But it's gotten much better.
4,910 karma · joined March 25, 2015
GitHub: https://github.com/jasonjmcghee
But it's gotten much better.
This model is specifically trained on this task and significantly[1] underperforms opus.
Opus costs about 6x more.
Which seems... totally worth it based on the task at hand.
[1]: based on the total spread of tested models
I’m thinking it’s still llama / dense decoder only transformer.
We could put them in a dedicated tag:
<script type="text/skill+markdown">
---
name: ...
description ...
---
...
</script>
For all the skills with you want on the page, optionally set to default which "should be read in full to properly use the page".And then add some javascript functions to wrap it / simplify required tokens.
Made a repo and a website if anyone is interested: https://webagentskills.dev/
I imagine this is before and after- not just after.
As in, they aren't just making lines long and removing whitespace (something models love to do when you ask it to remove lines of code)
As far as I can tell, this doesn't use it.
https://developer.apple.com/documentation/SwiftUI/Applying-L...
https://developer.apple.com/documentation/technologyoverview...
Also, for the vibe that this is going for, very surprised the title bar was left in.
---
Code has lots of claims that something is done and it isn't.
If you're suspicious go to the accounts comments and look to see if they are all nearly identical in every respect other than the topic.
Most are:
It's cool you did <thing you said in post>. So how do you <technical question>?
But what percentage of users of a document editor would say "don't install pdf stuff on my computer without asking, I don't need to export to pdf"
Installing dependencies for popular features is very much the norm. It's mainstream software.
The same complaint would be made for VSCode and Jetbrains - the most popular IDEs
> Dynamic code loading
This was only true for Claude Code for a while. Codex was poor and Gemini was unusable.
Since then Codex has gotten quite good.
Although i do use cmd+shift+r for global replace
Cmd+Shift+V - Stacked clipboard, you can start typing to search or hit a number to choose what to paste (keeps everything you've copied/cut inside jetbrains for a while)
Cmd+Shift+E - Recent locations, you can start typing to search - shows little buffers of where you've been recently
Cmd+Shift+A - Action tab of the command palette - fuzzy search for any command (really the only shortcut you need, other than maybe Shift+Shift for main command palette shortcut)
--- Through the Action bar...
Local History / Local History of Selection - you can start typing to search quite far back the history of all changes of the current file or selection - you can also right click a folder or the project and do the same. Much finer grained than git.
The general concept of being able to search for something and edit directly in the buffer of the search results.
Honestly amazes me you'd put so much effort into brand and not do copywriting yourself.
> It predicts which ones you meant.
> It ranks them.
> It understands you.
This is so good I want to know whether someone generated this or wrote it by hand.
What quant are you using? How much ram does it have?
It's amazing how much foundational prompting and harness matters.
So cool seeing these come to life.
Also a really well written piece.
I know the time it takes to get something to feel this good.
Really fantastic work.
The prompt processing sped up.
Not the output generation.
M4 was notoriously slow at this compared to DGX etc.