But its state-of-the-art models have improved significantly on mathematical reasoning. One of OpenAI\u2019s models recently won a gold medal <\/a>at the International Math Olympiad, a math competition for the world\u2019s brightest high school students. OpenAI believes these reasoning capabilities will translate to other subjects, and ultimately power general-purpose agents that the company has always dreamed of building.<\/p>\n ChatGPT was a happy accident \u2014 a lowkey research preview turned viral consumer business \u2014 but OpenAI\u2019s agents are the product of a years-long, deliberate effort within the company.\u00a0<\/p>\n \u201cEventually, you\u2019ll just ask the computer for what you need and it\u2019ll do all of these tasks for you,\u201d said OpenAI CEO Sam Altman at the company\u2019s first developer conference<\/a> in 2023. \u201cThese capabilities are often talked about in the AI field as agents. The upsides of this are going to be tremendous.\u201d<\/p>\n Techcrunch event<\/p>\n \n\t\t\t\t\t\t\t\t\tSan Francisco Whether agents will meet Altman\u2019s vision remains to be seen, but OpenAI shocked the world with the release of its first AI reasoning model, o1,<\/a> in the fall of 2024. Less than a year later, the 21 foundational researchers behind that breakthrough are the most highly sought-after talent in Silicon Valley.<\/p>\n Mark Zuckerberg recruited<\/a> five of the o1 researchers to work on Meta\u2019s new superintelligence-focused unit, offering some compensation packages north of $100 million. One of them, Shengjia Zhao, was recently named<\/a> chief scientist of Meta Superintelligence Labs.<\/p>\n The reinforcement learning renaissance<\/p>\n The rise of OpenAI\u2019s reasoning models and agents are tied to a machine learning training technique known as reinforcement learning (RL). RL provides feedback to an AI model on whether its choices were correct or not in simulated environments.<\/p>\n RL has been used for decades. For instance, in 2016, about a year after OpenAI was founded in 2015, an AI system created by Google DeepMind using RL, AlphaGo<\/a>, gained global attention after beating a world champion in the board game, Go.<\/p>\n Around that time, one of OpenAI\u2019s first employees, Andrej Karpathy, began pondering how to leverage RL to create an AI agent that could use a computer. But it would take years for OpenAI to develop the necessary models and training techniques.<\/p>\n By 2018, OpenAI pioneered its first large language model in the GPT series, pretrained on massive amounts of internet data and a large clusters of GPUs. GPT models excelled at text processing, eventually leading to ChatGPT, but struggled with basic math.\u00a0<\/p>\n It took until 2023 for OpenAI to achieve a breakthrough, initially dubbed \u201cQ*\u201d and then \u201cStrawberry,\u201d by combining LLMs, RL, and a technique called test-time computation. The latter gave the models extra time and computing power to plan and work through problems, verifying its steps, before providing an answer. <\/p>\n This allowed OpenAI to introduce a new approach called \u201cchain-of-thought\u201d (CoT), which improved AI\u2019s performance on math questions the models hadn\u2019t seen before.<\/p>\n \u201cI could see the model starting to reason,\u201d said El Kishky. \u201cIt would notice mistakes and backtrack, it would get frustrated. It really felt like reading the thoughts of a person.\u201d\u00a0<\/p>\n Though individually these techniques weren\u2019t novel, OpenAI uniquely combined them to create Strawberry, which directly led to the development of o1. OpenAI quickly identified that the planning and fact checking abilities of AI reasoning models could be useful to power AI agents.<\/p>\n \u201cWe had solved a problem that I had been banging my head against for a couple of years,\u201d said Lightman. \u201cIt was one of the most exciting moments of my research career.\u201d<\/p>\n Scaling reasoning<\/p>\n With AI reasoning models, OpenAI determined it had two new axes that would allow it to improve AI models: using more computational power during the post-training of AI models, and giving AI models more time and processing power while answering a question.<\/p>\n \u201cOpenAI, as a company, thinks a lot about not just the way things are, but the way things are going to scale,\u201d said Lightman.<\/p>\n Shortly after the 2023 Strawberry breakthrough, OpenAI spun up an \u201cAgents\u201d team led by OpenAI researcher Daniel Selsam to make further progress on this new paradigm, two sources told TechCrunch. Although the team was called \u201cAgents,\u201d\u00a0 OpenAI didn\u2019t initially differentiate between reasoning models and agents as we think of them today. The company just wanted to make AI systems capable of completing complex tasks.<\/p>\n Eventually, the work of Selsam\u2019s Agents team became part of a larger project to develop the o1 reasoning model, with leaders including OpenAI co-founder Ilya Sutskever, chief research officer Mark Chen, and chief scientist Jakub Pachocki.<\/p>\n OpenAI would have to divert precious resources \u2014 mainly talent and GPUs \u2014 to create o1. Throughout OpenAI\u2019s history, researchers have had to negotiate with company leaders to obtain resources; demonstrating breakthroughs was a surefire way to secure them.<\/p>\n \u201cOne of the core components of OpenAI is that everything in research is bottom up,\u201d said Lightman. \u201cWhen we showed the evidence [for o1], the company was like, \u2018This makes sense, let\u2019s push on it.\u2019\u201d<\/p>\n Some former employees say that the startup\u2019s mission to develop AGI was the key factor in achieving breakthroughs around AI reasoning models. By focusing on developing the smartest-possible AI models, rather than products, OpenAI was able to prioritize o1 above other efforts.\u00a0That type of large investment in ideas wasn\u2019t always possible at competing AI labs.<\/p>\n The decision to try new training methods proved prescient. By late 2024, several leading AI labs started seeing diminishing returns<\/a> on models created through traditional pretraining scaling. Today, much of the AI field\u2019s momentum comes from advances in reasoning models.<\/p>\n What does it mean for an AI to \u201creason?\u201d<\/strong><\/p>\n In many ways, the goal of AI research is to recreate human intelligence with computers. Since the launch of o1, ChatGPT\u2019s UX has been filled with more human-sounding features such as \u201cthinking\u201d and \u201creasoning.\u201d<\/p>\n When asked whether OpenAI\u2019s models were truly reasoning, El Kishky hedged, saying he thinks about the concept in terms of computer science. <\/p>\n \u201cWe\u2019re teaching the model how to efficiently expend compute to get an answer. So if you define it that way, yes, it is reasoning,\u201d said El Kishky.<\/p>\n Lightman takes the approach of focusing on the model\u2019s results and not as much on the means or their relation to human brains.<\/p>\n \u201cIf the model is doing hard things, then it is doing whatever necessary approximation of reasoning it needs in order to do that,\u201d said Lightman. \u201cWe can call it reasoning, because it looks like these reasoning traces, but it\u2019s all just a proxy for trying to make AI tools that are really powerful and useful to a lot of people.\u201d<\/p>\n OpenAI\u2019s researchers note people may disagree with their nomenclature or definitions of reasoning \u2014 and surely, critics have emerged<\/a> \u2014 but they argue it\u2019s less important than the capabilities of their models. Other AI researchers tend to agree.<\/p>\n Nathan Lambert, an AI researcher with the non-profit AI2, compares AI reasoning modes to airplanes in a blog post<\/a>. Both, he says, are manmade systems inspired by nature \u2014 human reasoning and bird flight, respectively \u2014 but they operate through entirely different mechanisms. That doesn\u2019t make them any less useful, or any less capable of achieving similar outcomes.<\/p>\n A group of AI researchers from OpenAI, Anthropic, and Google DeepMind agreed in a recent position paper<\/a> that AI reasoning models are not well understood today, and more research is needed. It may be too early to confidently claim what exactly is going on inside them.<\/p>\n The next frontier: AI agents for subjective tasks<\/strong><\/p>\n The AI agents on the market today work best for well-defined, verifiable domains such as coding. OpenAI\u2019s Codex agent<\/a> aims to help software engineers offload simple coding tasks. Meanwhile, Anthropic\u2019s models have become particularly popular<\/a> in AI coding tools like Cursor and Claude Code \u2014 these are some of the first AI agents that people are willing to pay up for<\/a>.<\/p>\n However, general purpose AI agents like OpenAI\u2019s ChatGPT Agent<\/a> and Perplexity\u2019s Comet<\/a> struggle with many of the complex, subjective tasks people want to automate. When trying to use these tools for online shopping or finding a long-term parking spot, I\u2019ve found the agents take longer than I\u2019d like and make silly mistakes<\/a>.<\/p>\n Agents are, of course, early systems that will undoubtedly improve. But researchers must first figure out how to better train the underlying models to complete tasks that are more subjective.<\/p>\n \u201cLike many problems in machine learning, it\u2019s a data problem,\u201d said Lightman, when asked about the limitations of agents on subjective tasks. \u201cSome of the research I\u2019m really excited about right now is figuring out how to train on less verifiable tasks. We have some leads on how to do these things.\u201d\u00a0<\/p>\n Noam Brown, an OpenAI researcher who helped create the IMO model and o1, told TechCrunch that OpenAI has new general-purpose RL techniques which allow them to teach AI models skills that aren\u2019t easily verified. This was how the company built the model which achieved a gold medal at IMO, he said.<\/p>\n OpenAI\u2019s IMO model was a newer AI system that spawns multiple agents, which then simultaneously explore several ideas, and then choose the best possible answer. These types of AI models are becoming more popular; Google<\/a> and xAI<\/a> have recently released state-of-the-art models using this technique.<\/p>\n \u201cI think these models will become more capable at math, and I think they\u2019ll get more capable in other reasoning areas as well,\u201d said Brown. \u201cThe progress has been incredibly fast. I don\u2019t see any reason to think it will slow down.\u201d<\/p>\n
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 27-29, 2025\n\t\t\t\t\t\t\t<\/p>\n![\"OpenAI](\"https:\/\/www.europesays.com\/uk\/wp-content\/uploads\/2025\/08\/GettyImages-1778704897.jpg\")
OpenAI CEO Sam Altman speaks during the OpenAI DevDay event on November 06, 2023 in San Francisco, California.(Photo by Justin Sullivan\/Getty Images)Image Credits:<\/strong>Justin Sullivan \/ Getty Images<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/uk\/wp-content\/uploads\/2025\/08\/GettyImages-515358462.jpg\")
South Korean professional Go player Lee Se-Dol (R) prepares for his fourth match against Google\u2019s artificial intelligence program, AlphaGo, during the Google DeepMind Challenge Match on March 13, 2016 in Seoul, South Korea. Lee Se-dol played a five-game match against a computer program developed by a Google, AlphaGo. (Photo by Google via Getty Images)<\/p>\n![\"Ilya](\"https:\/\/www.europesays.com\/uk\/wp-content\/uploads\/2025\/08\/019A7E29-3482-4F68-9A29-624BA22F0334.jpeg\")
Ilya Sutskever, Russian Israeli-Canadian computer scientist and co-founder and Chief Scientist of OpenAI, speaks at Tel Aviv University in Tel Aviv on June 5, 2023. (Photo by JACK GUEZ \/ AFP)Image Credits:<\/strong>Getty Images<\/p>\n![\"The](\"https:\/\/www.europesays.com\/uk\/wp-content\/uploads\/2025\/08\/img_5619.jpg\")
The OpenAI logo on screen at their developer day stage. (Credit: Devin Coldeway)Image Credits:<\/strong>Devin Coldewey<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/uk\/wp-content\/uploads\/2025\/08\/GettyImages-2197821080.jpg\")
AI applications (Photo by Jonathan Raa\/NurPhoto via Getty Images)<\/p>\n