01What GPT-5.4's Math Breakthrough Is and How It Works

GPT-5.4 solved a 60-year-old math problem that some of the smartest mathematicians on Earth couldn't crack. The problem is a conjecture about primitive sets, which are sets of numbers where no number divides another. Think of prime numbers: they are a famous example of a primitive set.

The conjecture was posed in 1966 by three legendary mathematicians: Paul Erdős, András Sárközy, and Andrei Šmeridi.

This wasn't a textbook exercise. It was a real research-level problem. For decades, mathematicians used a probabilistic approach to attack it.

They would convert the problem into a probability problem and work from there. That seemed like the natural way in. GPT-5.4 rejected that path entirely.

02Key Features of GPT-5.4's Reasoning Ability

  • GPT-5.4 doesn't just retrieve facts or pattern-match. It reasons through problems in ways that don't follow conventional paths.
  • The AI found a completely different proof path that the entire field had missed for 60 years.
  • It used the van Mangold function, a piecewise-defined function that most math students think looks weird and arbitrary. But this function encodes a fundamental identity about prime factorization that cracked the exact part of the problem everyone had been stuck on.
  • This is a major jump in reasoning capability over previous models.

03How GPT-5.4 Got Started on the Problem

  1. GPT-5.4 received the problem description and the existing knowledge about primitive sets.
  2. Instead of using the standard probabilistic approach, it kept the analysis in a different mathematical language the entire time.
  3. It selected the van Mangold function as its core tool, even though human mathematicians had largely ignored it for this type of problem.
  4. The AI then generated a proof that used that function to solve the conjecture.
  5. Stanford professor Jared Lichtman, who had worked on related problems for four years, reviewed and verified the proof.

04How to Use GPT-5.4 for Complex Problem-Solving

To use GPT-5.4 like this in your own work, start by defining a problem you've been approaching the same way for a long time. Identify the conventional wisdom in your field. Ask the AI to generate multiple approaches without telling it what the standard method is.

Lichtman compared this to chess: AI can find new opening lines that humans missed because of aesthetic bias. In your business, you can prompt GPT-5.4 to analyze workflows, marketing strategies, or client processes. Ask it to find a path that nobody has tried because it seemed too weird or too obvious.

The AI doesn't carry the same assumptions as human experts.

05Results and Comparisons from the Episode

The result is clear: GPT-5.4 solved a 60-year-old open conjecture. Terry Tao, widely considered the best living mathematician, had already suspected the standard probabilistic approach was unnecessarily complicated. GPT-5.4 confirmed that suspicion by going around it entirely.

Lichtman called the solution "surreal." He had spent four years on a related version of the problem and proved one part in 2022. The other part, the one from 1966, was still open until GPT-5.4 found a new path. The AI didn't just find an answer — it found a verifiably correct method that humans had completely overlooked.

06Tips, Limits, and Mistakes to Avoid

  • Don't assume AI can work alone. GPT-5.4 produced the insight, but Lichtman's domain knowledge was essential to verify it. Without his expertise, the output would have been meaningless text.
  • Beware of cognitive traps. The more you know about a field, the more certain routes feel wrong. AI doesn't have that bias. Use it to challenge your own assumptions.
  • Don't oversell the result. This wasn't a fully autonomous solution. It was a collaboration between AI and human expert.
  • Focus on method, not just math. The real lesson is how AI can find new approaches in any domain — marketing, law, engineering, medicine.

07Who This Breakthrough Is For and Why It Matters

This breakthrough matters for anyone who relies on conventional wisdom in their work. Business owners, marketers, engineers, lawyers, and creatives all have processes and assumptions that have been in place for years. Those are the places where AI can find a better path.

The episode emphasizes that your job isn't competing with AI at generating ideas. Your job is to understand your domain deeply so you can evaluate, verify, and apply what AI produces. Domain expertise still matters.

Lichtman's four years of work meant he could recognize the validity of the proof. Build that depth in your own field, and AI becomes your most powerful collaborator.

08What's Next After GPT-5.4's Math Solution

Lichtman announced on X that more information about this result is coming soon. If GPT-5.4's method generalizes into broader number theory, it could simplify a whole area of mathematics that's been tangled for decades. Terry Tao suspected the probabilistic approach was a detour — this proof may confirm that.

For practical use, the trajectory is clear. GPT-5.4 Pro is better at open-ended reasoning than anything before. The next version will be even better. Start identifying the Erdos moments in your own work — those problems where the field has accepted "that's just how things are." Those are the exact problems AI can solve with a fresh approach.