OpenAI shelves Arrakis AI: All you need to know about model named after desert in ‘Dune’

What was Arrakis AI?

The new AI model codenamed – Arrakis AI, was expected to make AI inference cost-effective. Reportedly, OpenAI halted its development due to some quality constraints. The now-shelved AI model was expected to be as powerful as GPT-4, but cheaper when it came to deployment. This was because it was partly developed based on the sparse principle.

Sparse principal is essentially a neural network architecture that uses sparse representations of the input data to improve the efficiency and accuracy of the model. In simple words, only parts of a neural network are used to process input, unlike traditional models that are dense models with their entire neural network active. 

Reportedly, OpenAI began developing Arrakis AI last fall and started training it sometime in the spring. According to reports, the team that was working on the model realised that the performance was not adequate. But they worked on tweaking it for nearly a month. However, their attempts were futile, and eventually, the model was killed. Inexplicably, the sparse principle worked wonders in the early tests but failed on the larger model. 

Although Arrakis AI has been shut down for now, OpenAI may use its work on the model in other upcoming models. It is learned that OpenAI later shifted its focus to models such as Vision and Gobi. Immediately after Arrakis failed, the researchers began working on updating GPT-4 into a version that is faster, possibly the GPT-4 version 0613.

It is to be noted that AI models that power ChatGPT are costly and resource-intensive. They require massive computing power from cloud providers like Google, Amazon, and Microsoft. According to a rough estimate, ChatGPT costs $7,00,000 to operate for a day. 

Other models in the pipeline

Even though OpenAI is the undisputed leader in AI at the moment, the company is bracing itself for some stiff competition from Google which is working on its AI model Gemini. Google is reportedly planning to launch Gemini this fall, and OpenAI has already enabled text-to-image capabilities to ChatGPT by allowing access to Dall-E 3 to ChatGPT Plus users.

Meanwhile, there are murmurs about the new AI model codenamed Gobi that OpenAI is working on. Gobi is expected to be multimodal from the start. Speculations are also rife that Gobi could eventually be GPT-5.