OpenAI is close to releasing new AI models that will be 100x more powerful than GPT-4, with OpenAI Japan teasing GPT-Next, and it's 100x more powerful than GPT-4 without wasting more computing resources.
The new codenamed "Strawberry" AI model was teased alongside "Orion", with Strawberry ending up as GPT-Next and expected by the end of the year, while Orion will be unleashed in 2025. OpenAI's new GPT-Next aka Strawberry has something researchers call "System 2 thinking".
What is System 2 thinking? This will allow GPT-Next to take the time to deliberate and reason through problems, versus just predicting longer and longer sets of tokens to complete its responses. System 2 thinking has impressive results: scoring over 90% on the MATH benchmark, a collection of advanced mathematical problems, reports Reuters.
Strawberry will also have the ability to generate high-quality synthetic training data, addressing the challenges across AI development: there's not many diverse, high-quality data for training models. OpenAI's achievement here could see a big boost in capabilities of GPT-Next, and gets ready for even more advanced models like Orion.
- Read more: Amazon announces Q just days after ChatGPT maker OpenAI announces Q*
- Read more: OpenAI made huge breakthrough before ousting Sam Altman, introducing Q*
- Read more: Elon Musk tweets 'Q*Anon' after OpenAI's new Q* platform teased, adds 'extremely concerning'
Tadao Nagasaki from OpenAI Japan unveiled the plans for GPT-Next, promising Orders of Magnitude (OOMs) leap of 100x more computational volume than GPT-4, while using similar computing resources. No additional AI GPU hardware will be required, as there have been optimizations to the AI model, and more from the OpenAI team.
- GPT-Next is expected to achieve a 100x performance increase without wasting significantly more computing resources.
- The improvement comes from better architecture and learning efficiency, not just raw computing power.
- GPT-Next will use a smaller version of "Strawberry", an advanced AI system OpenAI is currently developing.
- The slide mentions GPT Next 2024, which while unconfirmed, hints that the model may be released by the end of 2024.