Five breakthroughs that make OpenAI’s o3 a turning point for AI — and one big challenge

Artificial intelligence is facing a reckoning at the end of 2024, with industry insiders worried that progress toward smarter AI is slowing. But OpenAI’s o3 model, Just announced last weektriggered A new round of excitement and debateand suggests there will still be significant improvements in 2025 and beyond.

The model has been announced for safety testing among researchers but has not yet been released publicly, Achieved impressive scores on important ARC metrics. The benchmark was created by François Chollet, a renowned artificial intelligence researcher and creator of the Keras deep learning framework, specifically to measure a model’s ability to handle novel intelligence tasks. As such, it provides a meaningful measure of progress toward truly smart AI systems.

Notably, o3 scores 75.7% on the ARC benchmark under standard computing conditions and 87.5% under high computing conditions, significantly surpassing previous state-of-the-art results, e.g. Crowder scored 53% 3.5.

Chollet said o3’s achievement represents surprising progress. was a critic Large Language Models (LLMs) are capable of achieving this intelligence. It highlights innovations that can accelerate the advancement of higher-order intelligence, whether we call it artificial general intelligence (AGI) or not.

AGI is a much-hyped term and ill-defined, but it signals a goal: intelligence capable of adapting to new challenges or problems in ways that exceed human capabilities.

OpenAI’s o3 addresses specific barriers that have long hindered the inference and adaptability of large language models. At the same time, it also exposed the challenges, including the high costs and efficiency bottlenecks inherent in pushing these systems to their limits. This article will explore five key innovations behind the o3 model, many of which are based on advances in reinforcement learning (RL). It will draw on insights from industry leaders, OpenAI statementand most importantly Important analysis of Choletwhat this breakthrough means for the future of artificial intelligence as we move into 2025.

o3’s five core innovations

1. “Program synthesis” of task adaptation

OpenAI’s o3 model introduces a new feature called “program synthesis” that enables it to dynamically combine things learned during pre-training (specific patterns, algorithms, or methods) into new configurations. These may include mathematical operations, code snippets, or logical processes that the model encountered and generalized during extensive training on different data sets. Most importantly, program synthesis enables o3 to tackle tasks it has never directly seen in training, such as solving advanced coding challenges or solving novel logic puzzles that require reasoning beyond rote memorization of learned information. François Chollet describes programmatic synthesis as the ability of a system to recombine known tools in innovative ways, much like a chef using familiar ingredients to create a unique dish. This feature marks a departure from earlier models, which primarily retrieved and applied pre-learned knowledge without reconfiguration, the only viable approach to achieving better intelligence that Chollet advocated a few months ago.

2. Natural language program search

The core of o3 adaptability lies in its use of Chains of Thoughts (CoT) and the complex search processes that occur during inference (when the model actively generates answers in the real world or deployment environment). These CoTs are step-by-step natural language instructions generated by the model for exploring solutions. Guided by the evaluator model, o3 proactively generates multiple solution paths and evaluates them to determine the most promising option. This approach reflects the way humans solve problems, where we brainstorm different approaches before choosing the one that works best. For example, in mathematical reasoning tasks, o3 generates and evaluates alternative strategies to arrive at accurate solutions. Competitors such as Anthropic and Google have tried similar approaches, but OpenAI’s implementation sets a new standard.

3. Evaluator Model: A New Inference

O3 proactively generates multiple solution paths during reasoning and evaluates each solution path with the help of an integrated evaluator model to determine the most promising option. By training evaluators using expert-labeled data, OpenAI ensures o3 has a strong ability to reason about complex, multi-step problems. This feature enables the model to act as a judge of its own reasoning, bringing large language models closer to being able to “think” rather than simply react.

4. Execute your own program

One of o3’s most groundbreaking features is its ability to execute its own Chain of Thoughts (CoT) as a tool for adaptive problem solving. Traditionally, CoT has been used as a step-by-step reasoning framework to solve specific problems. OpenAI’s o3 extends this concept by leveraging CoT as reusable building blocks, enabling models to respond to new challenges with greater adaptability. Over time, these CoTs become structured records of problem-solving strategies, similar to how humans record and refine their learning through experience. This capability demonstrates how o3 is pushing the frontier of adaptive reasoning. according to OpenAI engineer Nat McAleeseo3’s performance in never-before-seen programming challenges, such as a CodeForces rating of over 2700 points, demonstrates its innovative use of CoT to compete with top competing programmers. A score of 2700 puts the model at the “Master” level, placing it at the top of the global programmer competition.

5. Deep learning guided program search

O3 leverages deep learning-driven methods during reasoning to evaluate and refine potential solutions to complex problems. This process involves generating multiple solution paths and using patterns learned during training to evaluate their feasibility. François Chollet and other experts point out that this reliance on “indirect evaluation” (judging solutions based on internal metrics rather than testing them in real-life scenarios) can be problematic when applied to unpredictable or enterprise-specific environments. may limit model robustness.

Furthermore, o3 relies on expert-labeled datasets to train its evaluator model, which raises concerns about scalability. While these datasets improve accuracy, they also require significant human supervision, which can limit the adaptability and cost-efficiency of the system. Chollet emphasized that these trade-offs illustrate the challenges of scaling inference systems beyond controlled benchmarks such as ARC-AGI.

Ultimately, this approach demonstrates the potential and limitations of combining deep learning techniques with stylized problem solving. While o3’s innovations demonstrate progress, they also underscore the complexities of building truly universal AI systems.

big Challenge o3

OpenAI’s o3 model achieved impressive results, but was computationally expensive, consuming millions of tokens per task—this costly approach was the model’s biggest challenge. François Chollet, Nat McAleese and others have highlighted concerns about the economic viability of such models, emphasizing the need for innovations that balance performance with affordability.

The release of o3 has attracted the attention of the entire AI community. Competitors such as Google and Gemini 2 and DeepSeek 3 and other Chinese companies Advances are also ongoing, making direct comparisons challenging until these models are more extensively tested.

Opinions on o3 are divided: some praise its technological advancements, while others point to the high cost and lack of transparency, suggesting its true value will only become clear with more extensive testing. One of the biggest criticisms came from Google DeepMind’s Denny Zhou, who implicitly attacked the model’s reliance on reinforcement learning (RL) scaling and search mechanisms. As a potential “dead end”,” on the contrary, it is believed that the model should be able to learn to reason Easier fine-tuning process.

What this means for enterprise artificial intelligence

Whether or not it represents the perfect direction for further innovation, for businesses, O3’s newfound adaptability suggests that artificial intelligence will continue to transform industries in some way, from customer service to scientific research.

It will take some time for industry players to digest what o3 has to offer here. For companies worried about o3’s high computing costs, OpenAI’s upcoming smaller version of the “o3-mini” model provides a potential alternative. While sacrificing some of the functionality of a full model, o3-mini provides enterprises with a more affordable experimental option—retaining most of the core innovation while significantly reducing the computational requirements for testing.

It may take some time for enterprise-level companies to be exposed to the o3 model. OpenAI said the o3-mini is expected to be launched by the end of January. The full o3 version will be released later, although the timeline depends on feedback and insights gained during the current security testing phase. It is recommended for enterprise companies to conduct testing. They want to build a model with their data and use cases and see how it actually works.

But at the same time, they already have access to many other capable models that are already out and well-tested, including the flagship o4 model and other competing models – many of which are already powerful enough to build smart, custom applications that provide practical value.

In fact, next year we’ll be operating in two gears. The first is realizing real value from AI applications and fleshing out models of what AI agents can do, among other innovations that have already been implemented. The second task is to sit down, eat popcorn, and see how the intelligence competition plays out—any progress will only be icing on the cake that has already been delivered.

For more information about o3 innovation, Watch the full YouTube discussion between myself and Sam Witteveen below, and follow VentureBeat for ongoing reporting on progress in artificial intelligence.