The leap between Opus 4.5 and 5.5 is actually absurd
I spent all weekend pushing it to the limit, and I had to fight the urge to stop it every time it started looking at random, unrelated code. “wth? Wtf?” I thought, on repeat.
Turns out, it wasn't random. It evaluated the impact of features that I forgot. The kind of obsessive deep dive that usually takes me two mugs of coffee and a full night of sleep to pull off.
It’s slow. It’s not fun to watch. But it actually does the job.
Just imagine where this will be when it’s running locally in two years. The future is wild amazing
@SuiBuilds It’s fun and fast.
They leverage existing connections to ur GPT account to deliver “hidden” value: mine is updating me of PRs progress or broken tests randomly (I’ve never asked for it)
> Post/surgery photophobia
can’t read claude’s session
Hmm let me tweak the screen brightness
Oh nvm sunglasses were made for this day
What r u guys shipping today?
I'll not put mascots in a b2b product
I'll not put mascots in a b2b product
I'll should not put mascots at a b2b product
💡 I could have a feature flag 'fun mode'
I've a hidden mascot in a b2b product now
@championswimmer Yeah, my favorite workflow right now focus on minimizing effort down to the first PR w structural domain and starting a new session. 'declutter this' on my eternal clipboard
Shaved for surgery.
My bank’s face validation doesn’t recognize me.
Can’t leave my house for 10 days going to an agency is out of the question
I need face validation to open a ticket about the face validation bug.
As a last resort:
“Claude, build a fake real-time beard for me. Make no mistakes.”
@maxleiter@GergelyOrosz You called on your own tweet,Max. You meant to write that blog, but never got to it
Others prob thought the same about building it. But now that’s packaged as “easy API”, the barrier is gone, anyone can use it, whether they understand embeddings tech or not
@michael_chomsky that’s one of the reason I stopped using code rabbit or similar products. You’re wasting token twice.
The fix isn’t at last mile (PR open). You need to tweak your SDLC to run deterministic tests that’re a proxy of a good review