OpenAI is reportedly planning to launch a new AI called “Strawberry” with advanced reasoning capabilities as early as this fall.
OpenAI is reportedly preparing to launch a new AI with advanced reasoning capabilities to boost chatbot activity and raise new capital. The new AI, codenamed Strawberry (previously called Q*, pronounced Q Star), is said to be able to solve math problems it has never seen before — something AI models can’t do reliably. It has also reportedly been trained to solve programming problems and answer non-technical questions. OpenAI, which is facing internal problems and massive expenses, is reportedly planning to launch Strawberry as early as this fall.
OpenAI is allegedly pressuring its teams to launch this new AI very soon.
This isn’t the first time the Q* project has been in the news. Rumors of the mysterious project first surfaced after OpenAI CEO Sam Altman was unexpectedly fired last November. Reports at the time indicated that Altman’s abrupt dismissal was due to disagreements between him and the Superalignment team tasked with ensuring OpenAI’s AI aligns with human values. While these reports have not been confirmed by the company itself, new reports are providing more detailed information about the technology.
A new report from The Information, citing two people familiar with the project, says that OpenAI could release a version of the project, now known as Strawberry, as early as this fall. It is said to be designed to solve novel math problems and optimize programming tasks. Its improved logic should help it solve language-related problems more efficiently when it has enough time to “think.”
In internal demonstrations, Strawberry reportedly solved the New York Times’ “Connections” puzzle. The model could also serve as the basis for more advanced AI systems that can not only generate content but also take action. According to a Reuters report last month, OpenAI has already internally tested an AI (though hasn’t explicitly stated that it is Strawberry) that scored more than 90 percent on the MATH benchmark, a set of math-proficiency tasks.
That would likely be Strawberry, which The Information said has also been shown to national security officials. Strawberry, however, is not expected to be limited to answering technical questions. According to the report, internal OpenAI documents outline plans to use Strawberry’s models for autonomous internet searches, allowing the AI to plan and conduct in-depth searches.
The report also notes that OpenAI is also working on an AI system called Orion that aims to surpass GPT-4’s capabilities with the help of Strawberry. Despite pushes for a quick launch, the report notes that it is not certain that Strawberry will launch this year. However, if it does, it would likely be a distilled version of the original model, offering similar performance with less computing power.
OpenAI has already used this technique for GPT-4 variants since the original model was released in March 2023. OpenAI hopes that this innovation will significantly improve the reasoning capabilities of its AI models. OpenAI researchers say that reasoning is essential for AI to achieve human- or superhuman-level intelligence. Strawberry’s primary goal would be to produce synthetic data for the Orion model.
AI Currently Performs “Mediocrely” When It Comes to Reasoning
While large language models (LLMs) can already summarize dense texts and compose elegant prose much faster than any human, the technology often falls short when it comes to solving common-sense problems whose solutions seem intuitive to humans, such as recognizing logical fallacies and playing tic-tac-toe. When the model encounters these types of problems,he often “hallucinates” wrong information. Chatbots hallucinateis a major challenge for businesses and is now one of the factors accelerating the degradation of the state of the web.
Many AI researchers argue that reasoning, in the context of AI, involves training a model that allows AI to plan, mirror how the physical world works, and reliably solve difficult, multi-step problems. Thus, improving AI reasoning is seen as the key to unlocking the ability of models to do everything (scientific discovery, planning, etc.).
Altman said earlier this year that in AI, the biggest areas of progress will be around reasoning ability. Other companies like Google, Meta, and Microsoft are also experimenting with different techniques to improve reasoning in AI models, as are most academic labs doing AI research. But the challenge is daunting, and there has been no notable improvement yet.
Additionally, researchers disagree on whether LLMs will be able to incorporate long-term thinking and planning into how they make predictions. For example, one of the pioneers of modern AI, Yann LeCun, head of AI at Meta, has often said that LLMs are not capable of human-like reasoning. Strawberry is a key part of OpenAI’s plan to overcome these challenges.
According to people involved in the project, Strawberry includes a specialized method of what’s called “post-training,” or adapting base models to refine their performance in specific ways after they’ve already been trained on generalized data. The post-training phase involves methods like “fine-tuning,” a process widely used in industry and that comes in many forms.
Strawberry is said to have similarities to a method Stanford researchers developed in 2022 called the Self-Taught Reasoner (STaR). It aims to enable AI models to achieve higher levels of intelligence by iteratively creating their own training data. According to one of its creators, Stanford professor Noah Goodman, in theory, STaR could lead language models to transcend human intelligence.
“Using Strawberry to generate higher-quality training data could help OpenAI reduce the number of errors its models generate, otherwise known as hallucinations,” said Alex Graveley, CEO of startup Minion AI and former chief architect of GitHub Copilot. However, previous studies warn that a model can collapse when trained on AI-generated data.
OpenAI Faces Massive Costs and High Turnover
The effort to launch Strawberry is part of OpenAI’s ongoing battle to stay ahead of other well-funded rivals vying for supremacy in conversational AI, or LLMs. The technology also has implications for future products called agents that aim to solve multi-step tasks. OpenAI and its rivals hope the agents could open up more revenue opportunities. Indeed, for these companies, an AI with advanced reasoning capabilities that solves difficult mathematical problems could be a very lucrative application.
But while OpenAI is already struggling financially, sources say Strawberry is more expensive and slower at inference. OpenAI’s business is growing at an incredible pace: its sales of LLMs to enterprises and ChatGPT subscriptions have roughly tripled to $283 million in monthly revenue over the past year, though its monthly losses are likely higher than that.
The company is valued at $86 billion. However, sources say that OpenAI’s prospects hinge in part on the eventual launch of the new flagship LLM Orion that it is developing. This model aims to improve on the existing flagship LLM, GPT-4, that OpenAI launched early last year. By now, other competitors have launched LLMs that perform roughly as well as GPT-4.
Altman is looking to raise more money for the company and find ways to cut its losses, according to the sources. OpenAI has reportedly raised about $13 billion from Microsoft since 2019 as part of a business partnership with the software giant that runs through 2030. The terms of the partnership could change, including how OpenAI pays Microsoft to lease cloud servers to develop its AI.
Cloud servers are OpenAI’s biggest cost. A report last month suggested that OpenAI couldlose $5 billion this year and run out of cash in 12 months. Regardless of the launch of Strawberry as a product, expectations are high for Orion as OpenAI looks to stay ahead of its rivals and continue its remarkable revenue growth.
Former OpenAI chief researcher Ilya Sutskever is believed to have provided the idea and foundation for Strawberry. He left OpenAI this year to start a competing startup. The breakthrough and security conflicts within OpenAI came just before the OpenAI board, led by Sutskever, fired Altman last November and quickly rehired him.
But since Altman’s return, a significant number of AI researchers have left the company. Some have been fired and others have resigned. According to analysts, it ‘s becoming increasingly clear that this has to do with Altman’s direction of partnerships and products.
ALSO READ: By 2045, AI Will Make Humans Immortal, Claims Former Google Engineer