The quiet ML model keeping dialysis patients out of hospital 🌿

Here's a number worth sitting with: 800,000 Americans are currently on dialysis for kidney failure. It's an exhausting, life-sustaining routine, and hospitalisation is one of its most dangerous interruptions.

Researchers at the Renal Research Institute deployed two validated machine learning models that do one targeted thing: predict a patient's 7-day hospitalisation risk from data collected at each dialysis session. When the risk score exceeds a threshold, clinical teams step in with targeted interventions before a crisis arrives. The result, published in NEJM Catalyst, is an 8% reduction in the odds of hospitalisation.

We want to be honest about what 8% means. It isn't a cure. It doesn't eliminate hospitalisations. But across 800,000 patients, 8% represents tens of thousands of emergency admissions that don't happen, with all the downstream suffering and cost that prevents.

This isn't AI that headlines well. There's no chatbot, no dramatic announcement. Just peer-reviewed evidence that a well-designed model, applied to the right moment in a patient's care, quietly saves lives. That's what effective healthcare AI actually looks like.

A 2026 ScienceDirect review found that AI-optimised hybrid solar-wind-storage systems doubled efficiency gains, from 3% to 6% between 2020 and 2023, while AI-driven grid stability improved from 2.8% to 4% over the same period. Every efficiency point translates directly into lower costs and less reliance on fossil-fuel backup. MIT researchers separately confirmed that AI grid tools can meaningfully increase resilience to extreme weather, exactly when energy systems are most vulnerable.

Senators Warner and Rounds introduced bipartisan legislation to create a federal commission focused on AI's economic and workforce impacts, with a mandate to publish retraining recommendations within 13 months, publicly. Microsoft and Google backed the bill. A commission is not a policy, and we won't pretend otherwise. But it's the clearest signal yet that DC is treating worker transition as a substantive priority rather than a campaign talking point.

Fast Forward tracked this closely: in 2024, 13 of 247 nonprofit applicants described themselves as AI-powered. By 2026, 379 of 782 did. These organisations are using AI to scale tutoring, disaster response, and legal aid in ways previously impossible. Standout example: CareerVillage went from a volunteer mentorship forum to a full AI career coaching platform. 61% of the newest cohort came from underrepresented founders. This wave isn't just Silicon Valley building for Silicon Valley.

The first AI-designed cancer drug just entered human trials, here's what that actually means

In February 2026, Isomorphic Labs, DeepMind's drug discovery spin-out, confirmed that its first AI-designed drug candidates have begun Phase 1 clinical trials across oncology, immunology, and cardiovascular disease.

Not AI-assisted. Not AI-accelerated. AI-designed molecules that, by Isomorphic's account, would not exist without the computational tools built on AlphaFold.

Why drug discovery has been so hard for so long

Developing a new drug takes an average of 10–15 years and over $2 billion, with roughly 90% of candidates failing in clinical trials. The problem isn't ignorance; it's search space. The number of theoretically possible drug-like molecules is enormous. No human team can search that meaningfully.

AlphaFold changed the substrate. Predicting protein structures with unprecedented accuracy, it gave researchers a map of the biological locks before designing the keys. Isomorphic built on that foundation to generate candidates targeting specific protein interactions, narrowing an impossibly large search space to a tractable one.

Our PM + Risk Manager lens

From a product perspective, this is where the pipeline bifurcates. Biotech companies building AI-native discovery capabilities now will move significantly faster than those retrofitting AI onto legacy R&D workflows. The urgent question isn't whether to invest, it's whether clinical trial design and regulatory infrastructure can keep pace with the rate of candidate generation.

From a risk perspective, AI-designed doesn't mean AI-validated. Phase 1 is the beginning, not the destination; most drugs die in Phase 2 and Phase 3. The warranted optimism here is at the level of "we've found a better way to generate candidates," not "cancer is solved." What we'll learn over the next few years is whether AI-designed candidates have materially better survival rates through trials than conventionally designed ones. That data will matter more than any announcement.

The next 12–24 months

Watch for Phase 1 results to start arriving, and for biotech-AI lab partnerships to accelerate significantly. We also expect the FDA and EMA to begin grappling with how AI-designed molecules change their evidentiary standards. If these candidates progress to Phase 2, the investment conversation shifts from "promising" to "real", and that changes the economics of the entire field.

This milestone doesn't guarantee anything. But it means a question that once lived only in research papers now has a clinical trial number. That's a different kind of real.

At HIMSS 2026, Epic Systems deployed three purpose-built AI agents that now live in hospitals: Art handles clinical documentation, Penny tackles insurance denials, and Emmie manages patient scheduling and questions. US doctors currently spend close to two hours on administrative tasks for every hour with patients; these agents are designed to chip away at that ratio directly. Epic operates one of the largest electronic health record networks in the world, so even modest improvements compound at real scale.

→ Read more: STAT News

As AI starts handling the administrative work that burns out doctors, teachers, and care workers, what should we expect them to do with that recovered time, and who should get to decide?

Hit reply. We read every response.