Highly misleading. They finetuned an existing model using a different existing model in a process called distillation.
The article is effectively saying “our model only cost $50 to make, plus whatever tens or hundreds of millions of dollars the models we stole from cost.”
That's rookie numbers I trained one in 1min with $1!
How long before congress bans this one too
The underlying research story is interesting, but the way it's written up actively makes it worse.
The researchers based s1 on Qwen2.5, an open-source model from Alibaba Cloud.
Watch me create a racing car for less than $50. Step 1: start with a Mercedes F1 racer...
Trained it to do very basic arithmetic tasks, not to rival OpenAI.