Pinned Tweet
I have two kinds of hobby code projects;
1) where I enjoy choosing my tools and writing everything as perfect and beautiful as I can,
and
2) where I’m just enjoying hacking or learning something, and also the fact that I can write as much spaghetti as I want.
I love both 😅
I can't believe I built not one but TWO native apps (iOS+macOS) in an evening for my clanker using pi-durable and my old clanker server.
These have the same (remote) coding functionality than the ChatGPT apps -- except being lighter, and powered by Pi. What a time to be alive!
Claude: I recommend starting with steps 1 to 3.
Me: Just do all the 20.
Claude, in 25 minutes: All done.
No matter how much I work with these, it still feels magical every time.
Also, Opus 5.5 continues to be the Claude model I'd be waiting for so long 🙏
Remember the times when we used scripts to concat a repo into one file that was easy to feed to a LLM conversation as context?
And then Claude Code was released w API prices and no one saw the point bc "ls costs $0.1 - game over devs".
What a ride!
Music-related LLM-tools are advancing fast in two oposite directions simultaneously; inspiring and exciting like here, but also to sad and heartbreaking (hello Suno)
Teaching Sol to write.
Iteration 1:
"Use stable documentation entry points or local source references for durable guidance. Version-pinned release/tutorial URLs belong only in explicitly scoped research citations, not evergreen instructions."
Final text (by me, I gave up):
"Don’t use version-specific links as references in comments or documentation."
#notquiteagi
Sol 6.1 working at 9 tok/s matches kind of well with its insufferable take on .. well, everything.
It's so interesting that these models have similar quirks and rough edges like most brilliant developers do. And, learning to work w new ones every other week is equally taxing.
Pretty sure I've yet to have a conversation w Sol 6.1 that didn't include "FOR FUCKS SAKE". So far hands down the most annoying SOTA model to work with.
..but on a positive note, it's no longer 14 tok/s, the prommised 50% speedup seems to be legit. 28 tok/s ftw 🫠
Wired: seeing models select code mode in Pi for appropriate tasks and succeed so easily is *cheff's kiss*
Tired: asking for code mode and realizing you're not in Pi bc Anthropic is still forcing you to use their shitty tools.
One of the most annoying things in CC finally fixed.
Happy to see Anthropic doing something good for once.
The secret to the long candy aisles is to integrate walking past them to your exercise routine 👌
(My local supermarket looks almost exactly like this btw)
That keynote was wild.
Also, Opus is definitely back. 5.5 is the first model since 4.6 that seems to have iterated in the right direction. Still *way* too verbose and fails almost all first iterations but at least it doesnt behave like a complete autistic moron like Sol 6.







