Selim Önal

HIMSS is almost here, and I’m looking forward to catching up with so many of you. I'm especially looking forward to our session with UC San Diego Health. Karandeep Singh and Jeffrey Pan will showcase how they're using Notable's Platform and AI Agents to act as a critical workforce multiplier, transforming patient access, care operations, and RCM across the enterprise. Learn more about everything we have going on at HIMSS here: https://lnkd.in/gQifSd2A. And definitely come say hello if you're attending!

29

2 Comments

I’ll be attending the Precision Medicine World Conference (PMWC 2026), March 4–6 in Santa Clara. At OmicXHealth, we’re building an AI/ML-driven computational medicine platform for early disease detection and stratification, using cfDNA methylation as a foundational signal. Our initial focus is liver disease and liver cancer screening, with the broader goal of enabling scalable detection across multiple diseases from a single sequencing assay. PMWC brings together leaders across genomics, AI, and translation, exactly the ecosystem needed to move computational medicine from promising models to real-world impact. If you’re working at the intersection of AI, precision diagnostics, or data-driven disease stratification for screening or drug development, I’d love to connect. @Tal Behar @PMWC – Precision Medicine World Conference #PMWC26 #ComputationalMedicine #PrecisionDiagnostics #AIinHealthcare #BiotechFounders #PrecisionMedicine #TechBio #MolecularDiagnostics #LiquidBiopsy

8

At Tempus AI, I’ve had the opportunity to help shape how we apply generative AI to one of the most complex, high-stakes domains there is: cancer care. This week’s announcement, our collaboration with AstraZeneca and Pathos to develop a multimodal foundation model in oncology, is an incredible milestone. Years of work have gone into assembling the right data, architecting systems that support AI at scale and laying the groundwork for insights that can lead to better treatments. It’s incredibly rewarding to see that vision come to life and even more exciting to know we’re just getting started. More here: https://lnkd.in/dJHHGn4T #AI #GenerativeAI #Oncology #FoundationModels #PrecisionMedicine #TempusAI #HealthcareInnovation

317

21 Comments

𝗢𝘃𝗲𝗿 𝗹𝘂𝗻𝗰𝗵 𝗶𝗻 𝗕𝗲𝗿𝗹𝗶𝗻, 𝗮 𝗳𝗿𝗶𝗲𝗻𝗱 𝗮𝗻𝗱 𝗜 𝘁𝗮𝗹𝗸𝗲𝗱 𝗮𝗯𝗼𝘂𝘁 𝗽𝗿𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗺𝗲𝗱𝗶𝗰𝗶𝗻𝗲. 𝟴 𝗵𝗼𝘂𝗿𝘀 𝗹𝗮𝘁𝗲𝗿, 𝘄𝗲 𝗵𝗮𝗱 𝗯𝘂𝗶𝗹𝘁 𝗮 𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝗱𝗶𝗴𝗶𝘁𝗮𝗹 𝗽𝗮𝘁𝗶𝗲𝗻𝘁 𝘁𝘄𝗶𝗻. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗳𝗼𝘂𝗻𝗱 𝘄𝗮𝘀 𝗺𝗼𝗿𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝘁𝗲 𝘁𝗵𝗮𝗻 𝗲𝗶𝘁𝗵𝗲𝗿 𝗼𝗳 𝘂𝘀 𝗲𝘅𝗽𝗲𝗰𝘁𝗲𝗱. My friend is a medical scientist and decades of experience in clinical diagnostics. I build agentic AI systems. Over lunch the conversation turned to personalized diagnostics, healthspan optimization, and the idea that every patient's biomarker profile tells a unique story — and most of it goes unread. He described digital patient twins — computational models that integrate a patient's full biomarker landscape and reason across all of it simultaneously. Not flagging what's high or low, but modeling how values interact, compound, and respond to interventions. I said: that sounds like exactly what Claude Code could do. Why don't we give it a try? Sitting in a café while it rained outside in Berlin, we began. ⏤ By that evening we had a working system. We validated the architecture against clinical datasets before running real patient data, then analyzed two complex cases known to us. The results were striking. It identified cross-biomarker interactions that single-value review doesn't surface. Values that, modeled as a system, revealed compounding effects — explaining clinical pictures puzzling for years. For each case it generated intervention scenarios — ranked by impact, weighted against cost and availability, with predicted biomarker trajectories. Not theoretical. Actionable interventions that could start tomorrow. Then we stress-tested it: independent blind analysis from raw data only. Same core findings. Same priority interventions. When real follow-up measurements came in, predictions held within stated confidence bounds. My friend, looking at the output: "𝗧𝗵𝗲 𝗰𝗿𝗼𝘀𝘀-𝗯𝗶𝗼𝗺𝗮𝗿𝗸𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝗿𝗲𝗰𝗼𝗴𝗻𝗶𝘁𝗶𝗼𝗻 𝗶𝘀 𝘄𝗵𝗮𝘁 𝘁𝗮𝗸𝗲𝘀 𝘆𝗲𝗮𝗿𝘀 𝗼𝗳 𝗰𝗹𝗶𝗻𝗶𝗰𝗮𝗹 𝗶𝗻𝘁𝘂𝗶𝘁𝗶𝗼𝗻. 𝗧𝗵𝗶𝘀 𝗱𝗶𝗱 𝗶𝘁 𝘀𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗶𝗻 𝗺𝗶𝗻𝘂𝘁𝗲𝘀." ⏤ I'm sharing this because it reframed how I think about AI. Not answering questions faster. Reasoning across complexity — simultaneously, with uncertainty made explicit. That architecture isn't specific to medicine. Anywhere the answer lives in the 𝘪𝘯𝘵𝘦𝘳𝘢𝘤𝘵𝘪𝘰𝘯 between variables — financial risk, legal analysis, engineering trade-offs, supply chain — this pattern applies. I didn't write medical algorithms. I designed a reasoning workflow — structuring clinical evidence so AI agents could reason with it like a specialist would. The domains where this hits hardest? The ones where experts say "𝘪𝘵 𝘥𝘦𝘱𝘦𝘯𝘥𝘴 𝘰𝘯 𝘢 𝘭𝘰𝘵 𝘰𝘧 𝘧𝘢𝘤𝘵𝘰𝘳𝘴." That's exactly where agentic AI belongs. What domain would you point it at? #TechTO #TorontoTech #AgenticAI #AI #GenerativeAI #ClaudeCode #SaaS

32

4 Comments

Thrilled to be attending the 25th Precision Medicine World Conference (PMWC 2026), March 4–6 in Santa Clara, CA, co-hosted by Stanford, UCSF, and Yale. At OmicXHealth, we’re building an applied AI/ML-driven molecular diagnostics platform focused initially on liver disease detection, risk stratification, and early hepatocellular carcinoma (HCC) detection in high-risk populations. We’ve completed preclinical validation using biobank cfDNA samples (described in a recent preprint) and are currently conducting a blinded validation study comparing HCC vs. non-HCC. PMWC is a fantastic forum bringing together leaders across precision medicine, biotech, healthcare, and technology to advance personalized patient care. Looking forward to learning, sharing, and connecting—if you’re attending and interested in AI-enabled diagnostics, liquid biopsy, or translational collaborations, I’d love to connect. @Tal Behar @PMWC – Precision Medicine World Conference #PMWC2026 #PrecisionMedicine #LiquidBiopsy #AIinHealthcare #Diagnostics #Biotech #HealthTech #Oncology #MolecularDiagnostics

20

LLMs are getting commoditized. The pace of core model breakthroughs has clearly slowed. We are no longer seeing step-function jumps from “just a bigger model”. For most real systems today: • architecture matters more than the base LLM, • planning, orchestration, memory, and verification dominate quality, • the model itself is increasingly a replaceable component. In my own code factory, I can swap GPT, Claude, or Gemini with minimal impact. There are differences in latency and edge cases, but they no longer define the system. The next real breakthroughs will not come from GPT-6. They will come from new architectures built around models, not inside them. LLMs are becoming infrastructure. Systems are becoming the product.

1

📣 Excited to share #Arbor, a framework for reliable navigation of critical conversation flows. When using LLMs to tackle complex tasks with large contexts, single-prompt approaches hit well-documented limits: context overflow and lost-in-the-middle effects. This compromises their use in healthcare where reliability is fundamental. By decomposing navigation into specialized, node-level tasks smaller, cost-efficient models using Arbor matched or exceeded larger models on single-prompt baselines. This demonstrates that thoughtful architecture can often trump raw model capability. Technical Report: https://lnkd.in/eAfrAAWv

34

1 Comment

Study Mode is a new feature in ChatGPT designed to help students learn more effectively. It offers personalized explanations, practice questions, and summaries to support focused studying. Just released by OpenAI, it turns ChatGPT into a smart learning assistant.

1

Exciting news: Our RFT-tuned model boosted ICD-10 coding accuracy 27% over expert clinicians—and it’s now spotlighted in OpenAI's Reinforcement Fine-Tuning use case guide! We're excited to see what the community builds with this powerful new customization technique. https://lnkd.in/gGUKiHqS Why is this a game-changer? ICD-10 coding is a maze for clinicians: ~70,000 codes, shifting payor-specific rules, & the risk of huge fines for errors. It's a critical administrative burden. We're aiming to simplify this. Our RFT-tuned reasoning system listens to visit audio, integrates relevant EHR context, & recommends precise ICD-10 codes in real-time. This supports clinicians at the point-of-care, improving reimbursement integrity and reducing compliance risk. This leap in performance is a testament to the power of RFT. Exceptional work by our friends at OpenAI in pioneering this new paradigm in fine-tuning!

173

7 Comments

LLMs are moving into personal health coaching—and the results are surprising. A new Nature Medicine paper introduces PH-LLM, a version of Gemini Ultra 1.0 fine-tuned for sleep and fitness guidance. Key takeaways: 🔹 Expert knowledge: PH-LLM scored above human experts on sleep and fitness multiple-choice exams (79% vs. 76% in sleep, 88% vs. 71% in fitness). 🔹 Personalized coaching: On 857 real case studies with Fitbit data, its long-form recommendations were rated nearly as high as domain experts—statistically *indistinguishable* in fitness. 🔹 Predictive insights: Using multimodal wearable data, PH-LLM predicted subjective sleep outcomes (like sleep disturbance) as accurately as traditional statistical models—while integrating results directly into conversational coaching. * What makes this work stand out is not just accuracy, but the integration of knowledge, personalization, and prediction into one system—bridging raw data, subjective outcomes, and real-world guidance. Yes, there are caveats: data biases, occasional factual slips, and the need for careful validation. But the direction is clear: LLMs are becoming more than assistants—they’re evolving into personalized health coaches #AI #HealthcareInnovation #WearableTech #SleepTech #Fitness #LLM #FutureOfHealth

1

Your generative application is only as safe, balanced, and effective as the data it is trained and evaluated on. Generating these datasets at scale? Very hard. Alisar Mustafa and Cherry Wu released a detailed whitepaper outlining Duco's 7-stage data refinement pipeline to deliver high-quality, human-generated datasets. A highly recommended read if you are working on this challenge. https://lnkd.in/eUnj3uDC

28

1 Comment

Ivan Zhao (@ivanhzhao), CEO of Notion, just dropped a brilliant essay: Steam, Steel, and Infinite Minds. Every era has its miracle material: Steel built skyscrapers. Steam powered factories. Now AI brings infinite minds to knowledge work. We’re still in the “waterwheel phase”—bolting chatbots onto human workflows. But soon: • Individuals: From bicycles to self-driving cars for the mind. . Organizations: AI as steel (scalable without degradation) + steam (reimagined structures). • Economies: From Renaissance Florences to megacities of continuous, leveraged At Notion, they’re already running 700+ agents alongside humans. The future isn’t copilots—it’s reimagining everything. Must-read for anyone thinking about AI’s impact on work. Full thread: https://lnkd.in/gv_ShfFn #AI #FutureOfWork #Agents #KnowledgeWork #Notion

1

1 Comment

Researchers at UC Berkeley and UCSF made it possible for someone who can't speak to talk again using only their brain signals. A person with paralysis silently tries to say something, and sensors pick up brain signals from the part that controls speech. Then, an AI model turns those signals into actual spoken words. And we're not talking robotic voice - it sounds like their voice. In the study, a woman named Ann who can't speak anymore was able to "say" sentences just by thinking them. The system used her past voice recordings to make the speech sound like her. What’s cool is it works with different types of devices - from implants on the brain to sensors on the face. It’s also super fast - almost real-time, like a conversation. Basically, it's like giving people their voice back, even if they can't physically talk. The source: https://lnkd.in/d3VCqydf

1

Andrej Karpathy on AI-assisted coding: "I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December. [...] It hurts the ego a bit" Full post with details and predictions on X: https://lnkd.in/dxP24SDf Andrej was recently on record (on Dwarkesh podcast) saying AI coding agents were not good enough for non-boilerplate coding tasks, which makes the new post so surprising (and yet not, to people watching closely what's been happening with the coding agents). Anecdotally, I've been hearing the same from *many* senior software developers who know their stuff (ie. aren't just blinded by hype). The development workflows are being structurally redefined very quickly. In my recent talk on AI-assisted software engineering I tongue-in-cheek joked that the information has a "best before" date of a couple of months. Seems right on spot ... I'll need to update the talk (again).

22

2 Comments

Scribes, AI or otherwise, are completely useless in healthcare. You CANNOT document patient encounters with "notes" in language. All you've accomplished is putting paper records behind the glass of the monitor. That's stupid and you can tell that bizdiot Judy Faulkner I said so. You need a data driven EMR that is searchable and relatable. Before you say "mine is" or "AI," no it's not. Sentia has provided a data driven EMR that relates symptoms to observations to clinical measurements to lab results to diagnoses to treatments to outcomes. You can literally pick a list of symptoms and get her back a collection of cases with positive outcomes. You can't do that with language. It's also faster, easier and cheaper than typing "notes." Stop. You are doing it wrong. Start over with a real EMR that is built the way it should have been in the 70s if Judy had a brain in her head. Read this: https://lnkd.in/gR3ArMD9

7

We were honored to see Plurai included as a Representative Vendor in a leading market guide covering AI evaluation and observability platforms. For a young company, this kind of recognition is meaningful - but not because of the logo. It’s meaningful because it reflects a shift many CTOs and VPs of Engineering are already living through in production: the hardest part of AI is no longer building it, it’s trusting it. As AI systems move beyond single prompts into agents, tools, and long-running workflows, traditional software assumptions break down. Behavior becomes nondeterministic. Failures are rare but high-impact. Issues are deeply coupled to product logic, policies, and real user context. In that world, classic testing is insufficient, and observability alone only tells you what happened after the fact. One of the strongest signals in the market guide is that evaluation and observability are converging into a new layer of engineering infrastructure. Not as a one-time validation step, but as a continuous discipline that connects development and production. The goal is no longer just measuring quality, but maintaining it over time as systems evolve. What stood out to me is that evaluation without a path to improvement doesn’t scale. Scores, dashboards, and alerts are necessary, but they don’t solve the core bottleneck engineering teams face: understanding why systems fail, reproducing those failures reliably, and fixing them with confidence rather than trial and error. This is why approaches based on realistic simulation, synthetic scenarios, and closed feedback loops are becoming central in this space. Not as academic exercises, but as practical ways to stress-test system behavior before users encounter failures, and to intentionally improve agentic systems using production-grounded signals. That perspective is also what makes Plurai’s inclusion in this category meaningful to us. Our work has been focused on system-level behavior rather than isolated model calls, on simulating real product scenarios rather than relying solely on historical logs, and on explicitly connecting evaluation to concrete actions like new tests, targeted guardrails, and ongoing optimization. From an engineering leadership perspective, this isn’t about better metrics. It’s about regaining control over systems that are inherently probabilistic. My takeaway for CTOs and VPs of Engineering: AI evaluation and observability are becoming as foundational as CI/CD and production monitoring once were. But their real value shows up only when they help teams move faster and safer at the same time. The question is no longer “does the AI work?” It’s “can we trust its behavior today - and keep it trustworthy as the system scales and changes?”

80

11 Comments

Beyond the hype, the true value of GenAI lies in how it can reimagine workflows—within our own organizations and across the industry. Discover how emerging tech transforms operations: https://1blk.co/4oKiYu4

2

It’s not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality). The key is to build intuition to decompose the task just right to hand off the parts that work and help out around the edges. But imo, this is nowhere near "business as usual" time in software. https://lnkd.in/gjGUr8ar

5