Two innovative models that could redefine the machine learning landscape are about to be launched by OpenAI. These projects aim to advance AI capabilities beyond current boundaries, to move one step closer to Artificial General Intelligence, or AGI.

Formerly known as Q* or Q-Star, Strawberry appears to be more than a chatbot; it’s focused on demonstrating a significant leap in AI reasoning. Sources familiar with the project have told various media outlets, such as Reuters or The Information, that it has demonstrated a remarkable competence in solving complex mathematical problems and in improving logical analysis.

Orion, meanwhile, is being positioned as OpenAI’s next flagship language model. It could potentially replace GPT-4. With the added ability to handle multimodal input, including text, images, and video, it is designed to outperform its predecessor in language understanding and generation.

Underscoring their potential strategic importance, both projects have attracted the attention of U.S. national security officials. This development comes as OpenAI continues to raise capital despite a significant increase in revenue, likely due to the high costs involved in developing and training these advanced models.

Strawberry and Reasoning Power

Despite the endless speculation on the Internet, OpenAI has not officially said anything about Project Strawberry. However, the alleged leaks focus on its sophisticated reasoning capabilities. Unlike traditional models that provide quick answers, Strawberry is said to use what researchers call “System 2 thinking,” which is able to take the time to deliberate and reason about problems, rather than predicting longer sets of tokens to complete its answers.

Strawberry will not only improve its own abilities, but also pave the way for more advanced models like Orion. In the same way that some users today train their own custom models using images generated by Stable Diffusion.

For real-time applications, however, Strawberry’s deliberate processing approach may pose challenges. OpenAI researchers are reportedly working on a “distillation” of Strawberry’s capabilities, essentially scaling it down so that consumers can make massive amounts of inferences at a low cost of computation.

Project Orion or GPT Next

The Orion project is being touted as the ambitious successor to OpenAI’s GPT-4o, with the goal of setting new standards in language AI. It may be called GPT Next, according to a recent presentation by Tadao Nagasaki, executive director of OpenAI Japan. Orion is designed to excel at natural language processing while expanding into multimodal capabilities, leveraging advances from the Strawberry project. And OpenAI claims the leap won’t be incremental.

It would be a technical advantage for OpenAI to train Orion with data produced by Strawberry. However, this technique should be used with caution. Finding the sweet spot where Strawberry can make Orion powerful without compromising its accuracy seems to be key for OpenAI to remain competitive, as researchers have already shown that models begin to degrade after being trained with too much synthetic data.

Another significant advance will be Orion’s native multimodal capabilities. As reported by The Information, this will open up new possibilities for ChatGPT users and put the company in competition with Google’s Gemini. When using ChatGPT or OpenAI’s Playground API, users will interact with this model.

Orion’s development fits OpenAI’s broader strategy to stay competitive in an increasingly crowded AI landscape. Orion is essentially OpenAI’s bid to stay ahead of the curve as open source models such as Meta’s LLaMA-3.1 and next-generation models such as Claude or Gemini advance rapidly.

By Audy Castaneda