Google has done it again.
In 2024, John Jumper and Demis Hassabis of Google Mind in London shared the Nobel Prize in chemistry with David Baker for, as the Nobel committee put it, succeeding “with the almost impossible feat of building entirely novel proteins,” and more to the point of this essay, “developing an AI model to solve a 50-year-old problem: Predicting protein’s complex structures.”
To meet the challenge, Jumper and Hassabis developed AlphaFold 2, a machine language model designed to figure out the 3D structure of proteins based on the sequence of the protein’s amino acids.
Google used data from over 200,000 proteins whose 3D structures were discovered using X-ray crystallography, nuclear spectroscopy, and more recently cryo-electron microscopy.
That database and the known 3D structures of evolutionarily closely related proteins, coupled with an iterative process in which protein models were progressively refined by running them through the program repeatedly until they reached near atom-level precision, paid off.
Just as impressive, they solved the puzzle quickly with near-perfect precision.
Now Google has developed a similar program, GenCast, designed to accurately predict weather forecasts for 15 days — a far more difficult task than 10 days.
GenCast ran circles around a 2023 version of the program as well as a program, called Ensemble Prediction System, designed by the highly respected European Centre for Medium-Range Weather Forecasts — hitherto the world leader in medium-range weather prediction.
Like AlphaFold 2, GenCast uses hard data — in this case, weather data collected over a 40-year period with known weather outcomes, rather than depending on, as the European program does, masses of now-time weather to predict what will happen over the next 15 days in the weather.
Extending precise predictions out to 15 days is difficult for programs based on real-time weather because so many factors are at play with weather.
Like AlphaFold 2, Google’s GenCast depends on knowns — known 3D structures for proteins in the case of AlphaFold 2 and documented past weather supplemented by recent data in the case of 15-day weather forecasts.
Like AlphaFold 2, Google published their key paper in what is the best general science journal in the world, Nature, and shared its data with others including the European Centre for Medium-Range Weather Forecasts from whom they received most of their high-quality past weather information about past weather.
What’s surprising about AlphaFold 2 and GenCast is how quickly they were developed and improved dramatically, the speed with which they can analyze data and the similarities in approach.
It suggests to me that what we are about to see will be applications of similar programs to solving other challenging problems in science and possibly long-term weather (climate) and other applications that I can’t begin to imagine now tackled.
And not just Google is involved in AI. There are other western and eastern countries working on similar AI programs.
If that is not enough, faster chips, larger memories and better problem-solving algorithms are in development.
However, if there is a giant in the room right now, it’s quantum computing.
It took 50 years to get to where we are now with silicon chip-based computers. Marvellous as they are, they will be superseded by quantum computers within the next few decades and possibly much earlier. The basic building block of the silicon chip computer is on or off or put another way, one or zero.
However, quantum computers can consider all possible values — not just two possibilities. That innate capability cuts two ways. Quantum computers can simultaneously handle much more data.
The downside, so far, is that quantum computers are unstable and make frequent errors.
That error rate is coming down and as it does, quantum computers will reach a level when they are as error-free as digital computers but much, much faster and capable of handling far more data simultaneously. That is the promise.
But they are not there yet even though Google recently announced that they have reduced errors significantly and their experimental quantum computers can handle far more complex problems, much more quickly than silicon computers can now or are ever likely to be able to achieve.
Welcome to 2025.
Dr. William Brown is a professor of neurology at McMaster University and co-founder of the InfoHealth series at the Niagara-on-the-Lake Public Library.