More details have been released about OpenAI’s secretive Project Strawberry, including its expected release date and the areas it will specialize in.
A recent report in The information quotes “two people involved in the effort” and goes on to say that Project Strawberry could launch this fall and could be better at math and programming than any chatbot we’ve seen yet.
Previously, OpenAI’s Project Strawberry was thought to be focused on “deep research,” the ability to conduct follow-up research independently, without human intervention. While that still seems to be true, the additional information that Project Strawberry will do math better than we’ve seen before is welcome news to many, as ChatGPT’s relationship with math has been, shall we say, fraught? There’s been plenty of that for a while now memes from screenshots ChatGPT shows that it solves simple math problems incorrectly, leading many to wonder why ChatGPT can’t do basic math. The reason for ChatGPT’s errors in math is that the training data does not contain enough mathematical information, which, as we will see, could be one of the improvements that Project Strawberry wants to implement. Whatever the reason, something was definitely wrong.
Improved ability to solve programming challenges is also welcome, but the scope of Project Strawberry extends far beyond just being better at math. In demonstrations to other employees, Project Strawberry workers have shown how the new AI is capable of more advanced levels of thinking, allowing it to solve puzzles like the New York Times Connections, a complex word puzzle.
Open AI CEO Sam Altman gave Project Strawberry rumors a boost when he tweeted an image of some strawberries growing in a pot on August 7 with no further explanation than the text, “I love summer in the garden.” Since then, there have been many rumors that OpenAI was working on a powerful new LLLM, and had demonstrated a version of Project Strawberry to national security officials.
It’s not yet clear when Project Strawberry will be released, but insiders believe it could happen as early as this fall (September or October), perhaps with a smaller version of it becoming part of the ChatGPT chatbot in ChatGPT 5. If Project Strawberry doesn’t become part of ChatGPT 5, its ability to produce higher-quality data could be used to help produce the massive amounts of training data that Open AI’s next LLM will need if it wants to reduce the amount of hallucinations (aka factual errors) it’s prone to.
ChatGPT recently, and quietly, released an improved version of its cutting-edge ChatGPT-4o model, which is much faster than the previous version, leading many to speculate that this might be what Project Strawberry was all about. Now it looks like the project is set to bear even more exciting fruit.