The new high-profile AI model has been released by OpenAI, the artificial intelligence research company and is named Strawbery. This new model brings a new approach in formulating the AI as it displays the capacity of ‘working its way reasoning’ through problems much in the same way as a human might. Strawberry instance also proves that the future of AI does not only depend upon making the models bigger, but also the probability of the creation of new methods to solve problems.
Strawberry is officially referred to as OpenAI o1 and it is at least nine times smarter than current models including OpenAI’s latest model GPT-4o. Two aspects make Strawberry stand out: the latter is a team of problem-solvers. Rather than giving an answer in one step/solve, Strawberry solves the problem step by step or, in other words, is ‘talk-aloud solving’.
A New Paradigm in AI Development
Mira Murati, the CTO of OpenAI notes that Strawberry stands a mirror of what she includes as a ‘new paradigm’ in AI models. This shift of paradigm is expressed by the increased capacity of the model to solve problems of reasoning that are beyond the scope of the previous models. The introduction of Strawberry in parallel with the creation of GPT-5, further master model of OpenAI, suggest that the firm is experimenting with the scaling strategy as well as this new reasoning-based method for pushing the AI progress forward.
In the development of Strawberry, there was the application of reinforcement instructions to the model as it allows positive reinforcement to the correct answer or negative reinforcement to the wrong one. This process helps the model to learn and therefore shape its own thinking and strategies and thereby improve the problem solving accuracy. There has been such successful applications of reinforcement learning in playing games and designing computer chips, to show that it can improve the capability of AI systems.
Impressive Problem-Solving Abilities
While performing a live demonstration of its problem-solving capabilities during one of its performances where Mark Chen, the vice president of research at OpenAI, was at strawberry’s helm. The model accurately answered chemical equations with a high degree of difficulty and other problems that I proposed before even GPT-4o could not solve. This performance of logic supports the notion of independent decision-making in the model as opposed to mimicking human thought processes.
On different problem areas such as coding problems, math problems, physics problems, biology problems and chemistry problems, Strawberry has performed far better than GPT-4o. For instance on the American Invitational Mathematics Examination (AIME) GPT-4o got an average of 12% of the problems right while Strawberry scored 83% correct.
Balancing Speed and Accuracy
Despite this, Strawberry performs exceptionally well in terms of reasoning but is slightly slower compared to GPT-4o. Moreover, the capability of the model might not always be superior to GPT-4o because the model cannot surf the web and it does not understand or parse pictures or sound yet. However, the possibility of paying for the speed with increased precision in problem-solving is an incremental enhancement in AI systems.
The formation of Strawberry fits well with the continuous endeavors to enhance the reasoning abilities of large language models (LLMs). Other competitors in the field like Google’s AlphaProof project are also trying to integrate language models with reinforcement learning for complex tasks. What makes OpenAI’s breakthrough is the development of the more generalized approach to reasoning than previous models, which is suitable in different fields.
Implications for AI Safety and Alignment
Strawberry has demonstrated beneficial attributes in areas concerning with AI safety and alignment beyond problem-solving in addition. Through analysing possible consequences of its actions, the model has shown enhanced capacity to create non-desirable or possibly unsafe output. It is all comparable to the process of indoctrination of the children in order to make them imbibe some norms, behaviours and values.
When AI tools are used in decision-making processes that impact huge masses of the populace, it becomes crucial to have a comprehension of why an AI model has pitched a certain verdict. Nevertheless, it is evident that Strawberry’s development has marked a certain progress in this respect, though the problems relating to facts vs hallucinations in AI-generated content have not been solved as yet.
