The fresh model, dubbed OpenAI-o1, can address problems that tree existing AI types, including OpenAI’s most prominent existing design, GPT-4o. It reasons through the problem, properly speaking out loud as a person may, before coming to the desired conclusion rather than rushing up and coming up with an answer in one step, as a big language model typically does.
” This is what we consider the new model in these types”, Mira Murati, OpenAI’s chief technology officer, tells WIRED. It performs “much better” at completing difficult reasoning problems.”
The new concept, which was given the code-name Strawberry within OpenAI, is not a GPT-4o replacement but rather a match to it, according to the business.
Murati says that OpenAI is now building its future king design, GPT-5, which will be significantly larger than its predecessor. GPT-5 is likely to include the logic technology introduced today, even though the company also believes that level will help wring novel abilities out of AI. ” There are two concepts”, Murati says. ” The scaling model and this fresh model. We expect that we will take them up”.
LLMs usually generate their responses from massive neural network and massive amounts of training data. They may exhibit remarkable linguistic and logical prowess, but they typically struggle with remarkably straightforward problems like rudimentary math questions that require reasoning.
In order to increase its logic process, Murati claims that OpenAI-o1 uses support learning, which involves giving a concept good comments when it answers questions correctly and negative feedback when it does not. She says,” The model fine tunes the strategies it employs to get to the answer” and” sharpens its thinking.” With enhanced understanding, computers can now play games with extraordinary skill and perform important tasks like creating computer cards. The method is also a crucial component in transforming an LLM into a valuable and well-behaved robot.
Mark Chen, evil president of study at OpenAI, demonstrated the new design to WIRED, using it to resolve many problems that its previous model, GPT-4o, cannot. A lady is as ancient as the prince when she is twice as old as the prince when she is the bride’s time is equal to the lady ‘ current age, according to a sophisticated chemical problem and the following mind-bending mathematical formula. What is the time of the prince and princess”? ( The correct response is that the princess is 40 and the prince is 30. )
” The]new ] model is learning to think for itself, rather than kind of trying to imitate the way humans would think”, as a conventional LLM does, Chen says.
OpenAI says its new model works greatly better on a number of issue models, including ones focused on programming, math, physics, physiology, and science. On the American Invitational Mathematics Examination ( AIME), a test for math students, GPT-4o solved on average 12 percent of the problems while o1 got 83 percent right, according to the company.
The new design is slower than GPT-4o, and OpenAI says it does not always do better—in component because, unlike GPT-4o, it may search the web and it is not bidirectional, meaning it cannot read images or music.
Improvements in LLMs’ reasoning abilities have been a hot topic for some time in research circles. Indeed, rivals are pursuing similar research lines. AlphaProof, a project that combines language models with reinforcement learning to solve challenging math problems, was unveiled by Google in July.
AlphaProof was able to pick up math reasoning skills by examining correct responses. There are not always the right answers for everything a model might encounter, which is a key obstacle to expanding this kind of learning. Chen claims that OpenAI has made it possible to create a more general-purpose reasoning system. ” I do think we have made some breakthroughs there, I think it is part of our edge”, Chen says. ” It’s actually fairly good at reasoning across all domains”.
The key to more generalized training, according to Stanford professor Noah Goodman, may involve using a” carefully prompted language model and handcrafted data” for training. He goes on to say that having the ability to consistently trade the speed of results for greater accuracy would be a “nice advance.”
Yoon Kim, an assistant professor at MIT, says how LLMs solve problems currently remains somewhat mysterious, and even if they perform step-by-step reasoning there may be key differences from human intelligence. As the technology becomes more widely used, this might be crucial. ” These are systems that would be potentially making decisions that affect many, many people”, he says. Do we need to be confident about how a computational model is making decisions, as the saying goes?
The approach that OpenAI has developed today may also aid in improving AI models ‘ behavior. Murati claims that the new model has improved its ability to avoid producing unpleasant or potentially harmful output by considering the results of its actions. ” If you think about teaching children, they learn much better to align to certain norms, behaviors, and values once they can reason about why they’re doing a certain thing”, she says.
Oren Etzioni, a professor emeritus at the University of Washington and a prominent AI expert, says it’s “essential to enable LLMs to engage in multi-step problem solving, use tools, and solve complex problems”. He adds,” Pure scale up will not deliver this”. Etzioni says, however, that there are further challenges ahead. ” Even if reasoning were solved, we would still have the challenge of hallucination and factuality”.
Chen for OpenAI claims that the company’s new reasoning strategy shows that developing AI should n’t require excessive amounts of compute power. One of the exciting aspects of the paradigm is that we think it will allow us to ship intelligence less expensively, he says, and I believe that is actually the company’s core mission.
