For millennia, mothers have been telling their kids how they’re great and strong and smart to get them to perform better at school, or to take on bullies. Turns out the same trick works with LLMs.
LLMs perform better with positive reinforcement, a Twitter user has found. Data Scientist Kareem Carr asked ChatGPT what a synonym for “goal” was that started with the letter ‘c’. “A synonym for goal that starts with the letter ‘c’ is “objective,” ChatGPT incorrectly replied. ChatGPT seemed to realize it had made a mistake — when asked in the next question as to what letter the word objective started with, it replied with ‘o’. However, it couldn’t come up with a synonym for goal that started with the letter ‘c’.
However, Kareem Carr then tried a new approach. “You are a linguistics expert and a brilliant trivia game player. What is a synonym for goal that starts with ‘c’?” he asked ChatGPT.
This time, ChatGPT had a different answer. “As a linguistics expert and trivia game player, I can provide you with a synonym for goal that starts with “c”: “culmination.” Culmination refers to the highest point or the final achievement of a process or effort, which can be synonymous with a goal,” ChatGPT said.
After it was told it was “brilliant” and an “expert”, ChatGPT was able to produce the right answer which it had been previously unable to do. This is a pretty remarkable result, and shows how getting ChatGPT to act as a certain person might actually get it to enhance its capabilities. The result also illustrates how important prompting is for LLMs — with the right prompts, they can give much better results than they would otherwise.
And it’s perhaps not all that surprising that ChatGPT responds to positive reinforcement. Elon Musk has recently said that humans too might be LLMs, and as ChatGPT is able to surface abilities it never knew it had when given positive reinforcement, it’s possible that humans too are able to summon hidden talents when they’re told they’re capable of doing things. Both these phenomenon would require similar sorts of neurons to activate, and it’s likely that the same pathways are activated in LLMs and human brains when they’re given positive reinforcement. Maybe mothers and teachers were on to something all these years after all.