But that's not really what is alleged though? You just ask 100k questions to Gemini 3 and then you make a dataset for your own RL out of it and use GPT as a judge (though imho llama3.3 is still a really good cheap judge too, due to great instruct)

Fully automated RL set by spending tokens on Gemini... This is what "AI building AI" means?

Google says attackers used 100,000+ prompts to try to clone AI chatbot Gemini

ch0k1

But that's not really what is alleged though? You just ask 100k questions to Gemini 3 and then you make a dataset for your own RL out of it and use GPT as a judge (though imho llama3.3 is still a really good cheap judge too, due to great instruct)

Fully automated RL set by spending tokens on Gemini... This is what "AI building AI" means?