An unfiltered conversation with Chamath Palihapitiya

in the war of open versus closed I think open has basically won which is just a simple acknowledgement that all of these models are converging much faster if you want your model to be used need to be

publicly available and explained and understood look to be blunt like of all the models that I've seen I think deep seek is really wonderful I think open AI is is pretty good you know I think llama

has tremendous potential but I think the Gemini models are the best the marginal cost of energy and the marginal cost of compute would go to

zero chimo um I'm super excited for this conversation just because I think you uh you have a bunch of incredibly hot takes but also like you're close to the action that's happening across a bunch of

things um I'd like to get your reaction to this to start off which is I feel like there's this weird cycle going on in AI where um what the sort of conventional narrative is like ends up

flip-flopping in some like reasonably short amount of timeline I think the two most pertinent ones that have been top of mind for around this is um you know if you go back to two years ago the

entire narrative was all these companies building on AI are just wers there's no value being created um and I feel like in the last like 30 days The Narrative has like entirely flipped to like all

the values in the application layer etc etc the second one is like 60 days ago the narrative was like models are hitting the wall there's no more you know Innovation happening and then all

of a sudden it's like reasoning models are the new scaling law like everything is up for grabs you know progress is going to continue indefinitely I'm curious to just like your meta reaction

to those those two examples and and other ones I think it's actually quite um emblematic of most nent markets when

they're emerging so there's only ever three constituents in any market so constituent number one is what I would

say are people that have uh essentially a huge sunk cost and so as a result they have to have a very specific Orthodoxy

that they believe because the scale of their decisions are measured in billions or tens of billions or maybe even hundreds of billions of dollars so there's very little room for them to

reerrr one and then cohort 2 are the people that in our economy are always trying to raise money

and then cohort 3 are the people that actually have a firm sense of what they're doing independent of what's popular and I think what happens is that

the pendulum swings back and forth between these three different cohorts and in any given moment one of them gets a little bit more attention than the other because they have a

better hot take in that moment but I think that this whole AI thing right now is largely going through a sorting function which all markets go through of

which people in each of those cohorts are kind of like reliable and consistent and which are talking their own book which doesn't mean that they're

wrong but it just means that they're only right in moments and then who are the ones that are really precient and see the future and then everybody else who's

just kind of a goon and ding-dong so we're in that pH of the market of um taking those three cohorts and then separating into those four archetypes that's what that's that's um that's I

think what's happened and it's happened in every Market before it so it's pretty it's pretty similar in that sense and and what's your like very uh very tactically like what's your advice for

people who are sitting in the shoes of like you know I the Persona that I care a lot about is like the founder who is trying to build something and like they're hearing all this crazy signal like do you have a a sense of like yeah

what are you doing to get information from the right people and like you know not listen to all the craziness that's being talked about I don't think that there's

anything you can do other than read the actual papers and find a community of people that are speaking about the the

facts I mean look most Founders are going to fail right 95% of them is it that they not smart no they're all smart

is it that they're not courageous no they they are all courageous to start the problem is they have poor judgment and the founders that fail in a critical

moment fell for something that just turned out to not be true right um there's a different way of framing this which is um you know Josh whiteskin

calls it the most important question how can you distill a moment to the most important question and the founders that build enormous

companies tend to have extremely good judgment in those narrow moments right where you can get to the most important question and then answer it with High

Fidelity so the thing with Founders is that most of them if they are not focused on their own emotional

management are just going to waffle and flip-flop and you know make a really bad decision but by the way I think that that's healthy and that's the Darwinism of our industry and that's how it should

be in many ways um so I think that the founder needs to go back to First principles so often used now but what

does it even mean I guess um like you know two weekends ago the founders that are one step closer to knowing what the most

important question was was the one that was actually uh downloaded deep seek implemented it

played with it read the paper then read about you know grpo then read about PTX then actually asked their team why you know should we be avoiding Cuda why did

these guys find these improvements by voting Cuda should we actually you know why did why did why are we only so

reliant on um you know uh po those are some of the right questions still doesn't mean you're going to be successful but then there's an entire cohort of Founders who didn't even know

to ask those questions those folks are they could still be successful but the propensity is that way of making

decisions will cause them just you know unnecessary waste and time and you know disorientation because in two weeks there'll be another deep seek moment and two weeks from that there'll be another

and then there'll be another and so if you're not building the Reps to try to make good decisions in a repeatable way you're you're just going to

fail yeah to to pull on this deep seek thread really quick actually um is your take that it's all of the craziness that

people are are making it out to be or like and and actually like maybe practically for like you all at 8090 like did you wake up in the last two weeks and like you know we are the the platform is different I know you've

talked in the past about like you have you know infrastructure to be able to like swap out models and all that type of stuff but like do yeah did you like actually transition workloads over to these new models or has it just been

like hey this is a useful information for us to keep in mind as we as we keep building the honest answer is it's something in between because

we did we play with them yes did we um did we realize that we weren't moving fast enough also yes you know when like

inside of 8090 there are these three layers of our cake one layer is just the application experience that we work with on behalf of our customers and

Enterprises one layer is all of the systematic automation that makes it go super fast which is what we call our software Factory but then the third layer is this

wedge in between that we call high performance compute my initial instantiation of that high performance compute team was just

wrong the people were excellent but what they were able to do which was extremely low-level Hardware oriented

manipulation was a little bit miscast in our organization because we were largely a software organization and we didn't have the justification to manipulate the bare

metal but then when I saw the approach that deep seek took to get some of these performance optimizations I realized actually hold on a second there was a different way of

writing to the bare metal that we hadn't seen and what we did in that Monday morning you know our first three hours of Monday are all just very intense

detailed product review it's kind of like the most important stuff we save for the first meeting of the week um before everybody just gets tired and

frustrated you know uh and the first three hours of Monday were just like pinpointing all of the places in which to be honest we had either gotten

a little lazy or we had made assumptions that were just invalidated or we had tried an approach that was just wrong and then to rebuild the product plan for

this next week that starts to embrace some of these changes so you know in a couple weeks I will I will tell you that we hopefully will have embraced everything but my big takeaway from Deep

seek is just that the resource constraints of a small Nimble team should never be underestimated and that the surplus of resources for large teams

is a horrible distraction that's that's the that's a rule that I have known for being in Silicon Valley for a quarter century and I always ignore it and it's

and it just it's always staring me in the face and it's a wonderful reminder of that thing it's small extremely resource constrainted teams that land

big things and when the resources get quasi infinite it all it all you you just make very Orthodox decisions that never that tends to not

yield in great outcomes trath I'm curious because I it feels like given your historic stance on Capital Moes and AI like do you believe

that these open source models can truly challenge proprietary models themselves like where yeah where's your stance on that I think in the in the war of open

versus closed I think open has basically won um and I think that you see that capitulation now because even you know as of this past week I think that open

AI is now considering some form of open- sourcing what they're doing um which is just a simple acknowledgement that all of these models are converging much faster that the

Innovations if you um if you want your model to be used need to be publicly available and explained and understood so that then the incremental changes

that one would want to make um make a lot of sense um I think that there's just too much vendor risk for using closed Source it requires

too much tooling it's way too expensive and it's unpredictable so meaning if you have some deterministic

workflow inside of a company and all of a sudden you're going to replace that with quote unquote AI you're going to introduce some

non-zero error rate at some point in that process right some model is going to barf at some point and do something that's unpredictable so how do you want to

manage that risk do you want to have it be a black box where all of a sudden your engineers are like trying to use some tooling can't figure it out so then they're calling their customer service

rep that then that's escalating and then you know 4 days later something is fixed well maybe that works if

you're um an inconsequential consumer application but if that kicks off a hippo violation or all of a sudden you have a sock 2 violation which then all of a

sudden triggers a banking violation there's many markets many regulated Industries in which that idea doesn't work so open source is just a

fast path to putting a bounding box on risk so I think like that's one big reason open source wins the second is just pure functional cost because when

you trade them off you can just see that the open source models in its implementation in part because of its Ingenuity are turning out to just be much much much

cheaper ultimately I think that the right approach is to be you know pretty promiscuous which is to say that there there's an intermediate layer that needs to be able to

understand what model is good at what and given a price tradeoff or a performance tradeoff be able to understand what to use at what

time and I think like when when our when this AI market quote unquote or this ecosystem or economy becomes more sophisticated that's what will happen

you know you'll you'll have um you know if if you think about um in routing right why do we have routing tables well you have routing tables

because AC for you know OPF or some protocol to work very efficiently you're essentially pre- caling basically the shortest path routes so that you have optimal delivery

of packets and that pre-calculation makes a lot of sense I think that there's an analogy it's not perfect but I think there's an analogy here which is that

models are effective the same thing right so if you're a hospital and you want to use a model for breast cancer prediction I mean you're Pro you you

want to run five of these things you know and you want to maybe do some mixture of experts maybe you want to do some Chain of Thought maybe you want to do both maybe you want to do some reinforcement learning maybe you want to

do some super fine tunes swap them around right um all of those things should be possible because maybe for them cost is less important but the quality

is that there can be no errors right we can never have a false positive in breast cancer detection or something today that's not really possible right you're very linear right

you pick one model and you go to the end that's just not how real world works if the answer is the patient should never be told she has breast cancer if she doesn't if the answer should never be

told not 99.9 which doesn't work if you're doing 5 million breast cancer scans a year year right that that you know 0.9 error rate is

ginormous you have to have more flexibility so I think the the pendulum will swing to open and I think the the the infrastructure that's really not there in production today but needs to

be there are these shims that allow you to you know eval and trade off and un you know effectively like like these these League tables for models for different workloads against different

price and performance trade-offs jamont the play Devil's Advocate on this like maybe this this is tough because I think like I agree with you in some sense about like The Innovation that's

happening in open models but is this not also like a bet against like all of the Innovation that comes out of these Foundation Labs like

do you not think that they'll be you know the GPT 5 level of progress and or is your take that like there will be that gbt 5 level progress but like the progress that we're going to get is

going to come out of like open source and not out of you know one of insert whatever lab is is your favorite well maybe I'll answer a different way which is where will all of

the investment that's being made by closed models pay off right I mean that's that's kind of I think what you're asking and I think that there's a there is a very large

economy for that um I think that you are going to find that narrowly tailored use cases

may not need this generalized approach that I'm talking about um and that the the incremental improvements from going across models

creates too much either latency or complexity or cost overhead that isn't justified by the quality and I do think that that there

are models and look to be blunt like of all the models that I've seen I think deep seek is really wonderful I think open AI is is pretty good you know I think llama has tremendous potential but

I think the Gemini models are the best um I'm not just saying that but I I I do think that they are the thing that now you know Google

needs to do is just to create that ecosystem where there are very narrow use cases that get to scale and that are in production I mean I think the biggest problem with our AI

economy to be quite honest with you guys is that we have a bunch of fake toy apps and and I don't really see anything of any scale or importance in

production um and so the longer that that is true the more the open source World wins because we're still in the toy phase the faster that the closed models

actually get to a handful of use cases that are legitimate and reliable by you know folks that are not just using experimental toy budgets but are actually building things that are

used that are felt by consumers or patient or customers that's where the lock into the Clos Source models comes in and that's what will justify all the investment and all the Innovations but otherwise I

think generally many of these Innovations are going to converge there's going to be a time lag six months nine months a year but that's not going to mean much

in 10 or 15 years I think off that too I mean kind of going back to the reasoning models and ingenic models as well like obviously there's a lot in production to your point maybe a

lot of fluff what do you think will be needed to require and unlock the full potential of those models is there like a benchmark as soon

as a great question your question is the million dooll question so like if you if you think about what like what are we really good at and this may be some very

heterodoxical statements that piss people off but this is what I believe like what are like repet and cursor co-pilot what are they really

good at I think they're good at supercharging an individual engineer and at the limit it'll make a a

1x engineer a 10x maybe it'll make them a 50x engineer I don't know but it's still only one person and so then you ask the question

well how do two people how do the three of us if we're working together actually use these tools together and it's kind of maybe figured out but not really and

if you actually apply the problem to a very very complicated repo for a big complicated product at a real company that has Real World um consequences like

look if if something gets screwed up in Google and you have to you know uh revert some push because it just like was totally borked there's a dollar cost

and somebody will show up and say hey this is how much that that stupidity just cost you obviously that's not true for a startup

so today we have these like really good tools that can Empower an individual person but the things that that individual person can create is

basically pretty nonsensical it's not a real app like if you could make Facebook blue okay now you're cooking with something right if you could make Google

search if you could make if you could rebuild GFS and big table just with a prompt okay great we've solved it all but the honest truth is that's not where

we are we all know that so I think that what needs to happen is the extremely and I'm not going to say this in a nice way this is tedious

unglamorous work to go from 99% to 99.9% that's going to take two years of

just grind it out unglamorous like you know you're just it's it's not fun work but what is that work it's about fine-tuning and training

these models to be useful and then really connecting to the connective tissue inside of an actual organization so that these so that this code is

reliable and and that that this code actually works as advertised and I think that people that think that that statement is glib and they don't understand the scope of the

problem because they probably never talked to a real customer and we talk to these folks every day we have so many regulated customers in energy healthare Financial Services

manufacturing uh and on and on life sciences and the biggest thing that I took a step back at 8090 I was like oh man this is a way harder problem than I

thought not having some guy use cursor to Jin up some code that's easy but all of a sudden now going through like 30 pages of security

protocols and this and that and everything else and then being able to attest that this now this deterministic system that we've replaced with a fundamentally probabilistic system at

some points along the way is going to behave as advertised at the level of accuracy that they're used to man that's a hard problem and again it is not a

glamorous problem it's not fun it's not sexy it's not what you go home and tell people about but at the end of the day if there's millions of patients on the other side are millions of banking

customers if you can't make it work for them AI is a toy economy do do you think you get um how much of that do you think you get for free like how much do you think like you

know the large model providers open source model providers are just going to like get you the last nine you know the last percentage of reliability versus like how much of that work are you trying to do yourself you have to I

think like that's where they can't go this is so my point is the models will get you to 99 maybe 98 99 99.1 but you know when you go inside of

a large regulated organization and try to convince them to do something the the amount of just like what I would call cartilage like not bone and muscle but all this stuff

that's required for an organization to run now maybe what happens is we just rip apart all the rules and regulations and none of it exists anymore and so and there's no liability from these

companies for violating any of these things that would change the posture where you can just ship code quickly move fast and break things but the real

world economy right the 100 trillion a year of global GDP is made up largely of companies doing company

things and those are all overseen and regulated in nonobvious and sometimes very busy confusing ways and if you're going to use AI to sort of like

streamline this entire world that's the body of work that's the hardest work and it just requires a lot of patience and blocking and tackling that's part of what you know we're trying to do I mean

could I have gone and built consumer apps absolutely I mean I've done it my whole career I just don't think that's the really big prize in AI if you're

trying to figure out how the economy will work in the future right you got to step into these millions of companies that exist and answer the question for

them how will I do work in the future and part of that is not some sexy demo which you can do it's the very gnarly part that explains to an entire group of people how they may or may not

lose their job and then to absorb all of the change management and then to understand all the rules and complexity and then to and then to do that in an automated way that

isn't people dependent so that you're not building building this kind of software on human scale but on AI scale at least the time scale it's just really hard so I think

the models are really excellent that's a very long way sorry Nolan I could have made that shorter the models are really good but the people who

win they'll be the end boss of grinding that's what it is this phase of like really building a successful company is just going to be grindy ticky tacky but by the way I love

that that's that's the organization I built at Facebook you know I took a bunch of really smarty pants whiz kids and I made them grind correct me if I'm I'm wrong here this is just a clarifying question so

kind of the argument is I don't necessarily need a model to quickly spin up an answer and give it to me what I want my model to understand is the steps and the workflow that goes into actually producing that output and trying to

retain it more to humans as far as your everyday work I'm a human I work in this job here is kind of what I do to get the output but is that what your kind of thought process is as far as training that next wave of right we don't need to

focus on necessarily the output but more so the workflow that goes into Crea that if you go into a company and try to observe a

process you're probably going to find multiple ways of doing it and so if you're going to replace that or make that more

efficient one of the most important things you need to do is to understand the requirements of that system so that whatever output let's just say that there's some magical software Factory that can take requirements and spit out

working production quality Enterprise code compliant secure all of that stuff how will you know it did what it did and maybe there's like some simple

answer but when you look inside of like these workflows what you see as years and years of humans manipulating badly designed

software so that if the CEO at the top thinks like something is being done by you know moving from A to B it's not what actually happens underneath like you

guys probably know this like I'm sure there's many of your peers at Google who do their job in ways that the Google execs who think they're you know it would be a huge gap between how they

actually do it some crappy tool that they have to live with but that's the functioning of the world economy surprise that's how works it's not Instagram you're not just like swiping

up you know it's not reals and all of a sudden trillions of dollars just appear in the in the economy it's horrible bad software sold by what I call this

software industrial complex where the the they they did a very good job they're very effective salespeople they're very effective at feature

design but there's been so much sprawl and so so many workarounds that have built up and so what are you supposed to do because now

like you know it's it's not just within a company it's across companies right it's how company a invoices Company B across the oceans so my my my point is that like

for AI to reimagine economic Prosperity once you throw that stupid buzzword statement out and you get into

the details the detail s are how the hell are you supposed to do invoicing for a company that has 8,000 employees and there's 50 people that literally are doing it 50 ways and you you can't miss

a payment or not receive Revenue you know because you have a 99.99% errror rate now because you replace some deterministic code with some model they're not going to care at that

point whether it's open source or closed Source they're going to think what the hell have you done right or you do it in a way which all of a sudden triggers a hippo violation and now you get an audit

request from CMS and that says hey we're pausing all payments until we figure out what the hell just happened what are you supposed to say oh but we use open source or you know we're really up to-

dat in our token payments what nobody cares so that's where we are I think to like really translate this into the real world now it can be done but this is what I mean by this is like a 10-year

journey and you got to start slowly and you just got to bite you know how do you eat the elephant just like one small bite at a time to me this this feels like you're

pitching why the value creation uh is potentially going to be done by startups like is is it not like you know all of the the cartilage example perhaps is

like all of this institutional sort of positioning weirdness that has accured over time and like there's no mechanism in large companies to like remove all that stuff and like the beautiful thing

about startups is you are Ground Zero you take the raw regulations you take the raw you know whatever it is operating manual and you get to start from scratch and it feels like maybe I'm

too optimistic it feels like AI has a chance of like really accelerating people who are trying to do that and I feel like this world of yeah I'm curious what's your what's your take on that I think you're I think you're right in

some places and I'm a little bit confused in others um like like if you go and you know if you're if

you're a startup obviously like if your business works I think what you what a startup should be doing

is you should be using AI to accelerate your ability to have your own custom what I call like a business operating system and the the analogy I would give for people that are

interested is what Tesla did with with warp drive or now what's called warp right so what did Elon do he hired a really talented

CIO and he basically told him I want an end-to-end system from raw material to the car being sold and even afterwards how the car is maintained and

serviced and what that team was able to do about a decade ago was build that endend system that's extremely custom and specific to Tesla

and what that allows them to do is just be very efficient so to your point if a startup that enters a business and is successful my advice to them is build

yourself that operating system now that requires some level of investment and commitment as well but these AI tools will at least shrink the Opex that you will have to

spend theoretically to get there but again it just brings up this case of like what is everybody else supposed to do and there I think that startups trying

to improve things for Enterprises are stuck in a bit of a dilemma which is that the traditional Venture model right like if you graduate

from y combinator or if you raise a series a from a well-healed investor what they want to hear is a very narrow pitch right here's this one thing I'm going to do this one thing

really well that's why you should give me 10 million bucks and then we'll go from there the problem with with this idea inside of an Enterprise is that all you're doing is you're replacing a million of these little things with a

different of these little things and and I don't think that that's going to work and the best example of that is paler because what paler says is screw

all of the software industrial complex this is all just worthless it's dumb and I'm just going to create a substrate and you can decide with us over time how you

want to build on top of this and hang you know like because at the end of the day like what are we talking about we're talking about a bunch of three- tier apps right we're talking a bunch of UI a bunch of apis right and a crud database

like it's this is not super complicated so I think that denovo startups that find a business model I would encourage them to build their

version of Tesla's warp I think the startup trying to build like an AI enabled replacement for some Enterprise tool I think it's a good thing to do the problem is you're going

to go up against a more sophisticated selling motion of companies that say look let's just start with an entire substrate and we'll hang different apps off of that many of which we will build

for you others which we can just enable via a platform and that sale will be done at the CEO or CFO level by somebody that's very credible

and so now the question is can this startup who's you know roughly run by an anonymous person compete against Alex karp when they're selling the same Fortune 500 company and I just

think that that's a that's a hard thing to assume that can happen well some some will and you know some of these people will become really exceptional but I just think that that's

unlikely yeah can we can we shift gear slightly to the um uh Ai buildout and like the hardware massive amount of capex I feel like that was a huge topic

of conversation last year it kind of felt like late 2024 or we were spending a little bit of less time talking about it everyone was sort of making their bet and then I feel like in the last 30 days

you know it's now reopened up on both dimensions actually this like question of the the massive buildout and like how much more there still needs to to be

done there pushing on this you you had a I don't know if it was a talk or what it was but this thread and nol and I have actually talked about this a long time ago um around just like the cost of

energy and I think you you had some bold takes like two three years ago about like basically the cost of energy going to zero potentially the the allocation and the cost of compute yeah yeah and

allocation of like resources in the Middle East because of how like much energy abundance they had and how manufacturing would sort of moveed to some of these places I'm curious now

like um for you to reflect on that and like how do you think this is all playing out at that broad sense but also like specifically in the AI ecosystem yeah in the broad sense I think that was

roughly directionally accurate um I think that so my statement a couple years ago was I was pretty sure that the marginal cost of energy and the marginal

cost of compute would go to zero and I say marginal cost because my focus at the time was you know look taping out a chip is still somewhat expensive depending on the process technology

there's still money right but meaning like like maybe like if you want to think about it like a like like the cost of an ec2 instance like is that just only going to get

cheaper yes is the cost of using a you know TPU going to get cheaper yes and specifically where I was focused was in the division between training and

inference and again just to put my bias on the table you know I was very invested financially and emotionally because I had helped start this company called Gro with a former coworker of

yours Jonathan Ross in a world of massive right so so I I I'm very biased in that point of view but intellectually I I

kind of came to this conclusion that when when I didn't know when training became a roughly solved problem you know so we were we were only working

on small amounts of improvements that most of the focus and energy would go towards um inference and that there would be so

much of it like 100x that it would effectively be zero and it would effectively have to be zero because again you know the difference between deterministic and probabilistic

code is that at some point you know probabilistically you're going to have some errors or you're going to have a bunch of run sometimes and so you're going to need it to be super cheap so that you can afford to make mistakes

right so that's that cost quality tradeoff the energy one was largely because I when you looked at the curves

of um how dependent the United States was on foreign energy it was a curve that was basically going to zero and when you looked at how much energy we

were exporting it was the exact opposite it was just sort of like inflecting up and that was when we had very nominal amounts of

specifically solar penetration in the United States and this was before we had any form of you know two years ago we didn't know that Donald Trump was going

to win so we didn't we couldn't forecast the Boon that will probably come at some point when he or they make it easier for you

know small modular reactors and you know or even traditional nuclear to get spun back up but he's already spoken about you know making it much easier to uh uh

you know extract and refine that gas and oil domestic domestically so that was one data point and the other data point was how quickly the Middle East was monetizing their oil

which is to say it's in the ground they were ripping it up and they were selling it and putting money in the bank right that's the process of monetization and when you put these two

things together it you could see that there was just this massive surplus of energy coming and with all markets sort of like the typical Supply demand imbalance

would that with so much Supply the marginal cost of an electron will effectively be zero and and I still very much believe it and I think

it's the only problem is neither of those two things have played out I feel even more firmly in it um now why hasn't it played out one is regulatory capture

in America um and the second is what we spoke about in that first section of our discussion which is that there really aren't any

real production apps at scale so there's no infinite need of of inference compute other than chat GPT other than Gemini

you know maybe a few use cases for like ads targeting at meta but we need 100x more traffic which means we need 100x more

winners and we're not there yet we're at the early phases where a few folks have a ton of usage and nobody else has any usage that's roughly the truth

but I think in the next few years you know again back to the first part of our conversation more open source Alternatives everything becoming cheaper

everything converging on quality more shots on goal right a few of them will get through a few of them will become winners that'll grow the

pie it'll drive more inference meanwhile the regulatory wheels will continue to turn and the Surplus that we have today so we're almost at the point now where

we are we are absolutely non-r on external energy Imports we're probably going to cross that curve at some point in 2025 which is an enormously important

geopolitical thing um tremendous for peace you know there's many positives there that'll tip and

then when the land leases on federal lands get accelerated and fast per and this Energy starts to come online probably two years later 2 and a half years

later it's just going to start a tital wave that you won't really be able to um stop and then if you compound that with SMR so all of these things are

converging by the end of this decade I think that we're going to be um a wash in inference that's

effectively zero and energy um so yeah I mean so that's sort of where my head is at in terms of specifically AI we're actually so if you if you think

of that's where the puck is going there's one very narrow problem in AI that is not solved yet very well which

is around Cooling and specifically the process architecture that we use creates some fundamental constraints that generate enormous amounts of heat so especially if like you know you're

you're using euv you're at 2 nanometer the reality is this stuff runs super hot and so all this liquid cooling stuff is very

heavy it's expensive and it's pretty gnarly like it's not very good so there are some movements of

foot to really think about cooling using inert substances using gas um being able to have an heat exchanger that's highly

efficient that will also be a limiter and the re and and in fixing that is in critical because you get into this very negative Loop otherwise which is if you have only like the most advanced

Hardware that then has the most complex and demanding cooling characteristics that then only has you know a really gnarly solution

where if like all of a sudden like you know I don't know just like a plane accidentally flies into this data center you have this like crazy noxious stuff

like leaking all over Tucson Arizona that can't be how you know like we we deal with this stuff but even if that that kind of an

event didn't happen I think that you have a constraint of energy and cooling these things so that it it dictates where they can be built and I think you want more flexibility than that like you want to have Edge Computing everywhere

so you want to have much much lower power profiles which you can solve in a manner of ways but the combination of a simpler process architecture I don't know if you guys saw but you know there was some

stuff that said that HX just actually found some workarounds to euv that they think is even more efficient so you're not going to need these super crazy asml

machines um to do um to do these like you know crazy crazy crazy extreme ultraviolet you know deposition schemes um and then there's some really

interesting Innovations around cooling um I've been working with actually a guy on on this for like the last couple of years we're sort of at you know

trl4 um which in that engineering parlons just means that you know we're almost basically at working prototypes of this thing which is completely inert and uses CO2

gas and part of where I learned learned that this problem existed was from grock because like when you're deploying inference it's like man this is you see it's it's just you can't expect that energy is going to be everywhere and

then if you want to have small data centers literally dotting every place in the world so that you can have hyper hyper like millisecond fast inference

everywhere you're going to have to solve these practical problems and so you know I think there's some very very minor things that we need to fix for the AI

scaling but the the infinite compute Infinite Energy thing is uh I think we're like three or four years away from just complete infinite abundance and I'm guessing these very

long answers I apologize for these very long no no no it's good and I think kind of you know hitting on global energy and more so hitting on social capital and again I think thesis is you know over the next decade Technology Innovation

will create multiple 10 trillion dollar markets I was looking at some of the different verticals or domains obviously deep Tech Global energy which we just talked about uh Creator economy but one that stood out and the actual wordage of

it stood out is two was life sciences um and I think it's quoted as where discover or drug Discovery is moving out of the realm of phds and chemists into the hands of computer scientists which

again I think is a very I I totally agree I think it's could be alarming for folks who might read that on a piece of paper they're like wait what um but would love to get your yeah reaction to whether it's organizations or

companies you're currently talking to but like is there a means for trying to you know marry the two Hand inand where you know PhD and a computer science degree is kind of the new standard for

folks trying to build in that space like yeah just kind of curious to see how that plays out great I wouldn't exactly say a PhD but I would agree with you

that if you could marry biology plus some form of machine learning or chemistry plus machine learning or physics plus machine

learning I think that's like the Dark Art ninja skill for like the next five or 10 years I'll tell you what we've learned from our life sciences customers

um they have a super complicated job so the first thing about life science is that I I guess I kind of knew but what I

really learned is man they have an extremely complicated what I call Capital allocation model so like if you think purely as an

investor you know if you think about like generating a certain return the next thing is you would try to figure out how many bets do I need to make what are the distribution of those

returns to generate that right it turns out that life science is roughly is that way so it's all it's a highly probabilistic system there's all these

chained probabilities and so they have a very sophisticated Capital allocation model so life sciences as a result of that business model

construction is completely open to the kinds of things that AI can do because it allows them to shrink rink

the error rate in any of these steps which then as a result Cascades through their business process and improves the likelihood of overall success so I'll give you an

example let's say that you know they they say I'm and I'm I'm not scientific enough to get these right so I'm just going to make these up to to to make the

point um we have a new Target for small cell lung cancer okay cool well we want to attack this specific protein okay great

right the minute that they maybe start with that because you know lung cancer incidents are high the reimbursement rates are very good they want to go after that market for whatever

reason at some point what they Greenlight is like a thousand different potential proteins and today what they'll do is

they'll run those thousand through a lab now they'll may be down select to the hundred most likely but it takes a lot of time and precision to collect

that data and then to analyze it the first place where it may and you have a an error rate even still right even you take those thousand and you down select to 100 then you down select to

10 now they're going to have to run some sort of initial clinical validation on those 10 to get down to three and then maybe they'll do an IND on those three and then start the FDA submission

process on one but then there's a you know 50% or 60% failure rate anyway so maybe they'll start two and run them in parallel or maybe they run three so they're guaranteed one to

work super complicated now you could take Ai and just say let's run that first really important phase in silicone okay now all of a sudden you

run a thousand you know we'll use Alpha fold or maybe you know we'll do a deal with isomorphic and we'll get access to Alpha fold 3 um and now we'll do all of

that so that now we're much we have a much higher probability of getting you know down selecting to the three that are really going to

work and then when you when you look at it through that lens and you follow it so then in the indd process there's a whole thing then

in Phase One there's a whole thing and there is just an enormous amount of room if the biologist and chemist knows just small mome of machine learning a simple

example would be you know across all the number of test sites how can you distill so that you can get all the submissions to be extremely accurate there was a retrospective study that was published

uh I I saw it an archive but it basically said that there was like a just in collecting the data and assimilating it for submissions uh there was like a point5 or 6% error rate just

mathematical errors now if you think about that over tens of thousands of studies and you know how much does the Pharma industry spend 50 billion 100 billion a year on R&D you know how many

billions of dollars that's flushing down the toilet which is like humans paper pencil you know so that's an example there are all these clinical submissions

that's an example if a drug gets approved there are all kinds of ways in which doctors need to uh interrogate about the features of the

drug and you know all of the things where it could go right it could go wrong the Pharma company has to have an infrastructure to answer those questions all of these things are things

at different forms of models whether it's protein folding on the way in to you know a simple llm that can do construction of documents to a finely tuned one that can

answer so I think that of all the categories I've seen I think that life sciences is the is the most obvious because they already

deal in probabilities right so going back to that other they already deal with the stuff that it's like yes most you know there's an error rate a bank is much harder like there's

you know I mean if the error is in their favor they like it but my point is like you know there can't be error rates on account opening but these guys are much more

open-minded to the fact that there can be an error rate in protein folding because they they live with that every day and now what you're doing is you're taking a 10% error rate and shrinking it to 1% because you're moving it from

humans and wet labs to Alpha 3 and and you know a bunch of computers that Demis manages like that's that's a huge win so yeah I'm a I'm a I'm really convinced in

life sciences you're going to see a lot of innovation I think what the first thing you see to be honest is you're going to see just a lot of drugs get through the cycle Faster by the way that's before the FDA says we are going to have like a

digital FDA agent that auto reviews these submissions and gives you a 20 right now you have a 6month turnaround time so you have a six-month

shot clock and depending on which pathway in the FDA you can get accelerance like you can have essentially these coupons that you can use to kind of get priority review get to the front of the line all of that

could get replaced by you just submit it and in 24 hours you'll just get an extremely detailed response from the FDA a digital reviewer so think the the cycle in life

sciences is is is one where the errors in probabilistic systems are probably well tolerated enough where you'll see you'll see commercial successes fast

yeah chimo this this has been awesome I want to be super respectful of your time the the sort of closing questions um you can go rapid fire you can noain as long

as you want to it's totally up to you uh something that you want to see happen in 2025 something you hope doesn't happen in 2025 and then actually really tactically like what does your AI

productivity stack look like like what are you using the tools and stuff like that that you're using every day to sort of you know do your job and and live your life okay so what do I want to have happen

[Music] um so I I'll I'll give you maybe a a broad macro sentiment and then very Tactical

the broad macro sentiment is just um I think in the we're in a very fragile point in the United States economy where

um we're either going to establish a new set of rules to govern the next 25 or 50 years and at the center of those new rules is going to be

this idea of transparency that is an enormous unlock I think for business for real business businesses I think it's actually very bearish for

assets but I think it's really good for businesses transparency is great like assets you can see through and notice that the building is not uh occupied that building is worth a lot less right

so transparency is bad for assets you know you have some shitcoin and all of a sudden you realize that it's literally nothing that's bad for assets but it's really good for

businesses and what's incredible is I think that there's things happening in the United States government that could really make that a an expectation not a

nice to have that's my big 2025 wish in terms of me very narrowly at 8090 I want to see the the first real moments where I can credibly look you in

the eye and say we've solved some of these second decimal and third decimal problems where

it's not just curs and instead like something is really inflecting in 8090 where now you know I

can scale to 500 customers next year and really like help them improve what they give to their customers and to the world right now I

can't say that so that's my very narrow wish for myself is and again it's just like it's just hard man just like grindy

thankless work what hope doesn't happen I hope that the market does not uh puke when the United States tries

to refinance 10 trillion doll of debt I think we need the I think we need the we need the economy to be healthy so that

we can get all of these things done uh and then my own personal stack um I use a mish mash of many

things so um I use open AI I use Gemini um I use deep seek for my

information gathering and assimilation process um when I'm hacking stuff together I'm actually systematically moving through all the tools so I

started sort of like last year with the obvious ones like repet and cursor now I'm in this kind of like uh long tail of stuff um I mean I can tell you some of

the I'll tell you what's on my hit list right now hold on um um hold on I just stopped using bolt and it's it's written here so I'm

am going to try pair uh and then after pair I'm moving

to idx so those are the next two kind of like tools that are on my um um hit list to try to be quote

unquote a more productive engineer now I'm a terrible shitty engineer so I don't I can't claim to to do anything well but um it just allows me to see

kind of like the state of play um but at the top of the stack I I I kind of use everything I I use clad as well um um I I mean for code generation I just think

like going directly is there is like really good um uh and then now that the the Deep research things are out I mean you guys

have yours in kind of um which has been very good and now because of open AI I'd like to play a little bit more with deep research there's some narrow things like you know I'm not a very Pro patent

person but there's a handful of things where I was like oh maybe we should actually just patent these things this is for the HVAC company and so I'm trying to figure out whether we can use

it there and anyways yeah so I I I I I use a mishmash but I wouldn't say that I'm particularly like good at it in the sense that I don't feel like it really change like I get more information like

it's the same magical experience when I started using Google I feel smarter but I personally can't say that it makes me 10x better at anything I don't know if that's if

that's wrong to say if that's I don't feel 10x better at anything I think there's an underlying truth and thread of of some of this of like how maybe this is our just like

human uh you know weirdness of how we see ourselves the the two tools that you should check out if you're done with bolt uh lovable people love lovable lovable I've I've done I've went through

that I like it yeah vzer did you try VZ tried it already yeah vzer is great um you're you're you're on the frontier I love this jamat this was a ton of fun thank you for spending the time uh and

and answering all these questions we we got a lot more hopefully we'll have you back next year we can follow up on a bunch of this stuff uh and thank you again thank you