Vincent Lyu

founder of PangolinLab｜AI Eval Platform
a Dentist in China；ALL in AI
Make HealthcareAI for everyone

[Vincent Lyu @LyuVincent]

To understand how AI medical data is commercialized, you have to look at a company.

[Tempus @TempusAI]

Today we reported our financial results for Q3 2026.

See the full earnings release and business highlights: https://t.co/cfuKwHsMr0

[Vincent Lyu @LyuVincent]

An excellent doctor must first have a mature personality, humanistic care and scientific literacy.

[Author of “State-Led Globalization” (forthcoming) @PingFu1999]

美国之所以把MD门槛抬得这么高，是因为美国医学会（AMA）和美国社会形成了一种共识：一个优秀的医生，不仅要有高超的医术，还要有成熟的人格、人文关怀以及广泛的科学素养。

18 岁的高中毕业生在美国被认为在心理和心智上还不够成熟，不足以应对生离死别、高压的医患关系和复杂的医学伦理。

经历过 4 年通识本科（可能学过心理学、历史、社会学）的洗礼，到 22 岁再进入医学院的学生，往往心智更成熟，职业目标也更坚定。

[Vincent Lyu @LyuVincent]

On October 22, Microsoft published MAI-DxO research in Science and open-sourced SDBench. Using multi-agent collaboration, MAI-DxO achieved 83.2% diagnostic accuracy on 304 complex NEJM cases (vs. 36.1% for doctors) and 80.3% on 325 emergency cases (vs. 46.8%).

[Healthcare AI Guy @HealthcareAIGuy]

Microsoft's AI outperformed doctors on 300+ complex medical cases.

2.2x more accurate, 72% cheaper.

[Baba @BabaHQ]

Today, Baba raised a $40M Series A.

We give each Medicare member and their family a dedicated team of advocates and behavioral health experts to keep every part of their care moving forward. In a world obsessed with AI personal assistants, it’s never been more important to have expert humans in the loop to go the last mile, particularly in healthcare.

Our advocates have moved mountains for their patients - finding safe housing for the near-homeless, recertifying Medicaid for those about to lose their benefits, and getting financial assistance for people cutting their pills to afford groceries.

In the process, we’ve used AI to scale the impact this special team can have. And to truly teach AI to reason and execute like the best care coordinators, we’ve built the largest learning environment for the type of work that we do everyday.

There is still a lot to do: we’ve served fewer than 1% of the 70 million Americans on Medicare.

Baba does not lack capital or customer demand. Our ability to seize this opportunity is now limited by the capacity of our team. If you’re an exceptional engineer or operator in New York, come build with us: https://t.co/iNUpe3Zd6U.

Our Series A was led by @FirstMarkCap , with @CVSHealth, @valtruis, @generalcatalyst, @geniusventures, and @GroundUpVC investing alongside them.

[Vincent Lyu @LyuVincent]

OpenAI在1月收购Torch，医疗数据公司；4月发布GPT-Rosalind，生命科学模型；10月收购Orchestra Bio，AI走进药物研发。

[OpenAI @OpenAI]

We’ve acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development.

Together, we’ll build on Rosalind, our AI models and tools for life sciences research, to streamline those operations and help new treatments reach patients sooner.

Welcome @ashokaraj_ and team!

https://t.co/vi51OhznQn

[Vincent Lyu @LyuVincent]

OpenAI acquired Torch, a healthcare data company, in January; launched GPT-Rosalind, a life sciences model, in April; and acquired Orchestra Bio in October, marking AI's entry into drug development.

[OpenAI @OpenAI]

We’ve acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development.

Together, we’ll build on Rosalind, our AI models and tools for life sciences research, to streamline those operations and help new treatments reach patients sooner.

Welcome @ashokaraj_ and team!

https://t.co/vi51OhznQn

[Vincent Lyu @LyuVincent]

@jselanikio @OpenAI @advocatehealth @becker sad to hear from you. AI company is becoming a giant monster to be a Entrance, encroaching on the medical and health market

[Vincent Lyu @LyuVincent]

sad to hear from you. AI company is becoming a giant monster to be a Entrance, encroaching on the medical and health market

[Joel Selanikio @jselanikio]

Incredibly, hospitals are now paying for AI companies to disrupt them. 🤦‍♂️

Patients linking their MyChart accounts to @OpenAI's ChatGPT rose more than 7,000% in a single month at 69-hospital @AdvocateHealth, its CIO Bobbie Byrne told @Becker’sHR.

Patients set this up themselves in ChatGPT, so that the AI can explain their lab results, scans, and doctors’ notes. You know, instead of one of Advocate's doctors.

To add insult to disruption, every time a patient asks ChatGPT a question, Advocate’s Epic system has to look up and send the records and pay for the computing resources that make that happen. AND — brace yourself — Federal law says Advocate can't charge patients, or their apps, for that access to the patient's own records.

So a health system is paying for a competitor’s access to its own patients’ records. The competitor takes away the followup visit, and the healthcare system pays for that to happen — both in lost revenue and in increased computing costs.

That’s Interpretation Migration, subsidized by the disruptee. https://t.co/GTj7Hzfd1d #DrYou #InterpretationMigration #Disruption

[Vincent Lyu @LyuVincent]

@Berci On October 15, JAMA published a study on the real use of OpenEvidence: From January 2024 to June 2026, about 952,000 American clinical professionals submitted a total of 271 million queries. Among inquiries, treatment and management accounted for 33.1%, and diagnosis for 13.9%.

[Vincent Lyu @LyuVincent]

On October 15, JAMA published a study on the real use of OpenEvidence: From January 2024 to June 2026, about 952,000 American clinical professionals submitted a total of 271 million queries. Among inquiries, treatment and management accounted for 33.1%, and diagnosis for 13.9%.

[OpenEvidence @OpenEvidence]

"A growing share of knowledge access is shifting from recall to retrieval." Clinical judgment stays with the doctor.

From a Harvard-led JAMA study, co-authored with OpenEvidence researchers.

[Vincent Lyu @LyuVincent]

Why is there no Ant Afu in the United States and No OpenEvidence in China?

[OpenEvidence @OpenEvidence]

Clinicians bring OpenEvidence the question that matters most once a diagnosis is made: what to do next for this patient.

New in @JAMA_current, from researchers at Harvard Medical School and OpenEvidence, including our CMO @TravisZack2: treatment and management questions outnumbered diagnosis questions more than two to one.

That should change how clinical AI gets tested. Most evaluations still grade diagnosis with board-style vignettes that have one right answer. Real treatment questions rarely do. They depend on this patient's kidneys and on the trial that changed practice last spring. That is the test we want to be held to.

Physicians also reached well beyond their own fields. ER doctors asked about topics outside emergency medicine 88% of the time. Medicine is wider than any one specialty, and good clinicians check before they act.

[Vincent Lyu @LyuVincent]

Baba's website is great. Everyone should have a look! It clearly and intuitively reflects the dilemma of medical care! And baba's distinctive demand response plan

@BabaHQ

[Baba @BabaHQ]

Everyone’s chatbot can book a doctor’s appointment.

The real test is whether you can earn the trust of a Spanish-speaking diabetic woman, walk her through a Medicaid application while finding her secure housing, and coach her into choosing a 12oz (instead of 16oz) soda along the way.

Today we’re sharing BabaBench, a benchmark measuring how frontier models perform the real work of healthcare coordination. Our synthetic worlds are built from annotations by Baba’s in-house team of hundreds of expert doctors, therapists, nurses, and social workers.

Turns out, AI alone isn’t ready to replace human care coordination.

[Vincent Lyu @LyuVincent]

@BabaHQ Baba's website is great. Everyone should have a look! It clearly and intuitively reflects the dilemma of medical care! And baba's distinctive demand response plan

[Vincent Lyu @LyuVincent]

@BabaHQ huge achievement. Interesting AI Agent evaluation

[Vincent Lyu @LyuVincent]

huge achievement. Interesting AI Agent evaluation

[Baba @BabaHQ]

Everyone’s chatbot can book a doctor’s appointment.

The real test is whether you can earn the trust of a Spanish-speaking diabetic woman, walk her through a Medicaid application while finding her secure housing, and coach her into choosing a 12oz (instead of 16oz) soda along the way.

Today we’re sharing BabaBench, a benchmark measuring how frontier models perform the real work of healthcare coordination. Our synthetic worlds are built from annotations by Baba’s in-house team of hundreds of expert doctors, therapists, nurses, and social workers.

Turns out, AI alone isn’t ready to replace human care coordination.

[Vincent Lyu @LyuVincent]

OpenEvidence（OE）以外界合作为核心扩张路径：个人端免费服务超110万美国临床医生；企业端与Mount Sinai、Cedars-Sinai、Penn Medicine、NewYork-Presbyterian、Northwestern、Johns Hopkins、MSK、Sutter、AdventHealth、波士顿儿童医院、UTMB等十余家顶尖医疗系统/大学深度合作。

[OpenEvidence @OpenEvidence]

Johns Hopkins established the model for American medical education more than a century ago. Today @HopkinsMedicine and OpenEvidence announced a collaboration to define what that education should look like in the AI era, and to ground clinical care in the latest evidence.

As AI learns to synthesize the world's medical knowledge in seconds, the clinician's role moves toward clinical judgment, first-principles reasoning, and the human work of caring for patients. The education model Johns Hopkins and OpenEvidence will design builds on those foundations, with a deeper grounding in pathophysiology and evidence-based medicine. Knowing why an ACE inhibitor works is understanding a physician carries through an entire career.

Together we'll also build human-first AI that puts the best available evidence in the context of each patient's medical history and Johns Hopkins' best practices, so clinicians can reason through that information faster and make the best decision for their patient.

Sanjay Desai, Vice Dean for Education at Johns Hopkins Medicine, says AI "invites us back to the bedside."

For the educators here: which skills do you want every trainee to keep sharp?

[Vincent Lyu @LyuVincent]

@cristinatejerap @NEJM Los estudiantes, incluidos los profesores, pueden aprender recetas mecánicas convencionales, pero la IA puede romper los límites de conocimiento de múltiples especialidades y proporcionar un diagnóstico más completo y basado en la evidencia.

[Vincent Lyu @LyuVincent]

@ChenAaron 写的很好，医疗和法律在AI应用有相似和不同的地方。我也base杭州，有机会线下交流呀～

[Vincent Lyu @LyuVincent]

这种现象可能在医学，尤其是内科，会越来越普遍。学生包括当年的老师自己可能学到的都是机械的诊疗方案、常规开方，AI却能突破多个专科的知识边界，提供更加综合全面、更有循证依据的完善诊断。

[晓辉博士 @xiaohuiyuan]

最近跟几个在国内985大学当老师的朋友聊，他们都说，现在真的不太情愿手把手指导研��生干活了，用AI做科研省事又省心，教研究生教半天一堆问题，有的时候还有各种情绪，想想何必呢？为难自己，为难学生。

我的第一反应是心疼这一波儿读研的学生们，没有了导师研究品位和经验的传递，没有了手把手指导，在科研上可能要走一些弯路了。

而且如果老师们都这样想这样做，那陶哲轩所说的学术共同体的建设，日后也会越来越松散了，学术经验的传承也会成问题。

但转念一想，当学生们没有了老师们手把手的指导，其实也更自由了，真正热爱学术的会跟AI学，跟网上最好的资源学，而那些混个文凭的，本来也没必要学，用AI做点自己感兴趣的事，好像也挺好。

那问题来了，为什么这些孩子还要再读研究生呢？只是为了个学位吗？学位在AI加速的今天还有意义吗？

[Vincent Lyu @LyuVincent]

@xiaohuiyuan 这种现象可能在医学，尤其是内科，会越来越普遍。学生包括当年的老师自己可能学到的都是机械的诊疗方案、常规开方，AI却能突破多个专科的知识边界，提供更加综合全面、更有循证依据的完善诊断。