The end of the physician era, e/acc, AI & more
DrGupta.AI – The case for Healthcare AI
Today I’m unveiling DrGupta.ai. Dr. Gupta is a virtual healthcare assistant that can help you access health information and simulate interactions with a physician. Our aim is to make healthcare more accessible and convenient for everyone.
In the past, I faced criticism for raising the price of medicines, and I became a focal point for debates on pharmaceutical pricing. That surreal to me: pharmaceuticals only account for 12% of healthcare costs in the US, but they receive most of the blame.
Most healthcare costs are related to physicians, and all healthcare costs stem from physician decisions! This raises important questions: Why are these costs so high, and what can we do to address them? As we move forward, it’s crucial that we explore innovative solutions to make healthcare more affordable and accessible for all.
My central thesis is: Healthcare is more expensive than we’d like mostly because of the artificially constrained supply of healthcare professionals. Unfortunately, it is not feasible to instantly produce a million more physicians for practical reasons. A working labor market for healthcare workers is hindered by the efforts of lobbyists and bureaucrats who artificially maintain a limited supply. Although the importance of proper healthcare training cannot be overstated, the escalating financial burden of Medicare and Medicaid is pushing our nation towards bankruptcy.
Artificial intelligence holds the potential to significantly increase productivity in numerous industries, including the one I work in. My team and I rely on AI to program more efficiently, but it’s disheartening to see that healthcare has yet to fully benefit from technological advancements. Despite the logarithmic drop in the price of computation, storage and communications, physicians seem to be stretched thin, while healthcare costs continue to rise, leading to a reverse in productivity.
Although there have been significant technological advancements in recent years, it is surprising to see that the healthcare industry has remained impervious to innovation and efficiency. Even with these advancements, the number of FDA approvals for medicines has remained stagnant since the organization's establishment, and the number of doctors per capita has not seen any significant increase.
The root of the problem lies with people, as it is difficult, if not impossible, to change human limitations. For example, a doctor can only attend to one patient at a time, healthcare workers can become exhausted, and language and financial barriers exacerbate inequalities. Moreover, human errors, which can sometimes be fatal, are unavoidable.
While we should express our gratitude to every healthcare worker for their hard work, we should also question whether we can improve the current situation. Is there a way to overcome these limitations and make the healthcare industry more efficient?
Enter AI. ChatGPT-4 has demonstrated an impressive performance on the USMLE, leading us to believe that Dr. Gupta will soon provide advice on par with the world's top medical specialists. Discussions I’ve had with physicians remind me that it is very difficult for even “key opinion leaders” to stay abreast of developments in their field. Between distant medical conferences and endless publications, it is very difficult to stay on top of your specialty and still see patients. The average doctor I know barely has the time to read the New England Journal every week. Computers, of course, have no restrictions on processing voluminous data in real-time.
I envision a future where our children ask what physicians were like and why society ever needed them. Although Dr. Gupta cannot currently prescribe medication or perform surgical procedures, our team at DL Software (the parent company of Dr. Gupta, Druglike, and the forthcoming financial terminal product Godel) is committed to exploring this future. With our continued efforts, perhaps this dream will become a reality someday.
Dr. Gupta may not be ready to replace physicians entirely, but I believe it has the potential to be a valuable resource for many individuals. Previously, seeking medical information online proved to be a challenging endeavor. However, with advanced language models, individuals worldwide can now engage in authentic conversations with chat agents, regardless of time or language barriers. This offers those with limited healthcare knowledge, as well as physicians seeking a second opinion, the opportunity to access easy-to-understand healthcare information. Even experts like me have found this clarity difficult to come by in the past.
Dr. Gupta is on the path to improvement and will soon be even better. While we've rapidly launched this product, our roadmap includes the development of a truly distinctive application that enables users to learn about and manage their health in a comprehensive manner. As a first step, we will implement pushing personalized health information directly to you, providing relevant news as it occurs. Dr. Gupta is adept at interpreting medical papers, textbooks, or news articles and presenting them in an accessible and meaningful way tailored to your needs.
I am personally inspired by the potential to provide medical assistance to individuals in remote parts of the world where access to physicians is severely limited. I remember my parents recounting stories about rural Albania, where seeing a doctor was a near-impossible task, and the quality of care was often questionable. Reflecting on the countless individuals globally who are not as fortunate as those living in the United States, I am grateful that our product can offer them medical guidance for just a fraction of a cent. This is the driving force behind making initial use of Dr. Gupta free for anyone, anytime. However, for extensive usage, payment is required. Try it at www.drgupta.ai.
Over the last decade, I’ve learned and reflected a lot on healthcare. One crucial lesson I've learned is that people tend to approach healthcare with a more emotional rather than rational mindset. Health is undoubtedly the most critical aspect of human life, and any potential threat to our well-being, no matter how small, can be alarming. Even the perception of a threat can lead to misunderstandings and concerns.
As someone passionate about improving healthcare efficiency, I believe that the industry can greatly benefit from the advancements made in hardware and software in recent years. Whether it is through automation or artificial intelligence, I am convinced that a significant portion of healthcare information requests and decisions can be streamlined and improved. Given the enormous size of the global healthcare market, even a modest improvement in efficiency could have a substantial impact.
Dr. Gupta uses a combination of GPT-3.5 and GPT-4. We have trained our own LLM which we are testing. Last year I stumbled upon the reflection and “auto” recursion-based methods many are employing under various names. For instance, I had ChatGPT write a 200 page book and have embarked on a number of other toy projects, most of which I believe are ahead of the disclosed SOTA. Dr. Gupta uses such tools to create, in my opinion, better performance than Chat GPT alone. Try and compare it yourself. This makes the speed of Gupta a tad bit slower, but not a lot more. GPT-4 text trickles out slowly while Gupta’s comes in one bolus, so there are some optical illusions there. Hopefully, GPT-4 speed will increase soon. We’d like to increase the amount of introspection Gupta does.
Investors & Cap Table Fever
DL Software is funded by investors. In attempting to earnestly preview Gupta to the media, I rediscovered my hatred for journalists and their editors. There was a perplexingly deep interest in who our investors are by the standard fare: The Information and Business Insider. One intrepid reporter was able to track down both investors we talked to and didn’t in, hunting down every investor. He could probably put the cap table together better than our own accountants!
The identity of your investors is something many startups tend to dwell quite a bit on. But this is a classic mistake. Spoiler alert: Who your investors are does not matter and never has. Can you tell me who was the first investor in Microsoft? Did you know the average VC loses money? The only thing that matters to me is how much I can motivate my team to push itself as hard as humanly possible to create amazing technology that makes our customers more happy and productive. An investment from Warren Buffett (or insert whatever name you want here) himself would not ensure, or even improve, the odds of success of our company. It is the height of stupidity to focus on this. When I raised $90 million for Turing Pharmaceuticals, a huge amount in an era before today’s mega-rounds, I didn’t disclose any investors. There’s no embarrassment, unwillingness by them or string to pull on to unravel a conspiracy. It just doesn’t matter.
To other startups out there: Imagine disclosing every idiot that wrote a $25,000 check for your company. Are these names supposed to give people comfort? Most of you will fail. The invitation list to your funeral is not something to boast about. The few who will make it will not have made it because their venture fund steered them to success. If that were the case, the venture fund would return 5,000x with every portfolio company. Focus on your job and eliminate any other noise.
I’ve become more active with the accelerationist movement on Twitter (twitter.com/marty_catboy). Accelerationism means different things to different people. To me, it means allowing technology to bring its benefits to the masses without questions and regulations. The progress of technology is impossible to stop once it has started and people will not stand for its interruption when it benefits their lives.
As artificial intelligence, “blockchain” and other technologies reach increasingly exponential heights, their benefits democratize power. This worries current power structures. We must make it clear we do not want the SEC’s “protection” from crypto any more than we want NTSB’s “protection” from self-driving cars.
Regulation of AI is a cripplingly stupid idea that will distribute American hegemony. Maybe that’s the goal. The irony of American (and Canadian) ingenuity being exploited by non-North Americans to raise their standard of living to, and above, ours, would be disappointing. Take a stand now, before someone tells you what software and hardware you are permitted to have.
How do we get AI to plan? This seems like the subject of the month. Asking Chat GPT to articulate a plan as mundane as ‘clean a house’ or grand as ‘solve global warming’ doesn’t help too much. You get a reasonable plan that you can store in a variable. The problem becomes operationalizing that plan. It’s not hard to ask the computer to performStep(plan[step]). So, if the first step is ‘write a provocative essay on why global warming must be halted immediately’, it isn’t awfully hard to get that essay written. You can make the plan more complex by atomizing it as much as possible. So why don’t these tools work?
Planning is not useful without beliefs and crystallized learning. And probably a lot of other stuff I am clueless about. Repeated introspection probably helps in every step, but what are we judging on? It has been hard enough for me to guide AI through a very narrow task: write a cohesive fiction story. It has been awfully difficult to do so at scale, regardless of quality. LLMs are stochastic parrots with some emergent behavior, which while tantalizing, does not quite substitute for the real thing.
I think the Broca’s area equivalent of the LLM requires additional equivalent brain anatomy modules to get real utility. While I don’t think this will lead to an AI winter, I suspect the rapid desire to push out autoGPT and its ilk do not help. As many know, I am working on HumE, a 2D-RPG substrate to enhance the emergence of AGI within an abstract/finite but sufficiently complex world. There is some interesting emergence of unique behavior, but we’re not quite there. https://reverie.herokuapp.com/arXiv_Demo/ is a similar attempt, but is not a live server, so we don’t yet know how contrived this example is. I wonder if we should lower our goals even further and try to get a good mimic of a dog, cat or small child instead.
I think it’s worth reading this interview with John Carmack who is investigating AGI: https://dallasinnovates.com/exclusive-qa-john-carmacks-different-path-to-artificial-general-intelligence/. I agree that AI’s past may shed a lot of light on the path forward to round out the tools needed to approach AGI/ASI. I’ll have a lot more to say on this as my research progresses.