The Hardest Problems — Signal

Abstract

For centuries, humanity's hardest problems shared one cruel trait: they were too big to search. Too many molecules to test, too many storms to model, too many variables for any team to try in a lifetime. Artificial intelligence changed that arithmetic — it can explore millions of possibilities in the time a lab runs one. This brief looks at four fronts — energy, food, disaster and water — where that speed is already producing real, measurable results.

The number

0

people now live within reach of free, AI-powered flood forecasting — across 100 countries, with warnings up to seven days before the water arrives.

The data

A near-decade of chemistry, from one search.

Over decades, science catalogued about 48,000 stable inorganic crystals — the raw materials behind every battery, solar cell and superconductor. Then one deep-learning system, GNoME, predicted 380,000 more — a nearly ninefold jump in the known total.

0

Known before

0

After GNoME

Stable inorganic crystals known to science

0of 1,320 targets where an AI weather model beat the world's best forecast system — up to 15 days out

0less lithium in a new battery material an AI found by screening 32 million candidates in days

0the accuracy of the farmers it serves — matching a trained field officer — at naming a crop disease from a photo

0earlier an AI flags a water-quality threat at a desalination plant than the plant's own instruments

The four fronts

Four impossible problems, all in play.

The same idea — let a model search a space no human could — is turning up at the same time on the problems that matter most.

01

Energy & climate

Fusion — the reaction that powers the sun — needs its plasma held in shape by magnets, adjusted thousands of times a second, faster than any human. In 2022 a reinforcement-learning AI took the controls of a Swiss tokamak and held shapes no one had before.

The same tools now hunt better batteries, solar cells and superconductors — screening millions of candidate materials in the time a lab tests one.

DeepMind × EPFL · Microsoft × PNNL

02

Feeding the world

A free, offline app reads a photo of a sick plant and names the disease as accurately as a trained field officer — in the farmer's own language — so smallholders can catch an outbreak before a whole harvest is lost.

PlantVillage Nuru · Penn State, FAO, IITA

03

Predicting disaster

AI now forecasts floods and weather days ahead — faster, and often more accurately, than billion-dollar supercomputers — and pushes free alerts to communities that never had them.

In October 2025 one such model, WeatherNext, helped the U.S. National Hurricane Center track Hurricane Melissa's record Category 5 landfall on Jamaica days ahead — the season's top-performing model for track and intensity, buying families time to reach safety.

Google Flood Hub · DeepMind GraphCast / GenCast / WeatherNext

04

Clean water

At treatment and desalination plants, AI watches the water in real time — predicting quality threats and pinpointing hidden leaks hours before human instruments would, so less is lost and less gets through.

ACCIONA · water-network ML research

The hardest one of all

Teaching a machine to hold a star.

Fusion is the oldest dream in clean energy: the same reaction that lights the sun, promising near-limitless power with almost no carbon and little lasting waste. The obstacle has always been control. The fuel is a plasma hotter than the sun's core, suspended inside a magnetic cage whose fields must be corrected thousands of times a second — too fast, and too unstable, for any human hand.

In 2022, researchers at Google DeepMind and Switzerland's EPFL handed those controls to a reinforcement-learning AI. Trained entirely in simulation, it took command of the tokamak's 19 magnetic coils and sculpted the plasma into shapes the machine had never made — including a first-ever “droplet” of two plasmas held at once. It was the first time an AI directly controlled the heart of a fusion reactor.

Two years later, a separate AI went further still. On America's DIII-D tokamak it learned to steer a plasma clear of the “tearing” disruptions that can end a fusion shot in an instant — heading them off before they form, even in the scenario planned for the giant ITER reactor.

The pattern is repeating everywhere. When Microsoft and a U.S. national lab went looking for a safer battery, an AI sifted 32 million candidate materials and surfaced one using 70% less lithium — a search that would have taken twenty years, done in under a week.

“There are some root node problems that, if you solve them, you unlock entire new branches of research.” — Demis Hassabis, Google DeepMind CEO & Nobel laureate — who lists fusion, weather and material design among them

The takeaway

The hardest problems were never truly unsolvable.
They were unsearchable.

For centuries the limit wasn't our courage or our care — it was our reach. There were always more molecules, more storms, more failures to try than any lifetime could hold. AI doesn't replace the scientist; it hands them a search that never sleeps — and runs a million experiments before lunch. Fusion, famine, floods, clean water: for the first time, the problems that outran us have help that doesn't tire.

The HardestProblems