What’s Good for the Goose, AI Training Edition
Stephanie Palazzolo, writing for The Information (paywalled, alas):
Researchers at OpenAI believe that some rival AI developers are
training their reasoning models by using OpenAI’s o1 reasoning
models to generate training data, according to a person who has
spoken to the company’s researchers about it. In short, the rivals
can ask the o1 models to solve various problems and then use the
models’ chain of thought — the “thought process” the models use
to solve those problems — as training data, the person said.
You might be wondering how rival developers can do that. OpenAI
has explicitly said it hides its reasoning models’ raw chains of
thought due in part to competitive concerns.
But in answering questions, o1 models include a summarized version
of the chain of thought to help the customer understand how the
models arrived at the answer. Rivals can simply ask another LLM to
take that summarized chain of thought and predict what the raw
chain of thought might have been, the person who spoke with the
researchers said.
And I’m sure these OpenAI researchers are happy to provide this training data to competitors, without having granted permission, in the same way they trained (and continue to train) their own models on publicly available web pages, without having been granted permission. Right?
★
Stephanie Palazzolo, writing for The Information (paywalled, alas):
Researchers at OpenAI believe that some rival AI developers are
training their reasoning models by using OpenAI’s o1 reasoning
models to generate training data, according to a person who has
spoken to the company’s researchers about it. In short, the rivals
can ask the o1 models to solve various problems and then use the
models’ chain of thought — the “thought process” the models use
to solve those problems — as training data, the person said.
You might be wondering how rival developers can do that. OpenAI
has explicitly said it hides its reasoning models’ raw chains of
thought due in part to competitive concerns.
But in answering questions, o1 models include a summarized version
of the chain of thought to help the customer understand how the
models arrived at the answer. Rivals can simply ask another LLM to
take that summarized chain of thought and predict what the raw
chain of thought might have been, the person who spoke with the
researchers said.
And I’m sure these OpenAI researchers are happy to provide this training data to competitors, without having granted permission, in the same way they trained (and continue to train) their own models on publicly available web pages, without having been granted permission. Right?