iiinova — where your digital self keeps evolving

The lesson: an LLM is only one component. Careful task definition, prompt design, evidence-linked output, random sampling, human adjudication, and prospective validation are what make the surrounding system trustworthy.

100

This was a small proof of concept—not a benchmark or clinical validation. But it strengthened expectations based on my previous experience using LLMs and older NLP tools to extract information from published literature and clinical data.

100

The larger model handled all these cases correctly, including one prompt that initially tripped up the small local model. Making that prompt more explicit allowed the small model to succeed as well. The other results were essentially concordant.

100

They were realistic synthetic examples designed to test structured clinical information extraction, including negation, temporality, patient-versus-family attribution, planned-versus-completed procedures, chronology, irrelevant distractors, and conflicting documentation.

100

The queries contained no actual clinical data.

100

⌯⌲ Installed Bionic from Element Labs and downloaded Gemma 4 E2B, the smallest Gemma 4 model. ⌯⌲ Ran a set of carefully constructed test queries locally on my laptop. ⌯⌲ Ran the same queries through OpenRouter using NVIDIA’s much larger Nemotron 3.5 Lightning model.

100

⌯⌲ A larger model may be more tolerant of ambiguous prompt wording. ⌯⌲ For bounded extraction tasks, prompt design and validation architecture may matter more than model choice.

100

⌯⌲ It is now possible to get a small LLM running locally on an ordinary laptop in under two hours. ⌯⌲ Even a very small LLM can handle realistic clinical NLP tasks when given sufficiently explicit prompts.

100

Specific things I learned:

100

Matt Healy @matthewdhealy.bsky.social · 21h

For well-defined information-extraction tasks, frontier-scale AI may be more power than necessary—especially when prompt development and output validation occur within a structured, auditable framework that includes human curation and randomized sampling.

100

(1) many people with relevant training consider the secret 48-page memo to be filled with dubious reasoning, (2) the administration asserts that its actions are legal, but (3) the administration is determined to prevent the public from seeing its reasons for considering its actions to be legal.

storage.courtlistener.com

Memorandum & Opinion – #72 in American Civil Liberties Union v. Department of Justice (S.D.N.Y., 1:25-cv-10189) – CourtListener.com

OPINION & ORDER re: 51 MOTION for Summary Judgment filed by Department of Defense, Department of Justice, 48 CROSS MOTION for Summary Judgment filed by Center for Constitutional ...

100

In the dream the building was much closer to the train station and trains were going right past the window. From my real Yale office, I looked down at Metro-North trains off in the distance.

000

That particular building doesn't exist anymore, Yale demolished it to make way for a new building some years ago.

100

In the dream I was a postdoc at Yale School of Medicine, which I was back in the 1990s, and my office was fairly high up with a view out the window of New Haven and Long Island Sound, which it was, I think on the 8th floor.

100

Matt Healy @matthewdhealy.bsky.social · 09/10/2026

Dreams can mix real content in unreal ways. Like one I had last night with trains high up in the air.

110

A very useful meta-prompt for any AI tool: "Below are (1) a brief description of what I want to accomplish and (2) my draft prompt. Please respond with your suggested improved prompt" Examine the response. It will tell you what it thinks you want in some detail. Edit that as needed then submit it.

000

Matt Healy @matthewdhealy.bsky.social · 09/10/2026

How to make an atomic clock radioactive? Put a banana in front of it! Truly: the potassium in a banana is much more radioactive than the cesium in an atomic clock. So are you: your neurons won't work without potassium.

000

Why Christians need to know about the Bronze Age Collapse

You cannot understand the Bible unless you realize much of it is about events before the Bronze Age Collapse as they were remembered AFTER the Bronze Age Collapse. The Bronze Age Collapse changed EVERYTHING about the Ancient world across Europe & Asia. martin-dove.medium.com/why-christia...

martin-dove.medium.com

A missing piece of the Biblical story

041

Margaret Hamilton by Moonlight: Honoring an Apollo 11 Icon

Margaret Hamilton, who led the team that wrote the code for the Apollo 11 onboard computers and popularized the term "software engineering" has passed away age 90. In 2019 the mirrors at a solar power plant honored her with a gigantic moonlight image. www.youtube.com/watch?v=B7Cn...

youtube.com

YouTube video by Google

034

Track the 2026 Senate election with a red/blue map of the US updated daily using the latest state polls.

My favorite political blog sneaked a Simpsons joke into a post about campaigns' advertising rates. And why the GOP is spending buckets of money trying to rescue a scandal-ridden Senate candidate. www.electoral-vote.com/evp2026/Sena...

electoral-vote.com

Electoral-vote.com

010

WHY the Apatosaurus/Brontosaurus problem happened in the first place #history

Scientists can be jerks. This summary of an epic feud is quite true; the saga is well known in the research community but probably not among the general public: youtube.com/shorts/sQ_T0...

youtube.com

YouTube video by Astrid Lundberg

110

PS: while an LLM helped me edit this post, I the human picked the ☸ emoji because that's a cool glyph 🙃.

000

This was a modest feasibility exercise with synthetic text, not clinical validation. But firsthand testing already informs my thinking about a pipeline built around narrow extraction tasks, deterministic processing, human curation, and independent evaluation.

100

☸The model, prompts, runtime configuration, and test cases together form a package that needs to be versioned.

100

☸Unresolved disagreement should remain visible.Model changes require renewed prompt testing.

100

☸Small, focused tasks make failures easier to inspect.

100

☸Classification, supporting evidence, and output format deserve separate checks.

100

☸Prompt wording is part of the system and needs testing.

100

Mixed notes tested date selection; conflicting records tested whether the model would preserve uncertainty.

100

The cases probed crucial distinctions: patient versus family history, an uncompleted test versus a negative result, a scheduled procedure versus a completed one, past versus current symptoms, and suspected versus confirmed cancer.

100

The first prompt failed: the model copied the entire list of possible answers. Clearer instructions fixed that. Ten subsequent cases produced the expected classifications with exact supporting quotations.

100

I ran a small Gemma language model locally on a low-end laptop in Airplane Mode, using simulated clinical notes. I saved the exact prompts, outputs, tool versions, and timings.

100

And today I tried something new!

100