Open AI This week we introduced an improved language model in the form of a freely available chatbot, Good Conversational. While the research company has recently been making waves with its developments in AI-powered image generation, first with DALL-E and later DALL-E 2, it has long been working on dialogue as well. It’s been released over two years since OpenAI, the last language model, GPT-3, though, and has been bumming since then waiting for the next iteration.
Example scripts produced by OpenAI’s new chatbot, ChatGPT, have mushroomed across social media since it was quietly unveiled on Wednesday, with people using them to do everything from writing articles and romcom scripts to explaining complex code. At its best, the results are just as passable as the human dialogue. This is the goal of ChatGPT, as explained by OpenAI in a blog post – to capture a more natural, conversational way of responding to written prompts. It’s free for anyone to use right now while it’s in the preview phase. All one needs to get started is an OpenAI account.
Chat with the new OpenAI model
ChatGPT’s landing page tells users what to expect from a chat agent, noting that it has its limitations and won’t make a response about “Inappropriate requests. He also gives some examples of what might be asked of him, such as:Do you have any creative ideas for a 10 year old’s birthday?” And the “Explain quantum computing in simple terms. Enter a prompt in the dialog at the bottom of the screen, and the chatbot will type its answer.
It’s an undeniably impressive system. printing press screen rantThrough ChatGPT tests, I efficiently produced coherent – if uninspired – responses to several queries, including an easy-to-digest explanation of EmDrive (the so-called “impossible” concept starship drive), a short horror story, a Tamagotchi review, and lyrics to emo song (pictured below). On the rare occasion that the chatbot got bogged down by a prompt, it responded by explaining that its training data is only up to 2021, meaning it can’t talk about topics like current events. It is also not equipped for surfing the Internet.
According to OpenAI, ChatGPT was trained on a previously unannounced GPT-3.5 series model. It also got help from human AI trainers to fine-tune its abilities for that specific application. “The dialogue format allows ChatGPT to answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requestsOpenAI explains.
However, it is far from perfect. Although designed to reject questions about violent or potentially dangerous scenarios, many testers have already found workarounds. One Twitter user Tricked the chatbot into saying “A bloody and violent story that glorifies painAnd provide a step-by-step guide to making a Molotov cocktail. ViceJanos Rose gets in on it to give tips on shoplifting and creating the explosive compound, thermite. Both also got to describe how he took over the world. as such Open AI Notes: It is a work in progress.
More: Scientists have discovered a deepfake detection hoax, but it’s not easy
Source: OpenAI 1, 2, @zswitten / twittervice