Oscar Q1 numbers are proof that it really is possible to build a profitable healthcare business ever more beloved by members: 3.2M members, $4.6B revenue, $704M earnings from operations in Q1 '26. But could YOU have done it? Now you can test yourself... 1/2

.@getpeid Massive fan!

Oh you're going with "technically no police died on exactly that day" when 1) getting beaten up with a flagpole on the steps of the Capitol isn't really all that great either and b) even your search screenshot below speaks to the 5 policemen who were affected right after. Sure

Must be lovely to be a guy like you w/o any real responsibilities beyond yelling performatively at real CEOs & builders in earnings calls. You think any of us in the real world have time to wait around for any of that? If I even organized my kids' birthday parties this unseriously, I'd be fired.

Life would be better if we all adopted this as Peterson's 13th rule for life: "Peterson, you idiot, stay out of this"

This is a very well-known consequence of the fact that the non-farm payrolls report comes from a different survey than the unemployment report. Hedge funds, traders, economists know this. It says that on page 1: "This news release presents statistics from two monthly surveys."

One of the insane knowledge management issues in an organization is that the longer you've been there, the more Google Drive document access you have. If I tell newbies at Oscar to search the g drive for something, they get few hits. Weird superpower of veterans

I still think that to be a real company, you have to be a public company. But I can confirm that losing 94% of your value is as soul-crushing as it seems (not to mention for our investors). I always felt confident that we had all the bones to get the company on track, but when you are plumbing your lowest stock price lows, you wonder: what even are the historical odds that a stock comes back from a 94% peak-to-trough drawdown, like the one OSCR experienced? Here is the answer: in a universe of 3,000 publicly traded companies with a trough valuation of at least $50M, there were 300 companies over the past 10 years that experienced a drawdown of 94% or greater and are still trading. Another 60-80 got delisted. 40 of those ~400 companies (including OSCR) had a rebound since the trough of at least 8x. So that's ~10% odds for such a comeback at the trough. Of those 40 companies, 55% are oil companies, 20% are biotech, and 10% are bitcoin miners. To be clear, we are not even close to done, and have to keep earning our way back up the top. We still have a lot of work to do. But if you can count, then you better never count us out. (GPT-coproduced numbers here if you want to follow along: docs.google.com/spreadsheets…)

The best way to name rooms is "15 Min", "An Hour", "Next Year", because then people can say things like "let's meet in 15 Min" and "let's use Next Year", and meetings are fun again

Oscar has joined the Fortune 500 for the first time. I know what you're thinking, and I agree: what, only now? But because vibe-analyzing now takes < 5 minutes, here are some stats: - Only newcomer in the healthcare industry category named to the list this year - Only company on the list founded < 10 years ago, one of 6 companies founded < 15 years ago. Making the list 13 years after founding puts us in the top decile in terms of speed. In healthcare, only 2 companies on the list are < 20 years old - 50%+ revenue growth rate vs. 4% median for healthcare companies on the list - We have 5x revenue-per-employee vs. the list-wide median - This last one is only 85% true for us (11/13 yrs) but close enough: 4.8% of Fortune 500 CEOs are founders of their companies On the other hand: the only market cap-related permitted reason for celebration will be the Fortune-50 @JoshuaKushner fortune.com/ranking/fortune5…

We @OscarHealth are publishing our use cases, prototypes, ideas & research notes on using LLMs in healthcare. AI research moves too fast for companies to do this work behind closed doors, so we’re sharing ours for others to read & comment: oscarai.notion.site. 1/

Yesterday, OpenAI dropped a preview of a new model, o1. Our initial observations are that the o1-preview model is more autonomous in generating steps to solve problems, more precise, and handles complex tasks with higher consistency. The more digestible formatting of outputs and its increased sense of ‘confidence’ relative to 4o help too. o1-preview will unlock new use cases & accelerate time to deployment for existing use cases in healthcare. Here are several healthcare examples. 1/6

Any natural phenomenon where repeated independent tries occur with same probability creates a normal distribution. Best way to visualize that physically: the Galton Board or quincunx - balls falling from the top down sequential left/right junctions

Oscar Health, for the quarter ended March 31, 2024: - Total Revenue of $2.1 billion, a 46% increase YoY - Medical Loss Ratio of 74.2%, a 210 bps improvement YoY - SG&A Expense Ratio of 18.4%, an 870 bps improvement YoY - Net Income of $177.4 million, or $0.62 of earnings per share, a $217.1 million improvement YoY - Adjusted EBITDA of $219.3 million, a $168.2 million improvement YoY - Reaffirming full year 2024 outlook => Individualizing health insurance with tech works. Just took a little longer than we naively thought

The individual market is the healthcare market of the future, and Oscar is building it. To that end, we are raising additional financing. We are thrilled with the support from old and new investors. This capital positions Oscar for long-term growth. We have many more products and people to reach, and the member experience and technology to make it happen. Let's go. ir.hioscar.com/news-events-p…

This is exactly right. @vkhosla and @SamirKaul1 invested in Oscar 10 years ago, and Vinod still shows up with a day notice at the office to quiz us on when AI is finally going to reduce our healthcare costs by 90% & offers to brainstorm. Legend. Thanks! theinformation.com/articles/…

For the past month or so, a lot of the most complicated @OscarHealth member questions about how the healthcare system works have been answered by @OpenAI models (with a final coding push that came from GPT-5). Some stats and technical details: 1/5

Yes, this all leaked out long before the articles. When Oscar Health started to expand, we went to see Elizabeth in early 2015. (The security circus around her even in her own office was a sight to behold.) She was incredibly gracious and took copious notes. But we wanted a lab test network and testing capacity from her, and she kept trying to convince us to rent her beautiful website to just visualize Quest and Labcorp lab results, instead of testing with her own tech. It was strange. Nevertheless, the week after back in NY, I excitedly told the company at our weekly all hands that we're talking to her. One of our engineers took me aside afterwards, insisted on going into a quiet conference room, and told me: my friend works there and says it's all bullshit. This was months before the story broke in the WSJ.

1) Interview is in German (which I speak) & you're badly mangling his words: 1) frontal massed attacks didn't work, 2) Ukraine already switched & is having success now, 3) more long-range weapons will significantly tilt balance. Also, Reisner is a favorite to quote on Russian TV

Google draws some low-stakes pictures wrong and the stock crashes; United crashes half the healthcare system and the stock is up. Couldn't find a better illustration of how little product performance and competitiveness matter in healthcare vs all other industries

Surprised @DavidSacks hasn't congratulated Putin to his hard-earned election victory. What a lack of civility

Large-scale AI models are a once-in-a-generation opportunity to improve healthcare. 28 of the most forward-thinking payers and providers got together to figure out how we leverage frontier AI models to drive the change we want to see in healthcare. Here are our commitments: @AllinaHealth @BassettNetwork @BostonChildrens @CuraiHQ @CVSHealth @DevotedHealth @DukeHealth @emoryhealthcare @EndeavorWNY @MHealthFairview @GeisingerHealth @HMHNewJersey @Health_First @MethodistHosp @johnmuirhealth @KeckMedicineUSC @mainlinehealth @MassGenBrigham @MedUnivSC @OscarHealth @OSFHealthCare @premera @RushMedical @SanfordHealth @TuftsMedicine @UCSDHealth @UCDavisHealth @WellSpan Thank you to the @WhiteHouse for the partnership. whitehouse.gov/briefing-room…

Technology grind in healthcare: we chose to build our own EHR for the Oscar Medical Group back in 2020. The OMG providers didn't like it for a long time (I guess that's par for the course for EHRs, but still, painful). Finally it's looking up. Lots of small improvements. Plus some LLM impact in here more recently (generating records from secure messaging consultations, lab test summaries etc.). Lessons are 1) don't build an EHR and 2) if you do, a short 4+ years later you will be very happy you did

Found this old email talking about our Theranos experience! We met the Apple folks on the same day and the contrast couldn't have been more hilarious

But keep tweeting your takes, you're around 100 putinophile tweets away from that Japanese soldier who hid in the jungle after WWII and was the last one to realize they'd lost

"I hate free trade so here are some tariffs" - Trump "He just wants others to reduce their own trade barriers" - you "Bring manufacturing jobs back to US" - Trump "But things will still be cheap because robots" - Lutnick I hope you see the issue here: 1% plan, 99% winging it

Oscar's growth to 1.3M members is a case study in the power of long-term trends and the fallacy of short-term blips. It always seemed obvious to us that a well-functioning U.S. individual healthcare market would eventually win out. But that it had to go through early legislative sabotage (risk corridor defunding), almost getting killed off in 2017, a pandemic, subsequent Medicaid redeterminations, red states' eventual enthusiastic embrace and now the unstoppable force of employer market individualization (ICHRA) would have been impossible to predict. Lesson for healthcare startups is to optimize everything for staying in the game longer. The obviously better thing (product, regulation, market trend) will eventually win out (even in as illogical of a system as U.S. healthcare) & you need to be around to benefit from that. fiercehealthcare.com/payers/…

Mark on life and @OscarHealth, a must-see!

In 1984, IBM became one of the first large employers to adopt 401(k)s. It became a bellwether for the individualization of stock ownership - with a pretty straight line from there to ETFs, commission-free trading and eventually even r/wallstreetbets. Gradually, then suddenly. Today, legendary employee-owned midwest grocery chain Hy-Vee is launching an individual employer (ICHRA) plan with @OscarHealth. We have a chance to make this a bellwether for the individualization of the U.S. healthcare system. ICHRA will create a totally different healthcare system, with so much consumer value and so many new companies yet to be built. We will continue to be a driving force in this change. Thank you to the innovative folks at Hy-Vee. Can't wait for 1/1/26 when the plans launch. forbes.com/sites/brucejapsen…

Baillie Gifford, great investor, said to us: in their 100+ yr experience, best stock performance short-term driven by business results, medium-term by highest product NPS, long-term by best organization culture

This is just a hilariously misguided understanding of macroeconomics. The US is at zero risk of defaulting on its debt (because it's denominated in its own currency), so why the f would the repayment schedule matter. This isn't a mortgage refinancing where you're trying to time the market on low interest rates. In the history of the world, countries have had 3 pathways to reduce their debt load: (1) inflate it away (1950s US, 1970s Italy), (2) default on it (Argentina, Mexico), (3) grow your way out of it (1980s Belgium, 2020s Greece). Treasury yields do not reflect some kind of short-term refinancing pressure (do yields pop on days when the treasury issues? no), they reflect long-term expectations of the value of the repayment. If the economy stalls for real, the economic machine that lets the US service the debt also stalls (tax revenues down etc.). If the markets were to sniff out that the government will inflate its way out, yields would spike. Sure, interest rates matter for the rest of the real economy - mortgages, corporate debt etc. But the markets aren't stupid. If today's falling yield meant that corporations would now all happily start borrowing to awesomely invest, then the stock market wouldn't have fallen, and yields wouldn't come down. Yields come down because the markets collectively think the economy is crashing, so nobody will borrow, and there is not even going to be any expected demand for credit.

Those two surveys have a decades-long history of co-existing in ways that are extremely well-understood. We live in an economy of 23 trillion dollars. Sure, it's complex to measure unemployment. Has never really mattered beyond small, short-term deviations.

To be fair, I think @RobertKennedyJr meant midichlorians, not mitochondria

The Lagrangian density of the electromagnetic field, with F the field tensor, and the greek symbols the Einstein summation notation (just means they run from 0 to 3)

I am with you. Hiring towards a more diverse team was never incompatible with also building a meritocracy. The world isn't that black and white, literally and figuratively.

Hard to keep up with putting out press releases when you launch a lot of stuff all the time, so here is some data on something we've had in production since Dec 2024: our intake bot that runs the initial conversation before virtual care consultations with the Oscar Medical Group. Our doctors can deliver more care: - 95% of providers reported time savings; with over 50% reporting saving >3 minutes per consult - Directional reduction in provider messages for most common RFVs like sore throat and UTI - Shorter consult durations in asynchronous messaging Patients stay more engaged: - 11% increase in consult completion rate - 66% of patients completed all questions (We have our own scribing as well, will talk about that soon) Also, if you want to get the details on how we do this including instructions, let us know, AI moves too fast to remain proprietary & we like to share

If this was the complex plane, the vector would be a + i × b, and its length would be a^2 + b^2, not a^2 + (ib)^2

U.S. healthcare is full of this type of financial doom loops: - Set a fantasy, artificial, inflated price - Invent a business that negotiates a bit of a discount - Charge for the imaginary cost savings. Another great example here, Multiplan & out-of-network pricing: nytimes.com/2024/04/07/us/he… All this works because those who are supposed to ultimately manage the healthcare dollar - self-insured employers - are just preposterously incapable of driving any kind of market efficiencies and doing this responsibly. (The Affordable Care Act marketplaces, in contrast, are way closer to driving these efficiencies: hard to overcharge members who can just quit your plan and go elsewhere if they don't like what you're doing)

This is an adventure game that uses GPT-4 to give itself formal structure and then executes the game. GPT-3 tended to get confused with it, GPT-4 does it impeccably well

Right, because if you light a hundred dollar bill on fire, the ashes are worth $100. I have an urn to sell you

That's the guy who complained that his website wasn't showing up in Google when he had a noindex directive in his site. So my advice here would be to maybe first check if his computer is turned on.

Some numbers on why you need LLMs behind the scenes in customer service teams: in the Oscar concierge teams, we store all of our knowledge in one (pre-LLM, simple search) system. When care guides can't answer a member question, they escalate it to a content expert. When the content expert closes the case, they evaluate whether the care guide could have answered the question with materials available in the knowledge management system: for 66% of escalations, the answer is yes. So the knowledge is there and systematically accessible all along, but people don't find it. Different perspective from a separate content expert survey: when asked where escalations come from, they say 18% actual information gap, 37% care guide is learning new workflow, 45% care guide isn't leveraging available resources. So, only in 18% of cases did the escalation require some piece of information that wasn't readily available in the systems accessible. Everything else was just a knowledge discovery issue. Retrieval-augmented generation, or long context windows, will make a difference here.

Many new & awesome launches at Oscar! If you're not in the market for an individual plan, you better convince your employer to give you money to buy one (and save), because you're missing out big time

We had an onsite with our engineering, product and design directors+ this week. To kick it off, we did a 20-min game coding challenge: use @OpenAI GPT-4o and @amasad's @Replit to write a game in 20 minutes. It is a super fun exercise, highly recommend it. Almost zero prep necessary, people loved it, and it's the most practical way to get a better intuition for everyone on where code generation intelligence stands these days. Here are the games (each is from a group of 3-4): oscar-tech-game-challenge.re…

To your earlier comment on "patriotic traders keeping a steady hand": why don't you show us your portfolio so we can all see which great American businesses you are supporting with your LPs' hard-earned money during these so unexpectedly uncertain times?

One of the more counterintuitive issues in U.S. healthcare is insurers explaining claims denials to providers. Counterintuitive, because in theory claims adjudication is 100% deterministic, so there should be no doubt about exactly why a particular claim didn't get paid. In practice, it is very difficult to parse raw claims system output and then explain it in a human-understandable way. So we created a GPT-4-based claims payment explainer. Here is more detail on how we built it: First, Oscar built its own claims system over many years. The system is based on our own domain-specific languages to allow for extreme flexibility and arbitrarily complex configuration and scripting. It also produces detailed claims traces of exactly how a claim adjudicated. The image below is an excerpt of such a trace. If you send this trace into GPT-4, it exhibits the usual LLM behavior: it explains simple traces surprisingly well! But completely bombs in medium to complex cases. The issue is that to include in the stack trace the necessary detail required to explain adjudication fully, you would blow through the GPT-4 token limit, by a lot. So instead, we need a multi-step process that lets the LLM zoom into the details it believes it needs. The chart below shows that sequence. We start with a "skeleton trace" that captures the rough adjudication journey. We teach the LLM how to find areas that are relevant to the adjudication decision. We then let it inject additional details from deeper in the adjudication trace into the prompt. Finally, we construct the adjudication explanation from these components. (It is a sort of chain-of-thought agent plus custom function-calling systems design.) See an example below for the difference between the "naive" approach, and the "call chaining" approach. This gets you 100% accuracy for various buckets of problems (though still only 80% for a few difficult problem categories, more work to be done). One key ingredient here is to have a claims system that produces technically beautiful output like our system does, which is mostly not the case for old-school claims systems. See the Oscar continuous hackathon section here for more detail: oscarai.notion.site/03-Makin…

Same is true for you, buddy. At least @Osinttechnical a) tries to get it right & isn't afraid to admit when he's off, and b) appears to be generally on the side of those appreciating human civilization, not the kind of nihilists whose bidding you too often seem to do, whether deliberately or not.

One of the biggest missed opportunities in U.S. healthcare is that as a society, we could be investing in better outcomes for everyone, which would save money in the long term - but often doesn’t happen because it costs more in the short term. Here is a hands-on example that Oscar has lived through in practice: PrEP drugs. PrEP (pre-exposure prophylaxis for HIV) refers to medications taken by people at elevated risk of HIV infection; they are an incredible scientific feat, quite effective and can nearly eliminate HIV risk for adherent patients. When the first drug came out, it was without alternative in preventing HIV infection. Many people meet standard guidelines for who is at risk (nearly 1 in 300, more than the number living with HIV), so as society, we’d want this drug in as many hands as possible. But, it is very expensive. At about $1500/month in 2017, it was one of Oscar’s most expensive drugs not subject to utilization management. So we felt the very real tension of the dreaded “actuarial death spiral” on the one hand, and investing in members’ future on the other hand: by 2017 PrEP was our #1 drug, making up 6% of our total RX spend. Our younger-than-average population at that time made our exposure particularly high. In retrospect, we had about 15x the number of people taking PrEP as you’d see in a random sample of Americans (see chart). In other words, we were helping keep a lot of people out of HIV harm's way, but it risked getting so expensive that the entire idea of it would just get priced out of the market, together with us. (1/3)

Here is another iteration of our LLM-based claims denial explainer. The root cause of claim denials can be notoriously intractable to understand for providers, causing unnecessary friction and overhead in the healthcare system. We created Ocebot, a system that uses GPT-4 to explain claim denials based on execution traces from the claims system we built at Oscar (called Ocelot). When we take the detailed output from Ocelot and feed it into GPT-4, it is able to explain what drove a claim denial, but only for simple claims. Previously, we improved on this performance by starting out with a "skeleton trace", a high-level summary of the adjudication path. With that approach, evaluations against human-provided ground truth reached scores around 7.9 out of 10. The post below describes the next evolution of the approach: by enabling the LLM to make function calls to our system, and by prompting it to act as an agent (see the first image), it is iteratively able to piece together a deeper explanation for claims denials. The second image shows how the LLM, in typical agent-like fashion, initially creates a plan for itself, and then requests more intricate details of the claims system execution trace from the system. The third image shows the kind of content it gets back from these calls. This approach reaches scores of 9.4 out of 10 against human-provided ground truth. The post is here: medium.com/oscar-tech/enforc… And be sure to continue to check out the Oscar continuous hackathon here: oscarai.notion.site/01-Overv…

More, awesome AI tools coming soon from Oscar thanks to @OpenAI and their excellent work on GPT-5

I will keep posting this chart everywhere. Putting people into a single risk pool (21 million & counting for the ACA), letting them choose a plan, and making sure that plans are entirely transparent creates a better-functioning marketplace. I agree that the CMS regulation forcing payers and providers to publish their contracted rates has been generally a great thing, but it has done virtually nothing to push costs down so far. (The opposite: it has driven costs UP in some markets.) The end user needs to put competitive pressure on this. The end user is all of us, and not employers' HR departments. Those have been getting bamboozled by payers and other middlemen for 8 decades.

LFG

Hahahahahahahaha

The fourth law of thermodynamics: when health insurers run out of ideas, they start merging with other health insurers

Good reminder of a crucial flaw in many people's minds: overestimating Russia's relative size. Countries like Brazil are now way bigger (210M) than the (shrinking) Russia (144M). If they didn't have the former Communist glory (and nuclear weapons), they'd be global also-rans.

Cool video that the great folks at @OpenAI shot at Oscar Health showing some of our use cases: piped.video/watch?v=rnC10yXh…

Chart from last year's investor day but I love it as the best illustration of what we have to do: drive down healthcare costs. In a stable individual market, there is way more competitive pressure to make healthcare affordable. All of US healthcare will go direct to consumer.

Reminds me of @nntaleb 13th rule for life: "Peterson, you idiot, stay out of this"

I used to think it's bad that videogames train kids on infinite iterations (save games & do over if you die). I think my kids' exec functioning is worse bc of that. However! Post-transformer creation is all about costless iteration, so maybe this is actually the future

David Sacks, proud democracy saver in the tradition of the My Pillow guy, Drunk Giuliani and squirrelly stapler guy from the Justice Dept

Pretty awesome how easy @OpenAI has now made it to create voice agents & great to be featured in this launch announcement with @OscarHealth openai.com/index/introducing…

I've been meaning to print out that guy's collected works, because I ran out of toilet paper. Incredible example of Silicon Valley's lottery winner phenomenon

Sure, just like $6B valuation on $4M of revenues is an appropriate efficient market valuation

Don't disagree w your point. But that chart has weird issues. Better to look at it in PPP. Also, ppl under-appreciate how much population growth matters for GDP. US still better in GDP per capita (= best intuition for how rich country feels), but way better in pop growth.

The other day I noticed something very cool: the most important equation describing particles in physics is basically the same as the most important equation describing risk in finance - with one crucial difference. Behold: the Schrödinger equation in quantum mechanics, and the Black-Scholes equation in finance. We will turn these into a quantum theory of finance. Includes vibe-coded simulation so read on 🧵

Data from the Oscar Medical Group's AI-written messaging summaries: the time it takes to document visits to virtual care providers via secure messages dropped by around 30% with automated summaries. The chart below shows how providers are using the AI-written summaries. Interesting: 47% of the time providers added data points - but most of those were specific data points that the model had no knowledge of, like specific lab results or the next follow-up appointment. Points to a clear opportunity for improvement (make the longer-term memory accessible - we have all that data already). Post is here: medium.com/oscar-tech/ai-use…. Also on continuous hackathon at hioscar.ai (which has been redesigned, check it out)

I zoomed in all the way for you! infinitezoom.replit.app/view…

This annual chart by @Rock_Health is always a great illustration why the employer-sponsored healthcare model is such an ineffective idea: nobody wants their employer playing a role in managing their healthcare. (Nobody wants the tech companies either btw.) ICHRA FTW

Jeff Bezos actually had a good quote on this one that has some merit: "I reserve the right to change my mind even in the absence of any new facts"

We've reached peak HLTH vendor spam, I just got this: "Let’s talk about the health of your critical data fax capabilities at HLTH"

We already have the solution for this: it's a competitive, transparent marketplace. For 7 years in a row now, employer healthcare cost inflation has vastly outpaced ACA inflation. ACA premium CAGR: 3.7%, employer premium CAGR: 5.6% since 2018. The employer market is the problem

In the category of "everybody knows this but every time you look at it it's still insane": price per colonoscopy in various Oscar markets. Always stunning how much higher outpatient prices are vs ambulatory surgery centers. But what's even more interesting is how ambulatory surgery centers actually seem to exhibit some characteristics of an open market: way lower price variation, for one. (This kind of data used to be guarded jealously by insurers, now you can just piece it together yourself from the cost transparency files we all have to generate. Great progress and we've only begun to see the impact of that newfound transparency) Finally, the limits of price transparency for driving consumer behavior are also obvious here: in the ACA, most plans' deductibles are now lower than the lowest-cost colonoscopy, so you don't have much immediate financial incentive to shop these prices

Good to remember that hospital EHR adoption really isn't very old at all (chart below has % hospitals with EHR). If you look at it this way, EHRs being widely used at hospitals are a more recent innovation than bitcoin

You should switch back to ketamine, the cocaine isn't doing it for you

One of our first investors tried to fire us when we said we want to do subway ads. But it's harder to build trust for someone's offline life with online ads, and a lot of healthcare still happens offline. So it worked for us. (We still do it outside of NY state)

There have been fewer deaths per capita from wars than any time in human history over the past decades, real interest rates have been lower than any time in the last 1,000 years (enabling Silicon Valley, to begin with), and the Marines' Hymn talks about Tripoli (the North Africa one, not the Wisconsin one) because the whole reason the US Navy exists in the first place is to protect global trade routes vital to US prosperity. And the reason most Western countries including the US went from conscription to purely professional armies is for the same reason - a liberal global order and US projection of power. Doesn't mean every war is just, doesn't mean everyone should enlist, doesn't mean NATO countries shouldn't pay more, doesn't mean dead soldiers don't suck. But we've tried all the alternatives and they suck a lot more.

Wrong. He has 6s in all the math and physics courses. 4s are in geography, "artistic drawing" and "technical drawing". "Naturgeschichte" is a 5, and that could be an old word for physics (Newton called himself a "natural philosopher"), but that's actually biology.

Hard to overstate how insane this is. I'm here on a greencard (and created plenty of jobs). Do I have to worry about getting snatched up & locked into a foreign prison? Really not much of a difference between me and some of these cases now coming to light.

The other place where that is clearly going on is Fox News. Entire homepage devoid of any Ukraine/Russia mention. And because consistency in ideology isn't their strength either, they do have a headline complaining about European withdrawals from Mali & ISIS

Right, and that's also what Jesus taught (given that you are a Roman Catholic per your bio). I think it's in the Gospel of Adolf 4:2

.@JoshuaKushner is a unique blend of friend, entrepreneur and investor, in that order. And the exceptional team at Thrive has built the fund into an organization that embodies exactly those values. They are an incredible partner to founders and startups, and few of those of us who built with them would be where we are without their enduring support. Congratulations on the raise & thank you for all the great work.

Why do you think @davidsacks is piping up? Probably picking up on discomfort from his Eastern friends who urgently need to buy some time. After all, can't happen every day to lose a ship to a country without a navy. (Actually HAS recently happened every day.)

Full view of Oscar's code base. Most proud of the Prolog code

The Oscar Medical Group (the virtual primary care group that operates our virtual care plan designs) has been using GPT-summarized lab test results for a few months. When lab results come back, the Oscar EHR creates an automatic summary. Some learnings: when it comes to the summary, provider behavior is fairly binary - either they delete the summary entirely, or they make minimal changes. When it comes to the patient follow-up instructions, providers tend to modify the AI-written summary a lot more - either to add personal style, or to add context that the AI was missing. (Lots more context we can give the AI, like the patient's state in a care journey - work in progress.) Interesting observation: provider edit behavior is the same across English and Spanish. Additional observations here: oscarai.notion.site/06-Enabl…. We'll keep publishing results on hioscar.ai as part of our continuous hackathon. Keep checking back.

Happy Lemonade customer btw!

Not that anyone is going public any time soon, but here is an important post-IPO learning that I bet most founders underestimate: turnover. Here is a chart of Oscar voluntary turnover vs the Oscar stock price (now back at $8.41, I think one of the best-performing healthcare stocks this year to date). 3 months after the IPO, turnover went up by 2-3x. (This is the whole company, so includes member services, ops etc.) Our stock price journey was particularly painful, so I don't know what this looks like if a stock just goes up... But anecdotally, many people who left after the IPO still loved the company, but just used the IPO as a milestone in their own personal journey and decided to move on. It's intuitive that this happens, but we could have planned more for this, and I think it ended up being harder to handle logistically than it could have been. (It did create lots of Oscar alumni startups, so that's very cool.) The other thing that's interesting here is that stock price eventually ceases to matter, for those who want to be at the company: I mean, it obviously matters and is on people's minds, but it's not the primary thing that drives your turnover anymore, as long as people are doing meaningful work and know it's going to take time to show up in results. So, for those planning to go public in 2025 (I guess that's now the earliest window?): plan for this.

Here is Oscar data from the last (partial) eclipse in NY in 2017. Conclusions: - The medical term for "staring at the sun too long" is solar retinopathy - Members were 30x more likely to seek care for solar retinopathy in the month following the eclipse vs. a regular month - Don't stare at the sun for too long

The 2024 Medicare Advantage enrollment period is the first time that the number of people turning 65 every year is stagnating, and from now on will only go sideways/down. (Medicare Advantage for 65+ has been the biggest profit drivers for most health insurers for the past 10 years.) So that annual second derivative boost is gone. Also funny how the number of MA plans available per potential customer was flat since 2010 but then doubled in 2018 because all insurers jumped in - right before the leveling-off was going to set in

Weren't you guys the ones who were all about being free speech maximalists? Has that position gotten inconvenient?

First: Klein-Gordon 2.0 Second: Riemann in 4D

Simple insight in how members select primary care physicians: first chart shows what happens when Oscar auto-assign primary care physicians based on various metrics like cost and quality - the farther away, the less likely the member is to actually visit. Second chart shows the same logic for members who select their own PCP (e.g., in the Oscar onboarding workflow): no relation to distance. Either a good sign of loyalty to PCPs, or a sign that those members see their PCP while at work or at a different address. (More on PCP loyalty in a different post soon)

Some notes from augmenting our 'find care' search with LLMs: the Oscar app has long had an omni-search bar that is able to distinguish between reasons for visit, providers, facilities and drugs. Way back in the days we had a SNOMED- and Unified Medical Language System-based search, but that sometimes produced funky results, so over the years we replaced it with hand-tuned, rules-based matching logic. Of course this should be an ideal use case for LLMs. So a subset of members now have a new LLM-based search enabled. The question as always is how to balance producing relevant results (precision) with not missing results (recall). Below is that curve from the team's testing. It produces neat results, for example when searching "dermatologist" (see image). But it's too general-purpose, and thus can miss searches that require more of a healthcare context, for example in searching for "runny" (see below). Further fine-tuning needed. More here: hioscar.ai/blog/related-cond…

That guy hasn't made money since 2008, when his intern told him about the housing trade. Why is anyone listening to him

How a company starts: mine were never a straight line. I have some data: since 2012, I've been tracking everything I do to the minute. So I went back and looked what % of time went into Oscar. Coffee with @JoshuaKushner with the idea in March, started reading about health insurance; Supreme Court upheld the ACA in June; brought on the first experienced insurance contractors in August; first regulator meeting on Oct 31, and only then we really got going. So many other things looked more reasonable than starting a health insurer (writing a data science textbook with my friend @sepkamvar who was at MIT at the time, a Facebook-based debit card in Brazil, etc.) I'm always curious how other founders do this stuff & if I'm just particularly disheveled

This article in the New York Times about record enrollment in the ACA has a photo of an enrollment kiosk from Las Madrinas ("the godmothers"). There is an incredible story here. Odalys and Mercy, the two founders, were individual AIG life insurance agents in Miami. They realized that Obamacare would enable them to get their clients health insurance coverage for the first time. So over just 10 years they built a juggernaut of an agency and have gotten hundreds of thousands of previously uninsured into coverage (see the Obamacare kiosks all over Miami). Oscar is proud to have been their long-standing partner (and of course I have to mention that they gave me a prize at one of their raucous agent kick-off events, back in 2019). There are many stories of incredible entrepreneurship in the building of the ACA which now will cover close to 20 million people who previously couldn't have gotten health insurance. This is public-private partnership policy at its best. nytimes.com/2023/12/21/healt…

Most consequential fun I ever had reading 906 pages. @JoshuaKushner, you're a legend, and thanks for always making everyone around you so much better

I think LLMs will clearly change children's education. Let me show you an example: a few months ago my daughter applied to a new school (she's 10). She built a graphical text adventure using @Replit and @Scenario_gg with @GitHubCopilot and a basic game structure from me. (This all got even easier recently with the new @cursor_ai.) Here she is explaining it: piped.video/watch?v=WZn5J2Cs… Here is the game (which is very meta because part of the phobias the kids in it have to overcome is that of going to a new school - she did get in btw): sienas-phobia-house.replit.a… Get your kids these tools and do some pair programming with them - schools still haven't figured this out

Generally true. With one big caveat: you seem to say employers are the solution. In fact, employers are by far the biggest enabler of this rent-seeking system. We're asking the least quantitative people in any organization to manage healthcare costs. On a more or less random and constantly overturning population. It's like telling your engineering team that some centralized non-technical procurement team will now select their programming languages, tools and libraries. (Actually that is how corporate software often works and precisely the reason why Workday and Concur are such terrible pieces of software.) Do everything you said, but put the risk and the health plan building in the hands of the providers, and individualize the purchasing of health insurance to put consumer pressure on the value chain, and you have a plan

The median speaker in meetings we record says 200 tokens/minute, so you get almost 3 hours of spoken human dialog without any indexing/chaining/... great stuff

So sad that the genius entrepreneur @elonmusk is blowing his chances to have a hugely positive impact on government by hitching his wagon to a lunatic he thinks he can control rather than putting together a bipartisan ticket over the next 4 years that many of us could get behind

Remarkable; but you also need this perspective: That bridge improvement supports a much larger economic multiplier than in 1937. Look at it in % nominal GDP: the bridge cost 0.037% of nominal GDP back then (inflation-unadjusted). Those nets cost 0.000002% of GDP today. We spend more, but we also generate more. Nets builders buy more dishwashers today with their salaries than the guys in 1937. Can't really look at those two independently.