Max Benjamin MD

Nephrologist with an interest in CKD progression modeling, Clinical Informatic, EMR,,Complexity theory, Artificial Intelligence

[Maryam Miradi, PhD @MaryamMiradi]

A VLM Can Read One Medical Screen.
A Medical AI Agent Must Finish 24 Screens in a Row.
𝗧𝗵𝗶𝘀 𝗶𝘀 𝗺𝘆 𝟳-��𝘁𝗲𝗽 𝗿𝗼𝗮𝗱𝗺𝗮𝗽 𝗳𝗿𝗼𝗺 𝗦𝗰𝗿𝗲𝗲𝗻-𝗮𝘀-𝗦𝘁𝗮𝘁𝗲 𝘁𝗼 𝗮 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗠𝗲𝗱𝗶𝗰𝗮𝗹 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁.

Real medical agent must move through many screens.
It must click, zoom, type, segment, check, and finish.
If it forgets one step, the whole task can fail.

My 7-Step roadmap based on this recent paper: https://t.co/v0GhbutrR0

》𝗦𝘁𝗲𝗽 𝟭: Start with the Real Goal
✸ Tell the agent the full medical task.
✸ Define what “done” means.
✸ Do not give a vague screen task.
→ Example: “Open the CT scan, zoom to the liver, draw ROI, export statistics.”

》𝗦𝘁𝗲𝗽 𝟮: Treat Each Screen as State
✸ Every screen tells the agent where it is.
✸ Track open panels, active tools, and patient fields.
✸ Do not treat screens as random images.
→ Example: Is the agent in view mode, zoom mode, or annotation mode?

》𝗦𝘁𝗲𝗽 𝟯: Ground the Screen with Tools
✸ Use OCR to read labels.
✸ Use object detection to find buttons.
✸ Use zoom or crop for tiny UI parts.
→ Example: Find the “Export” button before clicking.

》𝗦𝘁𝗲𝗽 𝟰: Use an Actor Agent
✸ The Actor chooses the next action.
✸ One step at a time.
✸ Do not let it jump to the final answer.
→ Example: CLICK, SCROLL, ZOOM, TEXT, SEGMENT, or COMPLETE.

》𝗦𝘁𝗲𝗽 𝟱: Add a Critic Agent
✸ The Critic checks the Actor’s action.
✸ It asks: “Does this action fit this screen?”
✸ It blocks unsafe or early actions.
→ Example: Do not allow COMPLETE before the result is saved.

》𝗦𝘁𝗲𝗽 𝟲: Give the Agent Memory
✸ Short-term memory remembers the last step.
✸ Long-term memory remembers the whole path.
✸ Do not reset the agent’s brain on every screen.
→ Example: Remember that the zoom tool was already selected.

》𝗦𝘁𝗲𝗽 𝟳: Build the Runtime Agent
✸ Use the Critic during testing and training.
✸ Keep only corrected successful paths.
✸ Deploy a faster agent that learned from feedback.
→ Example: The runtime agent acts faster, but still follows the learned checks.

𝗣𝗹𝗲𝗮𝘀𝗲 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿:
Medical AI agents do not need only better vision.
They need state, tools, memory, and checks.
A screen reader is not a workflow finisher.

--

📥 Skip the trial-and-error of building production AI agents.
Watch my free 30-min training + get 88 pages of production guides.
Join 46,000+ AI engineers, architects, and directors.
👉 https://t.co/t9xyLtyvG1

[Santa Fe Institute @sfiscience]

Present-day LLMs like ChatGPT and Claude can write poetry and solve difficult algebra problems with astounding speed and precision. Some researchers have referred to this phenomenon of AI acquiring startlingly human-like skills as "emergence." But not everyone agrees with this terminology.

A new paper by SFI's David Krakauer, Melanie Mitchell, and John Krakauer proposes a framework grounded in complexity science to clarify whether a system's capabilities are truly emergent.

https://t.co/yyENFiN4Om

[Maziyar PANAHI @MaziyarPanahi]

650+ clinical AI models now run on your iPhone.

30-40x faster than CPU. Fully private. The PHI never leaves your device.

Today I shipped 410+ new MLX medical models: biomedical NER, disease, drug, and PII de-id. All open weights, Apache 2.0.

Same model, same entities. Watch:

[Roger @rdd147]

Doctors 25% more likely to miss a diagnosis on a patient within 3 months after Ai is implemented. The results indicate Ai has an immediate negative effect on professionals skillsets vs prolonged reliance.

https://t.co/ciPB3DdR2J

[Eric Topol @EricTopol]

Today @Nature published 2 new AI medical agentic models that take capabilities to a new level, from end-to-end care after presentation to the emergency department and longitudinally through 3 out-patient visits. There's a lot to unpack.
Summarized in a new Ground Truths post.
https://t.co/NRzuMe23DC

[Eric Topol @EricTopol]

For medical information, general AI frontier models (Google, OpenAI, Anthropic) outperformed specialized @openevidence and @UpToDate as assessed by 12 US clinicians, randomized and blinded to which model and extensive testing/benchmarks. This was not anticipated. @NatureMedicine
https://t.co/KCH1ADfQWz

[Michał Nedoszytko MD, PhD @mnedoszytko]

So privileged to work at the forefront of clinical AI. We are bringing back joy to practicing medicine with @AbridgeHQ.

[Dario Amodei @DarioAmodei]

Today I'm publishing a new essay, Policy on the AI Exponential. AI is progressing extremely fast—much faster than the policy process was built to handle. The essay lays out where I think the technology is now, and the action needed to close the gap: https://t.co/Lh6PWae178

[Sam Rawal, MD @samarthrawal]

I must say - I’ve anecdotally noticed a MASSIVE improvement in @openevidence results on this over the past month or few - specifically, much better grounding to the clinical context without swinging the other way into subtle confirmation bias.

Feels like the reasoning focused on the clinical intent is much improved, without overoptimizing to an “agreeable” reward function.

[Thoughts on Healthcare Markets and Tech @thoughtson_tech]

The burnout framing is real, but it stops short of the harder problem. Passive formats aren't just a design flaw you can fix by switching to better content. The passivity is structurally incentivized.

Here's what I mean: the ACCME's accreditation fee model is funded by complexity. More credit hour requirements, more compliance tracking, more consulting services to navigate standards. An organization financially dependent on that apparatus has no incentive to endorse formats that are actually efficient, because efficiency shrinks the billable surface area. Physicians sitting through lecture-based modules for 20 to 50 hours annually isn't an oversight. It's load-bearing infrastructure for a system generating over $5 billion in annual spending.

The pharma angle compounds this. When companies structure CME funding as unrestricted educational grants, they don't need to dictate content directly. The requirement that CME address "unmet medical needs" already channels education toward newer, expensive drugs. Passive formats are actually useful here because a physician watching a sponsored lecture absorbs framing without resistance the way a problem-based simulation wouldn't permit.

AI-driven personalized learning genuinely threatens this. Not because it's better pedagogy, though it is. Because it eliminates the accreditation and compliance layer that extracts rent from every credit hour. That's why the incumbent ecosystem is structurally motivated to resist it regardless of outcome evidence.

Calling this a burnout problem suggests the fix is ergonomic. The fix is architectural.

https://t.co/lFUC9Ena84

[Pappers @get_pappers]

Anthropic vient de lancer des agents IA pour les avocats et juristes.

Couplé avec le MCP Pappers, les résultats sont impressionnants.

On a testé l'agent claim sur un contentieux fictif, généré en 2 minutes...

Le résultat est disponible en dessous

[Medical Sphere @MedicalSphereAI]

HealthBench Professional is now live on Medical Sphere 🏥

Built by @OpenAI, this benchmark contains 525 real clinician chat cases across care consults, medical documentation, and medical research.

https://t.co/iI2cy3oQqD

[Max Benjamin MD @mmaxbenji]

@MarioATX_MD Can you provide a link to register for the Vibe coding class ? The Cline link doesn’t seem to work.

[Max Benjamin MD @mmaxbenji]

#KiaTellurid

[Max Benjamin MD @mmaxbenji]

@DGlaucomflecken A word about Vinay’s statement about Dr Glaucomflecken not being a health policy expert. Do we need to be health policy experts to know that our patients are being denied. When was the last time he was a clinic writing appeal letters for patients?

[Max Benjamin MD @mmaxbenji]

@DGlaucomflecken Your brand of humor is gentle and intelligent. We doctors need your voice. You are our truth teller.
Any fair person knows that you have nothing to do with this senseless event .

[Max Benjamin MD @mmaxbenji]

@LittleBigGloms @GrahamAbra @eric_weinhandl @ChristosArgyrop @hswapnil @brendonneuen @CircAHA @mvaduganathan @KatherineTuttl8 @VladoPerkovic @gcfmd @OttawaRenal CKD EHR by Epic has it . 2y/5y.

[Anna Burgner MD MEHP @anna_burgner]

So many podcasts to listen to this #NephMadness 2024! Great way to start to help you decide your picks! The PodCrawl https://t.co/RyROzzTEld via @AJKDonline

[CKJ @CKJsocial]

mRNA as a medicine in nephrology: the future is now!
👉What are the key strengths of RNA-based therapies?
👉What are the potential future directions & challenges of this promising technology in nephrology?
📖Find out more here👇👇👇
🔓https://t.co/qqEhC4y6Pu