In October 2024, the Nobel Prize in Chemistry went to three people who work on proteins. Two of them, Demis Hassabis and John Jumper, work at an artificial intelligence company. Their model, AlphaFold2, predicts the shape of a protein from its sequence, and by the time the prize was awarded more than two million people from 190 countries had already used it [1]. Since then, machines have entered four other fields: materials, antibiotics, mathematics and the brain. The same pattern repeats in all of them. The machine proposes a huge amount, and someone has to check.
It is clear that AI is here to stay. How humanity uses it will portray our nature, and it is our responsibility to use it to improve society and move forward, not to attack or spy on one another.
In the science of the future, it is increasingly clear that laboratories will mix human work with AI. Most of the best advances and discoveries made with AI have had a human mind guiding and helping it.
Materials
In November 2023, Google DeepMind published in Nature a model called GNoME. It is a graph neural network: it treats each crystal as a network of atoms joined by bonds and learns to predict its energy. With it, the authors listed 2.2 million structures below the previously known stability frontier, of which 381,000 would be stable and new [2]. A crystal is stable if its energy is lower than that of any mixture of competing compounds, which means that, in principle, it can exist. Between 2021 and 2023, all other methods combined had added about 13,000 stable crystals (from 35,000 to 48,000) [2]. DeepMind described it as the equivalent of about 800 years of knowledge [3].
At Berkeley they built the A-Lab, a laboratory where robots mix powders, heat them and analyze the result by X-ray diffraction, using recipes proposed by a language model trained on the scientific literature. In 17 consecutive days it synthesized 36 of the 57 compounds it was aiming for [4]. The first count said 41 of 58. In January 2026, the authors published a correction: “new” meant new to their prediction platform, not necessarily new to science, and a reanalysis of the diffraction patterns confirmed 36 of their 40 declared successes, with four inconclusive [5].
There is another recent finding. The mathematician Vitaliy Kurlin, of the University of Liverpool, detected duplicates and near-duplicates in the large crystal databases. According to him, GNoME quietly removed more than 83,000 entries, over 20% of its database [7]. These are his claims, and I have not found a detailed public response from DeepMind.
Healthcare
In 2020, James Collins’s laboratory at MIT trained a neural network on 2,335 molecules whose ability to inhibit E. coli had been measured in the lab [8]. They then applied it to several chemical libraries. In the Drug Repurposing Hub, a catalog of already-known molecules, the model flagged one called SU3327. In another library of more than 107 million compounds (ZINC15), they tested 23 predictions and 8 turned out to be antibacterial [8].
SU3327 was not an antibiotic. It was known as a JNK kinase inhibitor, and the authors renamed it halicin. It killed Mycobacterium tuberculosis and carbapenem-resistant enterobacteria. In mice it treated infections by Clostridioides difficile and by Acinetobacter baumannii resistant to everything tried [8]. It works by dissipating the pH component of the gradient that the bacterial membrane needs to survive. In 30 days of serial culture no resistance to halicin appeared, whereas with ciprofloxacin it did [8]. Halicin is not in pharmacies and remains preclinical.
In 2023, the same group found abaucin, a narrow-spectrum antibiotic against A. baumannii that controlled a wound infection in mice [9]. That year another study measured 39,312 compounds, predicted 12 million and tested 283 until it identified an entire class of molecules active against Gram-positive bacteria, including Staphylococcus aureus [10]. In 2025, a generative model invented more than 36 million molecules that were not in any database. They synthesized 24 and 7 turned out to be antibacterial. Two worked in mice, one against resistant gonorrhea and the other against MRSA [11]. The funnel goes from 36 million to 7.
Mathematics
AlphaEvolve, a Gemini-based coding agent that DeepMind presented in 2025, writes code, runs it, evaluates the result and modifies it in an evolutionary loop. It found a way to multiply two complex 4×4 matrices with 48 multiplications, versus Strassen’s 49, which had been the record since 1969 [12]. A French team later derived a version with rational coefficients [13]. In computational complexity it improved the inapproximability bound for MAX-4-CUT, from 0.9883 to 0.987, and the machine itself rewrote the verifier of its results until it was, in some cases, 10,000 times faster [14]. A later work raised lower bounds on Ramsey numbers, such as R(3,13) up to 61 and R(3,18) from 99 to 100 [15].
The most famous testing ground is the problems of Paul Erdős, a Hungarian mathematician who lived out of a suitcase for decades and paid prizes out of his own pocket for solving them. The Englishman Thomas Bloom catalogued them on a website, and much of the battle has been fought there. On Christmas Day 2025, a user announced the first autonomous solution of a problem by a language model (number 333). Hours later someone found that Erdős had solved it in 1977 [16]. In January 2026, two young people, Kevin Barreto and Liam Price, solved problem 728 with GPT-5.2 Pro, and the proof was checked with Aristotle, a program that verifies it in Lean [16][17]. Lean is a proof assistant: if the proof compiles, every logical step is correct with respect to the statement that was written.
On May 20, 2026, OpenAI announced that an internal, general-reasoning, non-public model had found a counterexample to the unit distance conjecture, which Erdős posed in 1946 [18]. Nine mathematicians signed a companion text that summarizes and verifies the proof: Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang and Wood [19]. OpenAI had contacted them beforehand [20]. Gowers wrote that, had a human signed it, he would have recommended it to Annals of Mathematics without hesitation, and a few weeks later humans had already improved it [16]. Litt summed it up to Scientific American by saying the model was lucky to hit a case where experts had failed [20].
In August 2026, a team from Tsinghua University, with Damiano Testa (Warwick) and Shing-Tung Yau, presented FormaTheoria, an AI-assisted workflow that produced more than 994,000 lines of Lean across more than 850 files. It covers the Feit–Thompson odd order theorem, Glauberman’s Z* theorem and the Brauer–Suzuki theorem, and reaches the Bender–Suzuki theorem, pieces of the classification of finite simple groups [21]. According to Quantum Bit’s coverage, they finished in about seven months what the previous formalization of Feit–Thompson, with about 15 people, took six years to do, and added three more theorems [22].
A team with researchers from London, Oxford, Cambridge and Johannesburg trained neural networks to distinguish simple from non-simple groups from their generators. They were right about 97% of the time on simple groups and 86% on non-simple ones. Looking at what the network had learned yielded a conjecture, and humans proved it with known results. The authors themselves call it a toy theorem [23]. The machine found the pattern and humans proved the theorem.
Brain–Machine Interface
On March 13, 2026, the Chinese regulator approved NEO, a device from the company Neuracle, developed with Tsinghua researchers, for people with cervical spinal cord injury who have lost hand grip [24]. It is a coin-sized implant placed over the dura mater, under the skull, without penetrating brain tissue. It is combined with a pneumatic robotic glove that performs the grasp [25] and has eight epidural electrodes over the sensorimotor cortex [26]. In mid-July the first commercial surgery was performed, at Huashan Hospital in Shanghai, on a man whose hand was affected by a spinal cord injury from ten years earlier [27].
The registration trial included 32 patients at 11 hospitals, after a feasibility study with 4 [28]. According to the company’s IPO prospectus, 68.8% improved their grip score six months later without the device [29]. The rest of the public data are mostly preprints from Hong Bo’s group: in one, a patient with a complete C4-level injury used the system at home for nine months [26].
In the United States, a trial at the University of California, Davis decoded the attempted speech of a participant with ALS using a 125,000-word vocabulary. On the first day, with 50 words and 30 minutes of data, it was 99.6% accurate. On the second day, with all 125,000, 90.2%. It maintained 97.5% over 8.4 months [30].
References
[1] Royal Swedish Academy of Sciences (2024). The Nobel Prize in Chemistry 2024 – Popular information. NobelPrize.org. https://www.nobelprize.org/prizes/chemistry/2024/popular-information/ — Institutional source
Note: official press release; the AlphaFold2 user figure dates from October 2024.
[2] Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G., Cubuk, E. D. (2023). Scaling deep learning for materials discovery. Nature, 624, 80-85. https://doi.org/10.1038/s41586-023-06735-9 — With reservations
Note: computational study by Google DeepMind. Most of the catalog consists of predictions; I have not verified the number of structures actually synthesized.
[3] Google DeepMind (2023). Millions of new materials discovered with deep learning. https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/ — With reservations
Note: press release from the company that obtained the result.
[4] Szymanski, N. J., Rendy, B., Fei, Y., Kumar, R. E., He, T., Milsted, D., McDermott, M. J., Gallant, M., Cubuk, E. D., Merchant, A., Kim, H., Jain, A., Bartel, C. J., Persson, K., Zeng, Y., Ceder, G. (2023). An autonomous laboratory for the accelerated synthesis of novel materials. Nature, 624, 86-91. https://doi.org/10.1038/s41586-023-06734-w — With reservations
Note: robotic experiment at Lawrence Berkeley National Laboratory (USA), 17 days of continuous operation. The current abstract says 36 of 57 and the original said 41 of 58.
[5] Szymanski, N. J. and 15 more authors (2026). Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature. https://doi.org/10.1038/s41586-025-09992-y — With reservations
Note: published on January 19, 2026. Correction by the authors themselves, with a reanalysis of the diffraction patterns evaluated after publication. Volume and pages not consulted.
[6] Leeman, J., Liu, Y., Stiles, J., Lee, S. B., Bhatt, P., Schoop, L. M., Palgrave, R. G. (2024). Challenges in high-throughput inorganic materials prediction and autonomous synthesis. PRX Energy, 3(1), 011002. https://doi.org/10.1103/PRXEnergy.3.011002 — With reservations
Note: reanalysis of the products of [4] by chemists at Princeton and UCL. It is not an experimental replication.
[7] C&EN (2025, Dec. 16). Duplicate structures haunt crystallography databases. https://cen.acs.org/research-integrity/Duplicate-structures-haunt-crystallography-databases/103/web/2025/12 — Journalistic source (secondary)
Note: author not identified. Reports claims by Kurlin and Widdowson (Liverpool); I did not locate a detailed public response from DeepMind.
[8] Stokes, J. M., Yang, K., Swanson, K., Jin, W., Cubillos-Ruiz, A., Donghia, N. M., MacNair, C. R., French, S., Carfrae, L. A., Bloom-Ackermann, Z., Tran, V. M., Chiappino-Pepe, A., Badran, A. H., Andrews, I. W., Chory, E. J., Church, G. M., Brown, E. D., Jaakkola, T. S., Barzilay, R., Collins, J. J. (2020). A deep learning approach to antibiotic discovery. Cell, 180(4), 688-702. https://doi.org/10.1016/j.cell.2020.01.021 — With reservations
Note: preclinical. The figure of 2,335 training molecules comes from secondary sources. Groups from the USA and Canada.
[9] Liu, G., Catacutan, D. B., Rathod, K., Swanson, K., Jin, W., Mohammed, J. C., Chiappino-Pepe, A., Syed, S. A., Fragis, M., Rachwalski, K., Magolan, J., Surette, M. G., Coombes, B. K., Jaakkola, T., Barzilay, R., Collins, J. J., Stokes, J. M. (2023). Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii. Nature Chemical Biology, 19(11), 1342-1350. https://doi.org/10.1038/s41589-023-01349-8 — With reservations
Note: preclinical; culture and mouse wound model. Same ecosystem of groups as [8].
[10] Wong, F., Zheng, E. J., Valeri, J. A., Donghia, N. M. and 17 other authors (2024). Discovery of a structural class of antibiotics with explainable deep learning. Nature, 626, 177-185. https://doi.org/10.1038/s41586-023-06887-8 — With reservations
Note: 39,312 compounds measured, 12,076,365 predicted, 283 tested. Preclinical.
[11] Krishnan, A., Anahtar, M. N., Valeri, J. A., Jin, W., Donghia, N. M. and others (2025). A generative deep learning approach to de novo antibiotic design. Cell, 188(21), 5962-5979. https://doi.org/10.1016/j.cell.2025.07.033 — Recent
Note: more than 36 million compounds generated, 24 synthesized, 7 antibacterial; efficacy in mice.
[12] Novikov, A., Vũ, N., Eisenberger, M., Dupont, E., Huang, P.-S., Wagner, A. Z., Shirobokov, S., Kozlovskii, B., Ruiz, F. J. R., Mehrabian, A., Kumar, M. P., See, A., Chaudhuri, S., Holland, G., Davies, A., Nowozin, S., Kohli, P., Balog, M. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv:2506.13131. https://arxiv.org/abs/2506.13131 — Recent
Note: technical report from Google DeepMind. The mathematical results are verifiable.
[13] Dumas, J.-G., Pernet, C., Sedoglavic, A. (2025). A non-commutative algorithm for multiplying 4×4 matrices using 48 non-complex multiplications. arXiv:2506.13242. https://arxiv.org/abs/2506.13242 — Recent
Note: preprint by French researchers (CASC and CRIStAL); v7 of July 2026.
[14] Nagda, A., Raghavan, P., Thakurta, A. (2025). Reinforced generation of combinatorial structures: applications to complexity theory. arXiv:2509.18057. https://arxiv.org/abs/2509.18057 — Recent
Note: Google preprint. Authorship according to the preprint in [15]; results verified with problem-specific code.
[15] Nagda, A., Raghavan, P., Thakurta, A. (2026). Reinforced generation of combinatorial structures: Ramsey numbers. arXiv:2603.09172. https://arxiv.org/abs/2603.09172 — Recent
Note: Google preprint; constructions published on GitHub.
[16] Kakaes, K. (2026, Aug. 3). Why the legendary Erdős problems are falling to AI. Quanta Magazine. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/ — Recent
Note: journalistic source with interviews with Bloom, Alon and others. The original results are in preprints and OpenAI and DeepMind press releases.
[17] Sothanaphan, N. (2026). Resolution of Erdős Problem #728: a writeup of Aristotle’s Lean proof. arXiv:2601.07421. https://arxiv.org/abs/2601.07421 — Recent
Note: authorship according to [16]. Proof verified in Lean.
[18] OpenAI (2026, May 20). An OpenAI model has disproved a central conjecture in discrete geometry. https://openai.com/index/model-disproves-discrete-geometry-conjecture/ — With reservations
Note: company press release. The model is not public.
[19] Alon, N., Bloom, T. F., Gowers, W. T., Litt, D., Sawin, W., Shankar, A., Tsimerman, J., Wang, V., Wood, M. M. (2026). Remarks on the disproof of the unit distance conjecture. arXiv:2605.20695. https://arxiv.org/abs/2605.20695 — Recent
Note: preprint. OpenAI chose the verifiers and requested the comments.
[20] Scientific American (2026, May 22). AI just solved an 80-year-old ‘Erdős problem,’ and mathematicians are amazed. https://www.scientificamerican.com/article/ai-just-solved-an-80-year-old-erdos-problem-and-mathematicians-are-amazed/ — Journalistic source (secondary)
Note: author not identified.
[21] Nie, T., Zhang, A., Tang, Y., Testa, D., Yau, S.-T., Li, P., Zhou, Y. (2026). FormaTheoria: constructing large-scale Lean theories from mathematical literature — toward the formalization of the classification of finite simple groups. arXiv:2608.10894. https://arxiv.org/abs/2608.10894 — Recent
Note: preprint from Tsinghua and Warwick. Figures from a snapshot taken when the proofs were completed.
[22] Quantum Bit (量子位) / 36Kr (2026, Aug. 28). AI Completes 15 Mathematicians’ 6-Year Workload in 7 Months. https://eu.36kr.com/en/p/3959028620409988 — Journalistic source (secondary)
Note: the headline exaggerates. The comparison of about 15 people and six years refers only to the earlier formalization of Feit–Thompson in Rocq. The dates in the article itself (Jan. 22–Aug. 2, 2026) add up to about 6.4 months. The details about errata and translation review are attributed to the preprint; I have not checked them against it.
[23] He, Y.-H., Jejjala, V., Mishra, C., Sharnoff, E. (2025). Learning to be simple. AI for Science, 1(2), 025006. https://doi.org/10.1088/3050-287X/ae1d98 — With reservations
Note: accepted on November 7, 2025. Shallow neural network with 1,000 training points; authors from the UK and South Africa.
[24] BrainFacts (2026, Apr. 2). In a first, China approves brain implant for commercial use. https://www.brainfacts.org/neuroscience-in-society/neuroscience-in-the-news/2026/icymi-in-a-first-china-approves-brain-implant-for-commercial-use-040226 — Popular-science source (secondary)
Note: outreach piece from the Society for Neuroscience.
[25] Sixth Tone (2026, Mar. 16). China clears first brain-computer implant for commercial use. https://www.sixthtone.com/news/1018307 — Journalistic source (secondary)
Note: author not identified.
[26] Liu, D., Shan, Y., Wei, P., Li, W., Xu, H., Liang, F., Liu, T., Zhao, G., Hong, B. (2024). Reclaiming hand functions after complete spinal cord injury with epidural brain-computer interface. medRxiv. https://doi.org/10.1101/2024.09.05.24313041 — Recent
Note: preprint from the Tsinghua group. One patient with a complete C4 injury, 9 months of home use, F1 = 0.91 in grasp detection and 100% in object transfer. Eight epidural electrodes. Trial cited: NCT05920174.
[27] South China Morning Post (2026, Jul. 15). China completes world’s first commercial brain-computer interface implant. https://www.scmp.com/tech/big-tech/article/3360684/china-completes-worlds-first-commercial-brain-computer-interface-implant — Journalistic source (secondary)
Note: author not identified. Based on a press release from the Shanghai Science and Technology Commission.
[28] Chen, L., Qing, W., Lin, F. and others (49 authors) (2026). Clinical translation of brain-computer interface in China: a landscape analysis of investigator-initiated trials, registered clinical trials, and regulatory approval. arXiv:2607.07185. https://arxiv.org/abs/2607.07185 — Recent
Note: review preprint (Hainan University and others). Summarizes the 4 + 32 patient trial; it is not the primary source of the data.
[29] Neuracle Technology (Shanghai) Co., Ltd. (2026). 招股说明书（申报稿）[Initial public offering prospectus (draft), STAR Market]. Shanghai Stock Exchange. https://static.sse.com.cn/stock/disclosure/announcement/c/202606/002198_20260611_0RN0.pdf — With reservations
Note: document from the company itself, with a direct interest in the outcome and not peer-reviewed. I read only the first part of the document.
[30] Card, N. S., Wairagkar, M., Iacobacci, C., Hou, X., Singer-Clark, T., Willett, F. R., Kunz, E. M., Fan, C., Vahdati Nia, M., Deo, D. R., Srinivasan, A., Choi, E. Y., Glasser, M. F., Hochberg, L. R., Henderson, J. M., Shahlaie, K., Stavisky, S. D., Brandman, D. M. (2024). An accurate and rapidly calibrating speech neuroprosthesis. New England Journal of Medicine, 391(7), 609-618. https://doi.org/10.1056/NEJMoa2314132 — With reservations
Note: clinical trial with a single participant with ALS (UC Davis, USA); 84 sessions over more than 8 months.
[31] Vaccaro, M., Almaatouq, A., Malone, T. W. (2024). When combinations of humans and AI are useful: a systematic review and meta-analysis. Nature Human Behaviour, 8, 2293-2303. https://doi.org/10.1038/s41562-024-02024-1 — Reliable
Note: preregistered meta-analysis of 106 studies and 370 effect sizes, published between January 2020 and June 2023. Distribution by country not verified.
[32] Google Quantum AI and collaborators (2025). Quantum error correction below the surface code threshold. Nature, 638, 920-926. https://doi.org/10.1038/s41586-024-08449-y — With reservations
Note: experiment with the 105-qubit Willow processor (USA); a single group and a single platform. It measures a quantum memory, not logical operations.
Note on the labels. “With reservations” on [2]–[6], [8]–[10], [18], [23], [29], [30] and [32] indicates a single finding, lack of independent replication, a preclinical or disputed result, or publication by the company that obtained the result. It does not indicate doubt about the measurement. “Recent” on [11]–[17], [19], [21], [26] and [28] indicates an unreplicated result or a preprint. Journalistic, popular-science and institutional sources carry a categorical label because they are not studies.