Last week, OpenAI explained the five levels of AI which he believes is on the path to general artificial intelligence. level 1the lowest, are the chatbots that we currently know and of which ChatGPT is the best example. Level 2, are the so-called ‘reasoners’ that can solve problems at a human level and this is what the company seems to be on the verge of achieving. As Bloomberg reported at the time, OpenAI has already made internal presentations in which has shown a AI with human-like reasoning abilitiesOn the other hand, Reuters has published this Monday that the company has a secret project called Strawberrystill in development, with significantly more advanced reasoning capabilities.
The information about the project is quite vague and both pieces of information do not necessarily correspond to the same AI model, but could be different developments. In fact, Strawberry would not be so much an AI in itself, as we understand ChatGPT or Gemini, but a reasoning technology that improves other models.
According to documents and a variety of internal sources seen by Reuters, Strawberry wants to go beyond AI that simply generates responses to queries and It can plan ahead and navigate the Internet autonomously, without human intervention, to perform what the company calls “deep research.”.
‘We want our AI models to see and understand the world more like we do. Continued research into new AI capabilities is common practice in the industry, with a shared belief that These systems will improve in reasoning over time‘, an OpenAI spokesperson told the outlet.
The workings of Strawberry are a closely guarded secret even within OpenAI. It is the project formerly known as Q*which was revealed last November and was considered as one of the causes of the differences between the company’s board of directors and its CEO, Sam Altmanwho was briefly fired then.
9 months ago Q* was already able to Answer complicated science and math questions that commercially available models cannot reachMore recently, OpenAI has internally tested an AI that has achieved a 90% score on MATH testsThese are a set of math problems used as a benchmark to assess the capabilities of artificial intelligence models in solving advanced math problems. Reuters has not been able to say whether this is Strawberry or a different project.
OpenAI expects Strawberry improve dramatically the reasoning capabilities of their AI models and involves a specialized way of processing an AI model after it has been trained with data sets. It’s about a ‘post-training’ which has similarities with the method STaRacronym for Self-Taught Reasoner, developed at Stanford University in 2022.
STaR enables AI models ‘self-instruct’ to reach higher levels of intelligence through the iterative creation of your own training data and in theory could be used for Language models transcend human-level intelligenceaccording to the professor Noah Goodmanone of the creators of this method.
Another of Strawberry’s capabilities is that it can do what the company calls long horizon tasksLHT for short. That is, complex tasks that require a model to plan ahead and perform a series of actions over an extended period of timeTo achieve this, OpenAI is training it on a “deep research” dataset, according to internal documentation.