OpenAI Developing Advanced Reasoning Technology Codenamed 'Strawberry'

OpenAI working on new reasoning technology under code name 'Strawberry'

ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named "Strawberry," according to a person familiar with the matter and internal documentation reviewed by Reuters. The project comes as the Microsoft-backed startup races to show that the types of models it offers are capable of delivering advanced reasoning capabilities. Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said. The document describes a project that uses Strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms "deep research," according to the source. This is something that has eluded AI models to date, according to interviews with more than a dozen AI researchers. Asked about Strawberry and the details reported in this story, an OpenAI company spokesperson said in a statement: "We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time". The spokesperson did not directly address questions about Strawberry. The Strawberry project was formerly known as Q*, which Reuters reported last year was already seen inside the company as a breakthrough. Two sources described viewing earlier this year what OpenAI staffers told them were Q* demos, capable of answering tricky science and math questions out of reach of today's commercially-available models. A different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a MATH dataset, a benchmark of championship math problems. Reuters could not determine if this was the "Strawberry" project. On Tuesday at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. An OpenAI spokesperson confirmed the meeting but declined to give details of the contents. Reuters could not determine if the project demonstrated was Strawberry. OpenAI hopes the innovation will improve its AI models' reasoning capabilities dramatically, the person familiar with it said, adding that Strawberry involves a specialised way of processing an AI model after it has been pre-trained on very large datasets. Researchers Reuters interviewed say that reasoning is key to AI achieving human or super-human-level intelligence. While large language models can already summarise dense texts and compose elegant prose far more quickly than any human, the technology often falls short on common sense problems whose solutions seem intuitive to people, like recognising logical fallacies and playing tic-tac-toe. When the model encounters these kinds of problems, it often "hallucinates" bogus information. AI researchers interviewed by Reuters generally agree that reasoning, in the context of AI, involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions and work through challenging multi-step problems reliably. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of a "CUA," or a computer-using agent, that can take actions based on its findings, according to the document and one of the sources. OpenAI also plans to test its capabilities on doing the work of software and machine learning engineers.

OpenAI Working on New Reasoning Technology Named Strawberry: What It Does

OpenAI is building new AI use cases and help evolve ChatGPT further ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named "Strawberry," according to a person familiar with the matter and internal documentation reviewed by Reuters. The project, details of which have not been previously reported, comes as the Microsoft-backed startup races to show that the types of models it offers are capable of delivering advanced reasoning capabilities. Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said. The document describes a project that uses Strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms "deep research," according to the source. This is something that has eluded AI models to date, according to interviews with more than a dozen AI researchers. Asked about Strawberry and the details reported in this story, an OpenAI company spokesperson said in a statement: "We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time." The spokesperson did not directly address questions about Strawberry. The Strawberry project was formerly known as Q*, which Reuters reported last year was already seen inside the company as a breakthrough. Two sources described viewing earlier this year what OpenAI staffers told them were Q* demos, capable of answering tricky science and math questions out of reach of today's commercially-available models. On Tuesday at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. An OpenAI spokesperson confirmed the meeting but declined to give details of the contents. Reuters could not determine if the project demonstrated was Strawberry. OpenAI hopes the innovation will improve its AI models' reasoning capabilities dramatically, the person familiar with it said, adding that Strawberry involves a specialized way of processing an AI model after it has been pre-trained on very large datasets. Researchers Reuters interviewed say that reasoning is key to AI achieving human or super-human-level intelligence. While large language models can already summarize dense texts and compose elegant prose far more quickly than any human, the technology often falls short on common sense problems whose solutions seem intuitive to people, like recognizing logical fallacies and playing tic-tac-toe. When the model encounters these kinds of problems, it often "hallucinates" bogus information. AI researchers interviewed by Reuters generally agree that reasoning, in the context of AI, involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions, and work through challenging multi-step problems reliably. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. OpenAI CEO Sam Altman said earlier this year that in AI "the most important areas of progress will be around reasoning ability." Other companies like Google, Meta and Microsoft are likewise experimenting with different techniques to improve reasoning in AI models, as are most academic labs that perform AI research. Researchers differ, however, on whether large language models (LLMs) are capable of incorporating ideas and long-term planning into how they do prediction. For instance, one of the pioneers of modern AI, Yann LeCun, who works at Meta, has frequently said that LLMs are not capable of humanlike reasoning. AI CHALLENGES Strawberry is a key component of OpenAI's plan to overcome those challenges, the source familiar with the matter said. The document seen by Reuters described what Strawberry aims to enable, but not how. In recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company's pitches. They declined to be identified because they are not authorized to speak about private matters. Strawberry includes a specialized way of what is known as "post-training" OpenAI's generative AI models, or adapting the base models to hone their performance in specific ways after they have already been "trained" on reams of generalized data, one of the sources said. The post-training phase of developing a model involves methods like "fine-tuning," a process used on nearly all language models today that comes in many flavors, such as having humans give feedback to the model based on its responses and feeding it examples of good and bad answers. Strawberry has similarities to a method developed at Stanford in 2022 called "Self-Taught Reasoner" or "STaR", one of the sources with knowledge of the matter said. STaR enables AI models to "bootstrap" themselves into higher intelligence levels via iteratively creating their own training data, and in theory could be used to get language models to transcend human-level intelligence, one of its creators, Stanford professor Noah Goodman, told Reuters. "I think that is both exciting and terrifying...if things keep going in that direction we have some serious things to think about as humans," Goodman said. Goodman is not affiliated with OpenAI and is not familiar with Strawberry. Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks (LHT), the document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time, the first source explained. To do so, OpenAI is creating, training and evaluating the models on what the company calls a "deep-research" dataset, according to the OpenAI internal documentation. Reuters was unable to determine what is in that dataset or how long an extended period would mean. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of a "CUA," or a computer-using agent, that can take actions based on its findings, according to the document and one of the sources. OpenAI also plans to test its capabilities on doing the work of software and machine learning engineers.

EXCLUSIVE-OpenAI working on new reasoning technology under code name 'Strawberry'

July 12 - ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named "Strawberry," according to a person familiar with the matter and internal documentation reviewed by Reuters. The project, details of which have not been previously reported, comes as the Microsoft-backed startup races to show that the types of models it offers are capable of delivering advanced reasoning capabilities. Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said. The document describes a project that uses Strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms "deep research," according to the source. This is something that has eluded AI models to date, according to interviews with more than a dozen AI researchers. Asked about Strawberry and the details reported in this story, an OpenAI company spokesperson said in a statement: "We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time." The spokesperson did not directly address questions about Strawberry. The Strawberry project was formerly known as Q*, which Reuters reported last year was already seen inside the company as a breakthrough. Two sources described viewing earlier this year what OpenAI staffers told them were Q* demos, capable of answering tricky science and math questions out of reach of today's commercially-available models. A different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a MATH dataset, a benchmark of championship math problems. Reuters could not determine if this was the "Strawberry" project. On Tuesday at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. An OpenAI spokesperson confirmed the meeting but declined to give details of the contents. Reuters could not determine if the project demonstrated was Strawberry. OpenAI hopes the innovation will improve its AI models' reasoning capabilities dramatically, the person familiar with it said, adding that Strawberry involves a specialized way of processing an AI model after it has been pre-trained on very large datasets. Researchers Reuters interviewed say that reasoning is key to AI achieving human or super-human-level intelligence. While large language models can already summarize dense texts and compose elegant prose far more quickly than any human, the technology often falls short on common sense problems whose solutions seem intuitive to people, like recognizing logical fallacies and playing tic-tac-toe. When the model encounters these kinds of problems, it often "hallucinates" bogus information. AI researchers interviewed by Reuters generally agree that reasoning, in the context of AI, involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions, and work through challenging multi-step problems reliably. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. OpenAI CEO Sam Altman said earlier this year that in AI "the most important areas of progress will be around reasoning ability." Other companies like Google, Meta and Microsoft are likewise experimenting with different techniques to improve reasoning in AI models, as are most academic labs that perform AI research. Researchers differ, however, on whether large language models are capable of incorporating ideas and long-term planning into how they do prediction. For instance, one of the pioneers of modern AI, Yann LeCun, who works at Meta, has frequently said that LLMs are not capable of humanlike reasoning. AI CHALLENGES Strawberry is a key component of OpenAI's plan to overcome those challenges, the source familiar with the matter said. The document seen by Reuters described what Strawberry aims to enable, but not how. In recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company's pitches. They declined to be identified because they are not authorized to speak about private matters. Strawberry includes a specialized way of what is known as "post-training" OpenAI's generative AI models, or adapting the base models to hone their performance in specific ways after they have already been "trained" on reams of generalized data, one of the sources said. The post-training phase of developing a model involves methods like "fine-tuning," a process used on nearly all language models today that comes in many flavors, such as having humans give feedback to the model based on its responses and feeding it examples of good and bad answers. Strawberry has similarities to a method developed at Stanford in 2022 called "Self-Taught Reasoner" or "STaR", one of the sources with knowledge of the matter said. STaR enables AI models to "bootstrap" themselves into higher intelligence levels via iteratively creating their own training data, and in theory could be used to get language models to transcend human-level intelligence, one of its creators, Stanford professor Noah Goodman, told Reuters. "I think that is both exciting and terrifying...if things keep going in that direction we have some serious things to think about as humans," Goodman said. Goodman is not affiliated with OpenAI and is not familiar with Strawberry. Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks , the document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time, the first source explained. To do so, OpenAI is creating, training and evaluating the models on what the company calls a "deep-research" dataset, according to the OpenAI internal documentation. Reuters was unable to determine what is in that dataset or how long an extended period would mean. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of a "CUA," or a computer-using agent, that can take actions based on its findings, according to the document and one of the sources. OpenAI also plans to test its capabilities on doing the work of software and machine learning engineers.

OpenAI working on new reasoning technology under code name 'Strawberry'

ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named "Strawberry," according to a person familiar with the matter and internal documentation reviewed by Reuters. The project, details of which have not been previously reported, comes as the Microsoft-backed startup races to show that the types of models it offers are capable of delivering advanced reasoning capabilities. Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said. The document describes a project that uses Strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms "deep research," according to the source. This is something that has eluded AI models to date, according to interviews with more than a dozen AI researchers. Asked about Strawberry and the details reported in this story, an OpenAI company spokesperson said in a statement: "We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time." The spokesperson did not directly address questions about Strawberry. The Strawberry project was formerly known as Q*, which Reuters reported last year was already seen inside the company as a breakthrough. Two sources described viewing earlier this year what OpenAI staffers told them were Q* demos, capable of answering tricky science and math questions out of reach of today's commercially-available models. A different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a MATH dataset, a benchmark of championship math problems. Reuters could not determine if this was the "Strawberry" project. On Tuesday at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. An OpenAI spokesperson confirmed the meeting but declined to give details of the contents. Reuters could not determine if the project demonstrated was Strawberry. OpenAI hopes the innovation will improve its AI models' reasoning capabilities dramatically, the person familiar with it said, adding that Strawberry involves a specialized way of processing an AI model after it has been pre-trained on very large datasets. Researchers Reuters interviewed say that reasoning is key to AI achieving human or super-human-level intelligence. While large language models can already summarize dense texts and compose elegant prose far more quickly than any human, the technology often falls short on common sense problems whose solutions seem intuitive to people, like recognizing logical fallacies and playing tic-tac-toe. When the model encounters these kinds of problems, it often "hallucinates" bogus information. AI researchers interviewed by Reuters generally agree that reasoning, in the context of AI, involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions, and work through challenging multi-step problems reliably. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. OpenAI CEO Sam Altman said earlier this year that in AI "the most important areas of progress will be around reasoning ability." Other companies like Google, Meta and Microsoft are likewise experimenting with different techniques to improve reasoning in AI models, as are most academic labs that perform AI research. Researchers differ, however, on whether large language models (LLMs) are capable of incorporating ideas and long-term planning into how they do prediction. For instance, one of the pioneers of modern AI, Yann LeCun, who works at Meta, has frequently said that LLMs are not capable of humanlike reasoning. AI CHALLENGES Strawberry is a key component of OpenAI's plan to overcome those challenges, the source familiar with the matter said. The document seen by Reuters described what Strawberry aims to enable, but not how. In recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company's pitches. They declined to be identified because they are not authorized to speak about private matters. Strawberry includes a specialized way of what is known as "post-training" OpenAI's generative AI models, or adapting the base models to hone their performance in specific ways after they have already been "trained" on reams of generalized data, one of the sources said. The post-training phase of developing a model involves methods like "fine-tuning," a process used on nearly all language models today that comes in many flavors, such as having humans give feedback to the model based on its responses and feeding it examples of good and bad answers. Strawberry has similarities to a method developed at Stanford in 2022 called "Self-Taught Reasoner" or "STaR", one of the sources with knowledge of the matter said. STaR enables AI models to "bootstrap" themselves into higher intelligence levels via iteratively creating their own training data, and in theory could be used to get language models to transcend human-level intelligence, one of its creators, Stanford professor Noah Goodman, told Reuters. "I think that is both exciting and terrifying...if things keep going in that direction we have some serious things to think about as humans," Goodman said. Goodman is not affiliated with OpenAI and is not familiar with Strawberry. Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks (LHT), the document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time, the first source explained. To do so, OpenAI is creating, training and evaluating the models on what the company calls a "deep-research" dataset, according to the OpenAI internal documentation. Reuters was unable to determine what is in that dataset or how long an extended period would mean. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of a "CUA," or a computer-using agent, that can take actions based on its findings, according to the document and one of the sources. OpenAI also plans to test its capabilities on doing the work of software and machine learning engineers. (Reporting by Anna Tong in San Francisco and Katie Paul in New York; editing by Ken Li and Claudia Parsons)

OpenAI working on new reasoning technology under code name 'Strawberry'

Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said. The document describes a project that uses Strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms "deep research," according to the source. This is something that has eluded AI models to date, according to interviews with more than a dozen AI researchers. Asked about Strawberry and the details reported in this story, an OpenAI company spokesperson said in a statement: "We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time." The spokesperson did not directly address questions about Strawberry. The Strawberry project was formerly known as Q*, which Reuters reported last year was already seen inside the company as a breakthrough. Two sources described viewing earlier this year what OpenAI staffers told them were Q* demos, capable of answering tricky science and math questions out of reach of today's commercially-available models. A different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a MATH dataset, a benchmark of championship math problems. Reuters could not determine if this was the "Strawberry" project. On Tuesday at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. An OpenAI spokesperson confirmed the meeting but declined to give details of the contents. Reuters could not determine if the project demonstrated was Strawberry. OpenAI hopes the innovation will improve its AI models' reasoning capabilities dramatically, the person familiar with it said, adding that Strawberry involves a specialized way of processing an AI model after it has been pre-trained on very large datasets. Researchers Reuters interviewed say that reasoning is key to AI achieving human or super-human-level intelligence. While large language models can already summarize dense texts and compose elegant prose far more quickly than any human, the technology often falls short on common sense problems whose solutions seem intuitive to people, like recognizing logical fallacies and playing tic-tac-toe. When the model encounters these kinds of problems, it often "hallucinates" bogus information. AI researchers interviewed by Reuters generally agree that reasoning, in the context of AI, involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions, and work through challenging multi-step problems reliably. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. OpenAI CEO Sam Altman said earlier this year that in AI "the most important areas of progress will be around reasoning ability." Other companies like Google, Meta and Microsoft are likewise experimenting with different techniques to improve reasoning in AI models, as are most academic labs that perform AI research. Researchers differ, however, on whether large language models (LLMs) are capable of incorporating ideas and long-term planning into how they do prediction. For instance, one of the pioneers of modern AI, Yann LeCun, who works at Meta, has frequently said that LLMs are not capable of humanlike reasoning. Strawberry is a key component of OpenAI's plan to overcome those challenges, the source familiar with the matter said. The document seen by Reuters described what Strawberry aims to enable, but not how. In recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company's pitches. They declined to be identified because they are not authorized to speak about private matters. Strawberry includes a specialized way of what is known as "post-training" OpenAI's generative AI models, or adapting the base models to hone their performance in specific ways after they have already been "trained" on reams of generalized data, one of the sources said. The post-training phase of developing a model involves methods like "fine-tuning," a process used on nearly all language models today that comes in many flavors, such as having humans give feedback to the model based on its responses and feeding it examples of good and bad answers. Strawberry has similarities to a method developed at Stanford in 2022 called "Self-Taught Reasoner" or "STaR", one of the sources with knowledge of the matter said. STaR enables AI models to "bootstrap" themselves into higher intelligence levels via iteratively creating their own training data, and in theory could be used to get language models to transcend human-level intelligence, one of its creators, Stanford professor Noah Goodman, told Reuters. "I think that is both exciting and terrifying...if things keep going in that direction we have some serious things to think about as humans," Goodman said. Goodman is not affiliated with OpenAI and is not familiar with Strawberry. Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks (LHT), the document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time, the first source explained. To do so, OpenAI is creating, training and evaluating the models on what the company calls a "deep-research" dataset, according to the OpenAI internal documentation. Reuters was unable to determine what is in that dataset or how long an extended period would mean. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of a "CUA," or a computer-using agent, that can take actions based on its findings, according to the document and one of the sources. OpenAI also plans to test its capabilities on doing the work of software and machine learning engineers. (Reporting by Anna Tong in San Francisco and Katie Paul in New York; editing by Ken Li and Claudia Parsons)

OpenAI working on new reasoning technology under code name 'Strawberry'

(Adds details on AI performance on maths problems paragraph 11) July 12 - ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named "Strawberry," according to a person familiar with the matter and internal documentation reviewed by Reuters. The project, details of which have not been previously reported, comes as the Microsoft-backed startup races to show that the types of models it offers are capable of delivering advanced reasoning capabilities. Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said. The document describes a project that uses Strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms "deep research," according to the source. This is something that has eluded AI models to date, according to interviews with more than a dozen AI researchers. Asked about Strawberry and the details reported in this story, an OpenAI company spokesperson said in a statement: "We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time." The spokesperson did not directly address questions about Strawberry. The Strawberry project was formerly known as Q*, which Reuters reported last year was already seen inside the company as a breakthrough. Two sources described viewing earlier this year what OpenAI staffers told them were Q* demos, capable of answering tricky science and math questions out of reach of today's commercially-available models. A different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a MATH dataset, a benchmark of championship math problems. Reuters could not determine if this was the "Strawberry" project. On Tuesday at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. An OpenAI spokesperson confirmed the meeting but declined to give details of the contents. Reuters could not determine if the project demonstrated was Strawberry. OpenAI hopes the innovation will improve its AI models' reasoning capabilities dramatically, the person familiar with it said, adding that Strawberry involves a specialized way of processing an AI model after it has been pre-trained on very large datasets. Researchers Reuters interviewed say that reasoning is key to AI achieving human or super-human-level intelligence. While large language models can already summarize dense texts and compose elegant prose far more quickly than any human, the technology often falls short on common sense problems whose solutions seem intuitive to people, like recognizing logical fallacies and playing tic-tac-toe. When the model encounters these kinds of problems, it often "hallucinates" bogus information. AI researchers interviewed by Reuters generally agree that reasoning, in the context of AI, involves the formation of a model that enables AI to plan ahead, reflect how the physical world functions, and work through challenging multi-step problems reliably. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. OpenAI CEO Sam Altman said earlier this year that in AI "the most important areas of progress will be around reasoning ability." Other companies like Google, Meta and Microsoft are likewise experimenting with different techniques to improve reasoning in AI models, as are most academic labs that perform AI research. Researchers differ, however, on whether large language models (LLMs) are capable of incorporating ideas and long-term planning into how they do prediction. For instance, one of the pioneers of modern AI, Yann LeCun, who works at Meta, has frequently said that LLMs are not capable of humanlike reasoning. Strawberry is a key component of OpenAI's plan to overcome those challenges, the source familiar with the matter said. The document seen by Reuters described what Strawberry aims to enable, but not how. In recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company's pitches. They declined to be identified because they are not authorized to speak about private matters. Strawberry includes a specialized way of what is known as "post-training" OpenAI's generative AI models, or adapting the base models to hone their performance in specific ways after they have already been "trained" on reams of generalized data, one of the sources said. The post-training phase of developing a model involves methods like "fine-tuning," a process used on nearly all language models today that comes in many flavors, such as having humans give feedback to the model based on its responses and feeding it examples of good and bad answers. Strawberry has similarities to a method developed at Stanford in 2022 called "Self-Taught Reasoner" or "STaR", one of the sources with knowledge of the matter said. STaR enables AI models to "bootstrap" themselves into higher intelligence levels via iteratively creating their own training data, and in theory could be used to get language models to transcend human-level intelligence, one of its creators, Stanford professor Noah Goodman, told Reuters. "I think that is both exciting and terrifying...if things keep going in that direction we have some serious things to think about as humans," Goodman said. Goodman is not affiliated with OpenAI and is not familiar with Strawberry. Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks (LHT), the document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time, the first source explained. To do so, OpenAI is creating, training and evaluating the models on what the company calls a "deep-research" dataset, according to the OpenAI internal documentation. Reuters was unable to determine what is in that dataset or how long an extended period would mean. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of a "CUA," or a computer-using agent, that can take actions based on its findings, according to the document and one of the sources. OpenAI also plans to test its capabilities on doing the work of software and machine learning engineers. (Reporting by Anna Tong in San Francisco and Katie Paul in New York; editing by Ken Li and Claudia Parsons)

What Elon Musk and Ilya Sutskever Feared About OpenAI Is Becoming Reality

OpenAI is reportedly working on A.I. models with advanced reasoning capabilities under the code-name "Strawberry." As part of OpenAI's path towards artificial general intelligence (A.G.I), a term for technology matching the intelligence of humans, the company is reportedly attempting to enable A.I. models to perform advanced reasoning. Such work is taking place under a secretive project code-named 'Strawberry,' as reported by Reuters, which noted that the project was previously known as Q* or Q Star. While its name may have changed, the project isn't exactly new. Researchers and co-founders of OpenAI have previously warned against the initiative, with concerns over it reportedly playing a part in the brief ousting of Sam Altman as OpenAI's CEO in November. Sign Up For Our Daily Newsletter Sign Up Thank you for signing up! By clicking submit, you agree to our <a href="http://observermedia.com/terms">terms of service</a> and acknowledge we may use your information to send you emails, product samples, and promotions on this website and other properties. You can opt out anytime. See all of our newsletters Strawberry uses a unique method of post-training A.I. models, a process that improves their performance after being trained on datasets, according to Reuters, which cited internal OpenAI documents and a person familiar with the project. With the help of "deep-research" datasets, the company aims to create models that display human-level reasoning. OpenAI reportedly is looking into how Strawberry can allow models to be able to complete tasks over an extended period of time, search the web by themselves and take actions on its findings, and perform the work of engineers. OpenAI did not respond to requests for comment from Observer. Elon Musk and Ilya Sutskever raised concerns about Q* Altman, who has previously reiterated OpenAI's desire to create models able to reason, briefly lost control of his company last year when his board fired him for four days. Shortly before the ousting, several OpenAI employees had become concerned over breakthroughs presented by what was then known as Q*, a project spearheaded by Ilya Sutskever, OpenAI's former chief scientist. Sutskever himself had reportedly begun to worry about the project's technology, as did OpenAI employees working on A.I. safety at the time. After his reinstatement, Altman referred to news reports about Q* as an "unfortunate leak" in an interview with the Verge. Elon Musk, another OpenAI co-founder, has also raised the alarm about Q* in the past. The billionaire, who severed ties with the company in 2018, referred to the project in a lawsuit filed against OpenAI and Altman that has since been dropped. While discussing OpenAI's close partnership with Microsoft (MSFT), Musk's suit claimed that the terms of the deal dictate that Microsoft only has rights to OpenAI's pre-A.G.I. technology and that it is up to OpenAI's board to determine when the company has achieved A.G.I. Musk argued that OpenAI's GPT-4 model constitutes as A.G.I, which he believes "poses a grave threat to humanity," according to the suit. Court filings stated that "OpenAI is currently developing a model known as Q* that has an even stronger claim to A.G.I." Recent internal meetings have suggested that OpenAI is making rapid progress toward the type of human-level reasoning that Strawberry is working on. In an OpenAI all-hands meeting held earlier this month, the company unveiled a five-tiered system to track its progress towards A.G.I., as reported by Bloomberg. While the company said it is currently on the first level, known as "chatbots," it revealed that it has nearly reached the second level of "reasoners," which involves technology that can display human-level problem-solving. The subsequent steps consist of A.I. systems acting as "agents" that can take actions, "innovators" that aid in invention and "organizations" that do the work of an organization.

OpenAI may be working on AI that can perform research without human help - which should go fine

The 'Strawberry' project isn't ripe yet, but the seeds are planted OpenAI is developing a new project to enhance its AI models' reasoning capabilities, called "Strawberry," according to documents first discovered by Reuters. The project is a key element in the efforts by OpenAI to achieve more powerful AI models capable of operating on their own when it comes to performing research. According to the internal documents Reuters looked at, Strawberry is aimed at building an AI that will not only answer questions but search around online and perform follow-up research on its own. This so-called "deep research" trick would be a major leap beyond current AI models that rely on existing data sets and respond in ways that are already programmed. There aren't details on the exact mechanisms of Strawberry, but apparently, it involves AI models using a specialized processing method after training on extensive datasets. This innovative approach could potentially set a new standard in AI development. An AI that can think ahead and perform research on its own to understand the world is much closer to a human than anything ChatGPT or other tools using AI models offer. It's a challenging goal that has eluded AI developers to date, despite numerous advancements in the field. Reuters reported that Strawberry, which was then known as Q*, had made some breakthroughs. There were demonstrations where viewers witnessed AI could tackle science and math problems beyond the range of commercial models, and apparently, OpenAI had tested AI models that scored over 90% on a championship-level math problem data set. Should OpenAI achieve its goals, the reasoning capabilities could transform scientific research and everyday problem-solving. It could help plug holes in scientific knowledge by looking for gaps and even offering up hypotheses to fill them. This would vastly accelerate the pace of discovery in various domains. If successful, Strawberry could mark a pivotal moment in AI research, bringing us closer to truly autonomous AI systems capable of conducting independent research and offering more sophisticated reasoning. Strawberry is, it seems, part and parcel of OpenAI's long-term plans to demonstrate and enhance the potential of its AI models. Even after GPT-3 and GPT-4 set new benchmarks for language processing and generation, there's a big leap to autonomous reasoning and deep research. But, it fits with other work on the road to artificial general intelligence (AGI), including the recent development of an internal scale for charting the progress of large language models.

OpenAI's new 'Project Strawberry' could give ChatGPT more freedom to search the web and solve complex problems

OpenAI is constantly working on new models with several teams exploring different approaches to achieving the long-term goal of artificial general intelligence (AGI), and some of those ideas are working out better than others. According to Reuters, there is a new project code-named "Strawberry" that some on X predict could be a new version of the infamous "reasoning mode" Q* revealed through a leak last year. Project Strawberry seems to be a new model or system capable of improved reasoning, including going online to find information in preparation for solving a particularly complex problem. What isn't clear is whether Strawberry is the model OpenAI CTO Mira Murati was talking about when she mentioned the next generation AI would be as intelligent as someone with a PhD, or whether this is proof OpenAI has hit 'level 2' on its path to AGI -- the reasoners. The Reuters report, based on internal documentation and comments from an insider suggests this could be a significant upgrade in artificial intelligence capabilities. According to the contact, Strawberry is even a "tightly kept secret" inside OpenAI but it seems, that while it might be part of ChatGPT in the future its main purpose is to perform "deep research" beyond simple user queries. Some of the capabilities include being able to plan enough when responding to navigate the internet autonomously and reliably without having to be told to do so by the user. ChatGPT can already search the web for specific details but not reliably and often to get it to look for live data and not just look at its training data, you have to specifically tell it to do so. OpenAI hasn't addressed the reports on Strawberry directly but told Reuters it is common practice to carry out continuous research into a range of new AI capabilities. Project Strawberry is just that, a research project for OpenAI and as such it isn't clear if it will ever see the light of day. However, even if it isn't a product in its own right the research will likely feed in to future AI models trained by OpenAI or future product updates. I suspect any improvements in reasoning will be factored into training for GPT-5 and Project Strawberry could be like Omni, a new style model that comes out of nowhere to change the game. Unlike GPT-4, GPT-4o (Omni) was natively multimodal in and out. Strawberry could be a new type of model that is a native reasoner from the ground up without requiring fine-tuning or additional training. Very little is actually known but OpenAI did suggest in a comment to Reuters that AI understanding like humans was an important requirement.

OpenAI's project 'Strawberry' could power AI with super-human intelligence

The project, details of which have not been previously reported, comes as the Microsoft-backed startup races to show that the types of models it offers are capable of delivering advanced reasoning capabilities. Teams inside OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters in May. Reuters could not ascertain the precise date of the document, which details a plan for how OpenAI intends to use Strawberry to perform research. The source described the plan to Reuters as a work in progress. The news agency could not establish how close Strawberry is to being publicly available. How Strawberry works is a tightly kept secret even within OpenAI, the person said.

OpenAI developing secret "Strawberry" project to help ChatGPT outthink humans

Dubbed Strawberry, the project aims to help AI achieve reasoning capabilites on par with humans. As OpenAI's next-generation GPT-5 language model inches closer to release, we're finally getting our first glimpses at the potential AI breakthroughs in store. According to a new Reuters report, the Microsoft-backed startup is working on a new technique to improve AI reasoning capabilities under the code name "Strawberry." This news comes from a source familiar with the project and internal documents reviewed by the publication.