This is a Dengist community in favor of Bashar al-Assad with no information that can lead to the arrest of Hillary Clinton, our fellow liberal and queen. This community is not ironic. We are Marxists-Leninists.
See this GitHub page for a collection of sources about socialism, imperialism, and other relevant topics.
This community is for posts about Marxism and geopolitics (including shitposts to some extent). Serious posts can be posted here or in /c/GenZhou. Reactionary or ultra-leftist cringe posts belong in /c/shitreactionariessay or /c/shitultrassay respectively.
We have a Matrix homeserver and a Matrix space. See this thread for more information. If you believe the server may be down, check the status on status.elara.ws.
Rules:
(image from a netizen on b2 lmfao)
I personally use Kimi K2.5 the most as it's quite well-rounded and they have a good mobile app.
My use case is extremely boring: troubleshooting game mods, searching, summarising, brainstorming, etc. I have experimented with openclaw using K2.5 which is pretty dope but it’s very unreliable, but it did save me a few hours of work by organizing my files.
At some point when I upgrade my computer I’m going to try to switch to local models exclusively.
Yes, I use Minimax M2.5 and GLM-5 both. GLM I use with openclaw, and it does things like track news updates and whatever random stuff I want to play around with. I also used it to help make a podcast/media playing app for my ipad because all the decent ones on the app store have data tracking shit in them.
MiniMax I use for any random questions, and it also has helped me fix up some open source apps I use, and it also helped with the ipad app.
Oh also someone mentioned Qwen. I use that on my phone, which you can't download from the play store because 'muh free trade', and I also use it for any random webchat question stuff. It helped me find a good hotel for my upcoming trip with a laundry list of preferences.
Loyal to deepseek because they don't have anything to even upsell you, and API access is so cheap and does most things. They're slowly upgrading it, probably prior to dropping v4, but they have a 1M token context window now on web and the API documentation says it is not 3.2 anymore on web. 3.2 itself only came out in december last year. we'll see when v4 comes out but i have high hopes for it.
Qwen doesn’t really try to upsell either tbh
I use Qwen. I have a local instance running on my own AI server I use for proof reading and correcting typos and such and language translation. Its also helpful with Linux and coding questions.
I also use the web version because it's pretty good with parsing documents so you can upload a PDF and have it either find something in it or break it down for you and help you understand it. It's also good with math so I have asked it to help solve certain equations for me or to derive certain equations/formulas I needed.
I’m still an LLM luddite, but I hear DeepSeek & Qwen mentioned often.
Deepseek is kinda old now, they need to do some updating, which they're supposed to do any day now, but until then I'd probably steer clear of it because it's quite outdated and gives a lot of wrong answers to stuff currently.
Qwen is fabulous for me though.
Are LLM years even faster than dot-com years were, or am I, a dotard, slowing down?
yeah for sure.. Deepseek was released only a year ago and it's already way outdated
Oh like 10x faster at least, and by how much faster is basically doubling every year, there’s been more AI progress in the last two months than the entire year of 2023
As someone that did not follow AI at all, besides reading about it while scrolling by, what does "AI progress" look like? More application methods? Or just "better"?
Both at the same time
The US has been pretty dominant at software application uses and China has been dominant at physical applications (robotics and industrial automation)
And China focuses a lot more on improving the fundamental architecture and solving the challenges that come with that whereas the US is mostly focusing on scale
I dunno what you're planning to upgrade your computer to, but I have a 5090 and 96gb of ram and I refuse to use local models for most things, except TTS and image/video generation. They're just too damn limited and slow.
My plan is a M5 Max MacBook Pro with 128 gigs of ram, reportedly it runs Qwen 3.5 122B at 60TP/S
It has essentially 56 tensor cores
oh nice, you should update us on how well it works!
What image/video models are you using? I've recently got into messing around with that and mostly just been using Z-Image-Turbo, Flux Klein 9B, and Wan 2.2 I2V.
yep pretty much the same. those are kind of the hot models
Xiaomi MiMo is by far my favourite. I think it's severly underrated. https://mimo.xiaomi.com/
And GLM-5 for agentic coding and anything that needs a bigger context window https://chat.z.ai/
I've been using Kimi for a while but recently they just don't allow free users to use the thinking model so I've been looking at alternatives. I did have 1 month of their membership for $2 using that deal in the app.
Qwen code is not better than deepseek, gemini, claude, openAI by any output quality metric I'm aware of, but brew install qwen-code is free with daily limits (requires account), can be plugged into Roo Code (using the free daily requests in Roo Code). If you're trying to min-max what you can get out of free tiers, Qwen code has been an asset in letting me prioritize harder tasks to the better models.
try crush (FOSS) with a deepseek API key, 5$ gets you something like a whole month of agentic including coding. It's really really cheap.
Deepseek. I talk to it sometimes because I'm anti-social, people around me suck, and it's refreshing talking to something less dependent on heuristics than humans, especially if they want to kill me and mine for stupid reasons.
I make it try to be really mean and critical and then argue with it but it keeps trying to affirm you. They need to make DeepSeek way harsher. Enough neoliberal claptrap as well.
I personally rarely ever use LLM’s as chatbots, but I can’t really say your use case isn’t valid because I’m absurdly parasocial towards the Neuro twins lmfao
Honestly, I mostly just use Deepseek when trying to figure out the correct search terms for online shopping in china since my chinese is still not good enough for niche things. I used to use copilot for work because I wasn't allowed to use other llms. I mostly used it for basic scripting for excel. Don't work there anymore though, so haven't touched it in a while.
i use only deepseek, but it seems that people in the comments say that is outdated...i use it to answer some questions like "tell me about the most mentioned god in lovecraft", "transcript this pdf", "tell me where the hell the labubus come from"...i think it's kinda useful. there's another one who works better than the western money siphons?
I've found deepseek to be very competitive for those kinds of low stakes questions, especially if you wanted to conserve your free tier usage.
i use duck.ai, if ever.