OpenAI LP Chief Executive Officer Sam Altman has divulged that the startup is currently not training a new version of GPT-4, its most advanced artificial intelligence model.

Altman made the disclosure on Thursday during an event at the Massachusetts Institute of Technology.

GPT-4 is a large language model that OpenAI debuted last month. It can generate text, explain mathematical concepts and perform other tasks. It has also proven highly adept at test taking: during an internal OpenAI evaluation, GPT-4 successfully completed a simulated version of the bar exam.

At the MIT event, in response to a question about a recently published open letter, Altman disclosed that OpenAI is not training a new version of GPT-4. In the letter, tech executives and researchers called for a pause on the development of AI models more advanced than GPT-4. The signatories argued that governments should issue a moratorium if such a pause can’t be quickly implemented.

“The letter is missing most technical nuance about where we need the pause,” Altman told the MIT audience. “An earlier version of the letter claimed OpenAI is training GPT-5 right now. We are not and won’t for some time.”

Altman added that OpenAI waited six months after training GPT-4 to make it publicly available. In the open letter released last month, the signatories argued that development of advanced AI models should paused be for at least six months.

Altman also took issue with other parts of the letter. In particular, he stated that it failed to discuss potential AI safety concerns raised by OpenAI’s current GPT-4 model. “We are doing other things on top of GPT-4 that I think have all sorts of safety issues that are important to address and were totally left out of the letter,” he said. 

Last month, OpenAI announced plans to extend GPT-4’s capabilities with a new plugin feature. By default, GPT-4 can only access information that was included in its training dataset. The new plugin feature allows developers to supply the AI model with additional information that it can use to perform previously unsupported tasks.

According to OpenAI, the feature also enhances GPT-4 in other ways. Developers can now write plugins that allow GPT-4 to not only generate text but also take actions on a user’s behalf.

One of GPT-4’s flagship features its ability to generate software code in response to text prompts. Using the plugin feature OpenAI released last month, GPT-4 can not only generate code but also compile it. Compilation is the process of creating a functioning program that can run on users’ computers.

Although Altman dismissed speculation that OpenAI is training GPT-5, he didn’t indicate the startup is pausing work on other neural networks. Alongside GPT-4, OpenAI offers AI models for generating images and transcribing audio. The startup has also developed neural networks for other use cases, such as configuring robots to perform specific tasks. 

Image: OpenAI