Dana Fried on Nostr: npub1q9ma5…lzgdu it's interesting because LLMs self-poison if they ingest too much ...

npub1q9ma5xxhn8m69ztdj3lrefpnarszxgh3gk8wykkq53gfsl80mlssulzgdu (npub1q9m…zgdu) it's interesting because LLMs self-poison if they ingest too much of their output; also more people are using image poisoning tools.

So, there may come a point where we can no longer safely train models on new data.