Quantization is a massive efficiency gain for near negligible drop in quality. If the tradeoff is quantization for an 80 percent price drop I would take that any day of the week.
You may be right that the tradeoff is worth it, but it should be advertised as such. You shouldn't think you're paying for full o3, even if they're heavily discounting it.
I would like the option to pay for the unquantized version. For creative or story writing (D&D campaign materials and such) quantization seems to end up in much weaker word selection and phrasing. There are small semantic missteps that break the illusion the LLM understands what it's writing. I find it jarring and deeply immersion breaking. I'd prefer prototype prompts on a cheaper quantized version, but I want to be able to spend 50 cents an API call to get golden output.
seems like rosetta 2 will be around for a long time, especially considering they are still putting dev effort into game porting toolkit which is heavily dependent on rosetta 2.
A version that tracks the underlying distro better, or even closer to mainline. Current WSL2 kernel is 6.6, kernel is 6.12 or 6.15. Debian Trixie will be 6.12.
strace shows that the sleep program uses clock_nanosleep, which is theoretically "passive." However, if the host suspends and then wakes up after the sleep period should have ended, it continues as if it were "active."
I don't understand. Docker/podman/distrobox/lxc all allow you to do the exact same thing without the virtual machine overhead. I think the real win of WSL is that its a best of all worlds. You get to use Windows with access to every game ever made plus all of the proprietary apps everyone needs to use, with all of the upside of having a full and complete linux command line experience.
You get all of Windows telemetry, vulnerabilities and backdoors, the always fun game of spot the new Advertising opportunity, AI “copilot” spyware I mean feature, updates that reset your machine at will, a terrible UAC model that encourages “just click OK already!”, and dependence on a company that has gone out of their way to prove how much of an unstoppable behemoth they are; and best of all you get to pay for the privileges above.
I know… every year is the year of the Linux desktop… but seriously the AI spyware included was enough to get me gone for good.
It's hard to pick the Windows feature I hate the most, but floating around at the top is Defender. It can't be disabled, at least not easily, and it demolishes IO performance. And Windows update takes the computer hostage, and takes ages to do anything giving no feedback in the process, meanwhile APT can update to a new major version in like 5-10 minutes.
You can setup local and limited user accounts under Windows. Many applications including every development tool out there doesn't need any admin permissions.
Spyware and adware is a government policy / regulation problem. Thanks to GDPR and DMA, using Windows in EU is significantly better experience (try setting a Windows desktop with an EU image). You can remove almost all of the apps including Edge and Copilot. There are no ads in the UI. Neither in Explorer nor in Start menu.
The current process to install windows11 with a local account… is to, press SHIFT + F10 at a screen in the middle of install after the first reboot, enter into the command prompt: ODBE/BYPASSNRO, and disconnect from any internet options, and/or ipconfig disable your networking…
But guess what? Fuck You because that is the old way of doing it now, and now the new command is start ms-chx:localonly
Yes, you get Windows telemetry which enabled fixing bugs without a bug report, you get minimal ads in the start menu (if you're playing "spot the new advertising opportunity" I found it. It's in the start menu. You can stop playing now), AI "copilot" which isn't spyware just because you think it is, updates that ASK you nicely multiple times to update (I don't want to be ableist, if you suffer from a Christopher Nolan Memento-like disability where you don't remember the warnings, you might think it's "resetting at will", but I assure you, it isn't), a great UAC model that's a lot better than "just type your root password into this terminal already, and just hope the binary wasn't hijacked in some way to keylog you, because unlike UAC, there is no visual evidence that you're not getting hacked", and dependence on a company that SV_BubbleTime thinks "has gone out of their way to prove how much of an unstoppable behemoth they are" with no evidence or clarity so they must just be making FUD, and best of all the OS costs so little you can pay it in 8 hours of working as a software developer.
I think the whole Windows client is closed. On macOS though you can use it from the command line just fine (apart from a couple quirks due to a completely different VPN implementation [1]).
"This repository contains the majority of Tailscale's open source code. Notably, it includes the tailscaled daemon and the tailscale CLI tool. The tailscaled daemon runs on Linux, Windows, macOS, and to varying degrees on FreeBSD and OpenBSD. The Tailscale iOS and Android apps use this repo's code, but this repo doesn't contain the mobile GUI code."
and
"The macOS, iOS, and Windows clients use the code in this repository but additionally include small GUI wrappers. The GUI wrappers on non-open source platforms are themselves not open source."
On the other hand, while I suppose the Windows app is probably reasonably straightforward to replicate, I guess it would be much harder to produce an iOS or Android app because of the vagaries of mobile programming.
> I guess it would be much harder to produce an iOS or Android app because of the vagaries of mobile programming.
on iOS you also need a special entitlement that's only available on specific request and only to known developers, so practically impossible for any open source project to acquire.
Yep, Tailscale takes a pretty reasonable approach to that IMHO. Open source on platforms that are open source. I think that works out pretty well because it meets people where they are. For example the people who care about open source (and thus are running linux or android) get their open source needs met, and people who don't care about open source strongly or at all (as evidenced in part by them running closed/proprietary OSes) such as mac or windows users are also met where they are. Of course this also helps protect their business model because then competitors can't just take the open source versions and run off with them, and the number of linux users is quite small compared to mac and windows so it keeps the majority of the client closed while still providing the openness to those who truly care about it.
*In my perfect world everybody would care about open source, but the evidence is pretty clear that only a tiny minority of people actually do, even among engineers
with devices having unified memory now we are no longer limited to what can fit inside of a 3090 anymore. consumer hardware can have hundreds of gigabytes of memory now, is it really not able to fit in that?
6 * 16 is still nowhere near 512gb of vram. On top of that that monster that you create requires hyper specific server grade hardware, will be huge, loud and pull down enough power to trip a circuit breaker. i'm sure most people would rather pay a 30 percent premium to get twice the ram and have a power sipping device that you can hold in the palm of your hand.
According to WMT 2014 benchmarks, GPT 4o and Gemini Flash 1.5 are acceptable. I think Gemini Flash 1.5 8B is the most used right now due to price. In my experience, it is pretty good except in name translation and consistency across text which is why pros use translation dictionaries. It will mostly translate idioms which is nice.
as a person who thought they were arbitrary names when i first discovered them and spent an hour trying to figure out the difference i disagree. it gets even more confusion when you realize that opus, which according to their silly naming scheme is supposed to be the biggest and best model they offer is seemingly abandoned and that title has been given to sonnet which is supposed to be the middle of the road model.
