2025 Week 31 news

This week we look at:

• AI models achieving golds at the maths Olympics whilst knowing their own limits
• New research that reveals which jobs AI impacts most
• How AI infrastructure spending is now contributing more to US GDP than consumers

Self-aware AI climbs down from Mount Stupid

As we reported last week, both Google DeepMind and OpenAI achieved gold medal performances at the 2025 International Mathematical Olympiad (IMO), scoring 35/42 points. This week, Google released a de-tuned (with reduced thinking time) version of the winning system in the form of Gemini 2.5 Deep Think, which could usher in the next phase in the development of the reasoning model. As details of the IMO record-breaking attempts emerge, it appears the most striking aspect of these results isn’t the elite problem-solving. OpenAI’s model looked at the hardest problem in the competition, one designed to separate the very best human mathematicians and refused to attempt it. For anyone working with language models, this refusal is revolutionary. Previous generations of LLMs would have confidently produced a plausible-sounding but entirely incorrect proof. This new model knew that it didn’t know.

This represents what you might define as AI’s progression along the Dunning-Kruger curve. Early LLMs were perched atop “Peak Mount Stupid,” confidently hallucinating answers to questions far beyond their capability. They were the quintessential overconfident amateurs who didn’t know what they didn’t know. The 2025 IMO models, however, have developed epistemic awareness; the ability to assess their own limitations and the confidence of their own reasoning.

Noam Brown from OpenAI, a key contributor to the development of the o-series reasoning models, specifically highlighted this as a breakthrough, addressing a major complaint from experts that previous models would “output a very convincing but wrong answer” when stumped. The experimental OpenAI model expressed uncertainty naturally, using phrases like “good!” when confident and question marks or “seems hard” when unsure.

The technical implementations for achieving this “AI self-awareness” vary, offering insights into the future of reliable AI architectures. It is clear that progress is no longer solely about the raw capability of base models, but about the sophisticated systems engineered to manage the reasoning process.

OpenAI’s approach emphasises internal calibration. They have incorporated confidence scoring for each reasoning step, creating an internal “reinforcement learning signal” that guides the model’s exploration. If the confidence score drops below a certain threshold during the reasoning process, the model can halt, re-evaluate, or, as seen in the IMO, simply refuse the task. Google DeepMind’s Deep Think, by contrast, achieves reliability through a structured, specialist architecture focused on sophisticated search strategies.

Traditional reasoning models rely on a linear chain-of-thought, essentially a depth-first search where one flawed step early on can corrupt the entire outcome, leading to hallucinations. Deep Think utilises a “parallel thinking” methodology. Internally, this functions like an advanced, tree-search algorithm. When faced with a complex problem, Deep Think doesn’t just pursue one solution; it spawns multiple concurrent hypotheses and explores them simultaneously. This breadth-first exploration dramatically increases the robustness of the reasoning process.

Crucially, managing this parallelism efficiently requires more than just raw compute. Deep Think leverages novel reinforcement learning techniques trained on a “curated corpus of high-quality solutions.” This specialised training allows Deep Think to develop an internal intuition for recognising whether a specific reasoning path is likely to succeed. This enables the system to efficiently prioritise promising paths and, equally importantly, discard dead ends early, optimising its computational budget.

This combination of parallel architecture and learned intuition allows for rapid convergence on solutions; during our testing, Deep Think proved more capable on several complex tasks than Grok 4 Heavy. Furthermore, Deep Think maintains strict operational boundaries. The model card reveals it will sometimes “over-refuse” queries to ensure safety compliance, a policy-driven form of caution.

Ultimately these varying approaches are converging on the same outcome: systems that know when they are wrong. This is a critical necessity for future business value. This new level of reliability isn’t achieved by eliminating hallucination through perfect training data or larger parameter counts. It is achieved by building systems, whether based on internal calibration, parallelised search, or external verification, that can verify their progress, on hitherto hard to verify problems, as they think.

The implications extend far beyond mathematics competitions. In high-stakes fields like medical diagnosis, complex engineering, legal analysis, or financial modelling, the difference between a system that hallucinates confidently and one that appropriately expresses uncertainty could in the future be measured in lives or billions of pounds. An AI that can accurately assess the validity of its own reasoning process is the prerequisite for true autonomous agency in the enterprise. A model that can say “I don’t know” is paradoxically far more useful than one that always tries to be helpful.

The leap from models that could barely solve primary school maths in 2023 to achieving IMO gold with a few hours of reasoning represents just the opening chapter. We’re entering an era where AI reasoning time will stretch from minutes to days, potentially weeks, as these systems tackle increasingly complex challenges. The next frontier isn’t merely about solving competition problems faster, but about sustained reasoning that mirrors how humans actually work, iterating over ideas for months, pursuing dead ends, backtracking, and eventually arriving at genuine breakthroughs.

Takeaways: The 2025 IMO results suggest a promising transition from confident hallucination to appropriate uncertainty (epistemic awareness). Within the next few years, we’ll likely see models tackling far harder research problems, maintaining context over weeks of exploration, and perhaps even making inroads into problems that have stumped humanity’s brightest minds. The intelligence revolution won’t just be about speed; it will be about self-awareness, persistence, and the kind of prolonged contemplation that transforms fields of knowledge entirely.