Brahvim

Remember how Ollama makes a copy of the LLM first? LLaMA.cpp doesn't do that.

I love this lol

Yep!

Hmmm. How does ChatterUI compare?

This + that guy who did GLaDOS on the Pi :>

*Insanoty*.

Yeah, screw 'em for censoring us humans with BOTS. *BOTS!*

Ah, yes, *Benchmaxxing*.

I got nearly 6 tokens a second running Gemma 3 1b q4_k_m on my PHONE last night! (CPH2083, Oppo A12, 3 GiB RAM, some PowerVR GPU that could get 700 FPS simulating like 300 cubes with a Java port of Bullet Physics in VR. Not exactly amazing these days. Doesn't even have Vulkan support yet! Phone is a *SUPER BUDGETY*, like 150 USD, from 2020. Also by the way, Android 9.) Firefox had worse performance rendering the page than the LLM's own speed LOL. Did take nearly 135 seconds for the first message since my prompts were 800 tokens. I could bake the stuff into the LLM with some finetuning I guess. Never done that unfortunately. (On my 2021 HP Pavilion 15 with a Ryzen 5 5600H, 16 GiB of RAM, and a 4 *GB* VRAM GTX 1650 - mobile, of course, a TU117M GPU - THAT runs this model at 40 tokens a second, and could probably go a lot faster. I did only dump like 24 layers though, funnily enough.) Most fun part is how much this phome struggles with rendering Android apps or running more than one app in the background LOL. There barely is more than 1 *GB* of RAM ever left. And it runs a modern LLM ***fast*** (well, at least inference is fast...!).

Note: Experimenting to learn is *the* way!

XOR is used in assembly to clear registers, too!

Not really. I have never used Mimo, but long in the past, SoloLearn felt like a good option back then. IMO learning programming on the go has *never been a real method*. Best you can learn *that* way, are concepts. Or entire languages, considering you know a language or two already. Avoid *not* trying out things for instant feedback. "Tutorial Hell" is a frequent problem mentioned in failure stories on this subreddit!

W3Schools was good at SEOing their way to the top of the page on almost every search engine. The content on the site, though? *That* does suck. I personally recommend learning programming using LLMs, but people on this subreddit seem to often do *that* the wrong way... IMO W3Schools is the worst resource for anything unless you are **a very casual learner** who does not want *any* depth at all **or** does not ask questions, both of which, at least the latter, are *poison!* Their site serves as a good reference sometimes, but I'd never recommend learning there. # My recommendation ...would be: - To learn fundamentals ***very welll***, - To learn any libraries with your IDE's autocompletion and documentation, - **Only** *getting started* with LLM-generated code. *Have fun ^-^!*

I am sorry I give you no answer here, but if you're making a game engine, you might want to look into [ https://dataoirneteddesign.com/dodbook ]! IMO, an **extremely** important resource for low-level performance-cautious developers!

I moved from Ollama, to LM Studio, to Ollama, to textgen-web-ui, to llama.cpp itself, and then, to KoboldCPP. I like how it is more automated and *still* fast.

Yes, it **is** satire **till**... it's about NVIDIA's *server-space moneyz*.

Since everybody is offering their experience with Firefox, I choose to offer mine too: Chromium-based browsers like Chrome become *insanely* slow with ads and trackers for me *very quickly* (takes at most a month). Firefox's performance has always stayed extremely similar to what it was on the first day. The *Android app's* data and cache storage use remains *way* lower than Chrome's. Other than occasionally not supporting a few MIME types - always video formats, or videos that simply *refuse* to load (because the server does not provide them anymore) - Firefox has never given me any problems and has always been the faster (especially on slower machines, mind you!), more freeing, more private (after tuning), more stable, and almost as importantly - allows me to use it as an advertisement-free and tracker-free browser. Firefox's *Private Browsing*, sometimes (if not *often!*) manages to disallow even YouTube from what seems to be *identifying* me - YouTube being a site that takes very little *refreshes*, or *days*, since any access *over a VPN* (including Firefox Private Browsing access over a VPN!) to identify me, given that I am an Android user, accessing YouTube via Firefox, on a laptop running Debian, *via my phone's hotspot*. All my internet access on my laptop is over my phone's hotspot because I live in a rural area. ### Point is - Firefox Private Browsing *alone* seems better for my privacy on every site, even data-hungry ones like YouTube. Adding a VPN on top of it seems to magically ruin it. I am otherwise a very frequent user of YouTube, over Firefox, often logged in, often for hours, almost never on Tor - where I am not logged in, of course *By "identification", I'm referring to videos and channels accessed in Firefox Private Browsing, **painfully**, being recommended again-and-again, usually on the top of my feed, for **weeks**, when I access YouTube over an ordinary Firefox browsing session, where I'm logged in*. *My point here, is that Firefox's Private Browsing helps significantly delay, if not entirely eliminate such weeks-long, persistent, annoying video recommendations from YouTube*. *It probably helps to know the content I view here: it is always composed of either video game trailers, or heavily political news, or YouTube channels I dislike, but need some information about from the channel page, or new channels that I need to check out a video or two from.* Other uses of this Firefox feature involves accessing sites without logging in, or logging in with a different account temporarily without any care for tracking or establishment of relations.* *I'm never logged into a site under Firefox Private Browsing. Also note that I usually change IP addresses (thanks to mobile internet's heavy use of DHCP) between Firefox Private Browsing sessions and ordinary Firefox browsing sessions.*

...I mean, KDE Plasma and many F-droid apps convince me *otherwise*...

I'm still waiting for HTML to be the standard LOL.

But then... Xbox Game Pass, them combining Xbox and Windows, **buying so many studios**... ?!?!??

Oh no. *Mono* is a FLOSS implementation of dotNET's runtime.

*"Gee-gee-yuu-eff"?*

Me too. And I find it cute \^-\^.

...Yeah, so, essentially WSL, right?

Somebody would make a kernel module, right?

This, I agree with *the most*.

<:(

*Did you mean:* **Wayland and the desktop environment of choice** instead?

# **My point was about *familiar defaults* instead of *tried-and-tested* or *good* defaults.** The OSs are doing fine, we can say. So are desktop environments. They already have good defaults. Good, good enough, or "strong" familiarity **is possible**. 100% familiarity is of course, impossible. *We're talking about default settings for LibreOffice here.* Nobody is going to have problems with 2+ settings that are already easy to reach - please see the article. # Also, *switching an OS is already a big decision*. Peer-pressuring one into switching **is bad**. Not letting one know the consequences of a big decision is bad... If they are not prepared to configure *even this little bit*, (two settings, mind you! If I were to count the actual number of clicks with *no use* of the keyboard *and thus no shortcuts*, it'd be `9`!) they've made a decision in a hurry, or one that they did not intend to make - and LibreOffice literally shouldn't responsible. ~`10` clicks. Anyone who can read, aim with a mouse pointer, and click, can do this in *well under* a minute. If you installed LibreOffice freshly right now, and tried this according to the article, you'd see that these metrics *are similar* to your results.

I've heard this *everywhere* and *tried my best to rely* on ONLYOFFICE myself (how many Windows users will have our *"software **does** need maintenance!"* mindset...?), but doing this has failed. Yes, ONLYOFFICE is simpler. It works as Microsoft customers want it to. However, not everything is available... I don't think LibreOffice will be the problem if the themes and layout are customized to be in Microsoft's style. No lawsuits making their way in even if they copy the layout right? Like, these customizations probably exist, right? They're just not the default is all. If we as a community educate new users to *absolutely do all of this* customization for the Microsoft experience they expect, or *make the bold decision to change the default preferences* from what we've gotten used to / liked / *relied on* [since it was what we had] to *what the Microsoft customers expect*, we could win! After all, LibreOffice *is* the one with more functionality, right? ONLYOFFICE surely is simpler, and more Microsoft-like with their UI and better first-attempt compatibility - however - LibreOffice is the do-it-all that actually *achieves* it all. The UI (and possibly *UX*) is *the only* pain-point for Microsoft customers, really...

TBH LibreOffice *is* better given that ONLYOFFICE cannot play presentations. Cloud features et etcetera are awesome but LibreOffice also has themes and stuff. `libreoffice-gtk3`, `libreoffice-plasma`. Yay!...

*I, sir, love your username.*

Okay. Enough mentions of FreeBSD in my life. I need to say this instead now: *"All OSs not fully custom and not Windows are either openly-based on Linux or distributed in secrecy whilst basing on FreeBSD - Unix's modern forms run the world even today"*.

*Sofa! Made of **not!** Plastic bottles!*

"Kroot"?

Funny how it's always the Google Pixel phones that people buy to modify with some fork of Android. *Similar vibes here!*

> *`ssh git.orhun.dev`!* *I mean...*

"orhunp"? Your username reminds me of `nohup`.

Yes, it is true that you will be writing code on top of already-built APIs (such as `java.util`'s). Also, linked-lists of any kind are far less used than plain arrays or `ArrayList`s. It's better for CPU cache. Node-based data-structures are in general not. Trees and graphs can be awesome, though - especially graphs. ...Not something I should mention but I guess it's nice as at least a fact.

*Did you mean:* "AMD Radeon"?

Like `ip` did for multiple thongs? I got it.

This IS the **real answer**.

Reaction and... ***PFP***.

"Janky"? You're simply choosing to go *all "lo-tech"!* *(...Not as in using *only* stuff from lo-tech, the brand though, of course!)*

*Sorry, if this feels like the wrong place to ask, but:* How do you even run these newer models though? :/ I use textgen-web-ui now. LM Studio before that. Both couldn't load up Gemma 2 even after updates. I cloned llama.cpp and tried it too - it didn't work either (as I expected, TBH). Ollama can use GGUF models but seems to not use RAM - it always attempts to load models entirely into VRAM. *This is likely because I didn't spot options to decrease the number of layers loaded into VRAM / VRAM used, in Ollama's documentation*. I have failed to run CodeGeEx, Nemo, Gemma 2, and Moondream 2, so far. How do I run the newer models? Some specific program I missed? Some other branch of llama.cpp? Build settings? What do I do?

What? It's clearly Llama! ...Or Phi, if you prefer slight protection from societal issues - another way to say, well, "want alignment".

**Disclaimer:** Since you don't happen to know of software development terminology, I can only try offering what might feel oversimplified. Sorry if you don't "feel" any of it. Experience in designing as well as actually building and failing teaches developers these things. *Okay. So:* "API" is a quite generic term. It doesn't refer to just web APIs (which are built on HTTP and used to talk to a cloud service). Parts of an application can also be called "APIs". In my case, "API" refers to something like a library, framework, or similar - any code that has a frontal side that faces the programmer, and a hidden side that does actual work - a backend. Software is made with *modules* these days to encourage this kind of stuff. To allow a programmer to split parts of their code and let each part have a separate responsibility. Say you have an app you're building with the Android SDK. *An SDK* provides you with actual executable programs such as tools to convert source code into what your target platform can run, package it for said platform, testing tools (e.g. emulators), and finally, **APIs**, which give you a mind-map of the entire system's capabilities and what to expect them to do. **My point was about concurrent programming**, that if two APIs (no matter they're from the same SDK or not) were to be used in an application's code, an LLM may not be able to keep track of their usage, given that said APIs ran their backends concurrently - asynchronously, then an LLM *may not be able to provide robust safety features absolutely necessary for such a case*. ...And many bugs or inefficiencies are found in such cases ("cases of multithreading" which were not initially planned) after rigorous testing. Even if you find them, they are difficult to fix - *even as a human*, since they often involve changing your application's design to consider them - because you didn't consider them in the first place! Popular programming paradigms such as object-oriented programming suffer from such issues a lot and still get preference for their initial ease of use and thinking ease for certain regions of an application. An LLM wouldn't ask you to redesign your entire app in such a scenario. It can't design "pipelines" like an engineer. It'll follow your instructions *till the end* and provide immediate solutions. It can't suddenly ask you to rewrite your app with all of the code in mind. Even "simple apps" can be pretty hard to maintain! (PS >!Please *consider* sharing your app's source code so people can positively criticize it and stop your app from having any issues. **I'm sorry for sounding ANY AMOUNT of *offending* at all!** You see, bugs and design issues *creep in* over time. They aren't received as part of a trade or only "born and fixed". Things can go wrong anytime. Oh, and inefficiency is a thing!...!<)

Waiting for when they initially remove programmers from jobs, realize that prompting for code takes extreme effort and that LLMs can't solve problems with enterprise constraints, start hiring programmers to "fix" these issues that breaks code consistency with LLM-generated stuff that the AI can no longer make perfect sense of - all whilst hobbyists have the time to perfect their projects and give rise to a golden era for software development that no LLM might be able to learn from soon enough.

Imagine writing a highly concurrent system with two medium-size APIs - OO or not. Debugging that shouldn't be possible for an LLM, 0-shot, yet. And hopefully "simple apps" don't turn out to unacceptably unsafe practices like querying a DB from the frontend. Also, build systems! I wonder if LLMs can write build scripts well enough. Highly custom ones? Probably not too easily.

As an 18 year old I FULLY agree with the "2 hours of massaging"! And C. OH MY GOD, C! Asking any GPT model to write C in an established project is, well... And... contextual understanding kinda' sucks for older models. It can't work with multiple files reliably at all. A GPT is *awesome* as a rubber duck, an established-library explainer, and "non-toxic stackoverflow".