Most people I know ditched ChatGPT and Claude and are just paying for DeepSeek API nowThe reason? It's just (much) cheaper.
I think they use whatever is reasonable for the task at minimum costeasy task : chinkshard task : anthropic
that's a national security issue
>>109156756Just host it on your own hardware.
fucking chinks taking ERP work away from honest hard-working American AIs
>>109156727It also works, Deepseek knew what I wanted and coder knows security dangers. Literally a better product for less.
>>109156727I ditched gpt and claude for glm 5.2. deepseek is too retarded in comparison
>150 billion dollar investment>no moatdo techbros really...
>>109156756retard take. amerimutt companies are hosting deepsneedv4 on-premaws bedrock, oracle, azure are using it
>>109156791>>109156727Can you use Claude code with it?
>>109156727I called this from the startPeople don't need all the features of a frontier model AI. Even when the craze started, they were having difficulty utilizing all the features of a first-generation AI. If it can provide some reasonable level of conversation and results that aren't full-retard, that's enough for 99% of people. The market was decided before this fight even began.
>>109156756Orange shitter is part of the reason it became this way. The only reason people were interested in AI to begin with was because anyone could use it. Give it an email, and you could be dealing with your own virtual slave with the peak of uncensored data training. Then Orange shitter passed retard AI guardrail laws, then orange shitter passed chip embargoes driving up the price of RAM, then orange shitter began banning entire companies and models if they didn't suck his dick and bow to Israel. etc. etc. etc. He's the reason this is happening.
>>109156727Deepseek is still too dumb desuA bit more expensive, but GLM 5.2 is so much better
>>109156804They decided moats weren't their problem because anyone they were meant to keep out wasn't their customer. This is the actual downside of not providing an entry-level option for your product. Someone WILL provide one, and that someone will eat up the majority of your market.
>>109158913GLM is nice, I like minimax.
What’s the deal with retards who think the world of Chinese AI starts and ends at deepseek? I truly do not understand this phenomenon. It isn’t even that good.
Most people I know ditched their cars for a beater toyota from 1994.The reason? It’s just (much) cheaper.
>>109158937They are goyim who have never heard of anything else. American and European news outlets purposefully hyper-focus on Deepseek because it is the least capable of the Chinese clones. People will use it, notice immediately that it's not matching up even to Haiku from Anthropic, and immediately return to the plantation. Only people who research learn more about the other models.
>>109158937Deepseek is _extremely_ cheap. $20 will get you close to a billion tokens.Things like glm 5.2 are slightly better but nowhere near as cheap, so might as well sub to golemAI
>>109158937DeepSeek is extremely cheap compared to other Chinese models, and practically free compared to Amerimutt models.You can try messing with Kimi and GLM but they're like 10% better for 4x the price.
>>109158979Minimax is great and I will never stop shilling it
>>109158989>Minimax>and I will never stop shilling itwell go ahead, how good is it? it's 3x more expensive than deepseek so if it's only 10 20% better it's not worth it.
>>109156727Do you still have to give up your ID and other shit though?
Why is Qwen so underappreciated?
Chink shills are sort of like the tech equivalent of vegans.
>>109159038Benchmaxxed pile of shitnever had good results from it, be it the large models or the endlessly shilled 3.6 ~30b moe
>>109159038Qwen is useful for extremely tiny modelsMost non-LLM models, like Image gen, TTS, music gen, and video gen models nowadays use some sort of Qwen of various parameters for their text encoder nowadays
>>109158901As someone from /biz/, the winner is whoever can make a "just good enough" model that is integrated everywhere. Like in the 80's, Microsoft was a literal who company but they managed to sell their OS to various computer manufacturers, making them skyrocket to #1 by the 90's. It wasn't about being original or the best, just business logistics. For AI, we need someone who can build tools that will replace all wagies. These tools have to be easy to use and highly accessible. Currently, I would say this is impossible logistically due to the lack of infrastructure. If we could get more power and data centers, then we'll start seeing who can be crowned the AI king.
>>109158937it's the same as the drooling retards who've only ever experienced claude at the high end and nothing else, except here they also larp as if they know anything about the other players
Do I need to do anything before installing Gentoo?
>>109159469Sorry just realised I posted this on the wrong thread
>>109158989MiniMax is the only one that can consistently get work done without saying "actually" 50k times
>>109158876I think so. I've ben using Cline with it. Only thing you need to watch out for is API doesn't process images so you have to turn browser control off. I don't find this a problem in practice, if I see a visual problem I know what element it's in and can give meaningful instructions.
I don't want to be enticed by a false promise so only use the pay as you go API costs, it's no contest in that arena.
>>109156727If you use a Chinese AI model you support China and stagnation.If you use Claude and ChatGPT, you support the free world and progress.It's as simple as that. I know that it's cheaper to use slave labor, but that's how they get you.
>>109156756Oh no! The chinks are going to know that my company has a React app that shows what's available in the cafeteria this week. How will the nation ever survive?
>>109159038Alibaba is about one nanometer higher in reputation than Temu and Wish. Qwen is seen as a fleamarket scammer's LLM. Deepseek is at least up at Wal-Mart level.
>>109159936>free world and progressAhh yes, the american labs, that don't release open source models, that actively employ fear based marketing that has led to them being restricted, that have models that will actively lie to you a degrade their own performance without warningStop being indoctrinated, Anthropic is evil, openai is evil, Google is too retarded to make decent models
>>109156763what hardware?
>>109156727I don't even pay for API, because the chat is already good working with 2-4 source code units
>>109159078That's because your hardware specs are shit. Qwen 36 3.5 35 moe with MTP is required. Anything smaller is worthless. Anything less than Q4 with flash attn is completely lobotomised. For k, v cache you need at the very least k_8 for both, k_4 is completely lobotomised and k_8 is cope already. That means that at the very least you need a 20 gb model meaning 48 GB-64GB+ RAM and 8-16+ vram to see results Qwen is by far the best local model, it can do at least 80% of claude sonnet without costing a cent. Where it gets fucked is the 262k context window is tiny and it's very vram sensitive and gets lobotomised by having shite hardware. Going from q8 to q4 cache means increasing failed tool calls by 30% unusable and just using f16 means basically perfect convos without degradation. Certain online AI providers have been caught cheating providing only shit 32k context windows raping ai performance as well
>>109156727Have you managed to get it to reply in English without having to preface every prompt with "reply in English"?
>>109161452The rig you built 8 months ago using reasonably priced components.
>>109161674AI is honestly the greatest divider between rich and poor. The same model. The poor person gets shite unusable results or has to cope with constant rate limiting and the rich person gets great results or has no rate limits/quota
claude code for orchestration, deepseek for the grunt work. simple as.
I dont use any AII haven't even adopted an IDE yet. I do everything in Nano.It just works.
