OpenAI invites everyone to test a new AI-powered chatbot, with fun results

On Wednesday, OpenAI announced ChatGPT, a dialog-based AI chat interface for its GPT-3 family of large language models. It is currently free to use with an OpenAI account during a testing phase. Unlike the GPT-3 model found in OpenAI’s Playground and API, ChatGPT provides a user-friendly conversational interface and is designed to severely limit potentially harmful output.

“The dialog format allows ChatGPT to answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests,” OpenAI writes on its announcement blog page.

So far, people have put ChatGPT through its paces, finding a wide variety of potential uses while exploring its vulnerabilities. It can write poetryto correct coding errors with detailed examples, produce AI art prompts, write new code, expose on the philosophical classification of a hot dog sandwich, and explain the worst-case time complexity of the bubble sorting algorithm… in the style of a “wise-talking fast-talking guy from a 1940s gangster movie”.

ChatGPT also refuses to answer many potentially dangerous questions (related to topics such as hate speech, violent content, or how to make a bomb) because the answers go against its “programming and purpose”. OpenAI achieved this through both a special invite it adds to all inputs and using a technique called reinforcement learning from human feedback (RLHF), which can fine-tune an AI model based on how humans evaluate its generated responses.

Mastering the offensive inclinations of large language models is one of the main issues that has limited their potential usefulness in the market, and OpenAI sees ChatGPT as an important iterative step in the direction of providing a safe AI model for everyone. .

And yet, unsurprisingly, people have already figured out how surround some of ChatGPT’s built-in content filters using quasi-social engineering attacks, like having the AI ​​frame a restricted output as a fictional scenario (or even a poem). ChatGPT also seems to be vulnerable to rapid injection attacks, a story Ars revealed in September.

Like GPT-3, its dialogue-based cousin is also very good at completely making things up in an authoritative way, like a book that does not exist, including details of its contents. This represents another key problem with large language models as they exist today: if they can breathlessly compose compelling information, how can you trust any of their results?

Yet, as people have noticedChatGPT output quality seems to represent a noticeable improvement compared to previous GPT-3 models, including the new text-davinci-003 model we discussed on Tuesday. OpenAI itself says that ChatGPT is part of the “GPT 3.5” model series which was trained on “a mixture of text and code from before Q4 2021”.

Meanwhile, GPT-4 rumors continue to swirl. If today’s ChatGPT model represents the culmination of OpenAI’s GPT-3 training work in 2021, it will be interesting to see what GPT-related innovations the company has been working on over the past 12 months.