← This Week in AI
This Week in AI
Is Anthropic a Cult? AI Beats ER Doctors & Recursive Self-Improvement | This Week in AI E12
01:17:00
Loading summary
Transcript195 lines
- [00:00]
A
AI is going to be the last technological innovation. There's nothing beyond this.
- [00:05]
B
I'm terrified.
- [00:06]
A
Jack Clark, one of the co founders of Anthropic, said AI systems are about to start building themselves. The people building LLMs right now will be replaced by the LLMs they are building in but two years.
- [00:21]
C
The only definition of intelligence that matters is can a species build something that can learn how to build itself? Really the step from where we are right now to the next step is not that big.
- [00:30]
B
90% of software development will be done by AI. And that's like not even an insane statement to make anymore. It's already happening.
- [00:35]
A
Thanks to our friends at PayPal, the exclusive sponsor for this Week in AI. Try the payment and growth platform that's trusted by millions of customers worldwide. PayPal open start growing today@paypalopen.com hey everybody. Welcome back to this Week in AI. This is the new roundtable. This is it, folks. AI is going to be the last technological innovation. There's nothing beyond this because this one is the culmination of, of everything that we've been working on for five, six, seven decades in the technology business. Semiconductors, the Internet, everything in between. And so we need to focus on it every week and we need to bring people on the show who are building the future and then look at the most important topics and have a candid discussion about it. Naveen Rao is back. He's from Unconventional AI. He's the CEO and co founder. They're rethinking the computer basically from first principles. Welcome back. You were on episode three. This is episode 12 of this Week. Nai, I think Naveen, you were one of the first folks to come back and get a second shot at the round table. It's. I'm not saying this is like the Apprentice or anything like that, but we are looking at the audience's reaction to guests and you had a really strong appearance because. Credible, candid. Credible and candid. That you explain to the audience what you're working on and why you chose to work on that and who your customers are and how it's going.
- [02:01]
C
I saw that the biggest hurdle to scaling AI is going to be energy. This became pretty apparent to me over the last several years as we just start to see how applicable AI can be. And I don't think we've even scratched the surface there. We're still doing pretty stupid stuff with it. So as we scale it up and try to apply it to new things, the cost starts to become a big concern. The amount of energy becomes a Big concern. In fact, it becomes such a concern that we won't be able, able to solve it. There is no good solution. So really I see this as kind of my life's work. It's something I've been thinking about for the better part of 30 years on how we could start to approach biological scale efficiency. And really it comes down to really taking the computer apart and looking at the first principles of what does it really mean to compute. And there's some relevancy from neuroscience as well here.
- [02:50]
A
Yeah. And essentially everybody's talking about, hey, how do I get more tokens? We went from this crazy moment in time in the beginning when ChatGPT and obviously that wasn't the beginning of AI, but that was the beginning of the public's kind of perception of AI, I believe, you know, outside of science fiction and mass adoption happened with the, the ChatGPT starting gun. And we went from hey, this thing isn't too useful to maybe you could find a use for it to oh my God, I have these agents. Oh my God. Claude. Four point X is unbelievable. Oh my God. You know, Grok in X is keeping me up to date on trends in real time. And these agents and open claw, all of that has one shotted people's brains. And now in just the last six months, I think you would agree, the race for tokens is on and we've all of a sudden become massively supply constrained, that prediction came true. And the prediction that this is a toy and it doesn't work and it's hallucinating, it's not getting results. So that's, I think something we'll unpack today. You want to lower the cost per token from the hardware energy side by 2, 3, 4 magnitude orders of magnitude, correct?
- [04:01]
C
That's right, exactly. And if that continue scaling, if that
- [04:04]
A
happens while people are making more efficient models, while people are working on the transport layer, moving data around faster, you know, there's like five or six different ways to lower the cost of tokens and we're seeing it all happen at once. So it's really interesting since you're doing this very foundational thing to have Trey Halterman with us for the first time. You did a great appearance, by the way, on this week in Startup. So if you want a full interview with Trey, not just talking about the Trends, it's episode 2155. 2155. So you type this week in startups and Trey Holterman, you'll find it on YouTube very quickly. He's the co founder and CEO of Tener T E N N T E N N R. He couldn't afford the last E but he's saving up for it. And you're doing the application layer. And this is I think really great to have this tension between the two of you, one of you on the front line, one of you building the foundation. Tell the audience, Trey, what you're working on, why it's important who your customers are and what's the state of affairs in terms of product market fit with your product totally.
- [05:11]
B
And by the way, quick, quick, aside unironically, the, the reason we don't have the E was one because I grew up in the Silicon Valley and that was what you're supposed to do FL Flickr Tumblr. And by the way, tenor with an E was $300,000 and we were coming out of college. Tenor without an E was 1200. So you can obviously see which route we went down.
- [05:32]
C
The best reason in the world, 100%.
- [05:36]
B
It's called a real constraint. We get patients from A to B. So basically if you think about you go to your primary care, you go to the hospital, what happens next where you actually get treated as going to be a surgery, it's going to be a therapy or it's going to be a device. 50% of patients actually fall into what we call the US medical black hole which is they never actually get to the specialist because there's a six month delay to get into a neurologist. Never actually make it onto therapy, get denied with prior authorizations, complete end to end system for taking patients through that journey and getting them there. We're in over 90% of counties, 300,000 providers across the country. It's more than 20% of US healthcare at present. This is app layer but very much sort of like plumbing of how do you actually move patients through a system so that they get care and yeah. No going through Naveen's background. I am, I'm like holy. You know I gotta keep you in my domain because we go down too far your side. I'm. I'm terrified.
- [06:23]
A
Dorian Naveen can. He can work on the application layer too based on his last appearance.
- [06:28]
C
Yeah.
- [06:29]
A
Why does the customer journey tray start with the primary care physician? Do we see that changing in the A of AI when everybody puts all their symptoms in, they know exactly what the first 10 things to do are. I've been suffering. You can hear in my voice the last week. It's not just me being at the NIC so game screaming. I actually got some kind of a cold Ear infection, traveling, whatever. I just went online. I'm asking Perplexity computer. I'm sorry, Perplexity Council, where you can do three different LLMs and then it tells you what they each said, what the differences were, and then it gives you the, you know, why they might differ. And it just told me exactly what to do because of my ear ringing, etc. I don't need to go to a primary care physician when the Chat, GPT and Gemini all working in concert with each other gives you such a perfect answer that even a doctor, and certainly all the studies coming out of Harvard and MIT right now are showing, and you're well aware of them, and we'll put them in the show notes that the AI for the first consultation is the better solution. So tell us how that's going to change and is there like institutional resistance to it, et cetera?
- [07:42]
B
I mean, by the way, like as it not. Not even just that first consultation, but it's actually that triage moment of like, hey, you know, Jake out where you go next? And you know, one of the really fast growing segments for us is actually digital health primary care providers because they have that digital presence and they're creating all these crazy new ways to basically work direct with patients, go D to C and provide that experience. But then, hey, when it comes time to like, okay, you're going to need to potentially, you know, you might have something far more serious or you're going to need to go get into the real world to get a, I don't know, you have a good nose, I don't think you would need any rhinoplasty. But by the time it comes to actually needing to get work done in the physical world to treat whatever is clogged up, you're going to then go and get onto Rails and Tenor is going to process you and get you in. So it's that interface. But yeah, absolutely. The front door of healthcare is changing a lot right now and it's certainly freaky, but at the same time, primary care and ed, they're so overworked that they're looking for anything. So you kind of have the, you have tale of two stories, right? The reality, the present reality of how overworked they are is and then the Borg intelligence of, of being the new front door. And it's, it's interesting to see which one will shake out.
- [08:49]
C
Well, what's interesting is that, you know, I've actually kind of been in and around the healthcare space for a long time, like just personally interested in it. In fact, I even tried to go into it after, after I left intel before starting Mosaic and just dropped it because it was too complicated. But I think the AI, the technology part has actually been super intelligent for a very long time. We did a study at Nirvana with a partner who had mammogram data from, I think it was from Mexico where they actually had biopsy data all the way back through a longitudinal study of mammograms and found that we could predict cancer much quicker than a human could. This is in 2015.
- [09:31]
B
So by the way, this is what, this is when I was in college and everybody was saying that the radiologist was dead, Right. And so what did you have happen? Nobody went into radiology and now you have just a complete and utter horrifying shortage of radiologists. So, I mean, do you know why? I actually don't know why. Like, what was the last mile that couldn't be overcome?
- [09:50]
A
Right.
- [09:50]
B
Like, why is there still so much. Do you know, first everyone, we focus
- [09:53]
C
on the technology problem. It's like, can it work? And it's like, yeah, actually very clearly it can work. But I think the liability part of it, everyone was scared about, oh, is it going to make a decision for you? And I think we just had to get this framework around it. Like, no, it becomes a tool as an input to the physician. The physician's still the licensed individual, all of that. So just figuring that out, it seems very simple in retrospect, but I think in practice it was quite complicated to think through because doctors were kind of scared. They didn't really want to know. They didn't want a place where they don't have relevancy. And so they're the ones making a decision. So I think it just took a long time. And insurance systems, as you know very well, are extremely byzantine and complicated. So just getting all of this to work in a way that someone would pay for the technology and it would actually provide an uplift in healthcare and somehow that's realizing as a, as a financial benefit. Just took a while.
- [10:43]
A
The reality is patients are now, I think, Trey, you're seeing this trend. Maybe you could speak to it a little bit. They're now deciding, hey, I'm going to just, I'm going to lead my healthcare, especially young people. And it's happened for, I think, two reasons. One, all these new tools have come out and they're free or close to free. So why wouldn't you. I guess there's like a footnote there of quantified self people, weirdos who do peptides and whoop and you know, like myself and who are, you know, down the rabbit hole.
- [11:14]
B
Are you, are you deep in the Peptides these days?
- [11:16]
A
I'm deep into everything. I, I moved to Austin, you know, and I'm just like, what's off the menu? What's next? Just tell me everything. Well, I mean, I don't want to experiment on my own body too much, but swim, Swim. Someone who isn't me was telling me about his experiences on Reddit, True Tide, after having done GLP, the other GLPs that are already on the market and they had a good experience and did a cycle of them and that was accretive. Swim did BPC 157, the Wolverine Protocol on swim shoulder, which he damaged carrying his wife's 60 pound roller. He's not bitter about that, but he did tear IT and then BPC157 healed. It could have been psychosomatic, but I do think there's that group Trey. And so there's like, either you embrace the technology because it's free and everything and then there's people who just didn't have healthcare insurance so they had no choice. These are the people who go to, you know, an emergency care place and ask for the price before they get the service and are discerning and they're actual customers. So what, what are your thoughts on how that is impacting this relationship and how insurance might change? And are the insurance people excited about this or are they bummed? Because I know the gps are excited about it. Right. They want people to come in educated. They want people to, you know, to have the consultation go well. Or maybe you're wrong. You're. Yeah, maybe I'm wrong.
- [12:53]
B
Yes and no. I mean, my mom's in urgent care now and she would tell you there is frankly nobody more annoying that come than the patient that comes in having, you know, effectively a hypochondriac episode that's been validated by a language model. So like for every, you know, self, you know, self responsible adult that's like, you know, asking about the ringing in their ears that, by the way, is it ear infection or just too loud?
- [13:19]
A
It was some sort of inner ear infection that. Because it wasn't external. Right. I didn't have like irritation there. I had something internal. And I think it was because of. We have oak fever. You have cedar fever in the January and then oak fever in the spring in Austin. And I live on a ranch and the oak trees are like, I could rename my ranch Oak Ranch and The whole like three out of the five family members got taken out by this thing. Then I went on a flight, then I was run down going to the Knicks game in Atlanta versus the Atlanta Hawks. And I was, you know, excited all night, then all of a sudden ear ringing, 2am and then it's just slowly come down over five days.
- [14:02]
B
Yeah, I mean you have this just, it's just a weird situation basically where the younger, you know, commercial insurance folks that are on United B, like their employer is paying for it, they want this like lovely digital, elegant experience that guides them step by step by step. And the providers that are operating on that are just, you know, they're just decimating the competition. They're not always digitally native. A lot of these are brick and mortar. But they're doing an incredible job. I mean, take a thing like CPAP that still is almost always paid for by insurance and you know, the folks that like know how to guide patients through that insurance just see such higher conversion rates that it's like they're just dominating the market. Do you see that type of Persona? But then separately, right, like when you look at, you know, the business on our side where we're doing complex infusions and complex therapies and complex, you know, work of patients, the average patient is a 68 year old female who you could build the best user experience, the most educated, you know, online web app and still the primary means to engage that person is going to be the phone. And realistically that's going to be the case with an aging population for the next like 15, 20 years. So you kind of want to build towards the puck going. But frankly you're also going to miss out on like the largest population that's going to need, you know, the health care world for the next 10 years if you're not meeting folks where they are. So it's just like, frankly it's confusing and we tend to be way more in the like brick and mortar sort of hard schlep problem type of businesses that we work with. And those are the folks that see the most patients in healthcare right now.
- [15:26]
A
And that's our second topic. So there's this Harvard study. AI beat two ER doctors at diagnosis. Researchers ran OpenAI's 01 model against two ER physicians on real cases from initial triage to final decision whether to admit the patients to the hospital. And 011, the study pulled together 76 real ER cases pulled from Boston's Beth Israel Medical Centers. ER admission and then they fed it into GPT. Quote the model identified the correct or very close diagnosis in about 67% of cases, compared with roughly 5050 to 55% for physicians. In one task, 01 received a perfect clinical reasoning score for 98% of cases it examined. And so why on earth, why on earth would we let a human do this in the ER today if there is a 15 point gap, which represents roughly 30% in terms of performance? If this is the truth, we should be sitting there with an ER doctor and they should have a copilot in 100% of cases. Just like Naveen in self driving. Do you drive a Tesla, Naveen?
- [16:41]
C
I had one for years and I don't have one currently.
- [16:43]
A
But you have to get the new fsd. What are you pushing, Naveen? Everybody wants to know.
- [16:48]
C
Well, I actually like driving and I don't drive very much anymore.
- [16:53]
A
You got a VAT? What do you got?
- [16:55]
C
No, my everyday is a M5.
- [16:57]
A
Okay. Fun car. Well, anyway, you have to get FSD because at this point anybody who has the latest FSD says the same thing. Trey. Which is I trust FSD more than I trust myself. These ER doctors, have they gotten there yet or they still have the classic God complex.
- [17:16]
C
So, you know, it's kind of an interesting thing actually. Like I was just having this discussion with my team. Like language models can integrate more information and context than a human can. It is just a truth of the world. Human brains are not designed to do that very well and human doctors like to think that they can. We generally integrate that context through experience and that's kind of a long process. But if you have context in the LLM sense that can be useful for the current situation, you should absolutely use that. Now I think the challenge why we don't have that today is it's actually very much a can I get the right context into the model? So yes, this study is great, but it sort of assumes I get perfect context. And I think oftentimes in the er, especially when you're doing real time stuff, you have to make choices on incomplete context. And I think that's where it starts to break down. That's where LLMs actually don't do so well. If you don't have that and you kind of have integrated experience, LLM will not perform as well as a human.
- [18:19]
B
Trey, this is like the perfect LLM use case, right? A ton of unstructured documentation, synthesize, make a determination and all the information to make the accurate determination is in there. Can you find it? Not through business, frankly, just from friends and folks. I've met in the er. Right. It is so much about the questions that you ask. So I totally agree there's gotta be some sort of copilot experience that comes in with that pre synthesis so they understand the history. But I'm sure a clinician would tell you their body language, what they're saying to you. Those are all inputs that you have to somehow feed in and you know what to ignore and whatnot. So but I mean like if the take is that. By the way like we talked about radiologists earlier, every radiologist, their PAC system that they're using to read images has some sort of ML models underlying and helping them drive their decision. It just turned out there was going to be so many patients that needed health care right that we need a ton of readers even with those assisted tools. ER feels like a similar situation where you could walk in with 10 times the context, synthesize so much more effectively and then you're kind of doing that last mile of interpretation that feels really like, it feels like a hugely positive like case for, for healthcare and you could plug that in. Don't have to go, you know, don't have to change anything about your, your payer models or, or anything like that. You could just, just provide better healthcare. And you know folks that one of the interesting things that I would add to this by the way is that so many hospitals do not employ their own ER docs. So ERs are often these standalone independent businesses that sell to the major health systems. They say hey we're the best ER doc physician group in the state, employ us. So that's actually one market where you're already seeing competitive dynamics where they are selling to their hospital system partners who are their customers. Hey, we're using AI assisted co pilots. We're coming in and doing that. So I actually, I've seen that a little bit and that, that's why ER actually I'm, I'm very bullish on adopting this technology because it will, it will be used in board pitches for why you should partner with them for $100 million a year contracts because they're going to have the best, latest and greatest technology. So the incentives are actually really good there which is a big bull case.
- [20:21]
A
The thing I'm going to do Naveen from now on when I go visit a doctor is I have this thing Claude pin. It's a little pin. You both are shaking your heads that you understand it and you can wear it as a wristwatch, you can put it on your jacket or it can go on the back of Your phone. I'm just going to record every single thing and I'm going to just be very slow and clear. Here are my symptoms, here's where they happened, just like I did on my own. But then in real time, especially if you were going with, you know, my dad's been sick. That's why I'm here in Brooklyn for the week. Like as caregivers. I'm going to just give one of these to my mother and tell her to start doing it. Then the magic is you don't have to take your phone out and start recording after it syncs. It has templates and I sure they have a template for healthcare visits. Then you can hit the share button and instantly share it with the rest of the family. So let's say it's your dad, you got two siblings, you got mom and you know, you got two cousins who are all in the industry. You just share that document and the summary of it and now you've got the entire history. We have to flip the script, I think, and change it from the doctors and the healthcare system having all that data, having all that analysis and we have to pry it away from them. Now we can have it. You just bring the plot pin and may I record this? Or in some states you probably don't even need permission and it's your healthcare. So listen, I'd have to think about my judgment on that. I'm assuming doctors assume they're being recorded, right? At all times, Trey.
- [21:50]
B
They have an ambient listener in the room now. It's all digital now, but.
- [21:54]
A
So they are doing that. What's it called? What's that?
- [21:56]
B
A bridge is the number one, is the leader in the space for most your major health systems. So by the way, you'll notice actually now if you go to the by the way, I say no, you've been staying healthy, jcal. You haven't seen that. All these docs will keep a phone next to them that'll be running a bridge ambience. Suki or Microsoft as a competitor. And they'll literally be kind of holding it near them and narrating, describing it. You could do it right back at them, sort of a one off of holding it right near there.
- [22:23]
A
Do they have to tell you, hey, I'm recording this and I'm taking my
- [22:28]
B
notes here or gosh, what's the best practice? I'm going to assume a patient consent form because all it does, like all it does, right. There's billions of dollars going into this and it's deeply sophisticated. But all it does is it takes the notes of what's happening really and narrates it and dictates it into the emr into epic, really.
- [22:45]
A
Got it.
- [22:46]
B
And so I assume there's no different from a privacy concern perspective than the fact that they're already taking notes on you, which I'm sure you're signing off at some point either way, so.
- [22:54]
A
Got it. It might be in that input document you signed. Right on the way in.
- [22:57]
C
Naveen, my whole family are physicians and you know, there have been various forms of recording for 30 or 40 years. Right. Because doctors want to dictate back their notes and that. I mean this is years ago and they would actually do recordings back then. So I'm assuming this hasn't been challenged in court. So it's probably okay.
- [23:13]
A
I guess that is always the thing. We have such a litigious society. There are all these Lawsuits are like a 2, 3, 4% drag on GDP is one of the stats I've been seeing float by on social media. I don't know what ambulance chasing and all the litigation going on has caused, but I think it's something in the range of 1 to 3% of GDP goes to lawyers suing people, of which who knows what percentage is. Is actually valid. I mean, some percentage is valid, I'm sure. But if it's 25 or 50% or 100, I think we'd probably all agree it's towards the lower end number. But yeah, this to me feels like we're going to in short order be able to process many more people and get them healthier quicker and reduce mistakes. Yeah, that's. I mean, I think everybody's goal. The next question I have, Trey and Naveen is I have superpower function. Whoop 8 sleep. I mean, I am dialing all these things in. A lot of people use Aura and I'm AI first. Right. And I'm patient led. I'm a biohacker, I guess, at this point, but like a junior one at best. When is that data and these conversations with my physician going to connect Trey? Because that to me feels like the other piece of the puzzle. I've been studying Whoop and Function Health and all of these software practice products. Whoop just added the ability to do blood labs. Whoop added a fitness program in it. They obviously have other things they check. They have a. A approximation of your blood pressure. The FDA sent them a nasty gram about like, hey, you can't do that. They were like, we're doing it. It's got a disclaimer you know, sue us. And then you got rfk. Trey, you know, basically running amok at the fda. He's running amok. I'm happy about the dies. I'm not sure about his position on all the vaccines, which he seems to be. But Trey, when do, when do we get that connected? And is that something you're working on or is that something you're going to wait for the marketplace to tell you, hey, we're ready to get blood panels and connect it all. Because the er, the, the electronic record, what is it? Erp.
- [25:36]
B
The EMR, the electronic medical record system.
- [25:38]
A
EMR, yeah. And EPIC seems to be the 800 pound gorilla and the most hated company in the world. I. Why is that? Why? It's like a three part question here. Why are they so hated and what's the state of affairs there? And then when do I take my first person experience and plug it in with my doctor? Because in whoop, there's a feature to like create a report and print it out as a PDF or send it to your doctor. So they, that must be, you know, if, if we're building that proactively in the WHOOP for biohackers, I gotta think the doctors are gonna want to ask you for that at some point.
- [26:15]
B
So I mean, there's two kind of like exchanges, right? There's exchanges between providers and there's exchanges like that you're talking about, which I think is much more like a patient led. Let me fetch my records. So there's this thing called the FHIR protocol, F H I R. And for you to be a valid legal electronic medical record system or electronic health record system, ehr, which, which Epic is the largest and the most like, well known, you have to conform to this protocol. So you could start an app tomorrow that says, hey, when you log in and you give me consent, I can fetch your records via fhir. And so you can do that. And honestly there are good like record fetching mechanisms. And honestly there's even some company that does a really good job of like enabling other digital health companies to help the patient fetch the records. So this is like a pretty good ecosystem. It's not like a business that we're in by any means. I'm almost amazed, I was amazed when I did function, to be honest, that it didn't pull any of the records. I think the main reason I would have expected it to was because insurance basically covered none of my function, which sucked. Sorry. You know, but it just was like, really, you got to pay like 600 bucks out of pocket. And then my mom clowned me because showed me that basically if I had just gone to primary. So there's a, if I had gone to primary care it would have been free. So neither here nor there. But as soon as they start wanting to deal with insurance, they're going to need to know, okay, well did he get a lab last week? Because we're not going to, you know, insurance isn't paying for this if it's not happening next week. So I'm always trying to look for like what is the business driver of it. My guess is that function feels like they have an incredible product without a bunch of like garbage junk data that will come from when you found yourself in an ER eight years ago. That's probably the most realistic take. But specialists, right? If they, if you go to that hospital and they send you into rheumatology, specialists are absolutely fetching your records. In fact you'll get rescheduled if they don't have them on time. Right. And even though that's an API by the way, it's an API of a bunch of unstructured documentation that they have to fetch. They're using tenor for this to then parse all that information and say is insurance even going to let me see this patient? And that's sort of the battle and the litigation, that war is sort of the one that we are in the middle of. But yeah, you asked the question about epic. I can circle back to it. Naveen, you have thoughts on that side? Otherwise it's always fun.
- [28:27]
C
Yeah. I actually have a buddy of mine in Dubai in the UAE who has been, who's put together a set of clinics called. His clinic is called Metabolic and they've actually done exactly this where you kind of sign a consent form when you come in and become part of their system and they issue you like a whoop and several other things. And basically when you come in, everything is there downloaded. You come into the first office visit, there's a whole AI summary just on a screen that a doctor kind of goes through with you and kind of lets you interpret and then says this is the course of action we think you should take. We probably need to do this test or you need to do this better or whatever. Right. They kind of, it's sort of like assisting you in interpreting the results. But it's exactly what you're talking about. It's non clinical data. I think the problem in the US has been just the liability. Like if I start using, you know, oura ring data, like well, was there Some kind of user error. Do they put it on somebody else they put on their dog? I mean, where's the data coming from? And I think that's been the problem in the US is just this. Like he pointed out though, the litigious nature has created a very conservative culture in health care.
- [29:33]
A
I have an idea. When I run For President in 10 years, I'm going to come up with a new concept. It's called informed waivers. We're workshopping. There's a better name where everybody involved in going on a ski trip can sign this document and they actually don't get to sue. You know when you sign that document that says you can't sue us if you die doing hella skiing or whatever, backcountry skiing, and then everybody sues anyway. I want to create a document that actually does that. So if you go to one of these doctors and they're like, hey, you want to do peptides? Hey, you want to give us your data? It's going to be yolo. It's outside the system. We're just going to have like, speaking of like Dubai in the Middle east they have free trade zones like Shenzhen in China, they have a zone in the UAE where you're going to be able to gamble against Sharia law. But there's. They're building a win. They have an area for the news and the media to operate in Dubai without censorship. If somebody wants to in that area say something critical about the kingdom, I believe they can. Like if you're a Reuters journalist, you're kind of protected because they're like, listen, we, we want Reuters journalists and New York journalists to feel like they're not going to get imprisoned here. So there's like a little area, I think that like economic free zones like Hong Kong was back in the day. We need some thing like that for advancements like this to occur. Same thing with crypto, right? People are like doing Isle of Man or whatever for their crypto. Zerg and Zurich. We need something like that for healthcare in America. I just say we call it Austin. If you come to Austin, you land at the Austin airport. You've now come into Joe Rogan territory, Lex Friedman, Tim Ferriss territory. You can do whatever you want. It's a free tech bro healthcare service.
- [31:20]
C
Well, I think to make that really work. So that has to be paired with some kind of individual incentive. I think this is why these betting markets work so well, is that they're predictive because you want to make money so you have a direct incentive to be correct. And the same idea here, there has to be some direct incentive to not want to cheat the system and actually give the correct data to the doctor. You know, like, I think right now there's some weird incentives probably baked in because of a whole ridiculousness of insurance that people might cheat the system. And that's why you have all these liability rules. Right. I actually don't understand why you would want to do that, but I'm sure there's some scenario.
- [32:00]
A
All right, talk to me about electric medical records and if you want to follow up on V's comments, you can.
- [32:04]
B
Trey, you ever sued anybody for a ski trip?
- [32:06]
A
Jcal?
- [32:07]
B
No, no, I'm just like, I. Have you been around it? Have you seen it happen? Because I feel like it is one of these things where you hear about it all the time. Never actually seen somebody like go distance.
- [32:15]
C
I know someone who got sued for hitting somebody on a, on a ski slope.
- [32:19]
A
Yeah, actually I do know something. Gwenneth Paltrow, a friend of mine and Gwenith, I heard, you know, got hit in the back and the guy landed on her and like wouldn't get off her.
- [32:29]
B
It was a little bit weird.
- [32:30]
A
And then he decided he would like get into this long lose lawsuit. He lost, obviously. That was a weird one. But I've seen a couple of pretty hardcore collisions. But yeah, I, I think people do sue for this. I mean, I understand it, but I do think there's got to be some, or maybe there's got to be caps or something. I don't know. Just feels like we're far too litigious.
- [32:54]
B
Back to the, on the EMR side. Like, I think it's almost a, it's a very tricky problem. I think people take off so much of their frustration in general with the healthcare system with basically two people, the payers and epic generally because they're like the two things that are the largest, most money producing entities in this system. You see a lot of crazy stuff, but you also see a lot of crazy stuff on like mom and pop businesses that you would never even think. And honestly, sometimes you see more crazy stuff on that side where there's like real rackets going on. Is everything to do with like complex therapies. Like there you have the person that gets paid, the person that pays, and the person that's supposed to measure how much gets paid. All the same. All are basically all the same entity or they own the entity and they have to have a firewall, but they all know anyways. So that's where I think, you know, there's probably more ire that should be drawn the, you know, some of the middlemen that you're not really supposed to speak about or, you know, you'll end up getting a skier hit you in the back on the way down accidentally. Like honestly, even I'm a little scared to be honest, to speak about some of the, you know, that side of the business. But that's like where I think more public attention should be spent. It's like these insurance companies that own drug companies that own the sparrow, that only pharmacy benefit managers that decide how much to pay for a drug. And oh by the way, if you get sent to us to your script, your pharmacy that they want to go, they can actually just block you off and say, actually you can't go there. You have to go to our pharmacy because we make the drug and we pay it so we know that we can just pay ourselves more. It's an insane, insane racket. And that's where like more public attention should be. But you know, there's a bill in Tennessee that will hopefully get passed that's going to make that illegal. So let's watch closely and hope that becomes sort of a national thing.
- [34:40]
A
If we were going to disrupt Epic, and we'll go on to the other stories of the week in a moment but if we're going to use the venture capital playbook, how might we disrupt a company like Epic and make an AI first version? Naveen, if you put on your thinking cap, your strategy hat, what would be the, you know, obviously like bottom up SaaS is like a classic technique. You know, let the small players, you know, from a bottom up use a tool, et cetera. I don't know if that would work or not here, but what comes to
- [35:12]
C
mind, we're sort of hitting around this main point, but really all the data that you have about your healthcare is context that's useful for making decisions. And I think the way EPIC has been designed is actually for human consumption, right? It's like, okay, so I can ingest some documents, I put it into some table or some format that I can send it to insurance providers and then other healthcare providers. Great, that's fine. It existed for its purpose. But now what we want to do is actually make all of that data consumable by AI. And how do you do that? Well, I think there are actually we're learning a lot. I mean my time at databricks from before, like how do we make data consumable? Well, companies like databricks are doing this so I think actually making Ingesting that data, putting it into an embedded form, where now you can start to consume it as part of an LLM and make it really efficient and fast. Prioritization of that data. Maybe recency matters, but also relevancy matters. Like, okay, you had a chronic disease that you diagnosed with 10 years ago. That's important even today. Whereas, you know, maybe you broke your wrist 15 years ago, and that's not super relevant. So kind of some prioritization, but making it consumable by AI. And I think that's what Epic has sort of been tiptoeing into a little bit, but it's still tiptoeing. It's not good. And this is the classic case, right? We're seeing this in a lot of, like, legacy apps. They, like, make that one little incremental step better, but, like, the guy who rethinks the whole stack is the one who's going to win. And I think that's what you're getting at. Make it consumable as context. That's relevant to an LLM.
- [36:45]
A
I think, Trey, we're going to get you in trouble here because you got to work with these people. But I give you permission.
- [36:51]
C
I don't.
- [36:52]
A
Yeah, I give you permission to torch your entire business and all business relationships by being on a podcast.
- [36:57]
B
Go ahead. I. Look, the key thing is I won't. We're like the ultimate Switzerland with emr. So I will never say anything bad about them. What I will say, though, is how you disrupt it. You can almost ask them, and you can look at their terms of service based on the apps that they allow. So what are they most conscious of? Let's say they are most conscious of somebody that is replicating the data that they have and taking away the clinical clinician experience. All right, so if you didn't know what ambient listener was, I'm going to, like, clue you into a kind of interesting and fast story. There is, like, the most VC darling company, and they're just complete legends. Dr. Shiv Rao, their CEO. This is a bridge, okay? $5 billion company. Raise your latest round with Andreessen and IVP and a who's who. By the way, Andreessen also invested in their number one competitor, ambience, so they're both sides of the game. And it's like, all right, what the hell's happening here? Whatever. So. And Bridge is the VC darling and their epics darling, and they're doing such a good job. You got friends at the wazoo, obviously, that work there, and they're really like owning that clinician experience. Now the problem is you can't own it too much. And that's, I think what ultimately was felt was. Wait a minute, there's a scenario where Naveen, what you're saying is if we are pulling the records in this hypersynthesized modality so the clinician no longer even needs to look at some tabular system that's organized for their eyes, but is actually just an assistant that's co piloting, that knows everything, so they can quickly ask a question dynamically. It's like that's the ideal experience. Maybe, you know, they have one, they have one earbud in. They're talking to you, they're treating you. You know, they're trying to figure out why, you know, what ridiculous Peptide Jake Alamos killed himself with last week and they're, you know, trying to treat him while they're, you know, dictating it back to themselves. So Epic basically kicks them out of the marketplace, comes out with their own competitor, and now you have like the ultimate war of health tech right now, which is between, you know, these. Now wait, did that happen?
- [38:52]
A
They kicked out a bridge.
- [38:53]
B
They, they, well, they took them out of their Platinum Premier sponsorship. You know, imagine every adjective they used to do like a big rev share and it's. They, they broke that off. They broke that off. So they, they are speaking with their. With volumes, saying that experience is the most coveted. Like, we need to own this experience.
- [39:10]
A
Wow.
- [39:12]
B
And I think it's davine to all your points, like, that's, that's the best way is to own the. Either you have to own the patient experience through records or you have to own the clinician experience. And reality is you're better off owning the clinician experience.
- [39:22]
A
We had Shiv Rao on this Week in Startups episode. Let me look at it here. Episode 1784, the Next Unicorns that was released in July of 2023 if you want to see it, a couple years ago. And I just had him on a recent episode, which I think we published this week in AI feed. Yeah, so he was on this Week in AI and he is. Yeah, pretty formidable guy. Naveen, your thoughts?
- [39:49]
C
Yeah, I mean, I will say one thing. Epic has been a survivor. They have survived many different tech transitions. Like they're pretty tenacious and they do exactly what you just said. Like they'll come out, they'll come out with a product that's sort of good enough and leverages their channel that outcompetes the new shiny thing. So I will Say they have survived these kinds of things before. And what's her name? Judy. I think the CEO, she's pretty damn focused and tenacious.
- [40:17]
A
And it's like a, it's a very closely held company. They have 80% of records. They have a huge margin. I think they have about a 40% margin because it's closely held, it's not public. I believe it's very hard to know what's going on over there. I think the AI scribe plus an open source medical record system that could be super disruptive. So if you made an open source system that anybody could use, Epic is kind of like duct taped together from what I've been told, you have to do crazy integration. So the way they make their money is not just selling the software, but it's a, you know, 100 million, 500 million to implement this stuff at your hospital chain. So they're just pulling out massive amounts of capital. What if there was an open source project that Shiv and other note takers and everybody else said, hey, we're going to invest in this open source platform, open standards, open platform, and just make Epic's margin their opportunity. Just make the platform so cheap. And they have only 40 or 50% of the hospitals use EPIC. But those hospitals represent 80 or 90% of the patients. So they claim that, oh, we got tons of competition. That's because it's too expensive and too kludgy from what I've read. And there's people who are experts on this that aren't named Jason Caliganis who have said this. And you could just, man, sweep the table. I think on an open source version over the next decade.
- [41:46]
B
He's the guy to do it, by the way. Yeah, he was, because he was only two weeks ago, by the way. He's one of those guys, you know how Jensen's got the, you know, the leather jacket and that sort of hold bravado. I was at a conference the other week and I felt like a dork because I was wearing my blue suit with everybody else's blue suits.
- [41:59]
A
Yes.
- [42:00]
B
And I see Shiv walking up and I'm like, God damn it, he's done it again. Like samurai, black jacket and sweatpants surrounded by a sea of suits. And it's just like, that's love it. If there's somebody that's going to do what you're talking about, J Cal, I mean, send him your PDF.
- [42:13]
A
There you go.
- [42:13]
B
Get him to go pitch launch. He'll do it.
- [42:15]
A
All right, here's an interesting one. Let's go big picture and then we'll go to the hyperscaler capex. How close are we to recursive self improvement? For those who don't know this is hey, AI systems improve themselves. I have systems doing this right now. I have OpenAI. I have open claw agents who go out every week and they look for new best practices in a vertical and then add it to their skills MD file. Okay, fair enough. But what we're talking about here is LLMs doing it. One of the co founders of Anthropic Jack Clark said AI systems quote AI systems are about to start building themselves. He says 60% probability on fully automated AI R&D with no human in the loop. It's very important that we state that part. With the Frontier models training its own successor by the end of 2028 this would de facto mean Naveen that The people building LLMs right now will be replaced by the LLMs they are building in but 2 years 60% chance. Do you believe that you take the over or you take the under? If we were gambling, you take the over or the under? I'll take 29, you take the under. You think even 2027 next year we could be sitting here and no human in the loop.
- [43:42]
C
Yeah.
- [43:42]
A
Explain to me your thought process there when you were thinking in bets.
- [43:47]
C
Well, I think we're already doing that to some degree but I mean there's still, there's still some human in the loop for guiding things. But you know what that human does has already been. They're already using AI assistance to do that process. Like I will say personally when I do research right now I just tell it what I want in English and it sort of runs the loop for me. So I'm effectively doing that already. And really the step from where we are right now to the next step is not that big. We've sort of laid the tracks to get there because we already have the AI assistant that can go from this is what I want to. This is the output. We already have the data gathering, all of that being automated and now it's just connecting those pieces together. That to me feels like a year long effort.
- [44:33]
B
I mean just because I think, I mean look, somebody that's not doing research actively right now, I mean would be so curious to hear what does it look like to. And what does one of your prompts look like? What does that even look like to engage like a research level prompt? I mean is this like a page? Is it a paragraph? Is this sentence?
- [44:48]
C
I mean I think you know at first It'll be pages and then it'll get less and less over time and eventually we build the next version. So I mean, this is. If you look at my Twitter bio, I basically said I want to build stuff that learns how to build other stuff. To me that has been actually that recursive definition of intelligence is actually the only definition of intelligence that matters is can a species build something that can learn how to build itself? And I think we did that. Biology did that. I think we're building machines that are now capable of it. I think it won't be great at the beginning. It's going to be kind of shitty. So end of 28, I think we will have recursively self improvement. I don't think it'll be very good. I think it'll be another couple of years before it becomes really good. So call it 2030.
- [45:36]
A
Here is a very interesting chart from my friend Tim Urban. Wait, but why? That we were all talking about years ago. He wrote this blog post in January of 2015. 11 years old. This is now almost just 10 and a half years old. What does it feel like to stand here? Human progress time. And you have this gentle curve slowly going up. Humans are making a lot of progress. Industrial revolution, Internet, semiconductor. And there's a little human there. And then something boom happens. That boom that happens is AI and it's super intelligence. But here's what it feels like normally and I think now we are directly in this moment in terms of the prediction which makes it super hard. I think we now all think it's fade.com plea. Yeah, Trey.
- [46:22]
B
It feels really hard to ignore. I mean it feels really hard to ignore. And it feels like I remember thinking people were out of their minds. They were saying 2027 that like you know, 80% or 90%. I think there's some great Dario quote of like 90% of software development will be done by AI. And that's like not even an insane statement to make anymore, right? Like by 2027. And it actually looks like it might happen by the end of 2026 and it's already happening at like 80. We're like 82% right now, this last week. And we're like a 70 person ENG team building real industrial grade products, moving patients along and it's like 70%, 80% plus. So yeah, it feels not totally insane and feels exciting.
- [47:03]
A
Yeah. Dario, who is loves to make predictions, Dario Modi, anthropic, is the queen of predictions here. He is always, I'm sorry, king, queen, whatever. He's always out there, you know, with the scepter, you know, saying here you hear he quote, we essentially have Claude designing the next version of Claude itself. Not completely, but most of it, quote. I have engineers with Anthropic who say they do not write any code anymore. Yeah, I just let Claude write the code and edit it and look it over within Anthropic, that means designing the next version of Claude itself. And yeah, many, many Mac says M 2.7 is their first model deeply participating its own evolution. So you now have multiple people on the frontier saying, this is happening. And then even me, in my modest way, with agents, you know, me asking a human, hey, can you get better? Every week at this task, to me seems silly, like, the human's a human. It's just not going to remember to do it. If they do, they have to be motivated. They have to have initiative. My Lord, it's just so annoying to do it with a human. I just literally wrote a skill the other day where I said, every Sunday, I want you to go out and look at the best practices on social media, these six platforms, and how people are using social media more effectively to build podcasts and programs, et cetera, and do clips and write threads and gain followers and get likes and go viral. Tell me every Sunday, when you do this deep research on each of these six skills, tell me what changed and what you implemented and then go test it for the week.
- [48:53]
B
I mean, that's just mind blowing for
- [48:57]
A
somebody who's been building content for 30 years on the Internet that.
- [49:01]
B
Have you learned, like, has it taught you anything that you didn't know, like, recently?
- [49:04]
A
Yeah, the one shotted moment was it said to me one week. Somebody who works for Mr. Beast explained this week in a podcast on marketing that they use AI to make 50 different clips. Then they use this service that works with a panel of YouTube users, just, you know, humans to show them different YouTube pages with videos to click on. And then they put those 50 different thumbnails in there to narrow it down to which ones get clicked on the most. Then they take those five or six, then they a B test those on their channels, and then pick the winner. And I was like, wow, that's like. There's a tool where you can, like A B test two headlines and two or three thumbnails built into YouTube. And then it just goes with whatever's best. Just like Mailchimp did 10 years ago with, like, the headline of your newsletter, you'd give it an A B test. Here's my first headline. Here's my second. Whichever one gets the biggest open rate, you know, so you send to a thousand people, which everyone gets the biggest open rate. Send to the next hundred thousand. They're doing like a panel before that. Now they produce one video every two weeks. So they really have some thoughtfulness. But to then go spend, they must be spending thousands of dollars, maybe $10,000 running that. That panel. It's almost like they're running a test audience like they used to do for movies. And people would write on the card, I don't understand the ending. Or like, I don't understand what happened in Blade Runner. And then the studio was like, okay, Ridley Scott. Nobody understands this movie. Blade Runner. Can you please put voiceover in to explain what the replicants are and to explain the ending and to explain to people that, you know, Deckard is a replicant. Spoiler alert. And he's like, I don't want to explain that.
- [50:55]
C
That.
- [50:56]
A
So then I met Harrison Ford. And I'm talking to Harrison Ford. Sorry, I just have a little name drop. It's the greatest story.
- [51:03]
B
Harrison Ford. Yep.
- [51:04]
C
I heard.
- [51:04]
A
I heard. So I'm at this party with her. You heard the name drop?
- [51:07]
C
Boom. I did. I did.
- [51:09]
A
So I'm talking to HarrisFord. I said, Mr. Ford, I just want to thank you for my childhood. He goes, son, you're making me feel a little bit old here. And I said, no, you know, I'm just. It's. It's just really an honor to meet you. It meant a lot for me. Han Solo, but very specific. And Indiana Jones. My dad bought me a bullwhip when I was a kid. And I used to sit there in the backyard and in Brooklyn and hit grapes and, you know, try to smash things with this, like, actual bullwhip. It's just, like, such a big part of my. My youth. So thank you. So that's. That's incredible. I said, but I wanted to ask you about Blade Runner. He's like, okay, kid. And I said, if it's annoying, I would.
- [51:41]
C
He's like, no.
- [51:42]
A
Did he say kid? He kept calling me kid.
- [51:45]
C
That's awesome.
- [51:45]
A
It was the greatest moment of my life. It was five years ago. It was that, like, when they opened the Gigafactory at Tesla in Austin, he was there in the vip. I'm in the VIP hanging out, and somebody actually took a video of me talking to him. Like, I wonder what they're talking about.
- [51:59]
B
Here it is, folks. This is my Harrison Ford story.
- [52:02]
A
I said, I heard that Ridley Scott. And you were like, admonished by the studio that nobody understood the film and that Deckard was in fact a replicant and they were sending replicants to retire. Replicants. It adds like a whole level of nuance and that they forced you to, you know, in order to get paid or whatever that to do these voiceovers and that you then went in and read them in a sarcastic voice. The year is 2019. Los Angeles Replicants, human robots that nobody can explain. And that you did it sarcastically. It's like son, I'm going to stop you right there. It's a true story. However, you know more about this film than I do and I was the lead.
- [52:51]
B
It's like Harrison Ford Irish accent in there. Basically.
- [52:55]
A
It's the Harrison Ford Irish accent, son. I was the lead in this film and you know more about it than I do. Greatest.
- [53:03]
C
He just read his lines, I guess.
- [53:05]
B
Yeah. Anyway, it was pretty damn good.
- [53:07]
A
I had a point about that story somewhere along the way. Yeah. But I just don't know what it was.
- [53:15]
C
But you got to be.
- [53:15]
B
How was that going?
- [53:17]
A
I think it was going to replicates themselves somewhere.
- [53:21]
C
Yeah.
- [53:21]
A
Self improvement of replicants is just. There is that one shotted moment where you're like oh my God. It's, it's, it's obviously happening. A big part of that of course is now we had a moment in time where people realized throw more hardware, throw more data at a language model, get better results. This has led to a crazy estimate. The estimate now from Brian Nowak of Morgan Stanley. He raised his hyperscaler Capex forecast again. He said 805 billion in 2026. Now he says it's going to be $1.1 billion trillion dollars in 2027. So 805 billion this year, 1.1 trillion next year. This estimate grew, you know, by 100, 200 billion up. Up 40 billion for this year, up 150 billion. So up a total of 200 billion. Here's a graph showing historical Capex spending that went viral this very week. It is extraordinary to look at. And I guess my question, and obviously Sachs has been talking about. This is a huge tailwind for the US economy and this is only like 5 hyperscalers here. My question, Naveen, are we going to get to diminishing returns? Are the returns on capital even possible here? If you go, if you look at this, we're talking about going into the J curve with these four or five players. Five players. You're talking about going into a three, $4 trillion loss and then having to make it back. So I guess there there's two questions here. Diminishing returns coming because of folks like you making things better or language models getting self getting recursive to our prior story, then making the need less or is the, the demand curve which we're seeing right now where anthropic is going down, anthropic is going down and rate limiting people. Is this demand going to continue for some number of years and we never catch up?
- [55:33]
C
Well, I think the demand will continue, but I think there's actually a bigger point here that something interesting is going on. Is so all the things we talked about a moment ago where you're saying like, yeah, AI just does it better and you can throw more data at it gets better. It turns out a lot of problems in the world can be solved through gradient descent. Okay, so a lot of economic value of what we do is really just honing away at something. Think about it. I'm a guy who knows how to, I don't know, whittle something and I could make figurines. And so it's just this honing aspect which just got better and better over time. And that created economic value at one point. It created economic value and it created purpose for those individuals. Now we're in this weird place where we're almost solving that through capex. Okay, so the thing that kind of scares me right now, and I don't have a good answer to, is these things I feel are not just. It's not a game of return on invested capital. It's like I'm putting a trillion dollars in, I'm going to get 4 trillion out in some timeframe. I think we're reshaping the economy and how demand actually looks and how supply actually looks and it's going to create some weirdness for individuals and how we participate in that economy. I actually don't understand what it is. You actually said something right at the beginning of this podcast where you said, you know, AI is the last thing, last final boss. Is it the final boss? Like is that really true? Are we gonna, Are we done with a demand based economy kind of.
- [57:06]
A
I don't know. Unpack it, Naveen, unpack it, unpack it. Give me some real world examples here of that scenario where it is the final boss. And then give me. And then I'll trail have you give the argument for it not being the final boss.
- [57:19]
C
Well, okay, so actually since we've been talking about healthcare, I think we can stay in that one, in that realm for a little while. Like I think all the things that we do As a healthcare process, like there's a standard of care. The way a doctor has to treat a patient, it has to be defensible. You have to go through this kind of chart of like I looked at this, I looked at that, and that whole thing is automatable. There's nothing human about that. In fact, humans are kind of shitty at it. So let's just use machines to do it and do it perfectly. Okay, so now what? What would a human go to a doctor for? Why should I go to a doctor? I think actually humans might still go for it, go to a doctor. And I think there's some other aspect here that each human might be willing to pay for and that's to understand something about another human's lived experience. Like I want someone to empathize with me and feel something that I felt and I will pay for that. And I don't know if that's automatable. And so I think our economy might completely shift away from this sort of like utilitarian aspect of like honing something down and buying that honed product, because that's done by machines and it's free now. It's about like there's something that isn't about honing. It's about an experience of another human being in connection.
- [58:28]
A
So the value of the product is delivered not through the product because that will just be solved by machines all the way down at zero cost. Effective at zero cost. So then what is left? Well, based on what you're saying, what's left is the human intention behind the directive the machine was given and the final product experience. Trey. An example of this might be we look at calories. We live in a post calorie world.
- [59:01]
B
Look at me and you start talking about calories immediately.
- [59:02]
C
Damn.
- [59:03]
A
No, you look great. You look great. I lost 40 pounds. Ro co twist go. And they're not the sponsor of this podcast, but I'm a spokesperson for Road co twist. Get on your GLP and I lost £40.
- [59:14]
B
But if we look supposed to make me feel better? No, I'm just kidding. Sorry, what were you saying?
- [59:18]
C
I'm probably the skinniest one here.
- [59:19]
A
You know, I'm going to the Naveen and I have black T shirts and we got cannons. I don't know what's going on under your Carhartt to get jacket there, but we'll see.
- [59:27]
B
You know, fair enough, fair enough, fair enough. Little fat kid syndrome. Sorry, I.
- [59:32]
A
No, we're we. As part of this week in AI, we all lift at the end of the show. So we're gonna just get our kettlebells and do a little lifting at the end. But if you think about calories, we do live in a post caloric world in some ways there are more calories available. Yes, sometimes a dictator can block people from the calories but there's nobody starving in America unless they're mentally ill. There's some other aggregate aggravating factor. Now I know some woke people are going to be like, no, no, there's victims who can't get calories. Yeah, I call BS in the modern world there are calories everywhere and they cost pennies a day to feed people. And people take calories and throw them in the garbage at the back of grocery stores every day that settlers in the west would have think they would think the garbage at behind an H E B was the htb that to them would have been extraordinary to be able to just pick from the stuff thrown away every day in people's garbage. So let's put that aside. And if we look at what people do for calories, they're picking it based on the aesthetics. They go to a Michelin star restaurant, they pick it based on the vibes. They go to Starbucks or you know, flat white or off that pop up in London because they like that cheeky nature of ordering a flat white. We pick our calories based on aesthetics and vibes and virtue signaling and just flavor profiles. There's nothing to it other than the story. Naveen, Trey, what's your take here? We've already done it. Like we have an exact.
- [61:09]
C
We're already moving there. Yeah, we're already moving there.
- [61:11]
A
Yeah, I think we're there.
- [61:12]
B
I guess my hottest take on the subject is just that like we think about work so often from the perspective of this like white collar, high level, you know, Naveen style work where I'm, I'm thinking I'm jamming, I'm like moving and it's like all these, all this tooling is so good for that and yet like 98% of production system are like production systems are like just deterministic and like need to happen the same way every single time. And where my like skepticism often is like, like it seems like the bottleneck to a lot of growth and the bottleneck to be able to satisfy all the demand for all the goods out there and being a truly like post demand society, like that's what you're getting at JCALs like how do you get a post event? It actually seems like. And Naveen, you're focused on this I think on some of the like real world things it seems like the, it's like the physical stuff that will actually get us there. The you know, person that needs to carry the box from the warehouse into the truck. Because it's literally more cost effective to hire human doing that than it is to lay a belt that goes and does that. And really only Amazon operates at a scale where they actually can have it go right on the truck and things like that. That like, that's sort of where my like, you know, put the fries in the bag, bro. Like, you know, brain goes on. Some of this stuff where I'm like, is the demand really that we need more intelligence or is it just that we need like more factory belts? Or you're saying, hey, we'll get there because we'll have robots that actually can do that last bit.
- [62:28]
C
Yeah, I was going to say, aren't those one of the same to me,
- [62:32]
B
I'd be curious, I mean, I would be curious. Will you be able to get that so cost effective that it can literally do the pickup and move it? Is that better than just laying a belt? It will be cheaper than building a belt.
- [62:43]
C
Well, I think this is what the recursive self improvement's about. You're right. Today that is true because there's still a lot of humans involved in the supply chain of building a robot. But what if that starts getting automated? It comes down to really just the, the price of the machine to physically dig up the ore and then assemble it like that starts dropping really fast. Then I think you're just going to out price a human every time in 50 years. I think you're going to out price a human every time. So a human is not going to be the economically most efficient choice. But I think there still will be humans doing stuff.
- [63:17]
A
Yeah.
- [63:18]
B
The percentage of goods that J guy you're talking about where you'll pay extra for the vibe, the service, the feng shui or whatever it is, like that will be a very minuscule percentage of the actual goods that you consume. Like I'm probably not going to get my Dixie cup, you know, with its extra dose of microplastics on the top, I'm noticing, you know, I'm not going to need that to be like human welded probably right.
- [63:39]
A
So despite, I mean, if you look at, I mean, it's interesting, you could almost take the other side every time we have the discussion. Because humans are buying like women are obsessed with this Stanley mug.
- [63:53]
C
Oh yeah, I've seen those.
- [63:55]
B
God damn it, you Stanley. Stanley mugged. Me. That's right.
- [63:58]
A
And it's like, I'm like, guys, Stanley makes power tools. And they're like, no, they make mugs. And I'm like, is this like some kind of like, men are from Mars, women are from Venus, like, big put on or something, that Stanley made this to troll people's wives so that they let their husbands buy power tools because they have an affinity for the cup. I don't know the background of this, but pull up and show this producers, like, here's the Stanley Cup. What is this about? There's a pink cup with a thing. Women. This thing sells out. Women talk about it. It's like a.
- [64:32]
C
My wife has one. It's heavy.
- [64:35]
A
Is that why?
- [64:36]
C
I. I don't know.
- [64:37]
A
Wraps in.
- [64:38]
C
I. I don't know. Like, I asked my wife because I saw it too, and I'm like, stanley, what. What is that about? And she's like, oh, that's like the cool thing to have. And I'm like, why? It's just itches.
- [64:47]
A
Is.
- [64:47]
B
Do you remember when camelbacks with the little mouthpiece were like the big thing? Right? Isn't this just like that? Like, isn't this going to go away too?
- [64:54]
A
Right.
- [64:55]
B
Either way, to your point, you absolutely. Stanley mugged me. Maybe every piece. Yeah, maybe. Stanley mogged me. All right, there you go.
- [65:01]
A
I. Stanley mogged you got one. Stanley one shotted. Stanley shotted.
- [65:07]
B
The thing that. I'm very curious I need to get out of this time with Naveen here is. And you. J Cal is.
- [65:13]
C
Is.
- [65:14]
B
You've got the greatest marketing machine and anthropic. And the Mythos seems to be now like GPT 5.5 seems to be the same level or better of what they're purporting Mythos to be. Is the Mythos release. Hubbub. Oh, it's too powerful. Makes people too uncomfortable. The JCAL model is the only reason they're not releasing it because of. I caught it.
- [65:37]
A
I caught it stray.
- [65:38]
B
I get the little stray.
- [65:40]
A
It's way too dangerous.
- [65:42]
B
Is it just because they're GPU constrained and they actually are going down and we're seeing that left and right too? Or is it actually, you know, the fear longer.
- [65:50]
C
I. I think they have used fear as a very, very effective marketing tool. And I don't think this is an exception from every. I mean, I. I know a ton of people in. In the company. From what I can hear, it is just another really good model. And to be completely fair, like, I think some of the. The. The media treatment of it has been way more than what Anthropic has said. Okay, like Anthropic actually hasn't said like it's a danger to humanity. They said well there could be a practical danger because it could cause some cyber issues. Like that's really been their message but that kind of got blown out of proportion a little bit. So I think I just want to be fair to what Anthropic has said here now. But I think they have very effectively used fear and sort of like put themselves in this category of being the trusted ones. And it's been really great for them. I mean I think their models have been great, they've been very good. I like to your point though, I don't see them being that much better. In fact Quad itself at coding, we're seeing that Codex often is better at certain tasks than Quad is. But somehow Quad became the default programmer tool. And I think they've just been really good at saying this is what we focus on. Actually what's interesting to me just sort of zooming out, it goes to consumers buy a product because of why the person builds the product, not necessarily what the product is many times. And that's what I feel like is happening here.
- [67:11]
B
I mean look, my Twitter feed is still software engineers Twitter feed and it seems like Codex has taken over. But jcal, what was your anthropic? I need to get that out.
- [67:22]
A
I actually have a hot take. I think this has to do with the gravity of the entrepreneurs leading each of the frontier models. If you are a P Doom score of like 70, 80 or above, in other words, you're predicting doom and that AI is really dangerous Terminator, you know, delete all the humans because somebody asked me to cure cancer. Quickest way to keep humans from getting cancer is to get rid of humans. I just eliminated all the humans, right? That whole scenario. If you are P doom 60, 70, 80, you self select and if you're like liberal and anti Trump, you self select working for Anthropic. That's where you feel at home. If you're that Pete Doomer and you're anti maga, anti Trump, you know, Kamala Harris, whatever that vector of person, Berkeley, whatever. No offense to Berkeley but you know that, that sort of contingent, no offense the Berkeley stray. If you're the opposite of that. If you're. I think I'm, I'm low P Doom. I'm low P Doom, low p doom, under 20% p. Doom, you go work for Elon because you want to be multi planetary. You want to win you want to do those high level missions. I want to make humans multi planetary and I want to explore the most important questions and discover the answers to those questions. With Elon, I want to work 80 hours a week. I care about the work, I care about solving big problems. So you go to one of those two places and there's a third place you go. If you are a marauding capitalist who wants to write in their journal, will this get me to 1B? And I donated a hundred thousand, but I didn't actually donate it. And you're a marauding capitalist who wants to maximize the returns on your RSUs, on your stock options. You go work for Sam Altman. And it doesn't matter what your P Doom is. What matters is how much money did I make, how quickly did I make it? So we've seen all this dancing around and then there's also go work at Google with Sergey and that's like a. It's independent. You could be somewhere in the middle. You could be.
- [69:42]
C
I think it's more well lined, it's more well aligned with the Elon side.
- [69:46]
A
Yeah, I would say it's more Elon Musk, but still you, you gotta, you got a little researcher in you and it's like a safe place to work. They're not gonna run out of money. So that's like the safety school, just a safe place to go. It's like your Stanford, you want to go to Stanford, you want to stay in academia, you go there. So over the last five years everybody's just moved to where they belong. I think the people at Anthropic actually believe because they're so P doom 70, 80, 90% that they are these Jedi monks who are protecting society. They believe this. It's not marketing to them. For us it looks like marketing from the outside because we're like, okay, it's really dangerous, but we're using the other ones. And it does the world. The sky hasn't fallen. They believe Jesus is coming, that they are the chosen ones, that they will be saved because they've accepted AI Jesus into their hearts. And the rest of us are all going to die in brimstone and fire and burn in hell for eternity because we don't accept AI Jesus. They believe it. Now how do I know this is true? I've met people from homeless companies. Put it aside. If you look at this specific issue of cyber security AI there are 5 million like cyber experts in the world. Of those, some percentage are elite, let's say 10% of them, 500,000, maybe 1%, 50,000. There aren't enough of those people to go get all the errors and there's nobody paying them enough money to go find all the problems with software. People are too busy building software. They're too busy making money from software. That's where the money is. There's not enough money inside. But. And you have to choose that if you want to be like you're working on the wall, like the people who. What is the wall called in?
- [71:41]
B
Yeah, the Wall in Game of Thrones. The Wall. Yeah.
- [71:44]
A
Like you go to the north and you go work on the wall. Like that's another group of like nights Watch.
- [71:48]
B
The Watch.
- [71:48]
C
The Watch.
- [71:49]
A
Those are the Knights of the Watch. Cyber is Knights of the Watch. They just want to go there and wait for this terrible thing to happen. It's happening in the next hundred or two hundred years. So we just got to wait on the wall to protect society. That's cyber people. There's just a small number of weird people who do that now. What if you had an unlimited number of them? What if you could take those 50,000, 500,000, 5 million of those and you could just add two or three zeros. That's what an LLM does if you have an agent working relentlessly on the long tail of software. Because people only work on the high surface area. High surface area is like Oracle databases, Windows, you know, web servers, CRM systems. They're working on the high surface area where the money is. Nobody ever gets down the long tail.
- [72:29]
B
Well, arguably there's still the same people on the wall, but you just have millions of white walkers coming towards you that will explore every frickin crevice to. Correct.
- [72:37]
A
Yes. So now you just have in this Cambrian explosion of cyber experts. Yeah, they're gonna find all the bugs that nobody else had the time. There wasn't a big enough bug bounty to go find those white walkers and eliminate them. So it is true that that software did find more bugs. It's not true that. It's cataclysmic. That's right, cataclysmic.
- [72:57]
C
These are small white walkers.
- [72:59]
A
That's just like. It's like little zombies from Walking Dead. You just hit them with the pike, boom, you're done. You get a 10 foot sphere and you can just take them out from the top of the wall. They're not material, but the people at Anthropic, they're like the height of the High Council of the Jedis. They believe that they are bringing balance to the Force. It's just A different mindset. They're monks. So here we go. Okay, I gotta stop because it's making me all worked up. Sorry, guys. Thank you for coming to my TED
- [73:30]
C
talk, but I 100% agree with you. Actually, that that tracks pretty strongly with the people I've talked to now. So you're talking. There's two different things here. One is sort of the external facing marketing, which is what I think. I think they have done very well at, I think. Then there's the other side, which is the internal facing, the people who work there. So 100% agree with you on the people. Now, do you believe that Dario, in his heart of hearts, believes that he is a monk?
- [73:53]
A
Oh, yes. I think he believes more than anyone.
- [73:57]
B
I mean, I think we've both seen people have moderate success with way less success than that get to their heads at a hundred times the level. Right, sure.
- [74:06]
C
Yeah.
- [74:07]
B
And they're not ushering in super intelligence. Right.
- [74:09]
C
It's just weird because I've. I've known Daario for a long time, kind of from almost academic circles and stuff like that. And what do you think you would
- [74:17]
A
have a better perspective. Does he actually think he's the chosen one to bring AI Jesus to Earth and save humanity? I mean, has he been one shotted?
- [74:26]
C
He might have been one shotted. Now, I. I don't feel like he was that before. I think he was a smart technologist and he was passionate about what he was doing. But I think maybe you're right. Maybe he has gone down this path of like. No, I am actually the one who's here for a purpose. Maybe there's.
- [74:41]
A
He's AI Moses. Maybe he has the commandments and he's going to bring Jesus now he's the Ark of the Covenant. Anthropic is like the Ark of the Covenant. Lightning's gonna come, fire, brimstone. I think they actually believe it. I think they're all one shotted.
- [74:55]
C
Oh, they do.
- [74:57]
A
They all have these illusions of grandeur.
- [75:00]
C
Totally. No, it's very weird. There's people there that I have known for a long time, again from the academic world, like 10, 12 years ago. And I talk to them now, and they're like, you know, because I told them the plan of what we're doing, unconventional. They're like, well, why are you doing these things? It. I'm like, what do you mean? They're like, well, all that's going to be solved by AI in five years. I'm like, you legitimately believe that AI is going to build its own physical hardware and figure it all out in five years. They're like, yeah.
- [75:24]
A
They're like, is it nihilism or is
- [75:27]
B
it like, people have been like this, like, literally in the 1960s or 70s, Buffett famously said the best thing that ever happened to him was everybody believed in the efficient market theory. And then in the early 2000s, I remember everybody was telling me, oh, Trey, why are you building a Web2 startup? Like, trying to solve. You know, it's all. That's all going away, man. It's all going to be this different thing. By the way, this one might be totally right. And I could, you know, look like a dork, but, hey, nobody will be listening to it anyways if that happens.
- [75:52]
C
So.
- [75:53]
B
But the. It seems like people have been justifying inaction for, you know, effectively as long as. As long as humanity's been around. Because it's kind of hard. I mean, to be honestly. Why are you doing it, man? Like, aren't you tired? Like, it looks like you had a great deal.
- [76:06]
A
Why don't you just, you know, you've always got a little bit of cheddar. We should all just off and go to Kawhi.
- [76:13]
C
Yeah.
- [76:14]
A
Just take and surf and just accept the fact that AI Jesus is coming. And whatever happens between now and AI Jesus is irrelevant. It's irrelevant. We are all going to be dust in the wind. It's. He's coming.
- [76:30]
C
We should just all drink the Kool Aid then. Let's just drink it and, yes, commit mass. Right?
- [76:35]
A
That's what we should do. This should be Jonestown.
- [76:37]
C
It should be Jonestown at a global scale.
- [76:40]
A
We should all drink it on the count of three. Thank you for listening to this Week in Aipdoom 100 Nothing Matters. Drink the AI Jesus Kool Aid and we will be the chosen ones. Naveen, Trey, two amazing guests this week in AI episode 12 is in the can. We'll see you all next time. Bye Bye.