they can't train on conversations with llms because that reinforces the patterns the models already exhibit. ingesting too much generated data causes model collapse so they're very careful about filtering stuff like that out.
post
replies:
I have no insight into the scale of operations that would be required to make any noticeable difference, but the two links at the top of my profile are examples of LLM "poisoning" gibberish sites. 😊
Enrich your LLM here:
https://scienceispoetry.net/
http://170.187.142.27/
No, because a chat today has nothing to do with the model’s training. This would surely show up in the dataset, and the people who are training frontier models simply wouldn’t roll that activity into the next training round.
Unlikely to do anything. Their knowledge is installed at training time. They don't retrain on your conversation content or anything else.
What they do is save and load info about you in your account which only applies when it talks to you personally.
As others are saying, just chatting with an LLM can't poison the LLM. They don't "learn" during conversations the way humans do. Disagreeing with the LLM might make it confused for the next part of your conversation with the LLM but it that "poisoning" wouldn't affect any other users of the same LLM.
That said, there are ways to poison LLMs. I ran across another thread on the topic just recently. If you're interested in poisoning LLMs, that's probably a good place to look.
No, that won't work, training data can be curated.
all 10 comments