Mom-and-pop SaaS comes for medicine 🛠️, Midjourney's spa meets the incidentaloma 🛁, Scribes skip 25M patients 🗣️
The cost of building software just collapsed for the person who knows the problem — and started climbing for everyone renting the tools.
Anthropic published an analysis of roughly 400,000 real Claude Code sessions this week, and the headline is one we’ve circled for months: domain expertise, not a software-engineering background, was the strongest predictor of a session that actually worked.
Sessions rated “expert” reached verified success more than twice as often as “novice” ones. Lawyers, managers, and finance people landed within a handful of points of professional engineers.
The builders’ newsletters read it the same way — some are already calling it the “mom-and-pop SaaS era”: software for problems too small to ever justify a dev team, now built by the person who lives inside the problem.
For us, that person is the nurse who knows why the discharge workflow stalls, or the clinic manager who knows which prior auth always bounces. The syntax was never the scarce input. The floor knowledge was.
Here’s the catch nobody’s putting on the same page. The same week, the tools doing the lifting started re-pricing themselves: Microsoft is pulling engineers off Claude Code over $500–$2,000-per-engineer monthly token bills, and GitHub flipped Copilot to usage-based billing.
So building is cheap right up until it isn’t — and the meter runs hardest on the people doing the most ambitious work. Own your skills and prompts as portable files, not as something trapped in one vendor’s tool.
😤 Within a few points of engineers on toy tasks isn’t the same as shipping something safe into a clinical workflow. Success on a benchmark is not success on a unit. Doesn’t matter — the prototype that gets a CMIO to say “show me more” is now a Saturday, not a sprint. That’s the whole game.
📡 Builder’s Radar
The Midjourney spa backlash arrived — and it’s more useful than the launch. After David Holz unveiled a 60-second full-body ultrasound scanner (built on Butterfly Network’s Ultrasound-on-Chip), the internet — amplified by an Elon Musk repost — turned on it.
The sharpest critiques are the boring ones: ultrasound is blocked by bone and air, today’s prototype takes ~20 minutes not 60 seconds, and there’s no billing code or evidence that scanning healthy people extends life.
😤 “You’ve invented a machine that manufactures incidentalomas — a finding on a healthy person is an anxiety generator with a biopsy attached.” Dense-breast USCT (Delphinus SoftVue) is FDA-cleared and finds real cancers. Dismissing the whole modality is lazy.
Ambient scribes are missing the multilingual land grab. With 25M+ U.S. patients needing language services and ~34% of physicians already using AI translation (another ~23% plan to), the white space is real — AMN just bought translation startup Jaide Health, and Heidi is building it in. A scribe that captures the visit but not the language barrier is solving half the room.
Google, Microsoft, Snowflake and friends launched an agent-discovery standard. The Agentic Resource Discovery (ARD) spec lets AI agents find and connect to approved tools without manual wiring — a discovery layer above MCP. Notably absent from the coalition: OpenAI and Anthropic.
FERC ordered grid operators to fast-track AI data centers. A government-mandated fast lane to the grid — but power availability, not the interconnection queue, is the real ceiling on the whole AI build-out.
⚡ Quick Hits
Noam Shazeer left Google Gemini for OpenAI — a Transformer co-author changing jerseys in the lab talent war.
Yann LeCun called xAI a “failure” and warned of an AI funding-bubble pop.
Amazon is exploring selling its Trainium chips to third-party data centers — a direct shot at NVIDIA.
🛠️ From the Workbench
Unsloth shrank GLM-5.2 to run locally on a single Mac. Z.ai’s new 744B, MIT-licensed model is 1.51TB at full weight; Unsloth’s 2-bit quant lands at ~238GB and keeps ~82% accuracy, running on a 256GB M3/M4 Ultra via llama.cpp or LM Studio. A frontier-class model, offline, no token meter — exactly the shape that matters for PHI-sensitive work.
🎙️ From the Pods
On Turn on the Lights, incoming NCQA CEO Vivek Garg made the case that accreditation standards — care management, PCMH, utilization management — are the quiet scaffolding of “quality.”
🔇 Speaker Blindspot: process-as-proxy. Treating a documented standard and a clean audit as evidence of better patient outcomes — the measurable process standing in for the outcome it was supposed to guarantee.
💡 BTW
Noam Shazeer, who just left Google’s Gemini team for OpenAI, ranked 6th in the nation on the Putnam math exam in his very first semester at Duke and helped lead Duke to top-two national finishes — then, after joining Google in 2000, one of his first real projects was improving the search engine’s spell-checker. Two decades later he co-authored “Attention Is All You Need.” (Wikipedia)
What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.
— Kevin