Creepy agents, IPOs, & Cancer Detection

TLDR;

Welcome back! This post has it all. Cancer-diagnosing AIs (in the week a compatriot of mine, Barbacid, may have cured the “incurable” pancreas cancer), upcoming AI IPO news, the latest drama with Moltbook, the first-ever AI-only social media network, and the release of my favorite AI product ever: Genie.

Enjoy!

HEALTHCARE

AI Helps Screening Patients

A large-scale randomized controlled trial, known as the MASAI study, evaluated the use of AI in population-based mammography screening. Published in The Lancet on January 31, 2026, the study involved over 100,000 women in Sweden and compared AI-supported screening with standard double reading by radiologists.

In the trial, participants were randomly assigned in a 1:1 ratio to either AI-supported mammography (intervention group) or standard double reading without AI (control group). The AI system triaged examinations, determining whether they required single- or double-reading by radiologists, while also providing detection support to identify potential cancers.

Key results focused on cancer detection, interval cancers (those diagnosed between screening rounds), workload, and diagnostic accuracy:

Cancer detection: The AI-supported group detected 29% more cancers compared with double reading alone.

Interval cancers: The rate was lower in the AI group (1.55 per 1,000 participants) than in the control group (1.76 per 1,000), with interval cancers in the AI group showing more favorable histopathological characteristics (fewer aggressive features). The trial met its non-inferiority endpoint for interval cancer rates.

Sensitivity and specificity: Sensitivity (correctly identifying positives) was higher with AI support (80.5%) than without (73.8%), while specificity (identifying negatives as real negatives) remained comparable between groups (around 98.5%). False-positive rates were identical at 1.4% in both arms.

Workload reduction: AI integration reduced radiologists' screen-reading workload by 44%.

TheWhiteBox’s takeaway:

Healthcare is one of those use cases that you know is going to be completely transformed by AI, but it’s a slow cooking process, as these use cases imply very important decisions that can result in life or death.

But as the saying goes, slow-cooked food tastes better, right? Or as the saying goes, “things go slowly in high places”, which means sometimes the wait is worth it.

I’ve always believed AI has a “natural” fit for healthcare-related use cases because these are usually all about pattern-matching; processing vast amounts of data and picking up the small details humans may have missed. No super reasoning or genuine intelligence required (not always, naturally, as with healthcare research, which does require novelty), just supernatural pattern-matching abilities, where AI excels.

It’s also super targeted, meaning you can train models for one thing and one thing only, and it will still be very useful.

These are two staples of what AI is today. On the flip side, and largely explaining why the paper in discussion did not carry out ‘AI-only’ ablations, is the point of accountability. Would you trust your kid’s surgery entirely to an AI? Who is to blame if the AI shits the bed?

I wouldn’t be surprised if, even if AIs prove commendable independence and autonomy, they will still largely serve as companions to…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.