😸 OpenAI solved 5 of 10 "impossible" problems

2026-02-15 · Source: The Neuron · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Software Development & Engineering · Depth: Intermediate, long

Summary

OpenAI's unreleased model has reportedly solved at least five of ten "First Proof" research-level math problems, a set of unpublished challenges created by top mathematicians to test AI's raw reasoning without prior training data exposure. Publicly available models like ChatGPT and Gemini could only solve two. Concurrently, on February 13th, OpenAI published a physics preprint where GPT-5.2 proposed a formula for gluon particle interactions, a problem physicists considered impossible for decades, which was subsequently verified by Harvard and Cambridge researchers. This demonstrates significant advancements in AI's scientific reasoning capabilities. Separately, OpenAI has quietly removed "safely" and its "openly share" commitment from its IRS mission filings over the years.

Key takeaway

For AI scientists and researchers evaluating the frontier of AI capabilities, these developments signal a rapid shift from AI assisting science to AI actively performing breakthrough scientific discovery. You should anticipate AI models, even unreleased ones, to tackle increasingly complex, novel problems in mathematics and physics, potentially accelerating research timelines. Consider how your research roadmap might integrate AI for hypothesis generation or problem-solving in areas previously deemed intractable for automated systems.

Key insights

OpenAI's unreleased models are demonstrating advanced scientific reasoning, solving complex math and physics problems previously thought impossible for AI.

Principles

• AI can achieve breakthroughs in fundamental science.
• Unseen problems test true AI reasoning, not pattern matching.

Method

OpenAI's internal model tackled 10 unpublished math problems in a week, with limited human oversight for output review and expansion, but without proof strategy input.

In practice

• Use negative examples in prompts to improve AI accuracy by 20%.
• Provide output templates to AI for significant quality gains.

Topics

• OpenAI Research
• AI Agent Development
• Prompt Engineering
• AI Ethics & Governance
• Global AI Adoption

Code references

• google-gemini/gemini-skills

Best for: Machine Learning Engineer, NLP Engineer, AI Scientist, AI Engineer, Prompt Engineer, Tech Journalist

Related on AIssential

• OpenAI cracks an 80-year math belief — A general-purpose AI model achieved novel mathematical discovery, signaling advanced AI capabilities for scientific breakthroughs.
• An OpenAI model has disproved a central conjecture in discrete geometry — An OpenAI model disproved a long-standing mathematical conjecture using algebraic number theory, marking a significant AI reasoning milestone.
• OpenAI just SOLVED MATH.... — AI's breakthrough stemmed from its ability to connect disparate mathematical fields, revealing novel solutions.
• OpenAI makes breakthrough on 80-year-old maths problem — OpenAI's AI disproved an 80-year-old math conjecture, showing AI can find novel solutions in complex problems.
• Inference: When Will We Solve Intelligence — Solving intelligence requires a hybrid approach combining formal verification with large language models for robust reasoning.
• OpenAI’s math breakthrough played to AI’s strengths — AI disproved an 80-year-old math conjecture by applying broad knowledge and persistent strategy exploration, marking a significant autonomous contribution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Neuron.