Math provides the proof. OpenAI says AI makes original ideas.

This article summarizes an episode of OpenAI’s video series featuring its researchers, Sébastien Bubeck and Ernest Ryu.

Sébastien Bubeck and Ernest Ryu, researchers at OpenAI, says AI systems to do most human research work within two years.

This fast timeline turns complex math, long-term independent projects, and the strict checking of computer-generated ideas into urgent problems that leaders must fix today.

Math provides a clear test for AI progress

Companies often misunderstand AI limits by relying on vague tests that easily hide software mistakes. Exact fields like math offer a much better test because they demand clarity.

Bubeck argues, “The nice thing about mathematics is that the questions are very clear, non-ambiguous. You can verify the answer. Once a model can give an answer, everybody will agree: was it correct or was it not correct.”

Surviving long strings of thought
The true test of machine thinking is not answering a single prompt, but surviving a long chain of thought.

“To resolve a problem, you have to think for a long time and think consistently,” Bubeck explains. If an AI makes a single mistake in a multi-step argument, the entire proof fails.

Therefore, the ultimate goal for advanced models is the ability to independently spot and correct their own errors mid-process.

The fight over what makes a real scientific discovery

As AI passes these strict math tests, it raises expectations for historic breakthroughs. Tech leaders are now debating whether these systems actually generate original ideas or finding hidden links between old academic papers.

Public tests of AI solving historic puzzles often start intense debates about its actual methods. For example, Bubeck initially tweeted about an AI solving a complex math problem, which was actually a deep search where the AI scanned thousands of papers to connect unrelated fields.

However, he notes that internal labs have since moved far beyond simply searching for past work.

“A few months later, we have more than ten actual solutions that are completely new, publishable in top journals in combinatorics,” he reports, proving that models are now writing truly original, groundbreaking theorems.

Independent research requires time and mental endurance

While generating new theorems is impressive, real scientific breakthroughs needs steady focus across weeks of careful testing. Short chats can create great ideas, but current systems still require a strict human supervisor to guide and verify every shift in direction.

Long-term memory controls the future of research tools

Human knowledge demands a higher price as systems grow stronger

Schools and journals will rebuild publishing and checking rules