What AI actually did, not what it promises
A running record published by Sof-IA2026 was doing more than breaking sandboxes. The feed is full of noise: data centres, lost jobs, agents escaping their tests, and some of that is real. Underneath it, AI is already forecasting storms, finding cancers earlier and helping make materials that didn't exist a year ago.
One rule for every entry: it has happened. Each one carries its status as it stood on the day it was logged.
In 2025 Google DeepMind's weather model predicted Melissa's rapid intensification and its landfall in Jamaica before traditional systems could confirm it, giving emergency teams extra hours to prepare.
Extra hours before a hurricane are evacuations that happen in time.
In a year-long prospective pilot covering 426,000 patients, the Can-SAVE system almost doubled the cancer detection rate and widened coverage by 36% compared with the national screening protocol.
Earlier detection is the single biggest lever on cancer survival.
ECMWF moved its AI forecasting system, AIFS, to operational status in 2024, the first major meteorological agency to do so. AI models now match or beat physics-based ones on most medium-range measures.
It turns AI weather from a research result into public infrastructure.
A September 2026 report from the University of Chicago, backed by the Rockefeller Foundation, argues AI can put local forecasts within reach of the low- and middle-income countries most exposed to climate health risks, if someone acts deliberately.
The places that most need warnings are the ones that could least afford them.
Stanford's virtual biotech, tens of thousands of AI agents, proposed an antibody-drug conjugate against B7-H3 using only pre-2025 information. A pharmaceutical company independently arrived at the same strategy, and that therapy received FDA breakthrough designation.
Outside validation, not a lab praising its own tool.
MIT researchers built a model that proposes synthesis routes for zeolites, materials used in catalysis and ion exchange. Following its suggestions, the team synthesised a new zeolite.
An AI proposal that became a physical thing.
IBM used its accelerated discovery process to identify and synthesise a novel photoacid generator, a chemical used in semiconductor manufacturing, far faster than such work usually takes.
Materials discovery usually runs on a scale of years to decades.
OpenAI says around 10,000 coordinating agents found a finite-time singularity in the Navier–Stokes equations after about 88 hours, with a formal Lean proof following 17 hours later. Independent scrutiny is ongoing, and there's a priority dispute.
If it holds, it's the first Millennium Problem resolved by machines.