m_w_
I’m all for recreating software, especially to be faster, cleaner, smaller, and more bespoke - but trying to one-shot a whole suite will never work. Building these foundational office-y tools requires an immense amount of testing and attention.
OpenRouter is not remotely close to a representative source for inference distribution.
It would be interesting to see the reverse, how long does each model take to get >50% WR against bots of various tiers.
Fwiw, what it's trying to replace (GitKraken) is slow, clunky, and resource hungry, so it's not hard to beat.
The lights are generally not reliable on the 949 - particularly the radar display on the front-left (vumeter per channel, basically), so that’s also worth a service, if you can.
After that it’s the caps, as you probably know. The thing is with the quad ones, there’s literally two of everything in there.
Best of luck with restoring and selling it.
Listing “5” for an email address, and then "4" for DNS and "6" for a gender much later in the article is perfectly emblematic of all the nonsensical decisions therein.
[1]: https://learn.microsoft.com/en-us/answers/questions/349797/u...
Yeah, you said it. I've also got a QX-949, but luckily no vintage refrigerators.
As much as the over-built, last-a-lifetime quality is a big component of owning of these old appliances, I can imagine how that fades when it's an object that weighs 800 pounds instead of 80.
Pretty insane. I suppose it lends further credence to the idea that anything that can be shown to be correct can be done by a model.
More than that though, I completely disagree that anything I'd ever choose to listen to is elevator tier / background noise. Just because it is playing while other things are happening does not mean it never receives focus or active listening.
As far as lo-fi goes, I must recommend Leavv. Mind Garden and Tales of a Flowing Forest. Both pre-AI music albums that are lovely to work/study to.
Wish I had this on hand like this three or four years ago when I was working on VR avatar software - we had hard-coded skin tone options at the time - would have been a cool addition!
Given the extremely competitive releases of GLM 5.2 and DeepSeek V4 (both pro and flash), I don't think there'll be appetite for it.
It seemed for a time that Google had finally gotten the ball rolling, but I'm doubting that more and more as time passes. We'll see what happens with 3.5 pro I suppose.
Hard disagree here.
For new music, Discover Weekly is great, if you take some time to engage with it on a routine basis. Even better, if you have an artist/genre you already like, the Fans Also Like or Discovered On will link you to other artists and playlists. Super easy to go down rabbit holes of new artists and playlists.
As far as connecting with others, I do like the spotify DMs (in-app share), the friend activity tab, and particularly the share attribution. When you share a song via link (url with ?si=), you'll permanently be linked to it. For a number of my favorite songs, I see "From {friend}" at the bottom while listening. Makes me feel super connected to friends I've bonded with over music.
That said - I would (again, maybe naively) suppose it's not hugely different - much of the work I do occurs in code where many people have and will work on it, and where the size of the codebase dwarfs model context windows.
In that case, I feel the same - current frontier models, when properly oriented to a task, with some assist on the big-picture thinking - are more than capable of generating good code that can slot into big codebases with many moving pieces. Of course, I'd have to point to other people's work to defend this, but I think that's still pretty reasonable especially against the declared "LLMs are worse than useless for generating code".
I don't think my job was ever to type fast, nor do I make any claim that I'm ontologically better than someone writing code by hand - but for what I need to do, I'm way faster now.
You might think he's an AI shill, but I was pretty compelled by simonw's post and idea - "your job is to deliver code that works" [1]. I think I can deliver more code that I can prove to work, faster now. The productivity is nice, but as I've said on other parts of the thread, it's also just fun to spend more time thinking, less time hitting the semicolon on my keyboard?
[1]: https://simonwillison.net/2025/Dec/18/code-proven-to-work/
Edit: should have used [1] in the first place instead of an asterisk.
Sometimes, I will review every line, test the front-end in a staging environment, verify the backend contract, et cetera. Over time, though, I realized that many of these reviews just didn't result in any necessary changes. The current model (with guidance/claude.md/etc) was able to one-shot the task.
Not to overly personify, but imagine how you might treat a junior colleague. You start by reviewing everything they do with a microscope, later you review the broad-strokes, and eventually, for low-stakes or well-scoped tasks, you just play with the demo and the ticket and approve it.
Otherwise it's not materially different than a pre-AI world - you've got sample I/O, test cases, hand-review, look at the application on different screen sizes, contrive some edge cases, test against a spec if there is one - et cetera.
For me, it is not about the syntax nor the mechanics of typing though. My enjoyment comes from thinking about problems and breaking them down, or thinking about what architecture would best serve them. I guess I'm meant to be a systems design guy, so I'm lucky that AI-coding fits this well (AI models have limited context windows, relative to the size of codebases, so doing the big picture thinking is still important, and fun for me).
Library-type work has mostly been side/toy projects, although fwiw, with a standard/spec on hand (CommonMark for example), I'm also happy w/ the output. It's often possible to "close the loop" and have the coding agent autonomously iterate until the standard is adhered to.