TikTok’s parent company, ByteDance, has been secretly using OpenAI’s technology to develop its own competing large language model (LLM). “This practice is generally considered a faux pas in the AI world,” writes The Verge’s Alex Heath. “It’s also in direct violation of OpenAI’s terms of service, which state that its model output can’t be used ‘to develop any artificial intelligence models that compete with our products and services.’”
Training one AI with the output of another AI will just make an even crappier AI.
Ever since that paper about “model decay” this has been a common talking point and it’s greatly misunderstood. Yes, if you just repeatedly cycle content through AI training over and over through successive generations, you get AIs that lose “fidelity.” But that’s not what any actual real world training regimen using synthetic data does. The helper AI is usually used to process input data. For example, if you’re training an AI to respond in a chat-like format, you could take raw non-conversational text (like a book) and have the helper AI create a conversation about that content for the new AI to learn from. Or to take a real-world example, Dalle3 was trained by having a helper AI look at pictures and create detailed text descriptions of them to use as the caption to associate with the image when training.
OpenAI has put these restrictions in its TOS as a way of trying to “pull up the ladder behind them”, preventing rivals from trying to build AIs as good as the ones they have already. Fortunately it’s not going to work. There are already open LLMs that can be used as “helpers” without needing OpenAI at all. ByteDance was likely just being lazy here.
There is actually a whole subsection of AI focused on training one model with the output of another called knowledge distillation.
I’ve watched Multiplicity enough times to know you get a slightly less functional copy.
Not necessarily: there have been recent works that indicate that filtering effects of fine tuned LLMs greatly improves the data efficiency (e.g phi-1). Further, if you have e.g. human selection on top of LLM generated content you can get great results as the LLM generation can be used as a soft curriculum, with the human selection biasing towards higher quality.
