Yes I use it from time to time, in my repertoire of LLMs. I do not use the app, yet (just like i don't use tiktok), but maybe I will. I know it has limitations around chinese history, but know your LLM masters. It can be hooked into Cursor, as a model. I believe overall, its a good thing, to have more LLMs, and if they really did train it so cheaply, we are going to have an explosion of models, which can only be a good thing. Cheaper = good for us, the end user (and the app developers). And google is responding with gemini already, so the tl;dr is yes, use everything, and competition is awesome.
So long as you don't ask it about the Chinese government, or anything historical relating to China, it works well. Otherwise it will start militantly defending anything China related.
A lot of other models have a liberal leaning, but with R1 for certain questions there is zero nuance.
It's better to wait for a similar model that is trained with different safeguards. This project is trying to do that, but there will be many more github.com/huggingface/open-r1
I've been testing it against Claude sonnet 3.5, and it's bonkers how good some suggestions are, and it will lookup on the internet if it's not finding anything.
Specially neat for low documented libraries (e.g kamal).
Having the reasoning as well useful when you need to correct it a little.
I like it so far. The o1 equivalent R1 is free for 50 messages. But the V3 is still as good. I'm also using the open-source Ollama's distilled R1 model, which is free and easy to set up.
No, won't use it either. I don't use models with an obvious bias.
I operate #watchdog, a chat moderation bot, so I need a model that is mostly unrestricted. Users can and will talk about China, Xi Jinping, Trump, and whatever else and I need to moderate all kinds of content, not just what the CCP approves as "acceptable to ask about"
By the way, I don't just have this complaint about DeepSeek. GPT-4o sometimes rejects prompts for political reasons too or answers in a biased manner, though it's not as severe.
Grok is supposedly unrestricted so I may use it as a fallback in cases where GPT refuses to answer.
EDIT: decided to give it a try locally just for fun: x.com/ben_makes_stuff/status/… - won't use in prod because of the high likelihood of censorship, but I did manage to jailbreak it and get it to talk about Xi Jinping fairly easily haha
Yeah just tried it out - works very good for my task. I just took the same code but only replaced open ai API with deepseek API and got the same results.
👋 Join WIP to participate
(Accidentally posted without post body)
Has anyone been playing with DeepSeek yet? The Chinese AI model that is apparently pretty advanced and dirt cheap to use compared to OpenAI?
What are the pros and cons?
I’ve only seen it mentioned a lot online, I haven’t checked it out yet myself. (Working only on my marketing site redesign at the moment)
Our entire team of 40+ at Raftlabs switched to Deepseek.
Most of them switched from Cursor & Windsurf.
Yes I use it from time to time, in my repertoire of LLMs. I do not use the app, yet (just like i don't use tiktok), but maybe I will. I know it has limitations around chinese history, but know your LLM masters. It can be hooked into Cursor, as a model. I believe overall, its a good thing, to have more LLMs, and if they really did train it so cheaply, we are going to have an explosion of models, which can only be a good thing. Cheaper = good for us, the end user (and the app developers). And google is responding with gemini already, so the tl;dr is yes, use everything, and competition is awesome.
So long as you don't ask it about the Chinese government, or anything historical relating to China, it works well. Otherwise it will start militantly defending anything China related.
A lot of other models have a liberal leaning, but with R1 for certain questions there is zero nuance.
It's better to wait for a similar model that is trained with different safeguards. This project is trying to do that, but there will be many more github.com/huggingface/open-r1
I've been testing it against Claude sonnet 3.5, and it's bonkers how good some suggestions are, and it will lookup on the internet if it's not finding anything.
Specially neat for low documented libraries (e.g kamal).
Having the reasoning as well useful when you need to correct it a little.
I like it so far. The o1 equivalent R1 is free for 50 messages. But the V3 is still as good. I'm also using the open-source Ollama's distilled R1 model, which is free and easy to set up.
Pretty good, in most cases the difference it not noticeable. So much cheaper.
No, won't use it either. I don't use models with an obvious bias.
I operate
#watchdog, a chat moderation bot, so I need a model that is mostly unrestricted. Users can and will talk about China, Xi Jinping, Trump, and whatever else and I need to moderate all kinds of content, not just what the CCP approves as "acceptable to ask about"
By the way, I don't just have this complaint about DeepSeek. GPT-4o sometimes rejects prompts for political reasons too or answers in a biased manner, though it's not as severe.
Grok is supposedly unrestricted so I may use it as a fallback in cases where GPT refuses to answer.
EDIT: decided to give it a try locally just for fun: x.com/ben_makes_stuff/status/… - won't use in prod because of the high likelihood of censorship, but I did manage to jailbreak it and get it to talk about Xi Jinping fairly easily haha
Same here. I'm hesitant to use it for sensitive operations such as reducing as restriction as possible.
Yeah just tried it out - works very good for my task. I just took the same code but only replaced open ai API with deepseek API and got the same results.
I haven't tried the new model yet, but I'm planning to do so this week, primarily the API.