No, he’s challenging the assertion that it’s “trivially easy” to make AIs output their training data.
Older AIs have occasionally regurgitated bits of training data as a result of overfitting, which is a flaw in training that modern AI training techniques have made great strides in eliminating. It’s no longer a particularly common problem, and even if it were it only applies to those specific bits of training data that were overfit on, not on all of the training data in general.
Last time I looked it up and calculated it, these large models are trained on something like only 7x the tokens as the number of parameters they have. If you thought of it like compression, a 1:7 ratio for lossless text compression is perfectly possible.
I think the models can still output a lot of stuff verbatim if you try to get them to, you just hit the guardrails they put in place. Seems to work fine for public domain stuff. E.g. “Give me the first 50 lines from Romeo and Juliette.” (albeit with a TOS warning, lol). “Give me the first few paragraphs of Dune.” seems to hit a guardrail, or maybe just forced through reinforcement learning.
A preprint paper was released recently that detailed how to get around RL by controlling the first few tokens of a model’s output, showing the “unsafe” data is still in there.
I’ve been working with local LLMs for over a year now. No guardrails, and many of them fine-tuned against censorship. They can’t output arbitrary training material verbatim.
paper?
No, training data.
No, he’s challenging the assertion that it’s “trivially easy” to make AIs output their training data.
Older AIs have occasionally regurgitated bits of training data as a result of overfitting, which is a flaw in training that modern AI training techniques have made great strides in eliminating. It’s no longer a particularly common problem, and even if it were it only applies to those specific bits of training data that were overfit on, not on all of the training data in general.
Last time I looked it up and calculated it, these large models are trained on something like only 7x the tokens as the number of parameters they have. If you thought of it like compression, a 1:7 ratio for lossless text compression is perfectly possible.
I think the models can still output a lot of stuff verbatim if you try to get them to, you just hit the guardrails they put in place. Seems to work fine for public domain stuff. E.g. “Give me the first 50 lines from Romeo and Juliette.” (albeit with a TOS warning, lol). “Give me the first few paragraphs of Dune.” seems to hit a guardrail, or maybe just forced through reinforcement learning.
A preprint paper was released recently that detailed how to get around RL by controlling the first few tokens of a model’s output, showing the “unsafe” data is still in there.
I’ve been working with local LLMs for over a year now. No guardrails, and many of them fine-tuned against censorship. They can’t output arbitrary training material verbatim.
Llama 3 was trained on 15 trillion tokens, both the 8B and 70B parameter versions.. So around 1:1000, not 1:7.
I thought he meant LLMs shot out bits of paper like some ticker-tape parade.