OpenAI Research Chief: AI Breakthroughs Are Two Years Away, Polish Startup Proves It

GPT-5.4, paired with the autonomous lab of Polish startup Molecule.one, independently found a way to improve a difficult chemical reaction used in drug production. OpenAI published the results of the experiment on June 17, and the company's chief research officer, Mark Chen, said this week in Seoul that similarly Nobel-caliber discoveries will emerge within the next two years.

The reaction in question is Chan-Lam coupling, a copper-catalyzed reaction that forms carbon-nitrogen bonds found in more than 91 drugs approved by the US FDA, ranging from oncology to cardiology treatments. Combining sulfonamides with boronic acids has produced low yields for years and has been a well-known bottleneck in the early stages of drug discovery.

How the experiment worked

In March, OpenAI gave GPT-5.4 an open-ended task: find a way to improve an important reaction in medicinal chemistry. The model was paired with Maria, an agentic chemistry AI built by Molecule.one and integrated with a lab capable of automatically running thousands of reactions at microliter scale.

Molecule.one scientists wrote the prompts that steered and evaluated the system's work, while human chemists approved the top-rated proposals before they were physically carried out in the robotic lab near Warsaw. The model itself proposed adding TEMPO, a stable organic radical that has been sold for decades and is mainly used in alcohol chemistry.

A result that surprised chemists

Combining two distant areas of chemical literature that had not previously been systematically linked improved yield in most of the tested compounds. Higher yields than the standard reaction were recorded in 88 percent of boronic acid cases and 83 percent of sulfonamide cases.

In manual bench-scale validation, 11 of 14 substrate pairs produced a higher yield, and in 8 of 14 cases the yield more than doubled. Four external reviewers judged the result to be novel, and the full protocols and reaction condition tables were made publicly available in a preprint.

The use of TEMPO in Chan-Lam coupling is poorly studied and has for the first time been tested this extensively - Piotr Byrski, CEO of Molecule.one

A Nobel in two years

Mark Chen spoke about the future of AI in science during a visit to OpenAI Korea's office in Seoul's Gangnam district, on the sidelines of the ICML international machine learning conference held in the city. As chief research officer, he is responsible at OpenAI for the company's research strategy and the allocation of compute.

Within the next two years, AI will generate breakthrough innovations worthy of the biggest prizes - Mark Chen, OpenAI chief research officer

Chen cautioned that it's uncertain whether any specific discovery will ever be honored with a Nobel Prize, but that it will hold enormous value for the world. As reference points, he cited the 2024 Nobel Prize in Chemistry for AlphaFold, a Chinese system solving a decade-old mathematical problem in 80 hours in April 2026, and OpenAI's chatbot solving Paul Erdős's unit-distance problem, posed in 1946, in May.

Poland's role in the AI-for-science race

Molecule.one has for years been developing a combination of AI-based retrosynthesis models with real experiments in its automated lab near Warsaw. The collaboration with OpenAI fits into the California company's broader strategy: in February 2026 it announced with Ginkgo Bioworks a 40 percent cut in protein synthesis costs, and it had earlier launched a specialized model, GPT-Rosalind, and the LifeSciBench benchmark for life science tasks.

For Poland's tech scene, usually associated more with delays in adopting AI than with building it, a domestic startup's role in one of OpenAI's flagship projects is a rare instance of that narrative being reversed. Ahmed El-Kishky of OpenAI described the result as showing what becomes possible when state-of-the-art language models are combined with specialized scientific agents.

The project is still a demonstration: human chemists approved every step, and the yield improvement from 16.6 to 25.2 percent does not yet translate into a finished drug. OpenAI and Molecule.one announced further tests on other difficult reactions, treating this result as evidence that autonomous AI systems can meaningfully speed up the early stages of drug discovery.

Sources: Let's Data Science (letsdatascience.com), R&D World (rdworldonline.com), VnExpress International (e.vnexpress.net)