2222222222 on Nostr: It would appear as though: 1. #DeepSeek (probably not directly but via some ...
It would appear as though:
1. #DeepSeek (probably not directly but via some intermediary entity) used reinforcement leaning to train their own model on ChatGPT outputs, probably via their API.
2. OpenAI's business terms of use prohibit that.
3. DeepSeek claimed that their model required very few resources, but probably because they were using RL on ChatGPT outputs and still, probably used H100's from NVIDIA finding ways around export restrictions (likely via Singapore).
4. Gave the model to the world for free, open-sourced it, but added their own form of pro-CCP censorship/taint to the model.
5. Sneered on social media about it blowing up NVIDIA's value and making models that costs billions to train and develop become a commodity overnight.
6. Gave free access to western end-users to their own self-hosted version of the model to collect their data, which can be shared under their privacy policy with various third parties, presumably the CCP.
🤯
1. #DeepSeek (probably not directly but via some intermediary entity) used reinforcement leaning to train their own model on ChatGPT outputs, probably via their API.
2. OpenAI's business terms of use prohibit that.
3. DeepSeek claimed that their model required very few resources, but probably because they were using RL on ChatGPT outputs and still, probably used H100's from NVIDIA finding ways around export restrictions (likely via Singapore).
4. Gave the model to the world for free, open-sourced it, but added their own form of pro-CCP censorship/taint to the model.
5. Sneered on social media about it blowing up NVIDIA's value and making models that costs billions to train and develop become a commodity overnight.
6. Gave free access to western end-users to their own self-hosted version of the model to collect their data, which can be shared under their privacy policy with various third parties, presumably the CCP.
🤯