Why you should sell your Nvidia stock right now! #notfinancialadvice

Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
This is why you should sell all of your Nvidia stock right now.

(00:07):
#NotFiancialAdvice don't ask to listen to us, but before we tell you why, welcome to
the #1 generative AI Meetup Podcast.
In the world, with your hosts, Mark Anciasharc, straight from Silicon Valley, we are bringing
you the latest and greatest tech news and our general conversation.

(00:30):
Together we run a Meetup in the Bay Area in Silicon Valley.
We meet a lot of really cool people who are just making the coolest stuff and have the
best ideas of all around generative AI.
Together we are bringing all of those conversations to you.

(00:53):
This is basically like all we talk about, but a little bit of housekeeping before we get
into all of the Nvidia goodness, we are going to have a Meetup at this company which will
potentially take the throne of Nvidia at some point.
Later this week, on Thursday, the 24th at 6pm in Palo Alto, we are going to post all the

(01:16):
details on the Meetup page and in the description.
It will be great, it will be at Samba Nova in Palo Alto.
The spots are filling up fast, so we hope to see you there.
But anyways, yeah, that's enough to shank how you are doing today.
I'm doing good.

(01:38):
I'm actually just looking, so this is something Mark and I did a couple weeks ago and we went
and got our DNA whole genome sequenced and I'm just looking at my data and uploading it to
sequencing.com to hopefully get some more insights than what the service provider that I got the

(02:01):
sequencing from has given me.
This is like another pretty cool application for generative AI because we have Alpha Fold
which has been doing pretty cool stuff with drug discovery, understanding, obviously
solving the protein folding problem where we can figure out what structure these proteins

(02:22):
should have so they can design these drugs and pharmaceutical industries to create things
and remedies for ailments where the drugs bind to different sites in your body.
Along those lines, I've been thinking about what if you could make an agent that you
have an agent to understand all of your medical information including your DNA, genome and

(02:49):
have this agent go through research articles and then find therapies and solutions for whatever
things that you're at risk for.
Yeah, I feel like that'd be really cool to have an agent to do that.
So little story is I got my 23-Me data so I did that test a few years ago now, maybe three

(03:13):
or four years ago and that actually is not the whole genome I learned this.
It's like 0.1% of the actual genome but it's still pretty useful.
It can tell you your ancestry data and can tell you some things you're at risk for.
I took that to my doctor and I showed him the data dump.
I took it to a site, the name of the escaping right now, it's called not Polaried Biss.

(03:38):
It was a Polaried Biss.
Anyways, not important to the name of it but it was a site that would interpret all the
data and then I showed that to my doctor and he was like, "Oh, I mean this is cool but
what do you want me to do with this?"
This is way too much information.
He was like, "Yeah, we can run some extra tests for what it says you may be at risk for

(04:01):
but ultimately we don't know what to do with all of this yet."
I feel like a doctor, they're busy, right?
What you really need is some sort of data scientist or what is it called?
Bioinformaticist or something?
I think that's what you would really need to help interpret your results.

(04:21):
But then a bioinformaticist is not a doctor so they may not actually be able to do anything.
What you really need is like, you need somebody who has the tender love and care of a doctor
plus a bioinformaticist who understands all the statistics behind it.
Plus you need to know what's going on with all of the latest studies and rolling medical

(04:48):
procedures and all of those things together is a lot.
So I feel like an AI agent could do that maybe really well.
And the fact that a doctor can't follow you around 24/7 and help you take action in your
day-to-day life that will result in healthy habits and a good outcome for you long-term.

(05:11):
But an agent can, I think.
It can listen to all of your biometric data, look at your lab results and give you little
nudges along the day and give you recommendations as new research comes out or as there are new
suggestions from the FDA or other influencers that you follow that you may want to incorporate

(05:34):
into your regular routines.
Yeah, 100%.
So one thing that I was actually thinking about is the idea of doing gene therapy.
So, you know, like a crisper, right?
So it's like, my understanding is that crisper is like a thing.
I don't know exactly how it works, but it's like a thing that allows you to modify your
like little parts of like your DNA.

(05:56):
Now, because like, I guess like every single cell in your body shares the same DNA.
And then if there's any like, let's say mutations within that DNA that like potentially cause
like a good thing or a bad thing.
So actually, I feel like we're talking about this.
Let me tell you what I know about like DNA.

(06:18):
So basically it's like in every single cell of your body, there is the DNA, right?
And now the DNA is the thing that makes you you, right?
So it's like DNA is like just like a long string of like ATC's G's in like a different pattern.
And then based off of that pattern, that like makes you like, you know, that makes like maybe

(06:43):
your eyes like a certain color, your hair, certain color.
Maybe it would make you be more or less likely to like develop diseases in the future, right?
And sometimes like you have a mutation or like a difference in your DNA that could cause
like a positive effect, right?
But that also could cause like a negative effect, right?

(07:04):
So like some people they have like mutation that makes them more likely to let's say get maybe
certain forms of cancer or something like that.
So in theory, what you may be able to do is you could use the thing like CRISPR to modify
part of that DNA.
So like let's say there was a mutation which you knew like, oh, this mutation will make it

(07:29):
so that like I'm going to get cancer like really early on.
And like this is like the direct cause of that.
So hypothetically what you could do is you could go in and then like inject some CRISPR like
tell to identify a particular mutation and then like I guess you could just like either
like cut that mutation out or like go and then like make it so the negative mutation is

(07:55):
turned into like a positive mutation or something like that.
So I feel like that is possible with like gene therapy if I understand it correctly.
But and there are like different trials that are going on like all over the world that
are doing these types of gene therapy on particular mutations.

(08:17):
But the thing is is like there's so many like potential like mutations that you may have.
So for reference when we got when I got when I was looking at the DNA file, the total amount
of like DNA it was like a 10 gigabyte file and then the mutations from the reference genome.

(08:38):
So there's a reference human genome that I guess like all the scientists that came together
like okay this is like what a human looks like.
I don't know and then like all of like the differences from the reference.
So it's like let's say the reference I don't know like a white guy who had like blonde
hair and blue eyes right but let's say like you're like a Shashank well you're Indian right

(09:03):
so it's like you have like black hair and brown eyes I guess right.
So it's like your DNA would be different right.
So it's like you'd have like different patterns so like oh like my hair is black and then
like I've got brown eyes right.
So basically it's like there's there's like a file where it tells like all the differences

(09:26):
from the baseline right.
So then like you can go and then say like okay like this delta file that's 400 megabytes
roughly.
So it's a lot smaller but it's still like pretty big in a lot of data process.
And then like every single like potential mutation like adds like some sort of like risk
or benefits to you.

(09:46):
So hypothetically what you could do is you could have like an agent well first you'd like
label the data figure out like where all the things are and then you can have an agent
go and then figure out like exactly what you should do.
It's like oh the agent would be like hey I noticed that you have this particular mutation
that's going to give you cancer and I found this study that's going on in this other country

(10:11):
and don't worry like do you want me just like go and sign up for you like to enroll you
in the study and we can I can just send over the DNA to that like research or research
organization and then they can get in touch to you once they have their gene therapy or
even better.
It's like the agent might just be able to like send the information and then like they

(10:32):
could just like automatically send you the gene therapy to your door.
I mean that like that's a possibility that we could be seen like really quickly.
Yeah that would be really cool.
I think you touched upon a lot of different capabilities in that kind of like a hypothetical
agent.
One of the benefits of an agent is that it can do a lot of work for you, a lot of abstract,

(10:55):
ill-defined tasks and kind of figure out what needs to be done by itself.
But on the other hand there's also limitations because of the nature of LLM's there's a lot
of hallucinations, lack of consistency so every time you run the agent you might get
different results.
So I think I can see a combination of agentic behaviors and plugging into one of these services

(11:20):
that has like a concrete deterministic programmatic solution where it just does the same thing
over and over again really well.
So maybe like cooking up into one of these services that analyzes your DNA with like traditional
science and using an agent to plug into multiple services and look at your overall health data

(11:43):
from like a holistic point of view apart from just the DNA like your diet looking, hooking
up into these nutrition apps and seeing what you eat, how your body responds to it.
If you have like you know you monitor various vitals, blood tests, glucose, cholesterol, etc.
Being into that data may be getting information about your other kinds of reports x-rays,

(12:05):
MRIs or whatever else you may have.
And then try to construct like a broad overall picture that it's kind of hard to do for one
service at a time because each service has one snapshot of your data like my Fitbit.
It's really awesome, it's really you know insightful for day to day stuff like if I'm not active

(12:31):
it'll give me like a little nudge telling me to go get up and move around.
But imagine if that was hooked up into an agent and it was like aware of this little gene
mutation that I had that makes me more susceptible to I don't know let's come up with some random
example cholesterol or something.
So if my meal tracker says I'm eating this you know greasy cheeseburger and maybe I'll

(12:55):
give me a little nudge and say okay maybe you can replace that with something else or give
me helpful suggestions as I incorporate more data points.
Yeah 100% that makes a lot of sense actually.
And also like you could even like incorporate let's say like some sort of meal tracking along
with your DNA because some people just like process things differently right so I think I've

(13:20):
told the story before but a while ago like both my wife and I we got glucose monitors and
I ate like a big meal of like potatoes and she did two we ate like the same thing I think
I had like five potatoes and she had like one potato.

(13:42):
But my like I guess ancestors are from like the Poland, German you know kind of like Eastern
Europe area where like they eat a lot of potatoes but all of her answers ancestors are from
Japan because she is you know from Japan or like maybe I think like some of her ancestors
and maybe from like maybe came over from China to Japan or whatever but it's like all more

(14:07):
like East Asian right so like the potatoes they barely affected my blood sugar at all
like I had like a very like moderate spike but like her blood sugar you know you would have
thought that she'd just say like a big bowl of candy or something like that just like you
know shot up like like crazy so it and I feel like a lot of that probably it just has to

(14:32):
do with like you know like your heritage and then like I guess like ultimately your DNA
so like if you like you can solve all these things be like hey look like maybe like potatoes
aren't good for you or maybe they are good for you like it just depends on like you know who you
are and like how you process your thing so like maybe like these agents or I don't know some sort

(14:53):
of like fitness app could just like tell you like oh yes these are the types of foods
maybe you should avoid like oh like maybe you should avoid like glucose or like lactose because
like your body doesn't handle that well or maybe it's just like oh you know the covers you want
because your body handles that fantastic yeah so like with AI and medicine or health and wellness

(15:14):
I see two perspectives one is nudges suggestions and things that AI can help you do to change your
habits day to day because realistically I think that is the most important thing for your overall
long term health and wellness changing your habits so that you just do the right things from the get go

(15:36):
but oftentimes I feel like the you know media public perception and even maybe the stock market
focuses more on the treatments and pharmaceutical industry drug discovery and all the millions

(15:58):
billions poured into treating diseases one that once they've already manifested in your body
and I mean honestly you know but you need to tackle both problems one someone has an ailment it is
a traumatic event in your life and having things like alpha-fold alpha-proteo and all of these

(16:24):
different AI tools to help you discover drugs to target that specific disease that's awesome too
and along those lines like what do you think is the next frontier for things that we can generate
because looking at some of the research from deep mind alpha-fold is one of the more popular

(16:45):
well-known ones apart from you know Gemini and LLMs that we've been talking about
yeah I was also looking at some of the other generative generative models that are out there
and they have alpha chip another one for synthesizing new chemicals they have alpha geometry

(17:10):
and geometry yeah is that for like doing geometric proofs or something?
yeah to solve complex math problems wow so I mean open AI has been trying to consolidate all
these different modalities into one large model with text image video speech but deep mind is

(17:32):
trying to tackle really complex scientific problems by acknowledging that at this point in time
in the early 2020s it's hard to do everything with one big model and maybe getting the really smart
scientists to consolidate all their information and knowledge into hyper-focused models might be

(17:57):
worth it for that specific domain yeah I think that like some of these hyper-focused models to
generate things could be potentially super useful so we were talking a little bit before this but
I think like material research could be like super interesting as well as like chemical research too

(18:20):
right because the thing is is like I guess you think about like in the past the only way that we had
to kind of like make new materials was just by trying to like go and like mix material together
like melted yeah burn it heat it I don't know like freeze it and various chemical reactions

(18:40):
pop it right anything right so it's like I think a lot of that was just kind of like trial and error
just like oh let's like yeah see what happens when we do this and let's see if there's anything
interesting and I feel like most of the time there probably isn't something interesting but sometimes
there is something interesting but it might take you like 10,000 tries to get something interesting

(19:06):
maybe you could try like you know I don't know one try two tries per day if you're lucky maybe one
so it's like let's say you did like a try per day and if you 10,000 tries to do something different
interesting like you might go your entire career and then never find some sort of like breakthrough
material like you might just like try something every single day for an entire 40 year career and find

(19:30):
nothing of value I heard that there's a similar parallel in medicine to people spend decades just
working on figuring out the structure of this one protein and then alpha full came out and then boom
it just it's solved all those problems for all those scientists putting in decades of work yeah
so like no I don't exactly know like how like protein folding is like beneficial to scientists

(19:56):
but like if if a scientist like thought it was worthwhile spending like 10 years trying to solve
that issue it is very important it must be super important because I mean like protein like that
that's like part of your body building blocks of life yeah seems important right and like how
it's folded within like a cell like pretty darn important yeah but like you know in addition to like

(20:20):
you know health and stuff right like thinking of like just materials right like think of like how
much like materials have like changed our lives like oh my god this iPhone and MacBook that I have
right now is one of the most heavily engineered items on the planet and has like cutting edge
material science and like the deformation and how rigid yet light it is and how they shove intricate

(20:49):
circuitry into this to power amazing features yeah I mean like because the Mac is aluminum or
aluminum as like Johnny I like to say and you know it's like strong yeah like light I mean it also
cheap too like the fact that like we're able to make like aluminum like it used to be so expensive to

(21:13):
make aluminum and now like I mean aluminum we have aluminum foil people just like use it to like
cover their food and just throw it away like imagine you did that like gold or something like that
it would be like crazy and like I think it was like 100 years ago 200 years ago something like that
like aluminum was like just as value was like gold or something like that now it's like we have

(21:34):
like processes to be able to make it and we can literally just like use it and then throw it away
it's amazing it like think about it like plastic like think how awesome plastic is like I
it like can can hold your like liquids or your food and it doesn't leak and oftentimes like we'll

(21:57):
have like plastic and then it's so cheap we just throw it away like people like the plastic bags
plastic straws I mean not to say like throw it away these things aren't good like I mean I
mean obviously reuse it but like it's really like kind of like a feat of engineering to say like wow
like wow we actually like we're able to to make this like super useful thing and then like get

(22:20):
it so cheap if people just like oh we only need it that that is kind of a double-edged sword plastic
was one of the marbles of human innovation in like what is it 70s or something or before that
and now it's coming to bite us in the ass again where it's you know polluting the environment it
doesn't degrade that well which is both a benefit and a con and now we're finding micro plastics in

(22:44):
our entire food chain and our bodies and so the ability to create new chemicals and understand it
how these chemicals interact with us the rest of the world and different materials is really
important so deep mind published paper last year or published a research blog article where they have

(23:08):
a tool called GNOM that's like genome but like without the yeah but I'm not sure why they called
that because it's used to create or find 2.2 million new crystals and like hundreds of thousands of

(23:29):
stable materials that could power the future of technology so I think right now EVs are
have been getting popularity for the last few years and one of the challenges is making the lithium
ion batteries because there's like a limited supply of some of these precious metals which are in

(23:50):
conflicted territories in Africa and now China is starting to buy a plan in different places and try
to mind these things and being able to find alternatives or even improvements over this technology
with new materials would like unlock a lot of cool applications make things cheaper

(24:10):
have like more capacity for the same battery maybe charge faster
so like every single industry can benefit from these kinds of things 100% you know do you remember
it was like a couple months ago when we six months ago there was that like rumor about some people
who figured out there's like super capacitors something like do you remember that I feel like

(24:33):
there's always something every few months anyways it's not that important but I feel like all you need is
like one new material that like can literally change the world that like I can bring about so like
I mean we're talking about like lithium ion but like I feel like you know we're art technology

(24:56):
has improved a lot and I think like battery technology has improved a lot but like I feel like
it's still not good enough right like imagine you could have like a battery on your cell phone that
like lasted a year or something like that that'd be cool that'd be really cool right
like and like right now like a battery takes up like a really large percentage just like the size

(25:20):
of the phone right it's like but like what if it was a way to like make it so like maybe you could
just have like your screen be like a solar panel and then like it charges the battery of the phone
and then maybe you'd never have to charge it right like I feel like battery tech is something that like
could go like a long way right I mean like it's getting better but like I feel like maybe if there

(25:43):
was some sort of material that could like store like energy like that'd be amazing but I mean like
not to store energy but store energy is like really efficiently and then like you could also make
your phone like really tiny right like I mean like it could be like a little chipy put in your head
or something like that it could be like so small and even apart from batteries you know

(26:07):
coming closer to the theme of the podcast there is one special model just for chip design it's
called alpha chip it focused on like accelerated and optimized chip designs and I know that
Nvidia also uses a lot of AI to help streamline the process I remember Jensen talking about

(26:31):
agents that they use for everything throughout their pipeline and designing the chips and
optimizing the layouts and so having these things improve the chips which in turn allow us to run
these models faster to improve the chips faster it's such a fortuitous cycle that is just going to

(26:57):
explode I think in the next few years like the chips are just like everything this flywheel
yes that you know speaking of our intro and video has it's got its hands in every step of this
life cycle of chip design to manufacturing to and users using it and the entire developer ecosystem with Kuda

(27:27):
what do you think about their position in the market here so I think it's true I think that
Nvidia is like doing really well and they have like an incredible mode so
for those those I don't know Nvidia powers a lot of the LLMs and large like actually just not

(27:52):
even just LLMs but they just like kind of power like a lot of what is being used today so a lot of
people will use Nvidia's library called Kuda and if you want to use Kuda which most people will
for their machine learning tasks you have to use Nvidia GPU which kind of locks you in plus like

(28:16):
Nvidia is sort of like a general purpose so like there's a lot of like pretty much like anybody
who's anybody who's like making like an AI model it's going to be built on an Nvidia GPU it's not going
to be like AMD it's not going to be Intel it's going to be Nvidia because a lot of the software

(28:42):
libraries are specifically optimized for Nvidia now I think that everybody kind of realizes that
like Nvidia's in their really good position and I think the stock reflects that but I think that
slowly there is coming to be like a lot of chinks in Nvidia's armor I guess you could say

(29:03):
and I think that Nvidia is the top dog and right now but I think that it slowly being eroded by
a lot of different things right so as I mentioned like the software that a lot of people use
as Kuda right but you can bet that AMD is working really really really hard to be able to make it so

(29:27):
that you can replace Kuda and then use their GPUs because I think a lot of people would really like
to move away from Nvidia because one of the problems in Nvidia is it's one of the most expensive
GPUs out there right so like just because of the supply constraint too they're not able to make enough
for the amount of demand that's out there so it's kind of driving up the price yeah but like I was

(29:48):
looking and for like a similar price like erb I should say like an AMD GPU which has like similar
specs to like the Nvidia GPU was like a lot cheaper it was like half the price or something like that
so I don't know you could like look it up but like the AMD like is able to match a lot of the specs

(30:09):
for Nvidia it's just like it doesn't have like the same like ecosystem and software support but
that's being fixed so like uh uh olamma I think a few months ago they announced that they wanted to
have AMD support so like we're in the past like you could only run uh their models on Nvidia now

(30:33):
you can run it on an AMD GPU yeah um just looking at some news articles here apparently Nvidia's
H100 chips cost four times as much as AMD's competing MI 300X chips yeah so I mean four times that's a lot
that's that's big right and I think that like sure they can do it um but I think the moment that

(30:56):
and people are working really really hard to uh remove that uh dependency on kuda right so it's like
um like a lot of people use this library called pie torch pie torch works really really well with kuda
but it's like a lot of people are also working really really hard to make it so that like pie torch can

(31:16):
run well on AMD GPUs um uh I was talking with uh George Hots uh a couple months ago uh so George Hots
like the guy who is uh head of uh comma AI and now he's making a company called uh tiny corp so tiny
corp is basically just like uh leveraging AMD actually this was George Hots' whole thing he's like um

(31:41):
Nvidia is really expensive like I'm gonna make like really cheap uh hardware um and uh he's like I'm
gonna use AMD he's like I'm not I don't have he's like I don't have like that much hubris to think
that like I could go and create um some sort of company that like competes within video he's like
I'm just gonna buy uh AMD GPUs and then like make pre configured uh boxes and then he's also making

(32:07):
software that runs on top of that um so he made uh a thing called um he's working on a software called
tiny grad which is able to like go into those uh AMD uh work really close with AMD uh GPUs and uh
ultimately he wants to be able to make it so like you could almost like use tiny grad as like a drop

(32:29):
in replacement for pie torch um but like it's be like way faster so basically it's like you know
he's kind of like approaching it's like hey look like um Nvidia is super expensive like
like a lot of this is just like a software problem like uh I mean if AMD is able to figure that out
like they're gonna make like a lot of money so it's like lots of people working really really hard

(32:51):
to try to like take Nvidia as their own um so I think it's I think it's gonna happen and um I think
like right now like people are from like all sides just trying to like be down Nvidia and I think
that is going to reflect in the stock price therefore gonna this might be like the time to sell
your Nvidia stock bring it back full circle to the look sell it all sell it all now um this is

(33:19):
remember this is not the national advice do your own research we're just two guys uh sitting here
over the mic um but I mean you you do make good points I think uh there's a lot of other companies
trying to chip away because this is where most of the money is being made not a lot of companies
have figured out how to monetize um these LMs I think there's way too many companies providing

(33:45):
LMs for free um I'm assuming Facebook is like subsidizing these models heavily
feels like they're just providing it for free in all these uh in all of their products um I think
you know Google is trying to figure out the revenue model there um I've heard proplexities trying
to introduce ads in their LMs um opening eyes raising tons and tons more money um although I do

(34:10):
think opening eye does have one of the most popular consumer products where it has a lot of
paying subscribers but a lot of hardware companies are making a lot of money and yes Nvidia does have a
mode with kuda which is one of the most popular frameworks to accelerate um your ML workloads

(34:31):
which just makes it easier when people are building these models but once you've really built the model
you can deploy it anywhere you can use AMD you can use syribras which is another
um chip provider which focuses on wafer scale computing they just make like massive massive chips
which are 10 times bigger than the Nvidia chips um or uh sambanova or groc uh yes sambanova yeah

(34:59):
just uh we're gonna be chatting with them next to earth actually Thursday yeah 24th
palo Alto we'll see you guys there hopefully we'll have some uh good uh technical talks that we can
bring to you guys um we're gonna try to get a chat with uh someone from the hardware side someone from
the software side um to give you some perspective about how these smaller companies that are focused on

(35:25):
one specific niche for really fast inference can you know take a piece of this uh market because
as we plateau um kind of with the capability of these models um people are gonna want to just run
these models multiple times in like an agentic workflow to get more value out of it because if you

(35:52):
just ask chat gpt1 question you're gonna get a reasonable answer but if you have it think through
its steps um have some kind of a chain of thought reasoning um where it just spends more compute on
the problem because you know you ask a simple question um where is the where is mountain Everest
any tiny model low one billion parameter model can probably give you the answer really quickly

(36:19):
but if you want it uh for example understand your health from a holistic perspective and you dump
it all of your medical data blood test x-rays etc you're gonna want that model to think through its
steps and spend a lot more compute on that so as we build more and more complex agents um we're gonna

(36:43):
spend more money on inference where these smaller companies with their competitive edge uh what was it
like someone over is the fastest inference and uh it's uh it's uh it's like 10 something x faster than
video chips right something like that i don't remember the exact like multiple but massive massive
improvement yeah it's way faster and i think that um like because we're like the speed i mean

(37:09):
it goes to show that like it's also like oftentimes like cheaper as well right because the thing is
is like you in video is kind of general purpose right like it can kind of do anything and everything
and it does it like pretty well i mean it's not as general as like a like a CPU right but
in terms of like graphics cards like they're pretty like general purpose which i think is like a

(37:33):
double edge sword right so it's good because it allows you to uh use it for literally anything and
that's why you see um if there is so many models that it is able to that are running on in video
hardware right like i mean if you go on hugging face that uh and you see like what is being trained on

(37:55):
in video hardware it's like tens of thousands with like hundreds of thousands of models um it's so many
but like syribrus uh who does inference i think they have like i think two models that they run
right now it's like the llama seven and seventy b now it's the fastest it's the like arguably like

(38:18):
maybe the best implementation of like the llama seventy b but it's like they only have two models right
no i'm sure like their architecture is uh set in a way where they're going to be able to add more
models uh going uh forward but like you know in order to run like each model i assume that it maybe

(38:40):
takes like a handful of engineers so like go and then like put the model on test it make sure it works
well like deal with any errors like make it work on their architecture and specifically for syribrus
given that it's a wafer scale chip i feel like uh the best way to get uh the most out of that kind of
an architecture is to find a model that fits just barely inside that whole chip so that you're making

(39:05):
the most use of the computation there um and you don't have to like use multiple chips which result in
like uh moving some memory around and adding some more networking overhead then bandwidth issues
whereas nvidia i think that's that is one of the things that they focus on the nb-like

(39:26):
nb link infrastructure which has like a really high bandwidth um interface for story memory and
passing information around so nvidia is a little more flexible but that comes with the downside i guess
yeah but i guess like you know i would just kind of like arguing against myself saying like why
like nvidia is so great because they're flexible but at the same time it's like sometimes like you

(39:49):
don't care about that flexibility right so um if you're a company right and you want to have like
your agentic workflow you know you're not gonna be changing your model like every like two weeks
or something like that like oh it's like this new one came out i gotta change everything it's like
no like you're gonna probably like fine-tune your model for your particular use case um and then

(40:10):
maybe you're gonna have like your agent that's gonna do your thing right so it's like maybe you're
gonna go and then uh do the healthcare thing or maybe you're gonna automate like your sales like
for example like if i'm McDonald's right and i'm gonna have um an a i agent take my um orders when
i'm going to like the drive-through right um like i don't need to change my model over two weeks like

(40:33):
i might work with a company like sambanova like cerebris like rock uh get my fine-tune model run
out of their hardware uh have that up in like a server farm somewhere and then um use that from
my drive-through is once i'm confident that like that works well enough like i don't want to touch it
because if a new model comes in and then like uh when somebody asks orders like one big Mac and like

(40:56):
i give them a hundred like that's gonna be like a bad cx right like uh like it's not good right it's
not a good customer experience like i want uh like if i'm McDonald's i really want my models to work well
uh to work consistently um and like i want them to be fast too and also want to be as cheap as possible
and then like i'm gonna like probably work really hard to to make it so that like my models work um really

(41:24):
well and i'm not gonna change i'm not gonna touch it i'll probably leave it there for 10 years um
like once i get it like right now we're kind of like in the fine-team phase everybody's sort of
just like trying out different models like oh like what if i tried this this one slightly better
oh maybe i'll try this but like at a certain point like that fine phase is gonna be over
people are gonna pick their software and they're gonna maybe use it for 10 years i think that's like

(41:48):
highly like a big possibility now uh i'm sure some companies will continue to try things out and like
you know everybody's like kind of innovate but like eventually like there's this software that
people use for a really long time and uh it works and you just don't touch it yeah also regardless
of uh which model you pick i think this is my prediction here given that the chat GPT's

(42:13):
oh one model just beat every other model out there it beat uh clawed in the leaderboards um in most
of the benchmarks i feel like every other company with the big model is gonna try the same approach
have some kind of an agentic approach that thinks through its steps and just spends more
computation on the same problem to e-cat more performance in the benchmarks i think uh

(42:38):
anthop is gonna have like you know an oboe and equivalent google is gonna have a similar equivalent
where there's gonna be spending more computation on solving the same question uh and then hopefully
get a better result i'm sure so that's gonna take up a lot of computing too computational resources
100% 100% um and i think that uh that will make it so that like these things are good enough to

(43:04):
uh solve like a lot of particular use cases so i think that like if you're a company basically
to sum it up like you're a company once your use case is solved like you don't necessarily need to
upgrade immediately um and i think we're gonna have more and more of those companies like potentially
move away from in video and they're gonna move to companies like syrievris like some of the like

(43:29):
grok uh i i think there's like a perfect parallel um with the cryptocurrency industry so like at the
beginning uh when bitcoin first came out everybody was using just like their laptop right um they were
like uh just i think that like i first learned about bitcoin like 2009 or 2010 or something like that

(43:49):
and um i remember like i was like oh this is something we tried a little mining and then i like i
started mining and then like i was like oh my laptop is like so hot like and uh i kind of regret
like stopping mine i never i never actually got it in bitcoin from it but like i tried i thought
it was like too hot i was like i can't use my computer what is this um like at the time was with

(44:10):
like 10 bucks or something so i you know you'll you'll learn you know you never know what you sometimes
miss out on but i think that's a very similar parallel because uh uh the big mining rigs they
uh we're just operating at such a scale that the cost of using bitcoin the fees asymptotically

(44:32):
approach the cost of energy in those places and because like you know you have uh whatever x
number of chips but you were running them 24/7 constantly uh churning out whatever computation uh
running these hash hashing functions and uh you just need to find the cheapest energy because
chips uh the cost of chip ownership approaches zero over time yes exactly so uh to kind of finish

(45:01):
the analogy right so it's like i think that like with bitcoin or most of these cryptocurrencies it
started with CPUs right and then people would start going a little bit more specifics so they would
go from the CPUs to the GPUs so like people were just like buying like AMD and video GPUs or buying
like racks of them they were uh taking like airplanes and shipping these uh GPUs and you saw that like

(45:21):
Nvidia stock like actually like shot up a lot like during the cryptocurrency craze and then like
the prices went down but people still kept on mining bitcoin this old mind Ethereum but they stopped
using like pretty much nobody uses GPUs to mine cryptocurrency anymore uh maybe for a few
cryptocurrencies they do but like for bitcoin everybody uses like ant miner they use like hyper-specific

(45:48):
asics uh to go and uh mine these uh cryptocurrency so like asics what does asics stand for it's like uh
uh applications specific integrated circuit yes so basically like an asic is like uh a chip that
is made for a very particular use case so i think that once the finding phases over like you know

(46:13):
with crypto it took like a few years like you know in 2009 we didn't see any bitcoin asics um but like
quickly people started building them and then like if you wanted to actually like be profitable like
you needed to have those asics and um i used to work at a company that did some like bitcoin mining
and i think they had like over a hundred thousand like uh ant miners that bought that were just

(46:35):
sitting like somewhere with like really cheap electricity um so it's like they didn't buy a bunch of
Nvidia GPUs they bought asics they bought a bunch of ant miners um which is a company that makes like
uh bitcoin mining rigs so i think that like um that same pattern is gonna follow in the AI world like

(46:57):
people just trying stuff out now they're using GPUs but soon and very very soon once the inference cost
get really high uh they're gonna be switching over to these uh hyper specific companies that do
nothing but inference because they're gonna get the model for free like meta is like already making
this a commodity right like i mean they're gonna take uh they're gonna take mistral they're gonna

(47:18):
take llama they're gonna put it on these a uh they say six and they're gonna use that for their
particular use case i think that's the future and then um that is going to like erode
Nvidia's uh dominance right so it's like people are coming from all ends right like people are coming
from Nvidia like i'm a hardware side i think like people are gonna move away once they kind of figure

(47:43):
out their use case and they're also coming on the software side because like Nvidia's way too expensive
right now so i think that like Nvidia's kind of in this like Goldilocks period where like
they're the top dog but i don't see it lasting much longer maybe it'll last another year maybe
another two but like i feel like uh you know in 10 years or even sooner like uh Nvidia's gonna be

(48:05):
knocked off and like they're just gonna be like one other tech company that's a bull claim so i mean
to uh say some words in favor of Nvidia uh yes they are you know more general than some of these
uh hyper focus chips like San Manova or Grok or maybe even Syribras but you know they

(48:29):
like Jensen Huang has been touting the benefits of specialized computing or uh accelerated computing
and the point is GPUs were much better than CPUs at graphics tasks but that's not the only thing
they do they focus on um i believe like video processing chips specifically for like a snapchat and

(48:56):
other companies that process a lot of video so they just look at the overall industry and see
what applications the biggest companies are using and try to accelerate that specific workflow and
the amount of integrations that they've done with every single cloud service provider to make sure
like all their networking switches and stuff plug into AWS, Azure, Google Cloud and work really well

(49:22):
and like it's it's it's a lot of work it's a lot of business deals it's a lot of uh integrations and
that is not trivial that's a it's like a massive uh operation and i don't know if these companies
are gonna be able to integrate with all of these other cloud service provider and keep doing it
every single year whenever they update and uh have changes in their architecture but looking at

(49:48):
the inference speeds it seems like cerebrus and sambanova are kind of uh neck to neck in the 70B
parameter they have like 450 tokens per second on the smaller side cerebrus is much faster

(50:09):
but only sambanova does the biggest lama model of like 405b and it doesn't seem like any of them are
even trying to run these biggest of models like tygbt or um uh geminize uh 1.5 pro or anthropics biggest
model so i wonder if any of these accelerated to provide providers can even run like gpt4 at scale

(50:37):
well i don't know do you know how big gpt4 is uh i think like and uh the internet says about 1.5 trillion
parameters oh so but it's also a mixture of experts so like eight something uh 200 billion parameter
models so if it's eight 200 billion parameter models then sambanova can run like a 400 billion

(50:58):
parameter model and i would suspect that they likely could run chatchetbt but maybe just like
open a i isn't reaching out to them to run it but potentially i would i mean i don't know uh maybe we could
we could ask them if we are able to chat with them next week that's a question yeah like could you

(51:18):
hypothetically run like a trillion parameter model or maybe even something like even bigger the next
generation gpt5 yeah i think we should ask them like what is like the limit to like how big like
you imagine like a model could run in like your architecture but they won't tell us but i feel like
i'd like to like to know that yeah so i don't know um nvidia has a big mode they have the most integrations

(51:43):
they could buy out any one of these companies too or allocated number resources to tackle inference
specifically like i could see them doing that because uh it seems like this space is still relatively new
if we're looking at the big models we saw consolidation and models kind of plateauing and all of the

(52:04):
other companies all the big uh whatever big five or some companies reaching similar performance
and i wonder if we're going to see that in this uh accelerated chip space where there'll be a
few competitors who release products and they all near uh the performance yeah yeah i mean

(52:25):
there's a good chance um that they all become like somewhat of a commodity but i think that
like the pie is very big um so all of them could be potentially uh just incredibly successful just
doing that they're a little piece um and also there's so many players in this space that maybe
mistrolls model runs better on sambanova so they decide to go with the sambanova architecture or

(52:52):
facebook's uh metas models run better in cerebra so they pick that and uh likewise yeah for sure uh i think so
and i think that like um to i think we're giving Nvidia a little bit much credit when it comes to
like the integration so i think that like it's true they integrate pretty well with like a lot of
like the cloud platforms but where Nvidia famously doesn't integrate well is Linux um oh yeah it's

(53:19):
like uh there's like a famous video of like yeah yeah of uh people ask like Linus Torval like the
creative Linus like hey what do you think of this video and uh i won't put it here because it's not like
family friendly yeah some uh very choice words but what do you mean how does uh enterprise workflows
uh how do they run uh if they don't integrate Linux with Nvidia uh like most of enterprise

(53:44):
software runs are Linux right a lot of like web servers and stuff run on Linux but like Nvidia has been
like really really hard to like integrate your GPUs now there's some Linux distros that like
invite uh work with Nvidia like better than others but like Nvidia has been like notoriously uh

(54:04):
kind of like almost hostile to like the open source community like everything kuda right like kuda
is like a proprietary software i think maybe it's what one's what i don't know but like it doesn't
work well on low so this is like some people who like try to do it but yeah like uh it's been
famously difficult to get like Nvidia working well on like you're of Linux box at home

(54:24):
maybe maybe like servers are different but like yeah it's a thing so uh i mean i don't know like
Linux like runs on most of the servers i feel like Nvidia is leaving out like pretty big market there
no i i see a list of integrations with the various Linux uh distros and compatibility with kuda
and so it seems like they have uh wide support but maybe it's different for consumer Linux boxes

(54:48):
maybe yeah and this was like the case like a few years ago like exactly i was like trying to do like
my Linux uh setup i remember like it was really hard to get working um i think i was like running
some Ubuntu instance and uh it was quite difficult and then like um i ended up getting working
uh getting it working i had to like install all kinds of like esoteric software to like make that work

(55:13):
um and like i think for a little while like just what do you turn on just
because like it couldn't figure out how to like handle the Nvidia hardware now that may have changed
like maybe like Nvidia's down to 180 um i don't know i hope they have because like i would love to
see like more in Linux integration but um i don't know there's just like every company's got a little

(55:36):
bit of a blind spot so let's say um you get access to like a sambal nova chip in your hands
tomorrow uh what would you use it for well what i use it for um well super fast inference
assume me i knew it to do with it right let's say it's a plugin play and it just magically works
oh that'd be cool uh like uh well i guess the first thing i'd probably do is uh put it behind a like

(56:01):
uh like an api so i could like call it from all over so it's that you have that so um now uh i have
like a thing that i could and let's say like i can make the api like pre-low uh latency sure um
i let me think um well i guess the first thing i would probably do is just like um use it to do

(56:24):
assuming it's like cheap enough uh i would use it to like rag on my uh by like computer because like my
my MacBook uh it can do like some rag but it's kind of slow um when i try to run with the model locally
so i would like to use that uh chip to help me like uh build like a better um uh spot light tool

(56:47):
like you know like the search bar mac i should like do rag over your entire file system yeah let me try
that um i have tried to do some stuff with like models but it's been kind of slow i feel like
doing like but doing that over my entire computer i feel like that'd be pretty useful um

(57:07):
uh so there's like spotlight alternatives but i just installed one i forgot the name of it uh
it'll come to me alfred no not alfred uh it'll it'll come to me uh i just can't remember right now
i'm surprised there isn't an AI powered spotlight alternative
there probably is but i think there might be i i don't know of one and uh i feel like that is

(57:29):
something that i'd love to see like apple or something like integrate into the operating system
well the new apple intelligence is coming with those features well until it's watch it doesn't
i i would agree with you for most companies especially startups uh but apple has a good track record
no i i'm convinced uh it exists yeah 100 percent convinced uh i'm sure it's coming but uh you know

(57:56):
that's like one thing i would like potentially do and then i think like another thing uh i might
do with like you know really fast inference chip like that uh is uh i'd probably start building some
like agentic tools yeah i think like um so as we were talking kind of like earlier about like
the health um yeah interpretation i think that'd be something that like i'd want to build because

(58:18):
i think that would take a lot of back and forth like okay like um i mean think about it if you have
like a 400 megabyte file uh that had and then like that's all text right so that's like a lot of
data in there well so hopefully you have some programmatic tool that just understands and
parses the DNA data too like you don't want an agent an lalm specifically to do everything

(58:44):
no i don't think you want that but i think what you would want is um so what you can do in one of
these files is you can annotate the files with each of the uh known mutations uh so like i think
that like what you could do is you could almost like kind of brute force it where you like you say
like okay like these are like the 10,000 like known mutations that you have okay and then i would want

(59:09):
like the lalm to go and then like uh do research on every single mutation that i have and figure out
like what to do about it if anything and then come up with like uh actual insights for that and then
like let's say it's like it'd be like okay like we've analyzed all 10,000
shears are here are like the five things you should do uh like we forgot like we ranked it we

(59:31):
got what's the most important i think that like um that would be way too expensive like right now um
like that would cost a lot i think but if you had something like a samba nova chip that was like
really fast and cheap uh i think that'll you know actually be like uh the points where like you could
bring that down to like consumer level yeah i uploaded my gna a dna sorry dna uh

(59:57):
diff to sequencing.com at the beginning of the podcast and it's still processing the 400 megabyte
file uh the whole file uh whole sequence is roughly like 10 gigabytes and i have no idea how long
that would take uh how much of that cost upload the sequencing.com uh it's free really yeah oh wow

(01:00:18):
i should do that. So the way they work is uh they allow you to upload your dna sequence and they have
a bunch of um they have a marketplace where you can buy uh analyses on your dna so depending on
what you're interested in um you can shop the marketplace and then buy different kinds of uh

(01:00:39):
things that analyze your risk of uh diabetes or stroke or um let me see what else yeah there's
a bunch of different things. So how much does that cost? They're all different there's like hundreds
of different uh analyses that you can buy. Okay that that's cool i'm gonna have to check that out
because i have the this like dna dumps on you so figure out what to do with it. But sequencing.com

(01:01:02):
i actually haven't heard about that but it seems super useful. It was one of the ones uh uh
in an article that listed uh all the different places where you can upload your dna data and run
some analyses on it. Oh very cool. Yeah so shashank uh if i turn the question around on you if you
had a sambanova chip and like or some sort of really fast inference chip like anything you would

(01:01:26):
want to build? Definitely agents lots and lots of agents to be like my personal assistance my advisors
my coaches in every aspect of my life. Health is obviously one that i'm interested in and have
been interested in recently. Financial wellness. Just productivity like a personal assistant.

(01:01:48):
Maybe you have one manage my chores and yeah literally everything that i can think of.
Ideally work too but obviously there's restrictions in a corporate environment you have to uh
you know make sure your data isn't being leaked and shared with a bunch of other third-party services
and um sambanova never do that. Yeah um but yeah agents um i'm really excited about building more agents

(01:02:19):
especially since uh the the o1 model seems really capable um but i think it'll also be a little
expensive and i have yet to see a good agentic easy to use tool um that retains like memory because
you can ask it a bunch of questions uh throw a bunch of documents and then keep talking to it and

(01:02:44):
the only um uh former memory you have is the context window which has some limits but then what if
you want to selectively retain some bits of information not others so chat gpt has that kind of like
built-in to its uh product where it has this um concept of memory where it stores pieces of information

(01:03:04):
that it thinks is relevant to like the long-term uh relationship it has with you outside of a single
chat window so an agentic workflow that has memory would be really useful too. Yeah so you mean like
like an open source one that doesn't have the memory because like i mean like opening ad they have

(01:03:26):
like the memory but like i don't know do you mean something that well? I'm not sure if so that's
so i'm not sure if that would be available to an api if i was to build an agent with open
ais uh apis and i today it's not feasible to like build an agent uh a complex agent within

(01:03:48):
the chat gpt uh consumer UI. I see what you're saying um so like you would say like you could use
the api but you'd have to add like more. Exactly so like you want something kind of like out of the
box that it could just like handle like uh isn't that kind of like notebook lm? Yeah uh that's
fantastic that's what i've been using so far um i don't know how they handle memory and pieces of

(01:04:10):
information that are um you know more important than others uh but it is really it's like a really easy
way to use rag out of the box. Throw out a bunch of documents ask questions and it'll give you
specific points from uh the documents that it finds that is relevant to the question that you're
asking with citations that you can click through and then explore further if you want. So um

(01:04:37):
yeah notebook lm is a fantastic. So where do you feel like notebook lm lacks that you would want
like the the memory to be different? Um i mean i guess uh it's fantastic form of rag but it's also not
an agent like uh i can't have it uh take a complex um high level task break it down and then like keep
hammering at uh solving it. Okay so see what you really want is you want to say like you want

(01:05:02):
i want something that has rag i want something that has agentic workflow tool calling um ability to
retain important bits of information in some kind of a persistent memory um and add all that
together. Oh man that'd be a really cool like startup if you could go and then make sure there are
some yeah but like i mean i feel like there's like room definitely for like innovation there where

(01:05:30):
like if there was a tool where you could go and then um kind of uh have like a bunch of integrations
i'm thinking like kind of like zappier uh but like you know it has like so zappier for those who don't
know is like a tool that um can just like uh make it really easy to like integrate with like a
bunch of different apis and things. So for example i could do workflows like with like no code whatsoever

(01:05:56):
it's like oh like every time i get an email from this person like automatically um put the
metadata like the title and a description of the email into like a google spreadsheet and then
also like uh i don't know like send a tweet or something like that like zappier could like go and

(01:06:17):
you know do those ten of kind of workflows like you can connect like a bunch of like things that like
weren't connected in the past uh you can make them connected like oh i could like every time i got it
uh so for example like every time like we post a podcast like zappier could like see that rss
speed updated and like go and update like a bunch of things um like i could go like send an email

(01:06:38):
post a link to and i don't know do whatever uh now we don't use that type of workflow but like
imagine you had that but with like agents so like uh you could just like be so much more powerful
like i mean i can imagine you could go and uh use like a tool like this to say like hey like uh let's
integrate with the expedient API and then go like use it to like plan my entire vacation right

(01:07:02):
to say like i could use the agent to say like hey like i want to go to uh japan um and then like uh
like i want to go like at the end of the year and be like oh well i looked at your calendar and i saw
that you had um uh uh not many meetings or appointments on this week so i booked it from uh the week

(01:07:25):
after christmas and because i saw that like flights were slightly cheaper here and uh there was like
some events that you were interested in going to i booked the floor it's like here's the flights
here's the hotels here's some things you can do here's the itinerary and i an agent could just
figure all that out and just like send it to me um and then maybe i just like like look everything

(01:07:48):
and just like okay and then built it um and like if it was like a consumer thing that could like help
you build uh agents like that that have like integration like the the power that you'd be able to have
is just like unimaginable i mean you'd be able to like replace like entire divisions within companies
like imagine like um like if you had like agents going like programming for you like doing market

(01:08:14):
research for you doing customer support for you i mean like you could conceivably have it where
like you have in literally like entire companies that are run by just like one guy controlling a
bunch of agents and then maybe even after that you just have all entire companies of agents don't
you don't even mean that guy i think at some point it starts to break down at least today uh well

(01:08:36):
we're not there yet we're not there yet yeah but in the future yeah i can absolutely imagine
once we figure out the problem with hallucinations and figure out how to stop these agents from
having their errors keep compounding and could could run haywire um i can see a future with just agents
and humans all out of work trying to figure out uh what our purpose is in this world yeah speaking of

(01:09:03):
that did you see do you remember that like a Tesla announcement like last week where they had all the
the new like robots so they had a bit the robo taxi event yeah um so uh yeah it's like uh
once you have like once you give these things a body yeah and um it's really out of work for humans
like what what are we gonna do with our time i guess we'll just have to like

(01:09:24):
hang out like play games do art yeah uh enjoy the things that you know make life
better like go like maybe build the agents that uh that do all the work for us
yeah but like at some point it's like building themselves they might be yeah i mean they probably
will be they probably do it better and faster than we will um how far how far do you think that is

(01:09:48):
for agents to build themselves uh maybe the step before that where agents start replacing um a specific
role i mean on a company i think it's happening right now uh but like for like very specific roles

(01:10:09):
yeah um uh tasks within roles maybe uh yes so i think that like um the way i sort of look at it so um
at amazon uh they have this type of like and i'm gonna like butcher this really badly but um they have
like different levels and then like uh so like uh like a level one would be like i think like uh

(01:10:33):
like a warehouse worker like you just started out um maybe like you're a delivery driver um it's
gonna be like pretty uh uh i'm gonna put this like there's not much room for like creativity in that job
right so it's like in that job like they're gonna maybe you have like you know move uh boxes from like

(01:10:54):
here to there like they're gonna you're gonna be delivering specific packages and it's like
you know you're not like necessarily trying to like innovate in that role like you are like doing
like a particular job um and then like as every level kind of like increases like the level of kind of
specificity uh goes down like you're doing less of like a predefined task and then like the the role is

(01:11:19):
like more open right so it's like if you're level one like you're doing like something that's like
hyper specific if you're level 12 or 13 whatever like Andy jassie is like the CEO like i mean
he uh is doing all kinds of like general things that is uh like very abstracted way and then like uh

(01:11:39):
it's kind of on a spectrum between each level i think that's sort of how they um like approach like
their leveling and i think that um for agents like we're gonna see um that the roles kind of like the
L ones like the L2s like in amazon like those will be the ones that i think are
more likely be replaced first because like they're the ones you'd like to like do like a very specific

(01:12:04):
job so for example like if you work at McDonald's and like all you do is like let's say work at the
cash register and take orders like uh you don't need an agent for that like you're just like they're
maybe gonna have like a like a like a up screen like an ipad that you order on and uh they're gonna be
able to get your food right uh it's like oh we now no longer need like a person like doing that

(01:12:28):
particular role what what do you think for that specific role the front desk cashier uh who takes
your order at fast foods restaurants or a cafe when do you think that will be replaced uh maybe
it'll just uh very specifically drive through fast food um so i guess like drive through is a little

(01:12:49):
bit different because like it's a little bit hard to like oftentimes like order from a screen so
like oftentimes uh like when you're in your car so assume you don't have the app um and you want to
like use your voice i think that could be most likely i think that technology is there now yeah so
i think it's good enough but i think that um uh it's gonna take some time to get it working well

(01:13:16):
so i think that like you know typical software development maybe takes like say like six months to
like a year till like actually roll something out so i think that like the technology was there
maybe six to twelve months ago so i would start this thing that you'd start to see it now but like
companies won't feel like confident enough to like use it for maybe like another year uh or two

(01:13:38):
till they know it's robust so i would say like maybe 2026
we would start seeing that maybe 2026 2027 uh we'll start seeing that like at a lot of fast food places
um and then like the people who did that like full-time like just won't uh be used anymore and then like

(01:14:01):
oh like they just won't be able to take the cash and then like all you have to do is um just like
use your credit card or use samultman's eyeballs canning technology and pay in cryptocurrency
yeah or that i think uh one of these fast food chains did try implementing it and uh some guy came
along and asked for one million packets of ketchup uh which is which is complimentary

(01:14:26):
the attendant had to step in the human feeling outside you know can't do that
yeah so these are the types of things i think they're gonna have to like kind of work out
like they're gonna have to have an agent say like hey is this like this is a reasonable
about addition yeah it's gonna take uh sometime yeah uh to get it working um but i mean like

(01:14:48):
they've had like these things in Japan for like a really long time where like in Japan like
oftentimes like before you go into restaurant like you're gonna go and then uh just get there's like
a like a machine out front and then like you're gonna pay it the machine it's giving you a ticket
and you just like hand the the ticket to like the the cook who's making your food um and then like

(01:15:10):
they'd be like oh like ding ding ding it's like it's like ready and then like you just go pick it up
from like uh like the pickup area and then bring it to yourself so um i think that like we see less of
that in the United States just because the tipping culture because i think that like a lot of these people
they don't actually uh they want they want to get the those tips and also like the restaurant
orders they pay them like two dollars an hour anyway so it's like yeah like it's so good to kind

(01:15:32):
of the human around because like i feel like even if you could automate the restaurant like a hundred
percent you still may want to have a human just to like like uh like uh security or something i don't know
i thought you were gonna say probably like the human touch i have some socialization feel a sense
of belonging community interact with someone real as opposed to like a virtual embodiment uh

(01:15:58):
but i i am familiar with culture in japan and i've heard it uh called like an introverts
a heaven to not be able to not need to interact with a person for any of your needs he can go out
have a whole day of shopping and restaurants food and not have to talk to a single human being
that's probably true i i think that like uh i mean like you can if you want to

(01:16:23):
you can't do lots of people um but yeah there's lots of things so like uh for example they have um
these uh these hotels they're called like love hotels in japan um which are uh hotels that you
can get um actually love hotels are like really interesting so like most of them are built kind of like
uh they're oftentimes like very ornate so it's sometimes like uh it's gonna have like some sort of

(01:16:48):
theme like they're gonna look like a castle or like this is the late night section of the podcast
here they got like a pirate ship or um i think there's one that is like uh it's got like um
dinosaurs everywhere and then like typically you can do uh i think it's like a short stay or like a long
stay um so like long stay like you're gonna stay the night and then a short stay uh is you're gonna

(01:17:12):
stay like a couple a couple hours um but uh they try to keep uh inanimity like very high there so uh
if you go to some of these hotels like where so like if you're in Tokyo like people aren't gonna
drive there they're gonna walk there from the train right so you don't need this but like oftentimes like
in rural japan like people have cars so like actually what they'll have is they'll have this like little

(01:17:34):
like thing that covers the license plate of your car so people can't like figure out like who's
actually uh staying in there so it's like in the parking lot it's like a little like thing that kind
of like metal thing that it goes up and like covers flashlight um and then like when you when you're
in there they will uh go and then um they'll just have like a little touch screen where you can like go

(01:17:57):
and then like select the room that you want and then like you could go uh you know stay in that room
with like yourself or like with your partner or whatever and then stay there and then like uh in
the room inside the room that they have this like little machine that you can actually pay um
and uh actually won't let you leave until like uh you you pay so like the door won't like physically

(01:18:22):
lock um and then you'd be like you're trying to like get out and like it's like oh no you gotta pay
so i guess like if you forgot your wallet you'd be like in a way especially if you didn't know
the language uh you'd be like in a world of hurt um but uh yeah uh so those hotels are like oftentimes
like you you can't even make a reservation if you want you you have to go in like just see if it's

(01:18:44):
available and then um like typically like the check-in time is oftentimes like a little bit later than
like a regular hotel so like a regular hotel maybe you could check in like maybe four or five like
three o'clock in the afternoon whatever uh maybe four but like there it's like you know the check-in
time like you might not be able to check until like eight or nine p.m uh and then um but it's

(01:19:05):
typically gonna be like uh maybe like a little bit cheaper uh than like other hotels um but uh
yeah that that whole thing whole hotel reservation you don't have to talk to a single person so like
the technology's there it's in Japan like uh like we don't have that maybe because like as you say
people want the human touch but like uh these types of hotels like um they're fantastic and they

(01:19:29):
are specifically made uh so you don't have to interact with a single person maybe just because like
sometimes people do things that are uh no more private in those hotels um but but still uh like
it's there like we can have it and i think that like uh as new things get built uh

(01:19:52):
automation is gonna be like used more and more and more yeah i can see both of these kinds of worlds
be served by AI one where you have the introvert who just wants the thing doesn't want to interact with
any human beings and they have a very specific need and i can probably fulfill it really well

(01:20:14):
obviously we need to figure out edge cases make sure there's no uh nefarious uses and uh prevent
people from doing uh taking advantage of the system um once that's worked out i think the other
avenue is like use cases where people want the fully tailored concierge service like if you're going to

(01:20:34):
i don't know like biophirory or something you kind of want uh like some uh uh fuss and some
nice treatment with a human being who is um catering to your needs and has a more complex
interaction spectrum uh but eventually i could even see that be full of a by an AI where

(01:21:00):
it's able to take care of you so well that you're like why would i ever want a human being
doing this for me yeah uh uh a hundred percent i agree i think that like but i think that is much farther away
um yeah i i i think that like you know sometimes people just like you know like
shiny objects so uh just like oh wow like look at this like new technology look at this watch like

(01:21:23):
look at this jewelry it's great and then like people may just like go together and like look at this and
like kind of like appreciate together like you may want that but i think it is up back i think that like um
we kind of worry about like sometimes like these jobs being um replacing humans but i think that

(01:21:44):
is i would actually argue it's not that much of a problem um so i think that like sometimes like in
the short run like you know let's say you're like a truck driver you're a taxi driver and then like
you know waymo comes and then people use that and now you're out of a job so like i think that like
in the short run there may be some like people who are hurt and that's like obviously uh you know

(01:22:05):
really tough for that particular person right but i think that like uh at like the macro level
it could be a good thing uh because i think that like one problem that people don't really talk
about much is like uh in a lot of the developed world there's like decliny populations right so it's like
my wife's hometown in japan there used to be um a building that had like a nurse nursery school

(01:22:31):
like for like kids and then like over time like uh they're just like wasn't as many kids so they knocked
down and put like uh like an old folks home so it's like uh like in japan like korea like they have
they're like way below like the replacement right and um they're shrinking right so it's like um

(01:22:54):
where i think that like in certain countries like like i don't like like in india like there's just
like a lot of people who like are are there like help um and uh you know uh do things right uh but then
like in india's got like a billion people it's like most pop one of the most i think it is the most
popular in a country now isn't it yeah so it's like there's so many people like a lot of people

(01:23:15):
looking for jobs right but like japan is kind of like the polar opposite like they had like a lot of
people and now it's like starting to shrink so it's like they still haven't like a lot of like like
like roads that need to be maintained um people that need to be taken care of um like uh food that
needs to be farmed uh and i think that like there's so much like work that still needs to be done

(01:23:43):
just to like keep society running um that like you almost like need the automation and order for
us to like maintain our same uh standard of living right so if you think about like in the past like
maybe a hundred two hundred years ago like most people were farmers right it was like 90

(01:24:03):
plus percent of people like just worked on the farm in order for them to eat right now it's like
i don't know like one percent two percent like it's around there that's the percentage of people
who do the farming and so like now like the 98 percent of the population doesn't have to be worried
about like their next meal um to like aren't farm like the two percent of the one percent is able to

(01:24:25):
you can be like one guy on some like giant tractor that drives itself that can like make all the food
for all the people um and then like everybody else can like do other things uh which i think is
fantastic and i think that like um or at least uh i think that there's a high likelihood and my
hope is that like these two things these forces will balance each other out right where like you're

(01:24:50):
going to have the uh the growth of technology and AI like eventually like automate people's jobs
but then at the same time like um like they're just gonna be like less people so like uh overall like
these things these two things will probably like reach some sort of equilibrium and then like it'll

(01:25:11):
be fine it's like people like because you know and people just like get old they die um uh and like
sometimes like the jobs that people had like just aren't needed anymore and then like as uh like
AI just maybe like role replace people by attrition like when when like somebody retires like maybe

(01:25:31):
like the AI will replace them and like might be fine i mean like i think that like as long as this
transition happens like relatively slowly which i think it was a it'll be fine
given the accelerated rate of technological progress that is a tough tough one um but like just

(01:25:52):
because like the technology like uh is done exponentially uh it doesn't mean that it's gonna be rolled out
overnight because like think about like self-driving cars like the technology for self-driving cars
is basically there like i mean like way more is a driving around uh in San Francisco like you
hypothetically don't need any uh rumors in San Francisco but like uh there's other forces that

(01:26:19):
slow the rate of change that isn't just like the technology like you have the regulation you have
well i mean regulations a big one yeah right i mean like the law like i mean like a lot of companies
like companies won't make like autonomous taxis because like they might be the ones who are liable
so it's like i don't know there's a lot of reasons why you wouldn't just go and um uh

(01:26:45):
like roll out the technology immediately yeah i think uh i like the sentiment that you shared um
where AI will just maybe uh supplement the work that can no longer be done by an aging population
but um realistically i think um technology is moving very fast and even though it's not rolling out

(01:27:10):
overnight it is it has been rolling out over the last few years um tasks are getting replaced
slowly i think some roles um some knowledge work will also start to be replaced um and eventually i think
you know way out in the future uh there's not gonna be any work that can be done by a human being

(01:27:37):
better than any i can do it so i think there's gonna be a tipping point where we still have people
we still have people moving around doing things but nothing for them to do uh to make uh economic progress
and contribute to society in that sense but we're gonna you know have existential crises and

(01:27:59):
try to figure out a purpose and meaning outside of that outside of uh the rat race and this capitalist
world that we live in to figure out how we add value to uh society and it'll probably be in some more
form of the arts and um literature and creating content and um connecting with the rest of humanity

(01:28:20):
that way i mean maybe right i mean like uh if you think about it like in the past there was never like
youtuber as they could job right um i mean there was a similar profession a storyteller a court
jester or a singer uh dancer to communicate uh these things for sure and i mean like these things

(01:28:41):
existed right but i think that like uh the prevalence of it is like so much higher right like i mean
uh like look at us like we're just chatting about like all this like on our podcast right now like
it's like what value are we in it's a society i mean i don't know like and we're kind of like giving
like opinions about the world but like right now at the moment like we're not like building something

(01:29:06):
like uh we're not like saving lives i mean like well maybe like uh like indirectly right but it's like
this is like just the act of us like talking right now is kind of like uh like a sort of like an
art type thing in a certain sense right until nopokala ruthless to but this auto-generated podcast

(01:29:27):
you guys wouldn't want to listen to that one when you i tried i tried listening to uh uh there's a
lot of youtube channels that are popping up with a fully a generated content and they're actually
getting surprisingly good where they have speakers who talk and their mouth and uh lip movements are
in sync with the speech um and they have like nice backgrounds they have like whole personas 3D

(01:29:51):
characters that looks very real like 99.9% real um but just the the consistent intonation and like the
the dryness uh with which that they're talking makes it feel a little bit robotic and it is it is
entirely agitated uh if you're in the dead internet theory i know yeah so it's basically what you

(01:30:16):
described where like the theory is is that like on the internet like um it's mostly just like bots
yeah and like people just like won't know if they're talking to like a bot like a human on the internet
and uh it's just like the internet is just like mostly just like overrun by that and it's like

(01:30:36):
actually just like not even real just like people are just like interactive bots all day they don't
realize it so i think that like that that phenomenon is just increasing yeah i mean it's probably real
and i'm sure there always will be humans on the internet but like uh you may not know it
it's like uh there's like famous uh like i think it's the xk cd comic where it's like oh on the internet

(01:31:00):
nobody knows that you're a dog and they have like a like a picture of like a dog sitting in front of
a computer um but yeah i don't know it's just like um it's hard to say like what would happen
when we don't need to work but like i would imagine it's probably not going to be that much different
from today i mean like if you think about it like what do like most people like really truly need to do

(01:31:29):
to survive like uh i mean sure like we need money right to like buy things but like a good chunk of
like the actual things that like are needed um like other than like market forces saying like oh well
you know houses are like really expensive or whatever but like outside of that right like the

(01:31:52):
actual things that we need is like okay like you need like some sort of shelter some sort of shelter
you need uh some sort of uh like food right um like ideally like i think she'd be like relatively clean
like you don't get like sick like maybe like running water like a like tricity um

(01:32:14):
outside of that like you don't you don't really need much else and a lot of that stuff is like
pretty cheap well i'm not i'm not thinking about needs i think we've ways or past this point uh where
we have to worry about our needs as a society um i think they're still maybe like a third or a quarter
of the human population that is um trying to pass that barrier where they're not worried about their

(01:32:38):
basic needs i think the we concern that i have uh in some medium term future is once all of our jobs
are get replaced this big chunk of time this eight hours a day uh 40 hours a week sometimes more
sometimes less um it's just free for all of humanity and that's i think that's going to be met with

(01:33:02):
the equal parts like shock confusion to figure out okay what do we do now and also excitement and
just energy and enthusiasm to explore this new world where everyone is just free to do whatever they
want i mean that sounds fantastic uh like i mean it could be great i mean like uh i think a lot of

(01:33:24):
people uh feel like uh what was the book i think was like by calmy port like bullshit jobs or something
where uh i think i think it was calmy where he makes like the arguments like oh yeah like a bunch of
the things that people do like don't add value and um they're just kind of meaningless so it's like

(01:33:45):
a lot of people like they don't like what they do i mean like they go to work like uh i have a co-worker
he can use like uh i think it's like two hours one way was david graber david graber bullshit jobs
and what do calmy port write uh he commented on uh this i think is yeah anyways continue okay so um

(01:34:10):
it's like you know people they like go to their commute they're stuck in traffic like i mean like what
economic value is like being stuck in traffic give you it's like you know none but like people spend
like hours and hours every day just like computing to their office where they sit on zoom um it's like
i feel like if we got to a point where people like didn't need to work and they had that like time

(01:34:37):
like it might be great i mean like i and if you think about it also like back to by analogy like
people like hundreds of years ago where they would um like work on farm what was the time they're working
on with their own farm like you know it's like back like uh what was it like i mean nidesys is founded
like they said like oh well everybody gets like 40 acres in a mule uh you know what i'm talking about

(01:35:00):
it's like uh all the new likes but yeah but it's like uh it's like something you learn history class like
it's like uh every time like people would come to the united states like uh like this is back in like
oh it's 17 1800s or whatever it's like land was like a dollar an acre or something like that and then
like the government just like give you like 40 acres of land plus like a mule and you could like go

(01:35:22):
live on it so it's like when people are farming they're like farming like on their own land um so like
everybody was kind of like a business owner like i think that like the this sort of like um state
that we're living in now where people are like working for like a large multinational company um uh

(01:35:46):
and like uh like community to work every day like that's kind of like uh like a new like concept
throughout like history like you didn't have multinational companies like in the past i mean like
maybe there's like a few like uh examples of that like i don't even touch East India company or

(01:36:06):
something like that but like in the past like a lot of people just like own their own business kind of
like and you have like a lot of like small little like family run farms family run shops like
you know it's like oh like uh like the the the sun would do what like the dad did um etc like uh so

(01:36:26):
like maybe like i mean the world didn't like fall apart back then and it's like i mean uh i don't
think we're gonna regress to that former world i think the world today is radically different even
if we don't have to work the amount of uh stimulus and uh uh data that we're inundated with is just
massive at any given point we're like bombarded with notifications emails content and it's hard

(01:36:51):
not to get sucked into this scrolling world that is infinite and is targeted to um excite like our
attention and uh get us hooked so i think it's gonna be different uh i'm hopeful i think it'll be
interesting to see how people choose to spend their time uh but i'm also a little worried that uh

(01:37:17):
all of these companies getting more efficient at getting your attention um it's kind of it's
could it be harder for us to um figure out how to do something healthy and meaningful with that
wise uh a hundred percent i mean uh once you have the vision pro they're gonna know if you've

(01:37:39):
actually looked and watched the ads yeah and then like once we got the chips in our brain uh they're
gonna know like you truly not just look at the ad but did you like absorb the information do you
think about it did you feel that did you like ponder the ad and then like uh they're probably gonna
have some sort of like ad netters like oh like how much how many people thought about my company today

(01:38:02):
my hope is that uh we also have a rise in meditation the amount of people who are thinking about
meditation and looking uh into developing um healthy habits because you know we didn't need to do
this pre-self-phone pre-computing because there wasn't that much stimulus to distract our minds now there

(01:38:23):
is uh but at the same time the number of people exploring healthy alternatives is also increasing
so i'm hoping that with this chip in our brain we'll be able to have better control of our minds and
not get sucked into activities uh that uh in an unintentionally uh drain our time and attention

(01:38:44):
yeah i mean we kind of do what Brian Johnson recommends and have the algorithm control our life
yeah but in a positive way yeah maybe right so for those that don't know Brian Johnson we talked
about him a few times on the podcast he's this guy who is trying to live forever and actually at the
beginning of the episode and it's it's been a long episode today but uh it's been a lot of good

(01:39:05):
things to chat about uh we actually got our DNA test shankin' out we went together uh to Brian
Johnson's don't die summit and uh we shatted with him a little bit um just really briefly and um
yeah he has kind of like the the thought process that like hey like uh technology is coming uh really

(01:39:28):
really quickly it's rapidly evolving and uh it's going to make it so that like hypothetically like
you made uh like they may be able to solve all the things that cause death uh but uh the problem
that i could be reviving body is like when you die unless like you like freeze your body or vitrify

(01:39:48):
your body right but like uh as long as you can like stay alive like you may be able to live long enough
to live forever so like what he does is he um follows in a really kind of uh like uh strict
diet regimen where uh he does all types of things he like um we'll like eat all of his meals like

(01:40:12):
early on in the day uh he will um uh try to like get like ideal sleep every night um he like watches
like the food that he eats uh he um that has like horses like a lot of scientists and uh tries really hard
to like increase his longevity but i think all of that is kind of because like uh his sort of theory

(01:40:36):
is like hey in the past like death was inevitable um because like you know we didn't have the technology
now it's like AI, gen AI like all these things like uh it might just be that like we figure out how to
like conquer this like death and dying thing so like uh he has like in a certain sense like protocol

(01:40:59):
he calls his protocol the blueprint protocol that can uh make it so that like uh he can follow that
like positive algorithm to help him uh live longer um and i think that like uh you know there's a lot
of like negative forces out in the world right like you know social media um can just like suck all your

(01:41:22):
time like and take your attention away uh but like if we're able to like replace those social media
like maybe using AI to like help us and then like you know do you like good positive algorithm
like there's some sort of algorithm that will help you like reinforce like your good sleep or like
your positive like nutrition habits or their exercise is like oh like i see you didn't like uh hang out

(01:41:48):
with people like you must be lonely like uh here like uh here are some like meetups in the area
where you can go and uh chat with people um and like i don't know it could be like a really like
force for good um potentially yeah that's uh i'm excited for that that kind of a future um i'm

(01:42:08):
trying to automate some of uh my habits through agents uh mostly uh chat gpt windows that have
context about its specific area of my life and uh it's been helping so far it's a good sounding board
helps me clarify my thoughts and i'm hoping um there are actually a lot of agentic tools out there

(01:42:31):
um as we've been talking i've been doing a little bit of research um if one of our listeners
anyone knows a good agentic framework to use out of the box uh please let us know um but that is
uh an exciting future with agents dictating the blueprint of our lives i mean if for a better future

(01:42:53):
i'm ready for it and i i think that's like a positive way to to end the show on today like i mean
i think there is like existential risk but like i'm hopeful for the future i think it's gonna be
a good one yeah i'm excited i'm ready for it yeah very excited until next time all right see
everybody
you

(01:43:14):
♪
♪
♪

(01:43:41):
♪
♪

All Episodes

Episode Transcript

Popular Podcasts

Dateline NBC

24/7 News: The Latest

Therapy Gecko

.css-15opob5{left:0;position:absolute;top:0.8rem;} All Episodes

.css-14f5ked{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:2;overflow:hidden;}Why you should sell your Nvidia stock right now! #notfinancialadvice

Episode Transcript

Popular Podcasts

.css-r6mb8g{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:1;overflow:hidden;}Dateline NBC

24/7 News: The Latest

Therapy Gecko

All Episodes

Why you should sell your Nvidia stock right now! #notfinancialadvice

Dateline NBC