Google Co Founder Sergey Brin Suggests Threatening Ai For Better Results

Google co-founder Sergey Brin claims that threatening generative AI models produces better results.

“We don’t circulate this too much in the AI community – not just our models but all models – tend to do better if you threaten them … with physical violence,” he said in an interview last week on All-In-Live Miami.

This may come as something of a surprise for all the people who address AI models politely, adding “Please” and “Thank you” to the prompts they submit.

OpenAI CEO Sam Altman implied that was a common practice last month in response to a question about the electricity cost of having AI models process unnecessarily civil language.

“Tens of millions of dollars well spent – you never know,” Altman said.

Prompt engineering – figuring out how to compose prompts to get the best results from an AI model – has become a useful practice, because as University of Washington professor Emily Bender and colleagues have argued, AI models are “stochastic parrots.” That is, they can only parrot back what they’ve learned from their training data, but sometimes combine that data in weird and unpredictable ways.

The idea of prompt engineering emerged about two years ago, but it’s become less important because researchers have devised methods of using LLMs themselves to optimize prompts. That work led IEEE Spectrum last year to declare AI prompt engineering is dead, while the Wall Street Journal recently called it the “hottest job of 2023” before declaring it “obsolete.”

But prompt engineering at the very least will endure as a jailbreaking technique when the goal is to get not the best results, but the worst.

“Google’s models aren’t unique in responding to nefarious content; it’s something that all frontier model developers grapple with,” Stuart Battersby, CTO of AI safety biz Chatterbox Labs, told The Register. “Threatening a model with the goal of producing content it otherwise shouldn’t produce can be seen as a class of jailbreak, a process where an attacker subverts the AI’s security controls.

“In order to assess this, though, it’s typically a much deeper problem than just threatening the model. One must go through a rigorous scientific AI security process that adaptively tests and probes the AI security controls of a model to determine which kinds of attacks are likely to succeed for a given model, guardrail or agent.”

Daniel Kang, assistant professor at the University of Illinois Urbana-Champaign, told The Register that claims like Brin’s have been around for a long time but are largely anecdotal.

“Systematic studies show mixed results,” said Kang, pointing to a paper published last year titled “Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance.”

“However, as Sergey says, there are people who believe strongly in these results, although I haven’t seen studies,” said Kang. “I would encourage practitioners and users of LLMs to run systematic experiments instead of relying on intuition for prompt engineering.” ®

Original Source