AI acceleration: problems unsolved for years that are being solved today

In recent years artificial intelligence has stopped being only a tool for writing text or summarising documents. In some scientific fields it has started to answer problems that had been open for decades. This is not a promise: these are published results with verifiable data. In this article we describe five of them, stating for each what was shown and what was not.

Proteins: a fifty-year problem

A protein is a chain of amino acids that folds into a three-dimensional shape, and the shape determines its function. Since the 1970s researchers had tried to predict that shape from the sequence alone, with poor results. In 2020 Demis Hassabis and John Jumper presented AlphaFold2. According to the Royal Swedish Academy of Sciences, the model made it possible to predict the structure of virtually all of the roughly 200 million known proteins, and it has been used by more than two million people in 190 countries. For this work the two received the 2024 Nobel Prize in Chemistry, announced on 9 October 2024 together with David Baker.

Antibiotics: looking where nobody was looking

Antibiotic resistance is a well-known problem, and finding truly new molecules is slow and expensive. A group at MIT trained a neural network to recognise molecules with antibacterial activity. In 2020, in Cell, it reported the discovery of halicin: the model had screened more than 107 million molecules and, of 23 predictions tested in the lab, eight turned out to be antibacterial and structurally distant from known antibiotics. In 2023 the same approach led to abaucin, active against Acinetobacter baumannii, a hospital pathogen that is often multidrug-resistant. The results are preclinical: test tubes and, for abaucin, a mouse wound model. These are not approved drugs.

Weather: ten days in under a minute

Traditional weather forecasting requires enormous computing resources. GraphCast, a Google DeepMind model described in Science in 2023, predicts hundreds of atmospheric variables ten days ahead, at 0.25-degree resolution, in under a minute on a single chip. According to its authors it outperformed the operational deterministic system of the European centre (ECMWF) on about 90% of 1,380 verification targets, with better results on tropical cyclone tracking as well. This is a comparison on historical data: it does not replace operational systems, but it shows that a data-driven approach can compete with methods that need far more computation.

Materials: millions more candidates

Finding a new stable crystal, useful for batteries or semiconductors, has long been a matter of trial and error. With GNoME, DeepMind used graph neural networks to predict the stability of new materials. According to the paper published in Nature in 2023, the system identified more than 2.2 million structures that are stable relative to earlier data, of which 381,000 are new entries on the updated stability frontier, for a total of 421,000 known stable crystals. A caveat: these are computed predictions. Each material still has to be synthesised and measured in the lab before anyone can talk about applications.

Mathematics: proofs checked by the machine

On 25 July 2024 Google DeepMind announced that AlphaProof and AlphaGeometry 2 had solved four of the six problems at that year's International Mathematical Olympiad, scoring 28 out of 42, the level of a silver medal. They included the hardest problem of the contest, solved by five contestants. Two clarifications, also reported by the authors: the problems were translated into formal language by hand, and some took up to three days of computation, against the nine hours given to students. The advantage of formal language (Lean) is that every proof can be checked automatically.

What these cases have in common

In all five, AI works on a well-defined problem, with clear data or rules and an objective way to measure the outcome. It speeds up the search phase; it does not replace verification: experiments, peer review, human judges. This is also why these results are credible.

What it means for businesses

For a small business the lesson is practical, not scientific. AI makes work faster where there is a precise goal and a way to check the result: classifying, comparing, forecasting, searching among many options. It works less well where reliable data or clear responsibility is missing. Before adopting a tool it is worth asking which repetitive problem you want to speed up and who verifies the outcome. If you want to think it through for your own case, ElaraTeq can help you assess where AI makes sense.

Sources

• Royal Swedish Academy of Sciences, press release for the 2024 Nobel Prize in Chemistry, 9 October 2024: https://www.nobelprize.org/prizes/chemistry/2024/press-release/
• Stokes et al., A Deep Learning Approach to Antibiotic Discovery, Cell 180, 2020: https://www.cell.com/cell/fulltext/S0092-8674(20)30102-1
• Liu et al., Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii, Nature Chemical Biology, 2023: https://www.nature.com/articles/s41589-023-01349-8
• Lam et al., Learning skillful medium-range global weather forecasting, Science 382, 2023: https://www.science.org/doi/10.1126/science.adi2336
• Merchant et al., Scaling deep learning for materials discovery, Nature, 2023: https://www.nature.com/articles/s41586-023-06735-9
• Google DeepMind, AI achieves silver-medal standard solving International Mathematical Olympiad problems, 25 July 2024: https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/