ChatGPT’s Most Charming Trick Is Also Its Biggest Flaw

Like many others people over the past week, Bindu Reddy recently fell in love with ChatGPT, a chatbot who can answer all kinds of questions with astonishing and unprecedented eloquence.

Reddy, CEO of Abacus.AIwhich develops tools for coders who use artificial intelligence, was charmed by ChatGPT’s ability to respond to requests for definitions of love or creative new cocktail recipes. His company is already exploring how to use ChatGPT to help write technical documents. “We tested it and it works great,” she says.

ChatGPT, created by startup Open AI, has become an internet darling since its release last week. Early adopters enthusiastically posted screenshots of their experiences, marveling at its ability to generate short essays on just about any topic, craft literary parodiesanswer complex coding questions, and much more. This has prompted predictions that the service will make conventional search engines and homework obsolete.

Still, the AI ​​at the heart of ChatGPT isn’t, in fact, very new. It’s a version of an AI model called GPT-3 which generates text based on patterns it has digested from huge amounts of text gathered from the web. This model, which is available as a commercial API for programmers, has already shown that it can answer questions very well and generate text from time to time. But getting the service to respond in a particular way required creating the right prompt to power the software.

ChatGPT stands out because it can take a naturally worded question and answer it using a new variant of GPT-3, called GPT-3.5. This tweak unlocked a new ability to answer all kinds of questions, giving the powerful AI model a compelling new interface that almost anyone can use. The fact that OpenAI opened the service for free, and the fact that its problems can be a lot of fun, also helped fuel the chatbot’s viral debut, as did some imaging tools using the AI ​​did. proven ideal for creating memes.

OpenAI hasn’t released full details on how it gave its text-generating software a naturalistic new interface, but the company did share some information in a blog post. He says the team provided human-written responses to GPT-3.5 as training data, then used a form of simulated reward and punishment known as reinforcement learning to push the model. provide better answers to sample questions.

Christopher Potts, a professor at Stanford University, says the method used to help ChatGPT answer questions, which OpenAI has already shown, seems like a significant step forward in helping AI handle language in a more relevant. “It’s hugely impressive,” Potts says of the technique, despite thinking it might complicate his job. “It got me thinking about what I’m going to do in my classes that require short homework answers,” Potts says.

Jacob Andreas, an assistant professor who works on AI and language at MIT, says the system looks likely to expand the pool of people who can harness the language tools of AI. “Here’s something presented to you in a familiar interface that gets you to apply a mental model that you’re used to applying to other agents – humans – that you interact with,” he says.

Comments are closed.