I swear he got so much dumber in the last two weeks. Today, Opus acted downright lobotomized in certain queries, not following simple commands.
Dude. I hear you. I don't like to be a conspiracy theorist - and I'm pretty sure they don't materially change the models like that ...
... but *god dam* it feels like 2025 today.
I swear Ima about to throw my laptop out the window - it's been one of the most frustrating few days of my career.
I think this may be an ongoing problem: The AI is 'poweful but broken' - aka it does not feel like a working project by it's very nature.
It's like a fast car, with doors that won't close, gas pedal 1/2 works, brakes 1/2 work.
Is it faster than walking? Yes. But it's just crazy - and Im beginning to think this is just a property of the AI and that overcoming it woudl require something like 10x the compute .. which is 'not worth it'.
So - we won't get a smooth ride for maybe 25 years, when RAM/GPU become and absolute commodity, and the work we do now is goes through 10 filters/passes before you see an answer so its not 'full of **' 1/2 the time.
There is a new psychology here.
I love development, I love AI - but this is driving me completely crazy.
I always wonder if I am hallucinating and having a bad day or this is actually happening. Yesterday and today Opus 5.5 has been making some strange mistakes. I have a workflow for testing a series of jobs that run in SQL agent. Instead of just generating a script to test the SQL directly as we have done 100 times, it decided to cursor through job steps and attempt to inject the debug parameters into a SQL string pull from the job.
I have noticed a few other silly choices it as made as well.