Cameron Po-Hsuan Chen

It's refreshing to see AI being used to help heal. Mike Meaney was already building something special at One Small Step — connecting people in crisis with peer supporters who have lived experience in substance use and mental health recovery. On-demand, confidential, real support when you need it most. Now, they can take the impact to the next level. Leveraging AI to scale that human connection responsibly and building the infrastructure to make sure these systems actually protect vulnerable people. So proud of what you're building, Mike. This matters!!

6

1 Comment

In 2025, AI research took a hard turn toward foundational innovation. From modular sparsity to world models, here are 5 deep AI research developments redefining how AI thinks, plans, and aligns. 1. Modular Sparse Mixture-of-Experts (MoE) at Scale: This architectural shift breaks the linear scaling wall of dense transformers and supports the real-time, low-cost deployment of multi-trillion parameter general-purpose AI models. 2. Neural-Executable Reasoning (NER) & Program-Aided Language Models (PALM++): This involves large language models (LLMs) synthesizing and executing code internally as a form of thought, which fuses formal logic with neural fluidity to improve performance in math and software engineering. 3. World Models & Active Inference via Generative Simulation: Models learn internal simulators of the environment to enable agents to "dream" and plan possible futures, which is key for autonomous robotics and safe AGI exploration. 4. Alignment via Constitutional AI + Reward Model Ensembles: This approach uses self-improvement loops and value ensemble learning to formalize alignment as a scalable, bootstrapped, multi-agent control problem. 5. Universal Multimodal Tokenizers (UMT) and Cross-Domain Latent Representations: This breakthrough unifies different modalities (vision, audio, motion, etc.) into a shared embedding space, creating a language of generalized perception crucial for embodied AI and human-AI interfaces. What other AI developments happened in 2025 that may change the future?

16

As a physician, here’s what keeps me up at night about LLMs in healthcare. It’s not that the models are unintelligent. It’s that we are confusing benchmark performance with real-world safety. A recent study in Nature Medicine evaluated how large language models perform when real people use them for medical self-assessment. When tested alone, the models identified relevant conditions in ~95% of cases. When paired with actual users? Performance dropped dramatically. Users identified relevant conditions less often than those using traditional resources. There was no meaningful improvement in selecting the correct level of care. The failure wasn’t simply hallucination. It was the interface between human and model. • Patients provided incomplete information • Models misinterpreted nuance • Users over-trusted confident language • Red flags were missed or ignored Benchmarks did not predict this. Simulated testing did not predict this. Real-world interaction changed the outcome. As LLMs increasingly become a first touchpoint for patients trying to interpret symptoms, the risk isn’t just incorrect answers. It’s misplaced confidence. It’s acuity underestimation. It’s burden quietly shifting from system to individual. And perhaps most concerning: the belief that “human-in-the-loop” automatically protects against harm. It does not unless that oversight is structured, monitored, and accountable. This is not an argument against LLMs. It is an argument against premature confidence. Model accuracy is a laboratory metric. Patient safety is a systems outcome. Until we evaluate these tools under real-world conditions with real users, governance conversations are incomplete. You can’t govern what you don’t understand. Link to the full text article below in the comments👇🏼 #HealthcareAI #PatientSafety #AIGovernance #DigitalHealth #ResponsibleAI

14

9 Comments

In general AI is failing women’s health. ŌURA's announcement today shows they are trying to fix it. Most AI models are trained on a "default" male physiology. The result? False alarms and dismissed symptoms. Oura built a proprietary LLM specifically to solve this. It’s clinician-grounded, cycle-aware, and—most importantly—non-dismissive. This is what "Responsible AI" looks like in practice. It doesn't just mirror the past; it builds a more equitable future. As always, Alice Zheng, MD, MBA, MPH has the hottest take on what's happening in women's health. Congrats, Oura team!

24

1 Comment

Thank you to Real Chemistry and Mary Faith James for putting together such a compelling panel at #SXSW26. I had a blast representing OpenEvidence and speaking alongside Jessica Calef, Reddit, Inc.'s director of health, on the importance of where we are accessing medical information as patients and providers in the age of AI.

42

4 Comments

This new Nature Medicine study confirms what many clinicians have been sensing: LLMs perform well on exams because exams sit entirely on the language axis. But medicine does not. Clinical reasoning, grounding, context awareness, patient variability, and risk assessment all require a stable internal substrate—not just fluent text generation. When models are evaluated through language-only benchmarks, their real-world failure modes remain hidden. When placed with real patients, those hidden gaps become dangerous. Healthcare will need systems that can maintain coherence and safety even when language breaks. That requires architecture-level change, not scaling alone. This study is a turning point. It shows the industry exactly where the fault line is.

3

3 Comments

If you’re running ClinOps at a Series A biotech, there is one trap that will kill your pipeline before you ever reach Series B. The manual execution death trap. I see this constantly: A small biotech secures funding. They have one or two early-stage trials that are their absolute lifeblood. Because the stakes are so high, the team gets complete tunnel vision on execution. With runway on the line and investors watching every milestone, scalable platform evaluation feels like an unaffordable risk. Teams have no option but to brute-force. They manage everything manually and rely entirely on a CRO to "just handle it for now." They prioritize the comfort of familiar processes over the efficiency of scalable technology and infrastructure. This works, just barely, for a small Phase I safety study. But here is what happens when the clinical data is positive and it’s time to scale: You hit an operational wall. You have limited internal infrastructure, you never were able to optimize your trial design with RWD, and you're suddenly trying to run a complex, multi-center Phase II trial using the exact same manual spreadsheets and ad-hoc processes that barely survived your first study. Tufts CSDD's 2024 research puts the direct daily cost of a clinical trial delay at roughly $40,000 - with every lost day compounding against your runway. The operational debt crushes your timeline. The best ClinOps leaders don't just build for the trial right in front of them. They build for the trial they need to run two years from now. When you are in the early stages, you cannot afford to just outsource execution and look away. You have to build a foundation. Invest in a sponsor-controlled operational platform; one that keeps your clinical intelligence in-house, regardless of which CRO is executing on the ground. Select vendors that easily allow you to transition from Phase I to Phase II without ripping out your core systems. Implement technology that fits seamlessly into your workflows, so it can immediately benefit you with reduced CRO outsourcing costs today and scale with your headcount tomorrow. Your molecule might be what secures your Series B funding, but your operational infrastructure is what actually gets your drug to market.

6