eyegor

Don't waste time on bonsai unless you have like 16gb or less ram (ram, not vram). Although I'm pretty sure you could still run this model well on 16+8, and qwen3.6 35b-a3b runs well on ~32gb combined ram at q4. Just offload most of the experts to cpu/ram if you have limited vram budget. With the moe models it doesn't hurt performance too much.

But bonsai is basically trash on real agentic coding work. I've tried both bonsai models, all the quants, messed with the hyperparameters, tried to fine tune (didn't succeed), and it just kept looping or failing to finish tasks. Qwen3.6 35b-a3b is my current go to for agentic work for any systems under 24gb vram, although honestly I'm not sure the dense qwen3.8 27b is meaningfully better. The main difference I see is it does a lot more testing attempts and burns a ton of tokens + compactions, so every task it takes 1.2-2x as long as 35b-a3b does. But the quality of the results is comparable.

I'm surprised no one's complaining about it randomly switching to mandarin. It seems like no matter what I do once every few days it'll go back to mandarin mid thinking trace and then stays in mandarin until I remind it to use English.

This one isn't really a "burnout test" though, that's just what they call it to get clicks. They're just using an oled monitor as their primary for daily work. This represents a pretty normal use case, it isn't a torture test like what RTings does. I don't see a world where I'd buy a monitor and not use it for static content the majority of the time.

Spinning rust is still the only cost effective answer to redundant 10+ tb storage. If you have 4+ disks, use raid 6 or raid 10 and you get decent fault tolerance too. Raid 10 also increases write performance but it's a bit less robust imo and more expensive per gb.

As for speed it's mainly tied to rpm and the market decided no one wants higher failure rates / lower mtbf for marginal speed gains.

If you need speed and capacity, throw an ssd or two in front of a raid disk array for caching with primocache or whatever.

Might as well go back to ye olden days and put 100% of your memory in a giant preallocated block and instead of stack locals you just use the block. No allocation cost at runtime, cheese benchmarks by making the super arena.

On prem beats the heck out of github post Microsoft though... At least you know how to get it working again when someone breaks it. These days with github you expect a weekly 500, a rainbow unicorn error, build failures due to unavailable errors, etc. Last I checked the third party tracker github services were barely pushing one 9 of reliability.

"Chinese repos" is a very charitable interpretation of the Google drive links they used to distribute the os. It seemed like it was on the free plan too, it often didn't work because it tripped the maximum downloads per month limit.

As nice as it looks, I have a lot of trouble believing the "we have magic money, it's free because that's good for business" logic.

    PDFgear is free of charge, and we don’t generate income through any hidden means. We Do NOT misuse or sell user data and we Do Not display ads. Here’s how we keep operations running: 
    We’ve secured investment to cover operational costs, including team expenses and technology like the ChatGPT API. We’re also experienced in optimizing technology usage to manage costs more effectively.

Plans for language bindings? Should be trivial to whip up simpler ones like python or dotnet but I didn't see any official bindings yet.

> even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.

You are far from the embedded world if you think 1 minute here or there is long. I have been involved with many projects that take hours to build, usually caused by hardware generation (fpga hdl builds) or poor cross compiling support (custom/complex toolchain requirements). These days I can keep most of the custom shenanigans in the 1hr ballpark by throwing more compute at a very heavy emulator (to fully emulate the architecture) but that's still pretty painful. One day I'll find a way to use the zig toolchain for cross compiles but it gets thrown off by some of the c macro or custom resource embedding nonsense.

Edit: missed some context on lazy first read so ignore the snark above.

I've used them as a quick way to get rootless configured base images. Not sure if official repos provide those now, but it used to be a big hassle to get things like postgres images running without root in their containers. Although I often had to read through their dockerfiles to figure out the uid setup, where configs live, etc because they were not consistent between the various bitnami images.

If you really want the 80 col experience, I think that was the fortran 77 default. You'd even get compiler errors if you try to exceed it. Of course there were flags to increase the line length limit but don't tell your colleagues.

F# has been a third class citizen for a long time... Last I heard the entire f# team was ~10 people. Pretty sure "find references" still doesn't work across c# and f# (if you call a c# method from f# or vise versa). That also means symbol renames don't work correctly.