Now Chat GPT has emerged as a benchmark for his versatility and learning ability. Wrongly, many use it as a search tool instead of Google: it was not created for this and the results are visible. However, it would appear that the chatbot's responses are best produced for a fee. That's right, right with a tip (like the one we give to the restaurant waiter).
ChatGPT and virtual tips: truly fascinating how AI learns from reality
As we know well, artificial intelligence models like GPT-4, on which ChatGPT is based, learn from a huge amount of training data poco immense. This data, somehow, made the chatbot behave like a waiter, improving their answers if they were given a tip.
The hypothesis (and demonstration) that a virtual tip can influence ChatGPT's performance emerged almost by accident, but quickly gained credibility through experiments conducted by a few curious users. One of these, Theia Vogel, started as a simple test, offering a virtual tip to ChatGPT and observing an improvement in responses in terms of length and detail.
Read also: ChatGPT will go further with GPT-6 and GPT-7. OpenAI reveals plans for the future
Subsequently, other users, like Abram Jackson, replicated the experiment, confirming a similar improvement in ChatGPT performance. To find out more, Vogel structured a more formal experiment. He asked ChatGPT the same questions in several rounds, varying the offer of a tip: none, 20 dollars and 200 dollars. The results showed a proportional increase in the length of responses in relation to the amount of the promised tip.
This phenomenon raises intriguing questions about ChatGPT's learning mechanisms and its ability to adapt to external stimuli, even if of a non-concrete nature like virtual tips. The possibility that AI has learned the concept of tipping and its impact on human behavior from sources such as online forums and discussions opens up new perspectives on cognitive and adaptation of linguistic models.
Despite the interest aroused, OpenAI has not provided official explanations regarding this behavior. Industry specialists, in an attempt to provide an explanation, suggest that AI may simply reflect behaviors and trends observed when analyzing large sets of online data. This highlights ChatGPT's ability to 'learning' and adapting in ways that may not be fully within the developers' control.
