ChatGPT - If Stack Overflow and Reddit had a child
- Thread starter wtfNeedSignUp
- Start date
-
Want to keep track of this thread?
Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
Create account
- Joined
- Aug 28, 2019
- Joined
- Jan 28, 2018
I'd go a step farther and claim they're probably safer to use as you're outside the jurisdiction of any chinese court and the chinese company providing the LLM is outside the jurisdiction of any western court and the legal cooperation on such nonsense like copyright is nil. That that'll ever change is also highly unlikely. I'd rather have the chinese have my LLM output than western companies, as the chinese having it pretty much assures the various western agencies that can actually reach me will never have a peek. I'm not even doing anything illegal or really even interesting but such distinctions are more important than ever before the more western governments flirt with total mass surveillance where you're guilty until proven innocent regarding thought crimes.
Of course local is better and this is not ideal but this is the clear pragmatic choice.
Think the biggest AI safety issue that needs to be addressed right now is how laws need to be introduced yesterday so it can't be used by both employers and governments to create a surveillance state where your every eye movement is tracked. Of course that's probably not gonna happen.
- Joined
- May 12, 2017
You still don't wanna use deepseek.com as a general rule. DeepSeek seems like something that ought to be run as a local LLM, but hooked up to official servers? Absolutely fucking not. As a general rule: the Communist Party of China is infinitely worse precisely due to the sheer opacity of the Chinese government and the enshrined status of the Communist Party being superior to the government. You give away a ton of network data including metadata to the Chinese government, and LLMs are generally capable of mimicking any given writing style (however imperfectly at the moment). Chinese stuff is useful insofar as breaking Western conventions are concerned, but the moment you're on an official server breaking Chinese conventions, the draconian arm of the Communist Party will clamp down on you.
The difference between Western three-letter agencies and Chinese tech corporations is that the Chinese tech giants are legally beholden to the Communist Party. They have literal agents in every major corporate undertaking. Foreign companies doing business in China need a liaison with the Communist Party, where their liaison picks a local Chinese company, and grants them 51% ownership stake in the partnership. Technology transfers are also required by Chinese law, and don't think your option to refuse would be respected, considering the sheer volume of IP theft and copycattery that goes on in China. A Western three-letter agency, conversely, is nowhere near as hamstrung by the chains of commanding.
You, as an average, ordinary jack-off using deepseek.com to satisfy your LLM needs, maybe buying cheap PC parts off AliExpress, and occasionally ordering a Temu package once in a blue moon, are not subject to Chinese law. Yet the data you provide these platforms will be stored in a Chinese server, it will be used to train Chinese technology on your habits, patterns, and tendencies, and the Communist Party will know who you are in granular detail we can't fathom. Why? Privacy policies for stuff like DeepSeek, Temu, and AliExpress will always tell us foreigners what we want to hear: it'll tell us all the same boilerplate Western privacy policy garbage we've come to expect. In Chinese, however, the story's entirely different and they're far more invasive than they let on.
You might be able to live with your unique device identifiers, your network metadata, your geolocation data, hell even your address, email, and telephone number existing on a Chinese server. Hell, odds are 99% positive my shit's on a Chinese server because I've ordered an assload of crap off AliExpress pre-COVID. Even so, I don't think we should encourage the usage of Chinese network tools without some degree of pseudonymity going forward. I'd sooner trust a FOSS GreasyFork script that modifies ChatGPT's functionality written by some Chinese jack-off licensed under the GPL or X11 licenses than deepseek.com proper. But, I digress.
- Joined
- Jan 28, 2018
I played around with Deep Research & Gemini 2.5 Pro today (finding several low powered ARM boards with specific requirements, like supported well by mainline with no outside blobs, staying inside a specific power envelope of 12 Watts, usable GPU support etc.) and it did a really good job in finding the few candidates that cover all these requirements. I know this because I did this research myself a few months ago, just that it took me a lot longer. I won't speak for it for a specific use but the report it wrote at the end was pretty cool and covered all the bases and even found a few things (like people discussing on a mailing list that a specific GPU driver doesn't work well) I missed in my research.. I'm impressed. Not bad at all. Especially good that it offers all sources so you can double-check everything.
I don't know man, I tell these things nothing I wouldn't be worried about to say in a public place. That's my policy with all online stuff connected to my name. I'm really boring and I'm certainly not up to some of the stuff some people are with LLMs, lemme tell you. I can justify it for myself, YMMV.
I don't know man, I tell these things nothing I wouldn't be worried about to say in a public place. That's my policy with all online stuff connected to my name. I'm really boring and I'm certainly not up to some of the stuff some people are with LLMs, lemme tell you. I can justify it for myself, YMMV.
- Joined
- Aug 28, 2019
Good one, lol. In Palantir we trust.
Interesting story for you:
- Joined
- Aug 7, 2018
What's everyone opinion on ChatGTP 5.1? Supposedly it has 6-8 types of glazing options....
invisibult plus+
kiwifarms.net
- Joined
- Jun 3, 2024
I still find myself going to 4 for a second opinion since 5.1 is more vague. It doesn’t like reading too much into you for fear of going outside the information you give it, so much so it sticks to whatever text you give it religiously. Also the bullet points under bullet points can be annoying to read. But, GPT does take much larger documents than it used to, which is awesome since I don’t have to split them as often anymore.
Last edited:
- Joined
- Jan 28, 2018
Yeah, heh. Well, I felt it was formally correct to at least point it out.
I heard that Server RAM is pretty expensive and in high demand right now, I wonder if these will ever be obtainable for mere mortals. If the price is attractive *eventually* though, it could be pretty cool to run even quite big models at home, especially now that MoEs make CPU interference even with the very big models viable. I currently don't see the economic advantage for myself with the current costs of interference on APIs, but I'm sure that day will come.
I've been using Deep Research by google quite extensively in the last days. I always liked Gemini Pro because at least from 2.0 on, in my impression it was always a model that had an almost unusually low tendency towards hallucinations. I was quite successful with prompts where it basically just summarized me news about certain interests (like LLMs) from time frames of a week or so from sites like r*ddit or github, basically saving myself the googling and having the AI read and summarize everything for me. If you tell it to avoid obvious shilling or "start up news", it impressively usually does manage to follow that also. One problem I noticed is that it's prone to capture the tone of what it reads, getting stuck on specific topics and write a very opinionated piece. The podcast you can generate from the summary is quite good too, but has sometimes the tendency to go in circles. It's pretty cool frankly and although it's not perfect, I can totally see this boosting productivity by quite a bit, you just have to be careful not to fall for the bias of the model and explore the sources yourself when in doubt, which it helpfully links on a per-paragraph basis. Think this is the first normie-approved AI based thing I'd call actually useful.
Last edited:
- Joined
- Jan 28, 2018
Google has released Gemini 3.0
31% on Arc-AGI means they're either benchmaxxing or this is one smart fuck. As always when I post new models, it is too early to tell.
Gemini 3 Deep Think is even smarter.
These numbers only seem to go in one direction. BTW. The baseline human score on Arc-AGI 2 is 60%.
(EDIT: Yep, I think this is the smartest one so far, we're probably in the phase of doubling capabilities again, the exponential growth of LLMs remains unbroken, as per earlier cited paper)
31% on Arc-AGI means they're either benchmaxxing or this is one smart fuck. As always when I post new models, it is too early to tell.
Gemini 3 Deep Think is even smarter.
These numbers only seem to go in one direction. BTW. The baseline human score on Arc-AGI 2 is 60%.
(EDIT: Yep, I think this is the smartest one so far, we're probably in the phase of doubling capabilities again, the exponential growth of LLMs remains unbroken, as per earlier cited paper)
Last edited:
- Joined
- Dec 17, 2019
Almost all of the open source (especially Chinese) LLMs are uncensored. You can jailbreak the big corpo ones (Claude, GPT, Gemini) but if requires finagling and there's always the risk if them banning you for violating their ToS (rare, but not unheard of) since they do log everything you send them. The only major banwave I've seen was for API abuse from stolen keys. Paranoid about you giving the Chinese your data to train their models on? Just use another provider on Openrouter that doesn't log your inputs. Of course, nothing is as 100% secure and private as running a local LLM.
Backpack Knight
kiwifarms.net
- Joined
- Jul 4, 2022
I am trying to figure out how to use local LLM's but I am quite new to software and computers like this. Does anyone recommend any tutorials on how to set up one?
I usually used deepseek or perchance for RPG stuff.
As an uninformed user, deepseek, perchance and ChatGPT can give very similar responses at some points especially in RP stuff. But that may just be me not being used to it.
I usually used deepseek or perchance for RPG stuff.
As an uninformed user, deepseek, perchance and ChatGPT can give very similar responses at some points especially in RP stuff. But that may just be me not being used to it.
grump
kiwifarms.net
- Joined
- Apr 14, 2016
Don't any of these reddit coomers know how to set up a local llm or at least an online service from a smaller more specialized provider? Sure local llms aren't on paper quite as technically good as the major brands but with the stacks of censorship I suspect they've been better for nsfw for quite awhile now. Even before 5 came out.
Its just embarrassing with all the potential freedom and customization out there so many of these people are so technically illiterate and hopelessly dependent on Daddy Altman to throw them some scraps. Guess its no wonder why the Internet is centralizing more and more with how inept and incurious most people are.
Its just embarrassing with all the potential freedom and customization out there so many of these people are so technically illiterate and hopelessly dependent on Daddy Altman to throw them some scraps. Guess its no wonder why the Internet is centralizing more and more with how inept and incurious most people are.
- Joined
- Dec 29, 2020
Has anyone here used Mistral LLM? I think it is the only foundational model that I didn’t use yet, and I’m wondering if the Fr*nch have any competitive edge.
- Joined
- Aug 7, 2018
For those complaining about people not able to make their own AI machine, why isn't there an absolute idiot proof how to guide to make one so that even the most tech stupid can't code don't even know how to code can create one of their own?
- Joined
- Aug 21, 2024
As someone who does some level of research into AI for safety try to always run the model locally if possible. Plus I believe it's also cheaper in the long run.
- Joined
- Aug 7, 2018
How retard friendly is it and where do you get models for free and retard friendly install instructions for them?
Slurred
kiwifarms.net
- Joined
- Jan 29, 2022
Very IMO. Model downloading is handled inside LM Studio (under the hood, they're all downloaded from https://huggingface.co/). Go to the Discover tab, pick a model and click the download button. They're all free. DeepSeek-R1-0528 Qwen3 8b (distilled version of Deepseek) and OpenAI gpt-oss 20B (smaller actually open source version of ChatGPT) seem to be the most popular at the moment.
After your model has downloaded, go to the Chat tab, select the model at the top, and start talking to it.