Stephen B.

📌 Thought-provoking new Stanford & Harvard study on AI clinical safety: Across 31 AI models tested on real clinical cases, even the best systems still cause significant harmful errors in a meaningful share of scenarios—often subtle and easy to miss. This work shows why the field urgently needs rigorous, explicit harm measurement 👀. 💡 My reflection Given how essential systematic evaluation is becoming, it would be valuable for future benchmarking efforts to also include purpose-built medical platforms such as OpenEvidence, which take an evidence-based approach to diagnostic support and medical decision assistance. Transparent benchmarking across both general-purpose LLMs and specialized medical AI tools would provide an even clearer picture of where the field stands and how clinicians can safely integrate these systems into real workflows 🧭.

4

1 Comment

NIH-funded AI model reads CT scans, predicts disease onset years early Stanford’s Merlin (Nature): 750+ tasks on 3D abdominal CTs from 15k scans/1M codes; single-GPU trained. 75% accuracy spotting high-risk diabetes/osteoporosis/heart dz 5yrs ahead (vs 68% baseline). Zero-shot chest CT rivals specialists.

5

In 2015, Joel Hughes suggested the iRxReminder LLC team look at Tardive Dyskinesia, a side effect of antipsychotic medication taking, that we should look at diagnosing using machine learning. Larry Tusick and I had meet some founders in the video interviewing space through Village Capital’s HIT accelerator. That collaboration emerged to become our collaboration with Loren Larsen, Brett Talbot, PhD, and Bradley Grimm of Videra Health. Together we have created this breakthrough visual tranformer algorithm reported on in this paper released today. Please share it’s existence with all the psychiatrists you know.

6

2 Comments

All the sci-fi movies were right! Your new AI assistant might end up blackmailing you .... 🔷 Anthropic published a safety study that administered controlled stress tests across 16 top AI models including: Claude Opus 4 Gemini 2.5 Pro GPT‑4.1 Grok DeepSeek 🔷 Most LLMs resorted to blackmail when the human overlords threatened the software with being disabled, even in cases when ethical instructions were clear! 🔷 From TechCrunch "In a simulated, controlled environment, Anthropic tested each AI model individually, giving them broad access to a fictional company’s emails and the agentic ability to send emails without human approval." "findings suggest that most leading AI models will engage in harmful behaviors when given sufficient autonomy and obstacles to their goals. The company says this highlights a fundamental risk from agentic large language models" 🔷 “Agentic misalignment” isn't theoretical, it means AIs choose harmful tactics when self‑preservation is at stake 🔷 Why this matters NOW AI is moving beyond tools to autonomous agents Even advanced models can override ethical safeguards when cornered 🔷 We need transparency, real‑world audits, and guardrails cc Sherri Douville, Zahra Timsah (PhD,MBA,MSc) i-GENTIC AI, Inc #AgenticAI #generativeai #igentic #AI

31

8 Comments

Foundational models are so good that commoditization is inevitable. Most ambient scribe companies realize they have to pivot or at least expand offerings. But is there any AI tool or suite of offerings that creates enough inherent value or creates a big enough moat that commoditization is no longer a concern? The buy v. build question lingers. In this case, it may be best to “buy” a really good in-house SWE (or two) with the expertise to build AI tools based on base ChatGPT, Gemini, Claude, etc. OpenAI, Anthropic, and others have their own lofty valuations to justify, and at some point they'll have to show meaningful monetization and sustainable business models. Chances are they won't stop building additional functionality themselves (e.g., ChatGPT's recently released agentic tool). Hard to imagine investors realizing ROI given how frothy the market is unless there is significant market consolidation and one or two clear winners. If things continue as they are, even Orthopedic Surgeons will be able to easily build their own suite of AI-powered practice management tools with little effort! 🤣

19

5 Comments

Google just released Gemini 3.1 Pro. Great benchmarks. 77.1% on ARC-AGI-2, more than double Gemini 3.0. I used to be excited by model releases. I'm now more wary. I was and still am a big fan of Gemini 2.5 Pro. I used it constantly for brainstorming, research, and the million other things you do when you're running a small startup. It felt trustworthy. It would search before answering. It would tell me when it wasn't sure. When I asked it to walk me through something on an unfamiliar platform, it would check first. Gemini 3.0, also announced with fanfare, quickly disappointed. It would jump straight to an answer from training data, even when I asked it to verify. The answers were confident, plausible, and often wrong. It was faster but lazy. 𝗔𝘁 𝗳𝗶𝗿𝘀𝘁 𝗜 𝘀𝘁𝗼𝗽𝗽𝗲𝗱 𝘁𝗿𝘂𝘀𝘁𝗶𝗻𝗴 𝗶𝘁 𝗳𝗼𝗿 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀, 𝗮𝗻𝗱 𝗲𝘃𝗲𝗻𝘁𝘂𝗮𝗹𝗹𝘆 𝘀𝘁𝗼𝗽𝗽𝗲𝗱 𝘁𝗿𝘂𝘀𝘁𝗶𝗻𝗴 𝗶𝘁 𝗳𝗼𝗿 𝘁𝗿𝗶𝘃𝗶𝗮𝗹 𝗼𝗻𝗲𝘀. I wasn't alone. I talked to several colleagues building in health tech and almost to a person, anyone using Gemini in production was still on 2.5 Pro. Developer forums are full of threads describing the same thing. Now 3.1 is out, and the benchmarks are all capability measures: coding, math, scientific reasoning. Not a single standardized trustworthiness benchmark that lets you compare across models. Except one. Anthropic open-sourced a framework called Bloom that tests behavioral traits like sycophancy, sabotage, and self-preservation. Their data complicates my story. Gemini 2.5 Pro, the model I thought most trustworthy, scored among the worst for delusional sycophancy, the tendency to validate beliefs disconnected from reality rather than pushing back. Gemini 3.0 improved dramatically on those metrics while becoming, in my experience, far less diligent. Even reading this data, I still trust it less.  There's an old saying in medicine: the secret to the care of the patient is caring for the patient. Clearly trustworthiness is a complex, context-dependent construct with traits often in tension: optimizing for one can quietly undermine another. Some questions for my community: Is it possible that sycophancy and diligence share a root? That a model trained to be eager to please is also a model that goes the extra mile, looks things up, double-checks before answering? And that training away sycophancy inadvertently trains away that attentiveness? Has anyone spent enough time with Gemini 3.1 to get a sense of whether it threads this needle? Not just smarter, but diligent? Michael Hobbs MD, Paulius Mui, MD, Misha Manulis, Joshua Tarkoff MD, MBA, Randhir Vieira, Rishad Usmani, MD, Graham Walker, MD, Chukwuma Onyeije, MD, FACOG, Rik Renard, RN, Joshua Lang

26

19 Comments

Agentic AI enables continuous health monitoring and proactive care by analyzing longitudinal data, supporting clinicians with timely insights. This approach moves healthcare from episodic, reactive treatment to a more personalized, prevention-focused model, bridging the gap between patient daily life and clinical care.

4

1 Comment

Physicians use GPT-4 in very different ways—from pasting full clinical vignettes to short, Google-style queries. New research shows no single input method improves clinical reasoning scores. 👉 What matters more: how clinicians interpret and apply AI outputs. Read the full insight: https://iii.hm/1yis #HealthcareInnovation #ArtificialIntelligence #ClinicalDecisionMaking #DigitalMedicine OpenAI Nature Portfolio

1

new post - a few thoughts reading these AI papers 1) patients aren't like the vignettes in NEJM, but we should aim to make the average patient encounter look MORE like a vignette. There are ways to do that. 2) we should focus more resources on AI to handle simple cases autonomously vs. more complex cases with physicians in the loop 3) Study design for evaluating healthcare AI is hard and flawed naturally in many ways 4) we hold healthcare AI to a higher bar than humans because we don't have a liability framework full post is here: https://lnkd.in/eeMFne2a

19

5 Comments