According to rumors, Amazon wants to develop its own competitor to GPT-4 and offer it via the AWS cloud. The “Olympus” model could contain up to two trillion parameters.

OpenAI has GPT-4, Google has PaLM 2 and Anthropic has Claude. Until now, AWS has always taken a different approach with Bedrock, on which customers can build their own models, and as a result it lacks a flashy LLM in its portfolio to demonstrate its capabilities in generative AI. That could change: Reuters has learned from its sources within Amazon that the AWS cloud division is busy tinkering with “Olympus”.

According to Reuters, this model could have up to two trillion parameters. If this number is correct, Olympus has one of the largest and most comprehensive LLMs developed to date. For comparison, OpenAI’s GPT-4 has over a trillion parameters. The driving force behind the project is Rohid Prasat, who also oversaw the development of Alexa in the past.

The more the better

What does the number of parameters say about the capabilities of an AI model? In this context, a parameter is a set of configurations that determine how the model processes data. The parameters record, among other things, how each input a machine receives affects the output. The rule of thumb is: the more parameters a model has, the more it should be able to do. The fact that tech giants are competing with each other to develop models with more and more parameters is the AI ​​equivalent of measuring who has the biggest.

Amazon will most likely make Olympus available on the AWS cloud in addition to its own Titan models and those of third-party providers via Bedrock. Olympus must then become the new spearhead of this AI offering. All of this should be treated with caution. We haven’t heard anything formal about Olympus from the Amazon warehouse.