes. It's too large for me to use effectively. Please convert it to 160 byte integers laid end-to-end." AI: Reads the entire fucking file and blows up the conversation.

The talk about AI making professionals obsolete needs to stop. It is hype, dangerous hype that is causing unnecessary fear. I keep hearing smart people say that translators are made obsolete by AI. There is no evidence to support this claim. 1. Translators have been using

Update: o3-mini fixed the death spiral I was on with Sonnet 3.7 (although it got everything _except_ parent relationships in a LOUDS trie right ... on the fifth from-scratch attempt).

That’s because our implementation doesn’t take into account “random bullshit”

I see the problem now

You’re absolutely right

vibe coding data structures is a new hell

just got yelled at by wife for asking what her top three diphthongs are

Shart

(This is a subtweet of Docker on MacOS, if anything)

Spent a whole year at Render with an M2 and didn't hear my fan once. Join Atlassian with an M4 and suddenly my office is just brr BRRRR brrrrr rBRRRRRRRRRrrr RRRRRRRRRR im hot BRRRRRR BRRrrrrrrr rr RBRRR

In case you're the kind of person heading into the weekend and thinking to yourself "gosh, I wish I had a 15K+ word blog post on error handling to mull over", I've got just the thing for you. https://t.co/zcuZbbYY8r

Coders love vibe coding because it’s just gambling: You give Claude code something you worked on for a long time, it takes some money, the little Claude asshole spins, and either your repo got a shiny new feature or it is fucked beyond repair. But the next prompt could fix it

THANKS, MOM