Anthropic's latest blog post after releasing Opus 5.5 has a surprising amount of insight on how to make your work faster and better, lets check it out Thank you to WorkOS for Sponsoring! Check them out here: https://soydev.link/workos Source: https://claude.dev/blog/how-we-made-claude-ai-faster/ Want to sponsor a video? Learn more here: https://soydev.link/sponsor-me Check out my Twitch, Twitter, Discord more at https://t3.gg S/O @Ph4seon3 for the awesome edit 🙏
ADVERTISEMENT
I made theo vids 60× faster. Tell gemini to summarize with key points and examples for each
Obsidian didn't get 3x faster because it was already fast.
Tldr. They said Computer, make it 3x faster.
Opus 5.5 remains Theo's favorite model of the week! I think that is 2 weeks in a row now almost a new record for favorite model of the week! *edit* No wait, I lied, he went Opus 5.5 -> Sonnet 5.5 -> Sol 6.1 -> Cancel your Claude sub -> Opus 5.5 best model of the week. (This all happened within the past 5 days)
"Good Quality Engineering" - User of JS for Backends
Theo does the greeting to start the 5h window as early as possible too lmao
>really focused on performance >Uses Js, electron and react native 😂
Here before the obsidian jokes
Watching this video at 3x speed to make my Claude 9x faster
6900 React hooks for the friggin composer?! What are they even doing in there?
I really love the nerd-out videos. This one is great
Bro, you have to talk about Sol 6.1 persistent usage and token limit. It's mind-boggling.
I can tell this kind of topic is what makes you glow. Just look at the happy carefree boy in those old video clips!
Anthropic engineering team considering instructions but not cycles says a lot... Instructions per cycle (IPC) matters A LOT.
I've been doing deep performance work with agents since Claude 4.5, but primarily in systems and graphics programming. I have a pile of platform specific profiling tools (for GPUs) that I've refined during this work, and also the best general frontier models work really well with perf on Linux. The recent OpenAI models are particularly good at using AddressSanitizer and ThreadSanitizer... Which you will be using a lot if you don't have the luxury of working exclusively in Rust. 😂 I find that setting a concrete goal helps agents stay on task, and even just getting the agent to do some Casey Muratori performance napkin math (i.e. how much can this computer actually do, and therefore how fast *should* this be on this computer?) helps a lot. I have a good intuition for how fast things can be, and giving it a concrete goal like "sample to present latency p99 under 1ms, with none exceeding 2ms" causes it to read trace timelines very critically. P.S. it is hilarious that p95 is a "further work" item for them, since on a 120Hz display that is six missed frames per second, which is an outrageously low bar. Claude (CLI) is still an embarrassment compared to the Grok CLI
Should make a video about how shitty Cursor is as well, that piece of software is a joke
Thank you to WorkOS for Sponsoring! Check them out here: /workos
We build this intuition now by literally watching videos like this and then try to go and look for failures in our real codebases, with or without AI. There's no substitute for understanding architecture and enough CS to sniff out and address root causes. This is why we need videos like this. they're still really important
Hey - I LIKE em-dashes, always have, always will :P
Bro is defending himself