I’ve seen posts of people installing iPadOS on iPhones and it works pretty much as you describe, you just plug it in to a monitor and have a desktop computer
Example:
www.reddit.com/r/iphone/s/G...
I think the model labs are doing this to some degree. Claude definitely tries to maintain a high information density in convos. It leans heavily on terms and comes up with new terms to refer to things. Maybe less from a token efficiency perspective, but useful for readability I think.
I love the mention of subtly conflicting values! Having all the values aligned feels too easy and unlikely to be inspiring.
What is it?
The only failure mode I’ve experienced is having the wrong audio input device selected 😅
I use Apple voice memos on my phone which hasn’t failed me yet! It also has transcription built in which contains some mistakes but LLMs can often infer from context.
On my laptop I use Superwhisper which also saves transcription and raw audio so you can always transcribe it again if needed.
Hahaha, you mean the pink overlays?
Back on that virtual try-on shit ✨
This looks cool! I’ll be there 🌊
Interesting to see that Rust is still harder for LLMs than other languages, and this matters on hard tasks! It’s not automatically the best choice when you need something fast and reliable.
Yes, the performance gains are mostly Vite’s accomplishment. But even if this was a regular refactor (which I’m pretty sure it isn’t), that’s a lot of API to cover.
I don’t have a deep understanding of what Next.js does, but the message of “we reimplemented the entire API surface of this heavily iterated 10 y/o project and made it 4x faster in one week” is just absolutely wild
Wow, I didn’t know this exists! Can’t keep up with this stuff 😅
I love that this exists!
From subjective experience, Sonnet now feels just as slow as Opus. Maybe not in raw tokens, but in outcomes.
I’d love to see model labs start measuring this more.
It’s interesting to see the variety of ways you interpret this, but you’ve responded three times to this post now, FYI!
does this ever help you?
I'm afraid I'm going to be reposting this every day until the far side of the singularity.
Work on the things you want to see more of!
Work on flourishing!
Work on things that are freeing!
Maybe we should rename “man pages” to “hu pages” then 🤭
Wdym? Humans do too, right?
And that documentation can and often should be different depending on target audience: human or AI.
Humans need a story and intrigue, AI will grind through dry docs all day. But if it works for humans, it will almost certainly work for AIs.
This is so true and I'm happy it turned out this way. Humans + AI both benefit.
The reason why it's MORE important now is that LLMs forget everything while humans learn from trial and error. We need to compress all that learning into documentation.
Looks like Sonnet 4.6 is much less token efficient, which brings it close to Opus cost level. A bit disappointing!
This is so awesome. The fact that I can turn my expert coding agent into a really good teacher by dropping some files in a folder is just wild.
Sonnet 4.6 offers strong performance at any thinking effort, even with extended thinking off. As part of your migration from Sonnet 4.5, we recommend exploring across the spectrum to find the ideal balance of speed and reliable performance, depending on what you’re building.
Ooh, I like this direction. I’ve been experimenting with keeping thinking disabled for straightforward tasks, mainly for speed.
I really wish model labs would start posting benchmarks on speed and token efficiency. Can’t we just measure the time/tokens to complete existing benchmarks?
Especially interested to see if smarter models can be faster on certain tasks by needing fewer tool calls.
How can I see the lexicon for this? I’d like to use it to learn about the app’s features
Let’s make some virtuous software today
