To be fair, I am very wary of using the platform... and this isn't super suprising since it is not a real business model. But, hosting your own local LLM is a feasible alternative
Reddit users have discovered a 50 question limit against R1 if using their website. If you use your own front-end to make an API call instead, the per-million-token pricing model kicks in.
Have to wonder what language was in the prompt to trigger any sort of ban.
The training data, they say, was generated using other language models, which apparently is good and accurate enough to use for training. This will go far!
One of the interesting things is that they used OpenAI to help train the model... so OpenAI "stole" from the public to create their models, and DeepSeek "stole" from them to give it back for free.
There may be some truth largely because the models are distillations of previous LLMs, whether their own, Qwen, Llama, etc. It's not like they started from scratch.
I think it is just the media acting ignorant to what this product actually is. I was coding with it today, it was somewhat working okay. There were a few hiccups.
Then, I got banned from using DeepSeek after an hour of using it, citing I violated their terms of service. Huh? Building a react app?
TBH, ive been firmly anti LLM AI for the most part, but now seeing that there is an open source one that can be run locally, im....kinda tempted to try it out
From what im seeing, you can run the deepseek model locally without any internet connection and this also gets around any censorship people see when using the online model
Thing is, why would you now pay 200/month with openai for an OS "free" equivalent
You can bake censorship into a model. That said it will tell you how to make "firey things", but there are some questions it refuses to answer. But yes, running it locally means it can't phone home.
Some actually kind of are. Simon Willison corrected me on this yesterday. Some are published under and compliant with https://opensource.org/ai/open-source-ai-definition. I don’t know precisely which but I trust Simon on this stuff.
I don’t think OSAID is sufficient and agree with others that it was pandering to big companies who wanted a definition that would allow them to claim “open” with proprietary control
It would be nice to have all of the pieces required to produce the same set of weights that the company did, and the standard does stop short of that. But it goes further than I thought it did, which was a nice surprise.
Deepseek-R1 is awesome, but you will need insanely expensive hardware to run the large models. I have fun with the 7B version on consumer hardware ay work, could prolly do the 32B model on a GPU at home.
And the complaining, ‘They stole our data!’ oh, the sweet, ironic Justice. I hope all of their VCs feel the painful, anguishing sting. Maybe they can get back to you with the open-source model they were supposed to be building.
Comments
DeepSeek AI (from China) just released their latest model which is as good or better than OpenAi's latest... and they open sourced it.
Have to wonder what language was in the prompt to trigger any sort of ban.
Kind of a Robin Hood vibe.
Tbf the models don’t usually come with the training data
What are the implications of mass adoption?
It's hard to know what to believe but it does sound too good to be true!
Then, I got banned from using DeepSeek after an hour of using it, citing I violated their terms of service. Huh? Building a react app?
We don’t have a way to validate them (no training data, no reproducibility)
They’re more like “open to use”
Thing is, why would you now pay 200/month with openai for an OS "free" equivalent