Positive users back the builder's view that similar Qwen outputs reflect normal model behavior rather than sinister distillation, while negative users criticize Anthropic for complaining about a common and acceptable practice.
Based on 6 visible X reactions from 11 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@xlr8harder Legal or not, properly licensed or not, breaking the TOS or not, you're not going to find a lot of people that will sympathize with them over this. Training on data that isn't yours is built into the bones of this technology and the only way it is remotely economical.
@xlr8harder I agree distillation is fine! Which is why I think Anthropic should stop complaining about it. 2. it probably helps for Chinese. 4. I doubt this happened here, without some evidence to the contrary I think this points to including Qwen and Deepseek conversations explicitly.
@xlr8harder 4 is why the claims of distillation based on similar outputs and names are stupid. They all know how each other talk and nature mimic each other. You dont even have to train it. It's just inevitably going to happen
@xlr8harder I don't know why the community note wants it to be a "glitch" or something sinister - what you said is probably the truth; this is just the way it works.
@xlr8harder their barbarous distillation vs our noble known glitch
@xlr8harder fully agree and even funnier: 'that word, i do not think it means what you think it means'
1. Distillation is fine. 2. Anthropic is unlikely to need to distill from qwen. 3. Even if they did, qwen's license allows this. 4. training on broad internet data likely produces priors that Chinese speaking assistants are probably qwen.
Positive users back the builder's view that similar Qwen outputs reflect normal model behavior rather than sinister distillation, while negative users criticize Anthropic for complaining about a common and acceptable practice.
Based on 6 visible X reactions from 11 accounts; directional sample.
Ask a question below.
Published answers will appear here.