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

The end of 2024 has brought reckoning for artificial intelligence, as industry insiders fear progress towards smarter AI will slow. But the OpenAI o3 model, Announced just last weekRaised A A new wave of excitement and debateHe points out that significant improvements still need to be made in 2025 and beyond.

This model has been announced for safety testing among researchers, but has not yet been released publicly. Achieve an impressive score on the important ARC metric. This benchmark was created by François Cholet, a renowned AI researcher and creator of the Keras deep learning framework, and is specifically designed to measure a model’s ability to handle new and intelligent tasks. As such, it provides a useful measure of progress toward truly intelligent AI systems.

It is worth noting that o3 scored 75.7% on the ARC benchmark under standard computing conditions and 87.5% using high computing, significantly surpassing previous recent results, such as 53% scored by Claude 3.5.

This milestone for o3 represents a surprising breakthrough, according to Chollet, who had achieved it He was critical The ability of large language models (LLMs) to achieve this kind of intelligence. It highlights innovations that could accelerate progress toward superior intelligence, whether we call it artificial general intelligence (AGI) or not.

Artificial general intelligence is an exaggerated and ill-defined term, but it points to a goal: intelligence that can adapt to new challenges or questions in ways that go beyond human capabilities.

OpenAI’s o3 addresses specific hurdles in reasoning and adaptability that have long hampered large language models. At the same time, it reveals 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 rely on the insights of industry leaders, OpenAI claimsAnd above all Important analysis of CholletTo illustrate what this breakthrough means for the future of AI as we move to 2025.

O3’s five core innovations

1. “Program tuning” to adapt tasks

OpenAI’s o3 model introduces a new capability called “program synthesis,” which enables it to dynamically combine things it has learned during prior training — specific patterns, algorithms, or techniques — into new configurations. These things may include mathematical operations, code snippets, or logical procedures that the model has encountered and generalized during its extensive training on various data sets. More importantly, the software synthesis allows o3 to tackle tasks it has never seen directly in training, such as solving advanced programming challenges or tackling new logic puzzles that require thinking beyond the routine application of learned information. François Cholet describes software synthesis as the ability of a system to recombine known tools in innovative ways, such as a chef preparing a unique dish using familiar ingredients. This feature represents a departure from previous models, which primarily retrieve previously acquired knowledge and apply it without reconfiguration, and is also a feature that Cholet has been championing for months as the only viable way forward toward better intelligence.

2. Find a natural language program

The core of o3’s adaptability lies in its use of Chains of Thought (CoTs) and the sophisticated searching process that occurs during inference – when the model effectively generates answers in a real-world or deployed environment. These CoTs are step-by-step natural language instructions that the model generates to explore solutions. Guided by the evaluation model, o3 actively creates multiple solution paths and evaluates them to determine the most promising option. This approach reflects human problem solving, where we brainstorm different ideas before choosing the best approach. For example, in mathematical reasoning tasks, o3 generates and evaluates alternative strategies to arrive at accurate solutions. Competitors like Anthropic and Google have tried similar approaches, but OpenAI’s implementation sets a new standard.

3. The evaluator model: a new type of reasoning

O3 generates multiple solution paths during inference, evaluating each of them with the help of an integrated evaluator model to select the most promising option. By training the assessor on expert-rated data, OpenAI ensures that o3 develops a strong ability to reason through complex, multi-step problems. This feature allows the model to act as an arbiter of its own logic, bringing large language models closer to being able to “think” rather than simply respond.

4. Implementing its own programs

One of the most groundbreaking features of o3 is its ability to implement its own Chains of Thought (CoTs) as tools for adaptive problem solving. Traditionally, CoTs have been used as step-by-step reasoning frameworks for solving specific problems. OpenAI’s o3 system expands this concept by leveraging CoTs as reusable building blocks, allowing the model to meet new challenges with greater adaptability. Over time, these learning techniques become organized records of problem-solving strategies, similar to the way humans document and improve their learning through experience. This ability demonstrates how o3 pushes the boundaries in adaptive thinking. according to OpenAI engineer Nat McAleese,o3’s performance in unseen programming challenges, such as achieving a CodeForces rating above 2700, demonstrates its innovative use of CoTs to compete against the best competitive programmers. This rating of 2700 places the model at the “Grandmaster” level, among the highest levels of competitive programmers globally.

5. Deep search for learning-oriented programs

O3 leverages a deep learning approach during inference to evaluate and improve potential solutions to complex problems. This process involves creating multiple solution paths and using patterns learned during training to evaluate their feasibility. François Chollet and other experts note that this reliance on “indirect evaluations” — where solutions are judged on internal metrics rather than tested in real-world scenarios — can limit the power of the model when applied to unexpected or organization-specific contexts.

Additionally, o3’s reliance on expert-labeled datasets to train its evaluation model raises scalability concerns. While these datasets enhance accuracy, they also require a significant amount of human oversight, which can limit the system’s adaptability and cost-effectiveness. Cholet highlights that these trade-offs illustrate the challenges of scaling inference systems beyond controlled standards such as ARC-AGI.

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

The big one Challenge for o3

OpenAI’s o3 model achieves impressive results but at a significant computational cost, consuming millions of tokens per task – and this expensive approach is the model’s biggest challenge. François Chollet, Nat McAleese, and others highlight concerns about the economic viability of such models, and stress the need for innovations that balance performance and affordability.

The release of o3 has sparked interest across the AI ​​community. Competitors like Google with Gemini 2 and Chinese companies like DeepSeek 3 are also progressing, making direct comparisons difficult until these models are tested on a larger scale.

Opinion on o3 is divided: some praise its technical strides, while others point to its high costs and lack of transparency, suggesting its true value will only become clear through broader testing. One of the biggest criticisms came from Denny Chu of Google DeepMind, who implicitly attacked the model’s reliance on the scaling and search mechanisms of reinforcement learning (RL). As a potential “dead end”.“, arguing instead that the model should be able to learn to reason from it Simpler tuning Operations.

What this means for enterprise AI

Whether or not it represents the perfect direction for further innovation, for enterprises, o3’s newfound adaptability shows that AI will continue in one way or another to transform industries, from customer service to scientific research, in the future.

Industry players will need some time to digest what o3 has brought to the table here. For organizations concerned about the high computational costs of o3, OpenAI’s upcoming release of the “o3-mini” model offers a potential alternative. Although it sacrifices some of the capabilities of the full model, the o3-mini promises a more affordable option for businesses to try – retaining much of the core innovation while dramatically reducing compute requirements at test time.

It may take some time before companies can get their hands on the o3 model. OpenAI says the o3-mini is expected to launch by the end of January. The full o3 version will be released next, although timelines depend on feedback and insights gained during the current safety testing phase. Enterprise companies will be advised to test it. They’ll want to connect the model to their data and use cases and see how it actually works.

But at the same time, they can already use many other competent models that are already well-tested, including the pioneering o4 model and other competing models – many of which are already powerful enough to build intelligent, tailored applications that deliver practical value.

In fact, next year, we’ll be working on two speeds. The first is to bring practical value from AI applications, demonstrating what models can do with AI agents, and other innovations that have already been achieved. The second is to sit down with popcorn and see how the IQ race goes, and any progress will be just the icing on the cake that has already been delivered.

For more about o3 innovations, Watch the full discussion on YouTube between myself and Sam Witteveen Below, and follow VentureBeat for ongoing coverage of AI developments.

Daily insights into business use cases with VB Daily

If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from organizational transformations to hands-on deployments, so you can share insights to maximize ROI.

Read our privacy policy

Thanks for subscribing. Check more VB newsletters here.

An error occurred.