China just destroyed the AI market

[gorzek]

DeepSeek R1 is here: https://github.com/deepseek-ai/DeepSeek-R1

long story short: a small Chinese startup has released a reasoning model that performs as well as OpenAI's o1 model.

why is this a big deal?

• this model was built by a 100 person team (small for an AI company)
  
• it is completely open source so you can make your own version any way you want
  
• it requires only 1/50th the resources of o1
  

those last 2 points are basically the death knell for the trillion-dollar AI market. having a model nobody else has is the golden goose these days. ChatGPT dominates because nobody else has models like theirs. well, now anyone can have an o1-quality model... and run it for 1/50th the resources.

the other thing is that the pipelines used to build these models are close proprietary secrets, especially for companies like OpenAI. Deepseek just taught the entire world how to make their own reasoning models from scratch.

this is all fucking amazing. this essentially kills the financial viability of massive closed-source models. womp womp!

i haven't tried it myself yet but there are apparently some GGUF models available for my setup, so obviously i will be trying them.

[gorzek]

i just tried out a Deepseek-R1 distilled onto a Llama model and it's very impressive for something I'm running on my own hardware. It doesn't "sound" Llama-ish, and the reasoning it displays is very solid. I did spot a few hallucinations (which are very, very prevalent on small, local models) but it's at least an order of magnitude better than anything I've run of this size before.

[Aloria]

can it be used to spoof my own voice?
I'm thinking of using it to record cheap audio readings for youtube or something.

[gorzek]

You can use other tools for that, but yeah it can be done.