AI is making many people miserable, but it doesn't have to. Decades of psychology research and empirical results show the importance of nurturing your own skills & development.
youtube.com/watch?v=SUZwYV5J…
A non-AI model, created using ideas behind weight space programming for #thyroid nodule risk stratification. Runs locally, no #AI, no CNN - just #HTML and JavaScript is all you need. @java Built using @claudeai
Why it matters:
✅ Fully transparent — every weight is tied to a measurement a radiologist can check
✅ Tiny — ~300 lines of code + a few KB of numbers, vs. ~11 million parameters in a small CNN
✅ Adaptable — a new question means writing a new measurement, not retraining a network
✅ Private — small enough to run entirely in the browser, so the image never leaves your computer
✅ Guarded against self-grading — fixed scorers, a rule-checking auditor, and hidden test sets the agent never sees
We applied that idea to ACR TI-RADS:
🔹 Step 1 – Outlining the nodule: an AI agent writes classical image-processing code (no CNN, no U-Net) and is scored against open source nodule masks.
🔹 Step 2 – Describing the nodule: the agent writes named measurements that mirror the TI-RADS lexicon — "solid-part brightness ÷ surrounding thyroid" for echogenicity, height ÷ width for shape, boundary roughness for margins, bright-spot detection for echogenic foci.
🔹 Step 3 – Scoring: a tiny logistic model per descriptor, then the standard ACR point system gives the TR level.
Can you build a thyroid ultrasound AI with no neural network at all — and still read every step of it?
That's what we're testing with iWSP TI-RADS.
The inspiration: a paper posted this week from @MSFTResearch@JohnsHopkins@NYUDataScience and @Columbia - "Interpretable Weight-Space Programming Builds Transparent Transformers and Brain Encoding Models" (@y_mehtu et al.). Instead of training a model with backpropagation, an AI agent writes the model directly as readable code, tests it, and revises it over many rounds.
💡 Levothyroxine alone provided better quality of life than combination levothyroxine/liothyronine in several domains, including:
📌 tiredness
📌 cognition
📌 anxiety
📌 depression
#ATA2026@AmThyroidAssn@EndocrinologyMvist.ly/5sqsp
Stopping a GLP-1 after starting it was linked to about twice the rate of heart attack, stroke, or death in type 2 diabetes: 3.94 vs 1.96 per 100 person-years. All-cause death was 2.23 vs 0.48. Observational, so not proven cause, and reasons for stopping were unknown. Still, a reminder to plan the stop, not just stop. doi.org/10.1016/j.lanepe.202…
Early-career endocrinologists: Here's your opportunity to grow as a leader and help shape the future of our specialty!
The new Daniel L. Hurley, MD, MACE AACE Emerging Leaders Initiative Award offers mentorship, leadership development, and opportunities to become more engaged in the future of endocrinology.
Be the inaugural recipient of this exciting new award and submit your self-nomination by the December 11, 2026. Learn more about the requirements here pro.aace.com/about/aace-awar…
Don't miss the chance to invest in your leadership development and professional growth!
ATTAIN-OSA in JAMA: oral orforglipron (Foundayo) cut AHI by 12-13 events/h more than placebo at 52 weeks in adults with obesity and moderate to severe OSA, on or off PAP. Weight down ~10%. GI side effects were common. Not yet approved for OSA. jamanetwork.com/journals/jam…
AI-Generated Echo Reports Pass a Clinical Turing Test
We tested AI-generated echo reports in a blinded, randomized clinical trial. Clinicians could not reliably tell which reports were written by AI and which by a cardiologist.
1/n Excited to share AI ECHO INSIGHT, the first blinded, randomized trial of AI for the full echo report, presented today by Keane Lee at #ScientificSessions26 Late-Breaking Science.
EchoPrime and EchoNet-Measurements drafted the interpretations and measurements for cardiologist review. #echofirst