Let the good times roll

Written By Dale Leszczynski

Most AI coverage in the last year has been about job losses, existential risk, or regulatory battles. Dale and Nick dedicated this episode of Adjunct Intelligence to the opposite — real breakthroughs from the last 14 months that barely registered because the news cycle was too busy with whatever disaster was trending.

The range is striking. In weather forecasting, the European Centre for Medium-Range Weather Forecasts deployed its first fully operational AI model in February 2025, achieving a 20% improvement in tropical cyclone track prediction at roughly one-thousandth the energy cost of traditional physics-based models. Cambridge's Aardvark system runs an entire weather prediction pipeline on a desktop computer. Google DeepMind's hurricane model, trained on 40 years of data and 5,000 observed cyclones, gave Jamaica three extra days of warning before Hurricane Melissa made landfall as a Category 5 storm. And at the University of Washington, Dale Durran and graduate student Nathaniel Cresswell-Clay simulated a thousand years of climate on a single processor in 12 hours — outperforming the models used in the most recent IPCC report.

In medicine, a multi-agent AI system called Deep Red outperformed experienced physicians on 163 complex diagnostic cases, and has since been deployed across 600 medical institutions. Rentosertib, the first drug discovered, targeted, and designed entirely by AI, entered human clinical trials for idiopathic pulmonary fibrosis — results published in Nature Medicine showed measurable lung improvement. An AI-powered stethoscope trialed across 205 NHS practices in northwest London through the TRICORDER programme made patients 2.3 times more likely to be diagnosed with heart failure — then the same device was tested in Ghana with 96.9% sensitivity. Though 70% of GP offices quietly stopped using it, which Dale and Nick flag as the real education story: change management matters more than the technology.

In creativity, Casey Harrell — a climate activist with ALS — had 256 micro-electrode arrays implanted in his motor cortex as part of the BrainGate clinical trial, and deep learning now decodes his brain signals to produce speech in his own voice, published in Nature. Patrick Darling, an Irish composer also diagnosed with ALS, performed an original song live at the ElevenLabs Summit in London through AI-recreated voice using ElevenLabs' voice technology. And Jacob Adler, a classically trained musician with no film training, won the Runway AI Film Festival Grand Prix using tools that cost $12 a month.

In global access, FoondaMate — built by South African co-founders Dacod Magagula and Tao Boyle — runs AI tutoring through WhatsApp and Facebook Messenger using Meta's open-source Llama models. It operates in multiple languages across 30+ countries and has grown virally with zero marketing spend. A 24-year-old in Canada built a shoulder-mounted robot that speaks endangered indigenous languages using children's voices. And a researcher at Dartmouth trained an AI to translate into an endangered script created exclusively by women, using just 35 sentence pairs.

The throughline across every story is the same: AI compresses the distance between human intention and what's possible. The question for education is whether institutions are producing people who can run the model, or people who know what the output means.

References & Links

Weather & Climate

Science & Research

Medicine

Creativity & Accessibility

Global Access & Education

Dale Leszczynski
Next

The year of Opposites