These tests don't feel practical - That is, they seem intended to collapse the model, not demonstrate "in the wild" performance.
The assumption is that all content is black or white - AI or not AI - and that you treat all content as equally worth retraining on.
It offers no room for assumptions around data augmentation, human-guided quality discrimination, or anything else that might alter the set of outputs to mitigate the "poison"
Specifically, we're implementing AI culled training sets which contain some generated data that then gets reviewed manually for a few specific things, then pushed into our normal training workflows. This makes for a huge speedup versus 100% manual culling and the metrics don't lie, the models continue to improve steadily.
There may be a point where they're poisoned and will collapse, but I haven't seen it yet.
For example, "base" pretrained models trained on scrapes which include generated outputs can 0-shot instruction follow and score higher on reasoning benchmarks.
Intentionally produced synthetic training data takes this a step further. For SoTA LLMs the majority of, or all of, their training data is generated. Phi-2 and Claude 3 for example.
Even without a human, if a LLM has access to code execution it can practice solving coding tasks with runtime feedback. There are many ways a LLM could obtain useful learning signals. After all, we got all our knowledge from the environment as well, in the end there is no other source for knowledge and skills.
https://www.youtube.com/watch?v=HLRdruqQfRk
I love this guy so much and wish he made far more videos.
If most of it is bad but you can get a better AI to tag it as bad, then it's not necessarily a problem.
https://www.lesswrong.com/posts/JbE7KynwshwkXPJAJ/anthropic-...
I can't find a link to the actual clause paper to verify the above link but a few other places mention the same thing about the training data. We don't know if this improved performance is because of synthetic data or something else. I'm guessing even antropic might not be knowing this too.
Granted, one could argue that this only happened because the API version of Claude doesn't appear to use a system prompt. If that's the case, then the LLM lacks any identity otherwise defined by the initial system prompt, and thus, kind of makes one up.
Nonetheless, point remains, it's kind of interesting to see that in the years since the launch of ChatGPT we're already seeing a tangible impact on publicly available training data. LLMs "know" what ChatGPT is, and may even claim to be it.
but time flows like a river, and the more shit that gets into it…
poison does not need to be immediately fatal to be fatal. some take a frighteningly long time to work. by the time you know what’s happening, not only is it too late, you have already suffered too much.
does this sound like anything more than a scary story to tell around campfires? not yet.
you wouldn’t want to alert the victim.
Dude what? That’s a pretty absurd claim. Most generally available models specifically curate their inputs for the express purpose of avoiding AI garbage induced collapse. It’s literally on their cited reasons for avoiding ai generated data as inputs.
This is the part that I don't really understand. Isn't this basically an evolutionary algorithm, where the fitness function is "whatever people like the most" (or at least enough to post it online)?
People rarely generate 10 pieces of content with AI and then share all 10 with the world. They usually only share the best ones. This naturally filters for better output.
Are they saying that evolutionary algorithms don't work?
Precisely.
Whether content is AI-generate, ghostwriter-generated, monkey-on-keyboard-generated, etc...presumably it is implictly filtered by value/quality.
Garbage AI outputs won't be as popular as good AI outputs. (And the same is true of human ones!)
That this happens doesn't surprise me, but I'd love to see a curve of how each organic vs machine content mixe ratio results in model collapse over N generations.