Glass Health 2023 Year in Review
Dear Glass Health Team, Glass Health Clinicians, Glass Health Investors, and Friends,
We have had a remarkable year together. To capture the intensity, dynamism, thrill, progress, peril, and triumph of this year for Glass Health in words is beyond my skills as a writer. But as Walter Isaacson and Ashlee Vance have yet to take on our story, for now, you’ll have to settle for me.
We’d Better Build a Landing Page
On January 14, 2023, Graham Ramsey, my cofounder and one of my closest friends in the world, and I arrived in beautiful, sun-drenched Sonoma, California, for the first-ever Y Combinator (YC) batch retreat. Over the years, I’ve come to believe that the strongest reception for the love, art, and innovation that the universe sends down to us can be found in northern California. Specifically, the reception and connection to the universe is best in the San Francisco Bay Area. It gets even better if you climb to the top of a beautiful mountain, look out at a sea of emerald forest and blue ocean, cross your legs on the ground, still your thoughts, and let the breeze run across your skin.
When we arrived at the YC retreat at the Fairmont Hotel in Sonoma, the startup magic of Silicon Valley that has inspired generations of entrepreneurs was already swirling through the halls, rooms, and grounds. It was the first batch of YC in person since the beginning of the COVID-19 pandemic, and you could feel that intangible forces had been waiting for us to come and build our companies.
In the previous year, 2022, Graham and I had accidentally caught startup lightning in a bottle. In March of 2022, we mistakenly launched our first product, Glass Notebook, by posting a screenshot of the user interface on Twitter, with the goal of recruiting ten beta users for the prototype. At the time, we did not even have a site for signups and asked folks to DM us if interested.
It hadn’t occurred to us that the extremely early prototype of Glass, which simply allowed you to organize your medical knowledge by schema, script, or case, would get doctors so revved up. But the post went viral on MedTwitter—now MedX—and our inboxes filled with over five hundred messages from physicians who wanted access to the product immediately.
We were deluged with notes from eager clinicians who couldn’t believe that a company was going to build products directly for them rather than for the hospital administrator or insurance company. Graham and I were delighted and jumped into action. The messages were piling up, and we realized that we’d better build a landing page for Glass Health....
SpaceX for Medicine
Glass Notebook was loved so much by doctors that within a few months of launching, we had attracted the attention of multiple high-profile investors. This clinician and investor interest in Glass Health was a huge milestone for us. Graham and I founded the company in March of 2021 with a simple mission statement of leveraging technology to optimize the practice of medicine, and the first year of the company leading up to the launch of Glass Notebook in March of 2022 was extraordinarily rough.
Before founding the company, I was a medical student at the UCSF School of Medicine in San Francisco and found myself increasingly dismayed at how limited and analog the technology available to doctors was. I had tried previously to start one company and had been Chief of Staff at another before settling into medical school, but the startup bug was in me, and I thought there might come a time when I would try to build a company again. Graham was working at Modern Fertility, the second of two startups that he had helped build, the first being PlushCare, and he was looking to cofound a company that could have a tremendous impact on healthcare. We were a good match.
We met up in coffee shops to brainstorm, where we talked about building a healthcare tech company that delivered beautiful, deeply desirable, and highly functional products like Apple. What was stopping a company from adopting a product-led strategy—like Notion, for example—in healthcare? What were the barriers, for good or for bad, between medicine and the application of the world’s most cutting-edge technology? We were obsessed with the limitations and possibilities of healthcare technology and dreamed about a company that could metamorphose medicine in the same way that we saw the finance, space exploration, climate, energy, and transportation industries transforming around us.
Upon beginning our startup journey in March of 2021, we were immediately walloped over the head by the numerous mountainous barriers that lay between us and our dream. For ten months after the founding of Glass Health, we tried to sell our mission and early product mockups to health systems and were turned away by everyone, including my alma maters.
A leader at one health system told us that he wouldn’t support a pilot because our mockups weren’t impressive enough or, in his words, because “this doesn’t knock my socks off.” Another health system leader, who was in charge of innovation for the hospital where he worked, told us that we were “wasting our time,” that I “would graduate from residency before I ever saw our products integrated into the EMR,” and that “making life better for physicians wasn’t a priority of the health system.” He said he’d try to find someone who was more passionate about Glass to help us. We never heard back from him.
Because of the resounding “no” we’d heard from health systems professionals, we decided to build a product that required no health system pilot to launch. We created Glass Notebook, a personal knowledge notebook for doctors, and took it directly to the medicine community on Twitter. We thought that if we could earn love for our products and company from doctors, then down the line, we could execute a type of product-led growth that is almost never done in healthcare.
Product-led growth is a type of startup jujitsu that companies like Slack, Asana, and Atlassian use to leverage direct-to-consumer sales into enterprise sales. We figured that we would have a first-person relationship with doctors, serving them through our direct-to-consumer product, and that those doctors would then go to their chairs of medicine and hospital CIOs to advocate for our product to be integrated into the health system to access additional enterprise features.
Of course, the trick to doing all this in healthcare, where product-led growth is usually limited by HIPAA compliance and EHR integration, was that we had to build a product that slipped past the need for health system oversight. It was a bold plan, and a notebook for medical knowledge management and not for direct patient care was the perfect vehicle.
Investors noticed our entry into the field of healthcare technology and building for clinicians. Within weeks of launching Glass Notebook, five or six major investors reached out to Glass Health. Graham and I could not believe it. I told my program director, Joel Katz, MD, that it was possible Glass Health might raise a $500K to $1M pre-seed round of investment that year. He told me he’d support me and to give him as much notice as possible if I needed to leave my internal medicine residency to pursue this dream. He offered to make an important connection that day with the great Tom X. Lee, who is the founder of One Medical, Epocrates, Galileo Health, and a Brigham and Women’s, Harvard Medical School alum. Joel also told me not to get too excited about the investor interest. “These sorts of things rarely work out, and when they do, they happen slowly.”
Almost all Joel Katz’s predictions about how medicine works bear fruit, but his predictions about how fast things could move for a startup were a bit off in our case. Glass Health and the investor conversations were moving at lightspeed. Every day after shifts, I would meet with Graham to work on product design, pitch our product to health systems, or meet with investors.
But despite all the interest and buzz around Glass Health, there was one problem that kept coming up over and over in investor meetings.... We were making no money, and our company bank account had less than a thousand dollars in it. Graham was working full-time for Glass without any salary or benefits, and I was working every night and on weekends—coming home from caring for patients, sometimes after 12-hour shifts carrying lists of up to 80 patients overnight, sometimes after heartbreaking rapid responses and codes, sometimes covered in sweat from rounds of CPR, sometimes emotionally fractured from talking to the families of patients—to build our company. All the while, we funneled as much of Graham’s savings and my residency salary as we could into the Glass Health bank account.
One by one, all the interested investors who had come knocking when we launched Glass Notebook passed on Glass Health, except for the first who had ever reached out to us. Jim Breyer, a legendary investor who famously led the Series A in Facebook from Accel Ventures, had found out about our company through his sons and his confidant and family friend, a fantastic investor and soon-to-be physician, Morgan Cheatham. Morgan had fallen in love with Glass and was one of our early clinician users.
The Breyer Capital team had come to hold a deep conviction in our mission and were similarly adamant about the need for transformative change in medicine. Even though we were astoundingly pre-revenue, Jim and his team were believers. Like us, they watched the devastation of the COVID-19 pandemic on our healthcare system with deep concern and knew that the time to make bold bets on how technology could improve healthcare was now. So, despite us having only a few dozen power users, based on the love doctors had for Glass Notebook and our company, Breyer Capital wanted to make a deal to invest 1.5M in Glass Health.
I remember signing the papers and wondering if Jim knew that Glass was just two friends who cared deeply about healthcare and dreamed of medicine having its SpaceX. We were indeed determined to build a company that would roil incumbent players, break records, and push the healthcare industry to its frontier, but at the time, we only had a vague idea about how we would go about achieving this impact and our mission.
Over the years, I’ve come to realize that Jim Breyer is a master investor who knows how to pick winners before anyone else can see them, how to support them, and how to let them operate. He and his team knew exactly what Glass Health was then and exactly what we could become.
10,000 Hours of Pickup Basketball Will Not Make You Michael Jordan
As I look back now on that wild spring of 2022, I’m less surprised that our notebook product was so loved by doctors. It was, in a way, a love letter to the doctor who was on a journey to become a master clinician. Glass Notebook was a tool that Graham, had built for me to be the best physician I could be during my medical training at UCSF and then Brigham and Women’s Hospital. Graham was championing me on the journey to become a master clinician that I had started at UCSF when Gurpreet Dhaliwal, a world-renowned diagnostician and now close friend and mentor, had given a lecture to the medical students. When talking about the need for daily, deliberate practice on clinical reasoning, Gurpreet had ended a slide by casually dropping the killer line, “10,000 hours of playing pickup basketball will not make you Michael Jordan.”
That woke me up. I kept replaying it in my head. “10,000 hours of playing pickup basketball will not make you Michael Jordan.” It had that pure, golden ring of truth. Gurpreet was right. If I didn’t adopt a careful and deliberate approach to improving myself as a clinician, I could spend tens of thousands of hours of my life in the hospital practicing medicine and never really get better. Actually, if I didn’t practice medicine as a martial art, like Gurpreet did, I would probably get worse as a doctor as time moved me away from residency, with its morning reports and noon conferences that discussed the latest evidence.
Gurpreet, who has been written about as the real-life House MD, imparted an approach to medicine that continued to have a tremendous impact on me through my time at UCSF. Several years after that first lecture, our relationship had grown into a mentorship and friendship.
When the possibility of investment became real, and I was on the eve of pausing residency to take Breyer Capital’s investment and work full-time on Glass, I called Gurpreet from Boston. In addition to being one of the world’s best educators and diagnosticians, he had a background in economics and could help me think through just about anything. “Dereck, if you continue on the traditional path, your chance of success for becoming a hospitalist is just about 100%. If you embrace becoming an entrepreneur, take the investment from Breyer, and leave residency, your chance of success is the base rate of startup failure multiplied by an unknown factor, which is you. But if this is your dream, and if you don’t go for it, your chances of regret are 100%.”
Gurpreet was kind enough not to explicitly factor in the failure of the AI brain-computer-interface startup I had tried as a premedical student, although I now realize that this failure actually increased my chances of success.
The Glass Notebook Graham and I had designed, with close and tremendous input from our good friend Julian Genkins, MD, a Clinical Informatics Fellow at Stanford, was a tool to help me and my physician friends train our own brain-tissue-based neural networks.
We were using Glass Notebook to sharpen our minds for diagnostic pattern recognition and externalize our memory with a system for storing and retrieving information about how to approach undifferentiated diagnoses and how to provide the best possible evidence-based care when a diagnosis was made.
We loved Glass Notebook. We needed Glass Notebook. Or, at least, I certainly did.
I could not stand that I’d had several lectures and noon conferences on the hyperosmolar hyperglycemic crisis and still wasn’t ready to teach it to medical students. I was frustrated that my medical school notes were all over the place in paper notebooks. I was furious that I didn’t have an organized way to log my learning from unusual and important patient cases and that without one, the lessons of patient care could slip through my fingers. I was at my wits end about my own memory’s forgetting curve, its neural pruning of medical facts, and why my brain seemed so determined to forget the medicine I’d worked so hard to learn.
Also, why were we managing our services in those god-forsaken empty text boxes in the EMR sidebar?
Graham, because of his deep empathy for others and his unique ability to get into a user’s experience, was ready to answer the call from me and from medicine. He had mocked up new ways of optimizing tasking for medicine teams that mirrored Asana, new ways of team communication that mirrored Slack, and now this beautiful Glass Notebook that mirrored Notion, except that it was intimately crafted and tailored to physicians needs.
Of all the product directions we explored, the notebook emerged as the most important. We chose to focus less on carrying out tasks and more on diagnosis and implementing evidence-based treatment, and this focus struck right at the heart of medicine. We became obsessed with Glass, as did our doctor community.
Why Would Anyone Ever Pay for This?
Fresh from raising our pre-seed round from Breyer Capital, Graham and I applied to Y Combinator for the summer 2022 batch. We were interviewed by the YC healthcare partner Surbhi Sarna, who grilled us for 10 minutes on users, acquisition, retention, usage, and product vision before abruptly ending the interview.
Surbhi had built a life-saving biomedical company as a young woman of color without access to the venture capital world or connections in the healthcare industry, and she’d sold her company for 275 million dollars before coming to coach founders at YC. Surbhi is made of steel. She was a wartime CEO, and she did not bring healthcare companies into YC to play games.
At one point in the interview, she fell silent and looked down and then asked, more to herself than to Graham or me, “Why would anyone ever pay for this like they pay for UpToDate?”
That one really hurt. Graham and I had an inkling that someone’s personal medical knowledge was much more valuable to them than a static encyclopedia. We knew that we valued our own Evernote and Notion accounts, where we kept our personal knowledge bases as much as we valued Wikipedia. We knew we were onto something, and maybe Surbhi did, too, but we were rejected by YC the day after our interview.
I had applied to YC once as pre-med and twice as a medical student, and Graham had applied once as a teenager. The summer of 2022 was our collective fifth rejection from YC, and it stung so much that the next time the application opened, Graham and I didn’t move to put one in. We were too busy managing our operations and growing love for Glass Health.
In the months following the raise from Breyer Capital, we used our funding to bring on our original brilliant clinical team members, including Andrew Sanchez, MD, Gurleen Kaur, MD, Kannu Bansal, MD, and Madellena Conte, MD Candidate. This early team put brilliant pages into the notebook that covered key topics in medicine. It was astounding to watch as their pages emerged overnight as some of the most practical and useful resources for clinicians, particularly clinicians in internal medicine training.
Dozens of Companies Just Like You
Our initial team was phenomenal, and support for Glass Health was growing in the medicine community. But our operational costs were also growing as at the same time the economy seemed to be slowing down. Graham and I discussed our situation with our advisors and concluded that we were in danger of running out of money in the coming 12 months if we didn’t raise a pre-seed extension for Glass Health.
We set out to raise a pre-seed extension in 2022, which I can only describe as a definitively miserable experience that I would not wish on any other founder. We reached out to 40+ investors, most of whom either refused to meet with us, passed during our Zoom calls, or tried to get us to change the company’s mission.
One tried to force us to put pharmaceutical ads into Glass before they would consider investing. Another agreed to invest in us but then called the next day to say they were pulling out of the deal because there were “dozens of companies out there just like you.” I was livid about the demand that we do pharma ads and frustrated with the back pedaling of some investors. Graham and I were disappointed that the round was not coming together, but we had a very limited network, and what could we do except keep hustling for warm intros to investors and trying to raise the round?
With all the frustrations of the pre-seed extension and the increasing demands of supporting the notebook product and its growing user base, Graham and I had not made moves to apply to YC for the W23 batch. It was only after getting a note from the admissions director telling us that, despite being rejected, we had been in the top 10% of applicants in Summer 2022, that we sat down and agreed to allot a single day for throwing together a new YC application.
This YC application was shorter, less polished, and more to the point. And in October of 2022, we were again invited to interview with Surbhi Sarna. This time her questions were harder. She pushed us on the fact that we hadn’t monetized and called us out on a lack of progress on user growth and retention since our last interview.
Graham and I left the interview with Surbhi convinced that we hadn’t gotten in once again. We told ourselves we wouldn’t worry about YC’s decision and tried to move on with our day. I tried to watch a movie, and when that wasn’t distracting enough, I turned to video games.
But then the startup magic started swirling again. Surbhi Sarna called my cell phone at around nine PM that night and asked if I had a moment to talk. I asked her to hold so I could patch Graham in. She told us that she was glad that we applied again. She believed in us as entrepreneurs, but she wasn’t sure about the notebook product. She expected us to start charging for Glass as soon as possible to ensure we were on track to find a product-market fit. And, she told us to show up in San Francisco in January.
We’d gotten into the W23 Y Combinator batch.
Permission to Dream
Back to beautiful and sunny Sonoma in January of 2023. Graham and I were walking through vineyards, biking through hills, and staying with other YC founders at a hotel that had the vibe of a middle school summer camp. Being part of the YC batch was incredibly fun and so very surreal.
The YC partners had accepted a select set of founders, 1.5% of 20,000 applications, to the batch for three months of intensive coaching. Joining YC granted access to a community of extraordinary founders that included the founders of Airbnb, DoorDash, Instacart, Reddit, Dropbox, and many others. I wrote in my journal and told my partner, Niralee, that Garry Tan, the President of YC, had said to us in his welcome talk, “We give you permission to dream as big as you can. We will support you.” I had not had an adult say that to me since I was in elementary school.
Y Combinator wasn’t all wine tasting, live jazz, Graham cleaning up in poker, and me pacing around the hotel lobby contemplating the future of the company. It was also a tremendous amount of work. Before Graham and I had even checked in to the hotel, we were confronted immediately by Surbhi when we arrived without small talk or introductions.
Her first words upon meeting us in person were, “Have you launched your Pro subscription yet?” Over the past few months, she’d been pushing us to fully test whether our product at the time, a medical knowledge management system or “notebook for doctors,” was ambitious enough to meet our mission of leveraging technology to optimize the practice of medicine. She strongly suspected it was not.
We had launched the Pro version of our notebook, albeit only a few hours before arriving. We had slow-walked this because we didn’t understand why Surbhi was so insistent on us charging. There weren’t enough Glass users for me and Graham to buy a pizza, much less pay our salaries. But we did it anyway, rather than showing up without having done what we’d agreed to do in October. On the way to Sonoma, I texted my family and Niralee to tell them proudly that two people had subscribed to Glass Pro. “We officially make $24/month!”
You’re Going to Die, Anyway
What I came to understand through the experience at Y Combinator was that Surbhi wasn’t concerned about us making money for the purpose of having revenue; she was concerned about clarifying our product vision, refining our mission, finding a product-market fit, pivoting if we had to, giving us a metric to measure our growth by, and doing all this as fast as possible so our company would not run out of money and die.
It was her guidance and teaching about the importance of tempo, execution, and deal heat, along with the teaching of our other close mentor at YC, Michael Seibel, that put us on track for phenomenal success in 2023.
I have long looked up to Michael Seibel, Managing Director of YC and cofounder of Twitch. His brilliance and ability to tell the hard truths to founders and the world with a smile, a laugh, and a do not f*** with me determination to solve the world’s hardest problems had been an inspiration. Not only is Michael one of the highest horsepower minds in technology, an innovator who has coached hundreds of companies to astounding success, but he is also one of a few Black men in technology leadership. Before I knew him personally, he was a role model for me at a time when I was constantly battling internal self-doubt about whether the world would take someone of my age, ethnicity, and limited experience seriously as a healthcare CEO.
So, when I arrived at YC and learned that Michael Seibel already knew about Glass and cared about the future of our company, I was thrilled. I felt like I was in a startup dream when Michael and I walked circles around the Sonoma hotel’s ornate water fountain, arguing about our product roadmap. He believed our roadmap, which was focused on improvements to Glass Notebook, could not achieve our mission without significant change.
“This is not going to work if you don’t actually help doctors treat patients,” Michael said to me. I tried to defend our oblique approach to improving the practice of medicine through personal knowledge management, and he cut me off in frustration. “Dereck, tell me about the company that could actually disrupt healthcare.”
I was quiet as we continued to loop around the fountain, looking down at the ground and turning over his challenge in my mind. Michael broke the silence, telling me that he expected Glass to achieve the greatness of Uber, which was the biggest transportation company despite not owning any cars, or Airbnb, which had become the biggest hospitality company without owning hotels. He believed that Glass Health could become the biggest company in healthcare without owning any hospitals or clinics. He also told me he was ready to stop walking around the fountain and go inside.
We sat down in the hotel’s lobby, and Surbhi Sarna came over and stood above us, listening to our conversation. I admitted to Michael and Surbhi that though Graham and I truly wanted to build a clinical decision support (CDS) platform rather than knowledge management tools for doctors, we were apprehensive. I thought knowledge management was important, and it was true that our ultimate objective was to provide CDS—but was this the right time? I worried that pivoting the company into this arena at such an early stage could be risky, if not outright dangerous to our company’s survival.
Michael laughed and said, “Are you afraid that you’ll mess it up, that you won’t have enough money to execute, that you’ll move into a high-stakes use case, and that the company will die?”
I was silent, and he was ready to hit me with one of those company-changing Michael Seibelisms. “Hate to break it to you. But if you don’t solve a real problem for doctors, you’re going to die anyway.”
The Art of Medicine
“If you don’t solve a real problem for doctors, you’re going to die anyway.” You pretty much couldn’t say anything to me for the rest of the Y Combinator retreat, because those words were banging around so loudly in my head. I stopped going to most of the scheduled events and started spending time in the hotel room thinking about Glass and how to leverage the opportunities and advantages of YC to build a company that could transform medicine. How could we have an impact on patient care as such a small clinical team with so few resources? What product was needed to achieve our mission? What were the legal and regulatory implications of moving into CDS? Could we actually pivot into the higher stakes use cases around patient care? And if so, how would we even do it?
The answers to these questions were not immediately apparent, but we were committed to moving the company in a new direction and open to inspiration on what this new direction would be. Luckily, the inspiration was clamoring loudly all around us. As we contemplated our next move, ChatGPT and OpenAI were taking the startup world and the larger world by storm. Posts were going viral as doctors tried their hands at making ChatGPT work for the medicine use case.
Graham and I closely and quietly studied the medicine community’s posts and explorations into AI. Doctors were trying to use ChatGPT to do research, obtain prior authorizations for insurance, and create documentation for billing.
To be honest, I was disinterested in most of these use cases. I found the needs of the insurance companies, billing, and hospital administrators to be a distraction from the art of medicine. In other words, I hadn’t been in medicine long enough to give a damn about billing. Actually, I didn’t give a damn about anything other than improving patient care and achieving health equity, and I didn’t see how ChatGPT was going to help us with either of those parts of our mission. But both my medical student sister, Jennifer, who had studied nutrition in Blue Zones, and my medical student friend and investor, Morgan Cheatham, did.
Jennifer and Morgan had been repeatedly encouraging me to look at what they thought were emergent clinical reasoning abilities coming from ChatGPT, sending me screenshots of how they were using ChatGPT to learn and perform on clinical rotations. Morgan had even written what is now a landmark paper on the topic.
With their encouragement, I continued to play with ChatGPT during my daily medicine study in those early weeks of YC. Gurpreet had taught me to practice medicine as a martial art, so even on leave from my internal medicine residency, I studied medicine daily. During these study sessions, I made a very perplexing discovery about ChatGPT. Despite my substantial efforts and study, ChatGPT was a much better diagnostician than I was.
As a medical student, I’d told our Dean of the School of Medicine, Talmadge King, MD—who had pioneered our understanding and treatment of interstitial lung disease—that I wanted to become the best internist I could be. He, in turn, told me that I should read every New England Journal of Medicine Clinical Problem Solving case that came out and that we should schedule another meeting about “writing textbooks.”
Talmadge also told me that I needed to get a personal life and stop working so hard. He told me frankly that he believed I wouldn’t make it through a career in medicine without taking more vacations and getting a support system. I am still working on this.
In addition to picking up the practice of reading the New England Journal of Medicine cases in each issue at his encouragement, I became an avid listener of the diagnostic reasoning podcast, The Clinical Problem Solvers. I loved the CPSolvers, which was founded by Reza Manesh, MD and Rabih Geha, MD, two of Gurpreet’s closest mentees. I would often call the podcast founders to discuss how we should think about complex identities and social inequities in our clinical reasoning.
Eventually, they charged me and my good friend Utibe Essien, MD with creating a podcast series within The Clinical Problem Solvers podcast on antiracism and health equity in medicine and clinical reasoning. Reza and Rabih directed me and Utibe at that podcast series because they knew there was much to be untangled about how we should think about race, ethnicity, bias, and identity when it comes to the practice of medicine, improving patient outcomes, and accelerating the achievement of health equity.
All of this is to say that I thought about diagnosis and the implementation of evidence-based care day and night. Why was this maddening little R2D2 of a chatbot knocking out complex cases and the board exam questions I was feeding it?
It was even better than I was when equipped with my Glass Notebook, where I had stored dozens, going on hundreds, of approaches to undifferentiated clinical presentations and evidence-based treatment plans for various diagnoses.
Over time, I came to understand that through its generalized training on petabytes of text, OpenAI’s GPT models had seen many more cases than I had, and my human brain was unlikely to catch up, even if I did review cases for an hour a day.
I also came to realize that while researchers had been trying to train large language models (LLMs) on medical texts for decades by restricting the pre-training data to academic or clinical literature, they had never provided the models with enough language to gain the emergent reasoning and clinical abilities of ChatGPT. All of that seemingly irrelevant-to-medicine text that ChatGPT had seen on the internet was actually an essential ingredient in its performance in the medicine use case.
This is when the plan to pivot Glass Health to AI CDS began to materialize.
I began to wonder if, instead of having thousands of doctors individually create their own notebooks, we could have hundreds of the best doctors train a single LLM model to help millions of doctors practice medicine.
My partner, Niralee, who is a brilliant doctor and former executive director of an international non-profit, is training in pediatrics and child psychiatry at Brown. In addition to her incredible clinical skills and deep compassion for families and children, she has a background in theoretical math and was a key counsel and sounding board as ideas for the pivot developed. She and I traded my phone back and forth across tables at our favorite San Francisco restaurants so that she could look at my page-long LLM prompts that tried to get a GPT model to simulate the thought processes of master clinicians.
I saw great promise in harnessing foundational models for CDS, but there were serious challenges to work out with LLMs in the clinical-use case. There was limited determinism in outputs, and the models could not be trusted to provide factual information out of the box. They could not handle complex cases that weren’t represented in their pre-training. The LLMs were not HIPAA compliant and could not be integrated into the EHR or clinical workflows. They hadn’t been assessed for bias and had almost certainly picked up biases from their pre-training based on my and Graham’s early investigations. Their clinical reasoning was weak when multiple comorbidities were present.
When it came to suggesting treatment recommendations, limitations of the publicly available LLMs were even more serious, as black-box pre-training could lead to clinical plans that included outdated or even incorrect treatment recommendations. On top of this, I knew that if there was a Zika outbreak tomorrow, there would be no way to update the model instantly to consider an epidemiological shift in the prevalence of the disease or reflect new recommendations from the Centers for Disease Control (CDC), the National Institutes of Health (NIH), or the World Health Organization (WHO).
But there was something real here, and my doctor instincts told me that I would provide better care equipped with an LLM than without it. I was starting to believe that, as a medical community, we needed to go on a journey to understand how to utilize this technology properly. In order to meet our mission of doing the best by our patients, we had to figure out how to harness AI for the medicine-use case, and we should not reject, resist, or ignore it.
Cookie Cake
Throughout January 2023, Graham and I covertly explored a pivot to an LLM-based application for AI clinical decision support. We didn’t even tell our five-person clinical team, because we knew the product would be controversial in the medical community. The first version of our AI system was built by the two of us alone over many long days in the Russian Hill apartment I had borrowed from another YC founder. We looked out at the Golden Gate Bridge and, using our best guesses, built what we believed was a strong prototype for AI CDS. At the tail end of January, we launched a beta of our AI CDS feature, and it went viral with 100x the force of our notebook a year earlier.
Unfortunately, while our first AI launch was met with incredible excitement from doctors, it was also met with rampant criticism. Many responses to the launch were negative, including the tweet, “Okay, well, I think you’ve had this horrible tool up long enough.” That one still hurts. Others attacked me directly, putting patient scenarios into Glass like, “A 30-year-old with a history of building startups believes he can use AI to diagnose people” and posting screenshots of the less-than-perfect outputs that resulted.
It pained me deeply that the potential wasn’t seen by everyone, but it was seen by my loved ones and many, many physicians. Almost none of the attacks against Glass Health at our AI launch were from doctors but rather folks outside of medicine who had a hard time understanding why using AI in this way could be so powerful.
My family, who had through their deep love for me and belief in Glass, invested almost all their savings, supported me through my disappointment and frustration around the mixed reviews. They comforted me in my embarrassment about the limitations of the initial product and listened to me as I wondered if we had gone down the wrong path.
My sisters, Cathryn, Jennifer, and Elizabeth, and my brother-in-law, Greg, in particular, were there for me on the day of the launch as I read tweets attacking the company and held back my fury about the fact that our use case had been so misunderstood. Glass AI was meant to be an adjunct to clinicians, not replace them. It was meant to suggest a differential diagnosis and draft a clinical plan, not diagnose patients and decide on treatment. It was meant to be supervised like a medical student or trainee and never replace the judgment of an attending. No one had read the fine print or, for that matter, the very big print in our disclaimers.
My family baked me the now-famous Glass AI cookie cake, which indeed cheered me up and which I now have a picture of on my desk. I swallowed my frustration, and Graham and I screenshot every criticism about Glass that was posted on Twitter, put them in a initiative memo, and began to write a plan on how we would fix each deficiency in the product.
Will It Transform the World of Healthcare?
Through February, I continued to field skepticism about our vision for AI-powered CDS, including that from folks who were outright angry with me at the mere existence of our AI product. But as the weeks passed, glimmers of hope that the world would come around to our foray into AI CDS started to appear. One morning, I woke up to find that Glass Health had been featured during UCSF Medicine Grand Rounds: “ChatGPT: Will it Transform the World of Healthcare?”
Aaron Neinstein, MD, VP of Digital Health for UCSF, had demoed Glass for the Department of Medicine, saying, “You’re sitting down, you’ve got your one-liner in the box there, and along comes the GPT model, and it gives you a differential diagnosis. It tees up your diagnostic and treatment plan. Of course, again, the accuracy here is not perfect, and this would require a lot of editing. But if this is starting the work for you, you can think about all of the effort that it can potentially save.”
I was ecstatic. Aaron understood exactly what we at Glass Health were trying to do with AI CDS.
Knowing that some of the world’s best clinicians at UCSF saw our potential if we could continue to break through the limitations of LLMs in the clinical environment drove my energy. I called my good friend, master educator, master neurologist, previous Vice Chancellor of UCSF, and now professor emeritus, Dan Lowenstein, MD, who had also presented during the grand rounds.
Dan and I had met many times in his office at UCSF over the previous years. Dan’s office was my favorite place in all the medical school. It was filled with art, vibrant color, ancient and new books, and his strong and gentle spirit. When I called him after the ChatGPT Grand Rounds, where he, in his wisdom, had focused on protecting the humanism in medicine from AI, I worried that he would try to discourage me from our quest to pivot into AI CDS.
Together we discussed a recent article about the limitations of LLMs that had been written by Noam Chomsky earlier in the week. He could tell I was worried about the challenges ahead, but he didn’t discourage as I feared he might. He told me that he had practiced medicine through the arrival of the internet and the cloud and seen the ways doctors had been scared to move from paper charts to EMRs, or move knowledge from textbooks to the mighty CDS companies that had dominated the 90s. He told me that doctors and the world would come around.
At this point, the Glass Health Team was all in on AI CDS, but we knew that the first version of our AI system wouldn’t be enough to meet our mission of empowering doctors. Glass could not rely on the black box pre-training and prompt engineering using the GPT model APIs alone to generate outputs; we needed to figure out how to reach a level of determinism that matched that of the static CDS tools—but the LLMs were new and moving extremely fast, and the way forward was not clear. That is, until the universe saw fit to send another bolus of startup fairy dust to help us on our way.
Only Doctor in the Room
As Y Combinator (YC) founders, Graham and I were able to leverage the strong relationship between YC and OpenAI to learn what would be possible for founders who were building AI application layer companies for their industries. As part of the YC batch, we were given opportunities to visit OpenAI headquarters, but more importantly, the opportunity to learn from fellow YC founders who were trying to harness AI for their own use cases.
Through these relationships, the cutting edge of what was possible was constantly in our view as movements to develop and harness AI surged through Silicon Valley. I distinctly remember the first time I saw a demo of an AI using retrieval-augmented generation (RAG). It was an application that used the OpenAI APIs and a RAG system to answer questions about YC and startups.
I looked around the auditorium and thought, “Good god. I’m pretty sure I am the only doctor in this room right now.”
The implications of RAG for the medicine use case were immediately obvious, and several ideas were banging on the door of my mind. I typed a message furiously to Graham, who was sitting next to me. If RAG could be done with articles about YC, it could be done with our Glass Notebook pages.
We could build an AI system the architecture of which was the reverse of our notebook product. Instead of individual physician users building external brains in Glass for themselves, Glass would connect a single high-quality external brain that would be created and curated by the best and brightest doctors among us. This would create an AI system that leveraged RAG of physician-validated context to assist doctors.
Graham and I discussed how big a project it would be to equip an LLM with physician-validated context via RAG and incorporate physician fine-tuning, N-shot learning, and reinforcement learning to achieve a level of clinical excellence that would empower clinicians worldwide. The plan was ambitious, and having only the theoretical understanding that it should work, moving forward and devoting resources to it was risky. Graham prototyped a first version using prompts I had created and sent it to me while I’d taken a weekend away from the high intensity of YC to hike in Yosemite, to clear my head, to hear whatever it was that universe wanted to me know about AI architecture, and to spend time with Niralee.
Graham never really took a break during YC, or at least not one I would consider a break. From a cabin outside of Yosemite, Niralee and I watched an AI-generated clinical plan draft derived from a Glass Notebook page come streaming down the screen, flanked on either side by a wall of Graham’s code.
Niralee and I were floored by how insanely great the clinical plan was. The plan for Glass 2.0, which would incorporate RAG to overcome the limitations of Glass 1.0, was going to work.
Our founding clinical team is made up of some of the most brilliant, kind, and hardworking people in the world. Glass Health never would have gotten the traction it did without this team, and Graham and I will be grateful to them forever. One thing you should know about Glass is that we only hire doctors that one of us would trust to take care of our families, and those who we’d want to consult overnight from the ICU. That means being anything less than kind to medical students is pretty much an automatic out at Glass. At Glass Health, we select for compassion and authenticity and dedication to our mission, as much as we select for intellectual ability, clinical prowess, and the ability to accomplish a mission at hand.
Graham and I presented our plan for 2.0 to the team, and they immediately got to work on execution, following us bravely into the pivot. They began to generate physician-validated context for high-yield topics that we had previously studied to create our notebook Community Library pages.
Soon though, it became clear that together with our original team of five, we would not be able to ship 2.0 on our own. So, I shifted to all-out recruiting mode to expand our team, bringing on physicians and leaders who are now critical to Glass Health. There were many core team members who joined us at this inflection point, but I want to send special recognition to Alex Spacht, MD, Amita Nawathe, MD, Evonne McArthur, MD, PhD, Maxwell Coll, MD, Jacob Kirkpatrick, DO, Ryan O’Keefe, MD, Alex Lupi, MD, Annika Belzer, MD, Braydon Dymm, MD, Breanna Hansen, MD, Gurbani Kaur, MD, Chelsea Tweneboah, MD, Emefah Loccah, MD, Meera Sury, MD, MTS, Catherine Wang, MD, Emma Levine, MD, Breanna Hansen, MD, Noah Rosenberg, MD, Priyanka Athavale, MD, Chris Rees, MD, Gurbani Kaur, MD, Isabel Sunshine, MD, Kody Harvey, MD, Justin Morrison, MD Candidate, Shivani Reddy, MD Candidate, Hina Iqbal, MD Candidate, and Zachary Cross, MD Candidate. These are the clinicians who put the most time and effort into building version 2.0 of our model and its subsequent updates, executed our pivot to AI CDS, and changed the trajectory of Glass Health.
On March 23, 2023, we launched Glass AI 2.0. This time, we heard none of the detractors and criticism that had come our way with Glass 1.0. Our combined strategies of prompt engineering, N-shot learning, and AI architecture customized for RAG with physician-validated context proved to be exceptionally powerful. There were still many pieces of the puzzle to work out, but we were onto something, and the world knew it.
We were achieving clinical excellence in AI using the same methods that we use to achieve clinical excellence in ourselves, the insight being that physicians are, in part, LLMs who do N-shot learning. The only difference is that medicine calls N-shot learning “see one, do one, teach one.”
Some Companies Claim AI Can Help
Unfortunately, just when everything seemed to be going right for Glass Health, the startup world started to fall apart around us. In March of 2023, we learned just how dynamic and perilous the startup journey could be. In the midst of our Y Combinator batch, Silicon Valley Bank, the bank used by most of our fellow startups, collapsed, threatening to wipe out all the money they had raised and shut down their companies. SVB’s implosion sent shockwaves through the industry and ground startup fundraising to a halt.
Glass Health had our money from the pre-seed in another bank, but it was a small one, and I thought there was a chance we’d be affected by a domino bank failure. I knew that the SVB situation was an existential risk to our company and team. I called Surbhi to tell her about my worries, not just about the safety of our resources but also that it seemed like our chances of raising a seed round were dampened, if not ended, by the macro conditions in the Valley. This was the part where we were eating glass and staring into the abyss.
Surbhi was busy helping the other companies who were directly affected save their teams and their organizations. She told me that she trusted me to do what I thought was best to protect Glass and achieve our mission.
Luckily, the industry instability abated, and in March and April 2023, Glass Health continued to grow at an astonishing pace as tens of thousands of doctors signed up, and the world continued to carefully watch our endeavor to make LLMs work for medicine. Skepticism remained about our mission and product, but the Glass Health Team knew that once again, we had captured lightning in the bottle with Glass AI 2.0 and moved forward with intense determination.
At YC’s end-of-batch Demo Day, we drew inbound interest from more than 80 institutional venture capital firms. Many passed on Glass, believing that focusing on diagnosis and treatment plans was not a viable path to success. Most didn’t understand the deep dedication we already had at Glass to ensure AI CDS was done safely, with the highest level of clinical excellence, and with a commitment to accelerate health equity.
But other investors, especially the best ones, were intrigued and moved forward in conversations with us. On the day we graduated from YC, we were covered by NPR in their article “Doctors are Drowning in Paperwork - Some Companies Claim AI Can Help.” The other healthcare founders at YC thought that I had somehow coordinated the article release with demo day to juice the Glass Health seed round. This certainly was not true. I did not know the article was coming out that morning. But it helped us, for sure, and I may have conveniently forgotten to correct the rumors that Glass Health had such PR power and prowess that we could steal the show on demo day with an NPR article until this very moment.
The YC batch was over. We had made it through, and our seed round was coming together. Surbhi Sarna told me that the YC leaders had discussed who the “decacorns”— companies that would one day be worth 10 billion dollars—were in the batch. Glass was her pick.
The Best by Far
In May and June of 2023, Glass Health was inundated with requests from individual doctors, private practices, community hospitals, tertiary care centers, and large health systems who wanted to know how they could integrate Glass into their workflows. The tides had shifted, and it was no longer in vogue to be a detractor about the use of AI in medicine. The world was seeing that a fundamental technology inflection point was happening, one that was as important as the computer, the internet, the cloud, or mobile. The AI revolution meant that new companies would be born, and for medicine, it was clear that doctors were going to use AI whether health systems figured out their strategy to incorporate it or not.
Glass was spreading like wildfire throughout the medicine community and getting rave reviews. No one had been able to catch us or our product architecture, our tempo was fast, and every week we continued to outdo whatever we had done the week before. I smiled reading the words of Dr. Anthony Chang, pediatric cardiologist and Chief Intelligence and Innovation Officer at Children’s Hospital of Orange County, who was quoted in Becker saying, “Glass is really best in terms of a clinician looking for a differential diagnosis in the clinic setting…when it comes to a difficult case in the clinic, Glass Health has been the best by far.” In the months that followed, Anthony became a friend and collaborator, championing the need for pediatric-specific context development and fine-tuning at Glass Health.
AI CDS for Health Systems
In July and August, we had no fancy releases. While Glass AI 2.0 was incredible, and our cadre of power users was growing, we saw major challenges to the widespread adoption of our product.
While early adopters of our AI CDS loved Glass and some had it open on a second monitor next to their EMR for the entire day, the early and late majority of users on the user adoption curve for our product were not going to adopt AI CDS unless our product was better designed for their workflows. We needed to make major changes to our product so that it supported clinical and inpatient workflows or our growth would flatten out.
During these months, we urgently prototyped an EHR-integrated version of Glass that could meet the needs of health systems CIOs who were bold enough to empower their clinicians with AI CDS. Our prototype was built by a special small team at Glass Health that included our fearless Enterprise and Partnerships Lead, Amita Nawathe, MD and our brilliant AI Associate Managing Editors, Alex Spacht, MD, Madellena Conte, MD Candidate, and Evonne McArthur, MD, PhD.
When we demoed Glass for one of the largest integrated delivery health systems on the East Coast, I was uncertain about what the outcome would be. Graham and I had been banged up by the way health systems had turned us away one after another in 2021. So, when after the demo the health system asked for some time to consider our pilot proposal, I expected that another “no” was coming.
The day after the demo, the head of the health system, who was himself a physician and specialist, asked me to meet him for dinner. We talked together about the challenges his health system faced and the many ways Glass would be positioned to help in the coming years. I realized that I was talking to a rare visionary health-system leader who had put other rare visionary executives in place at his organization to usher in the future of healthcare. After sharing our thoughts with each other about where we saw the healthcare delivery and healthcare technology sectors going, he told me that he had used Glass to draft a plan for a patient that morning, and that he’d been stunned at how good the draft was. There were many details he wanted to work out, but he put out his hand across the table.
We had our first health system deal.
Transform Our Healthcare System for the Better
In September, we began to work furiously on our plans for Glass 3.0, which will launch in the first quarter of 2024, and our EHR-integrated enterprise offering, which will launch later in the year. We upgraded our product design, sharpened our regulatory strategy, and built new internal tools that allow physicians at Glass Health to architect AI systems and engineer AI prompts. And meanwhile, the world has continued to talk about the work our incredible team is doing to achieve our mission at Glass Health.
We were featured in TechCrunch in their article, “Glass Health is building an AI for suggesting medical diagnoses,” which highlighted the huge potential our products have to help decrease diagnostic errors and save lives. We then publicly announced that earlier in the year, we had raised a pivotal $5M seed round led by Initialized Capital, joined by some of the greatest CEOs and founders in healthcare, including Tom X. Lee of One Medical, Heather Hasson and Trina Spear of FIGs, Connor Landgraf of Eko, Othman Laraki of Color Genomics, as well as employees and investors from OpenAI and other technology companies.
By the time we raised the seed round, it was obvious to us and our investors that Glass Health was a remarkably valuable company. So, we only took investment from Initialized, health systems, and CEOs and operators who had specific expertise that would help us to achieve our mission. Parul Singh from Initialized Capital announced our round in the press saying, “Glass Health is leading the way on how artificial intelligence can be leveraged to augment physicians, create unprecedented value, and transform our healthcare system for the better.”
AI Is about to Completely Change the Way You Use Computers
In November, we stayed laser-focused on building our soon-to-be-released newest version of Glass, Glass 3.0, and our EHR-integrated enterprise offering. The best clinicians in the world were joining our team, and our ranks grew to a mighty team of 150 of the most brilliant and innovative clinicians in the world, working part-time, as they might at a journal or for a traditional CDS company. The only difference is that at Glass Health, instead of editing static encyclopedia pages, doctors edit AI prompts, create, peer-review, and validate AI context, and curate datasets for AI fine-tuning and N-shot learning. Our AI Clinical Fellows also join sub-teams dedicated to parts of our larger mission of empowering clinicians with the best-in-the-world AI CDS, including our Enterprise, Community, Health Equity, Clinical Informatics, Operations, and Clinical Validation teams.
Glass is a remarkably exciting, unbelievably fun, and joyful endeavor, and both medicine and the larger world are watching. As a week of high-tempo context generation at Glass Health came to an end, one of our AI Clinical Fellows called out that we had been mentioned by Bill Gates in his recent article: “AI is about to completely change the way you use computers. And upend the software industry.”
The article had come out the week before, and none of us had noticed. I showed my mom the article, and she was so excited that she took my laptop and went outside to wave my dad down off the lawnmower. He came over to take a look, smiled, and said, “Now, this is pretty cool.”
Did I mention that impressing your first-generation immigrant parents is extremely hard?
Reflection, Rest, and Readiness for 2024
December has been a month of reflection, rest, and readiness at Glass Health. We have been looking back and celebrating all the progress we’ve made toward our mission. We’ve been carefully analyzing what barriers lie between us and achieving our mission. We’re planning bold action at the individual, initiative, team, and company levels to ensure we meet our mission of empowering clinicians with the best-in-class AI CDS in 2024.
We will be successful as a company only when we can show that we improve diagnostic accuracy, increase the implementation of evidence-based medicine, accelerate the achievement of health equity, and improve patient outcomes worldwide.
Even when we achieve these goals, because of our commitment to being best-in-class, our evergreen mission will not be complete. We’ll have to continue to innovate and deliver to ensure that technology is leveraged to optimize the practice of medicine, which will require different products and different initiatives across the coming years and decades.
There is no way to reflect on 2023 without thanking Y Combinator, and particularly Surbhi Sarna for her mentorship, Breyer Capital and our good friend Morgan Cheatham for his advice and support, and Initialized Capital, especially Parul Singh for believing in us and leading our seed round.
I am personally grateful to our friends Tom X. Lee, Heather Hasson, Trina Spear, and Connor Landgraf for their mentorship. Thank you to Julian Genkins, MD, our long-time friend and clinical informatics advisor and co-designer of our original notebook product for your support and counsel. Thank you Reza Manesh, MD, and Rabih Geha, MD, for your friendship, support, diagnostic reasoning advising, and expertise.
Glass Health Leadership Team, I will never be able to thank you enough for what you’ve done for our company this year. Thank you Alex Spacht MD, Andrew Sanchez MD, Gurleen Kaur, MD, Madellena Conte, MD Candidate, and Evonne McArthur, MD, PhD, our AI Associate Managing Editors. Thank you to our AI Section Editors Maxwell Coll, MD—who also serves as Cardiology Editor and Clinical Validation Lead—Braydon Dymm, MD, our Neurology Editor, and Michael Fishman, MD, our Deputy Section Editor for Pediatrics. Thank you to Amita Nawathe, MD, for leading one of our most mission-critical teams and initiatives this year as our Enterprise and Partnership Lead. Each of you is carrying an important and precious part of our mission. I know that in addition to being leaders at Glass, you are friends, spouses, parents, and doctors to patients. I know that you make sacrifices in your personal lives for our mission. I am determined to ensure that our work together will pay off for medicine and for your families.
Thank you to several recent additions to our team who have come onto the field with enthusiasm and extraordinary skill to help us achieve our mission as part of a newly established Office of the CEO at Glass Health—Scott Giberson, MD, MBA Candidate, Jane Chuprin, MD, PhD Candidate, Alex Lupi, MD, Olivia Reszczynski, MD Candidate, Katie McLaughlin, JD, MD Candidate, and Jermaine Blakely, MPA, MD Candidate. Though we have only worked together a short time, I am quite sure that each of you will be exceptional founders, CEOs, or executives, and you make up the strongest and most effective Office of the CEO in healthcare technology at this time. Thank you for everything you do to support me.
Special thanks go out to my family and Graham’s family and friends for supporting us in 2023. Thank you to Graham’s family and his brother Liam Ramsey, who designed the original Glass Health logo, and my best friend and partner, Niralee Shah, MD, MSc, who has supported me tremendously and, through me, Glass Health, in 2023. Thank you to my sisters, Jennifer Paul-Quinn, MD Candidate, Elizabeth Paul, Cathryn Paul, and Shakkaura Kemet-Rosenwhol, MD, my brother-in-law Greg Paul-Quinn, my brilliant and perfect 18-month-old nephew, Theo, my future brother-in-law Adam Jackson, my close friend Matthew Ryan, MD, PhD Candidate, and Niralee’s family, for your love and support.
Thank you to my strong, supportive, and loving parents, Dereck L. Paul and Sharlene J. Paul, who were among the first believers and investors in Glass Health.
Thank you to my extended family of wonderful aunts, uncles, and cousins for your support, and thank you to my grandfather William A. Paul, who funded my first attempt at a company as a premedical student with $4,000. He wrote the check from the living room of his house in BedStuy, Brooklyn, and reminded me that before coming to the United States from Guyana, in the village he had grown up in, the career options were harvesting coconuts, harvesting mangoes, or laboring to build the local railroad. He emphasized that it was the hard work of our family, a dedication to education, and the bravery needed to immigrate to a new country that created the opportunities that I have been so fortunate to pursue.
That my family, loved ones, and friends are active members of the medicine and healthcare community as either providers or patients is the reason for my insistence on intensity and exigency in our execution to achieve our mission.
Clinicians, I care deeply about you and the quality of your lives as you do the superhuman feat of practicing medicine in the modern age. I know that you will not stop in your quest to achieve the best possible outcomes for every patient, whatever that might mean for each individual patient’s life and story. And I know that we must empower you with the tools to meet that commitment to patient care and protect you from burning out for yourselves, your loved ones, and your patients.
Before ending, I must send my deepest thanks and most sincere gratitude to Graham Ramsey, our cofounder and Head of Product, for his herculean efforts and masterful delivery of the Glass Health product roadmap in 2023. Graham, you are a once-in-a-generation product leader, and I’m extraordinarily lucky to have you as a friend and cofounder.
Thank you, Glass Health Team and Glass Health Community. I am sending my appreciation to each and every one of you for your support in 2023. This year, you built a healthcare company that is on track to become as impactful as Epic, founded in 1979, or UpToDate founded in 1992. Happy New Year. See you in 2024.
Cheers,
Dereck