Meta declined to comment on the leak, but company spokesperson Ashley Gabriel issued the following statement: “While the model is not accessible to everyone, and some have attempted to circumvent the approval process, we believe the current release strategy allows us to balance openness and responsibility.”
ChatGPT is based on next-generation technology that has been available for some years and learns to mimic human language by identifying patterns in vast volumes of text, the majority of which was scraped from the web. OpenAI discovered that by adding a chat interface and offering an extra layer of machine learning including human input on the bot’s replies, the technology became more competent and eloquent.
OpenAI may gain a significant benefit from the data supplied by users engaging with ChatGPT or services built on it, such as Microsoft’s new Bing search interface. Yet other businesses are striving to replicate the fine-tuning that made ChatGPT possible.
Stability AI is presently financing a project named Carper AI that investigates ways to teach similar chatbots. Alexander Wang, CEO of Scale AI, a business that performs data categorization and machine-learning training for several technology companies, said that many clients are requesting assistance with fine-tuning similar to how OpenAI created ChatGPT. “Demand is quite overwhelming,” he says.
Wang believes that the ongoing efforts will result in the emergence of many more capable language models and chatbots. “I believe there will be a thriving environment,” he asserts.
Sean Gourley, CEO of Primer, a firm that offers AI tools for intelligence analysts, including those employed by the U.S. government, and advisor to Stability AI, anticipates that numerous projects will soon develop systems similar to ChatGPT. “The rumor around the water cooler is that this required roughly 20,000 hours of training,” he adds of OpenAI’s bot’s human input process.
Gourley thinks that a project requiring several times as much training would still cost a few million dollars, making it reasonable for a well-funded startup or a major technological business. OpenAI’s fine-tuning of ChatGPT, as described by Gourley, is “a miraculous advance.” But, it is not something that cannot be recreated.
What transpired following OpenAI’s April 2022 announcement of DALL-E 2, a tool for producing sophisticated, visually beautiful graphics from text, may foretell the future of ChatGPT-like bots.
OpenAI established protections on its picture generator to prohibit users from creating sexually explicit or violent images, as well as those including recognized faces and made the tool available to a restricted number of artists and researchers out of concern that it may be exploited. Yet, because the concepts behind DALL-E were well-known among AI researchers, comparable AI art tools quickly emerged. Four months following the release of DALL-E 2, Stability AI launched Stable Diffusion, an open-source picture generator that has been integrated into several applications and modified to produce images disallowed by OpenAI.
Hugging Face is a startup that hosts open-source AI projects, some of which were developed by Stability AI. Hugging Face’s CEO, Clement Delangue, believes it will be able to reproduce ChatGPT, but he is unwilling to estimate when.
He states, “No one knows, and we are still in the learning process.” He says, “You never truly know that you have a good model until you have a good model.” That may be next week or it may be next year. Both are not too distant.