February 26, 2025 • 45 mins
Can small, agile teams outpace massive, well-funded engineering orgs in AI and open-source innovation?
Co-hosts Alex Kehaya & Bidhan Roy, Founder of Bagel Network joins Greg Osuri, Founder of Akash Network, for a deep dive into the open-source AI stack, decentralization, and the engineering principles behind lean, high-impact teams.
Greg shares how Akash is revolutionizing cloud computing with decentralized infrastructure, the power of Zero Knowledge Proofs (ZKPs) for AI model validation, and why small, focused developer teams consistently outperform bloated, overfunded projects.
Key Dev Insights:
✅ Scaling open-source AI with decentralized computing
✅ ZKPs & AI security—why cryptographic proofs are the future of model validation
✅ Building with constraints—why limited funding fuels better engineering decisions
✅ Community-driven dev—how Akash leverages contributors for rapid iteration
Join us for an engineering-first discussion on the future of decentralized AI and why lean, open-source teams are leading the way.
Website: https://akash.network/
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:06):
Welcome to the index podcast hosted by Alex Cahaya.
Plug in as we explore newfrontiers with entrepreneurs,
builders and investors shapingthe future of the internet hey
(00:33):
everybody and welcome toEverything.
Speaker 2 (00:35):
Bagel, I'm your co-host, alex Cahaya, and I'm
joined by Ben Enroy, founder ofBagel Network.
Today, we're going to explorethe future of open source AI
with our special guest, gregOsuri, who's the founder of
Akash Network.
Today, we're going to explorethe future of open source AI
with our special guest, gregOsuri, who's the founder of
Akash, which is a decentralizedcompute marketplace.
Greg, thanks so much for beinghere today.
I appreciate you joining us.
Great to be here, alex.
Thank you so much for having me.
(00:57):
So for some context, I've knownGreg for a while since, I think,
before Akash actually launched,because we were pretty early
investors back in the day, andone of the things I love about
both Greg and Bidon is they'relike open source purists, which
is, you know, I got red pilledby open source by my partner,
brian Fox, who's the, you know,the author of the Baschel Rola
(01:19):
GPL licenses, and both Greg andBidon have been massive
advocates of open source and arean inspiration to me in that
realm.
Greg, I know you've beenfollowing Bagel, but I'm sure
one of the things you love aboutthem is like they ship really
innovative software technologyaround AI open to everybody and,
like most recently they figuredout how to use ZKP zero
(01:40):
knowledge proofs to prove whocontributed what to an AI model,
and that is what enables themto power monetizable open source
AI, which is huge, and theyopen source the white paper and
all the code that shows itactually working and they like
just to give you like an idea ofthe order of magnitude here,
they reduce the time that ittakes researchers to do the same
(02:03):
thing from like 1000 thousandsof hours to like two seconds
with a team of like threeresearchers Like there are nine
people on their team total.
So I've been like, yeah, reallyinspired about that.
We've been talking a lot aboutthat on the show, Just like our
last episode that we justrecorded was a decent.
Part of the episode was likewhat's the insight?
Like how do you get to thoseinsights that lead to these
(02:23):
kinds of big innovations whenthere's teams that have, you
know, 10 times the amount offunding and 10 times the
headcount?
But then this little team, thislike underdog, the Dave and
Goliath kind of story, shipsopen source, kind of like the
deep seek thing that justhappened.
I really feel like that whatthey did is very similar to that
where it's way cheaper, wayfaster and came from this like
way less funded, smaller team.
Speaker 3 (02:43):
It's funny how that works.
I wrote something in 2019 or Ithink 2018, a post that went a
little viral, that talked abouthow the amount of funding is
inversely correlated to progressin early stage, not later stage
, especially seed stage, and Iwrote this saying I mean, there
(03:06):
has historically hasn't been asingle success story that raised
or over fund, that's overfunded, that actually innovated and I
don't want to pick names incrypto, because if I say
something it's going to cause alot of scare, but classic.
Do you remember?
There used to be an app calledColor.
Speaker 2 (03:25):
I vaguely remember this, like vaguely 2011?
It was like a social mediathing.
Speaker 3 (03:31):
Yes, it was all social media back then and we
had companies that raised $50million.
I think they raised $15, $20,$30 million Back then.
It was a lot of money.
There were companies thatraised $100 million.
I even don't remember some ofthe names.
There hasn't been a singlesuccess story that came out with
companies that raised millionsand millions of dollars In
(03:52):
crypto very classic.
Now.
Solana is a classic story.
Ethereum is a classic story.
Pretty much any teams I meanthat are considered top in terms
of usage.
Forget the market cap, becausemarket cap can be may not be the
actual way to estimate progress, but in terms of actual usage,
solana raised what?
30 million before they launched.
(04:13):
I mean it's significantly lowerto actually build.
I remember because I was therethe first, like the first round,
ethereum raised what?
18 million in their icosignificantly lower to actually
get started.
Atom similar right and 15million or so just enough to get
you started.
And most of the successfulteams always have significantly
lower funding.
(04:34):
Because my thesis is once youhave money, you have
distractions.
Now you're under the radar tospend the money because people
are not going to give you moneywithout you know either, having
conviction that you're going tospend the money, because people
are not going to give you moneywithout you know either, having
conviction, that you're going todeploy the capital because you
know the opportunity.
Cost of money is fairly high,right?
So you can't just like sit andhave money waiting in the bank.
(04:56):
You get a lot of pressure frominvestors and usually they have
controlling.
You know authority to a degreeand then you're screwed.
So screwed.
But teams that have less funding, smaller sizes, your
coordination cost is much lower.
Similar to how I believe JeffBezos has a saying that
(05:16):
apparently in Amazon there is nosingle team that is too big for
a pizza to share.
If you cannot share a pizzawith a team, that team is too
big for a pizza to share, and ifyou cannot share a pizza with a
team, that team is too big towork.
So the pizza team size is anoptimal way of thinking like
small teams actually make a lotof progress versus large,
humongous teams with overfundingusually typically don't make
(05:39):
progress because they have toomuch distraction.
That's a thesis.
But here's a very classicexample here of how Bagel is
innovating with less funding.
Because they're moreresourceful, they're more
ruthless when it comes to focuson the user and the market In a
cash ride.
We raised $1.8 million tolaunch what is a billion-dollar
chain.
Right, but a lot of that comesfrom our hyper-market focus and
(06:04):
hyper-ruthless execution, andthat comes with less money.
Speaker 4 (06:08):
So, greg, as we were talking about this already,
about Akash, some insight.
I'd be interested in hearingwhat worked in the early days of
Akash, what kind of problemspace you were exploring or
usually that's called likeproblem maze, idea maze you were
exploring, and how did youstumble upon this exact problem
space that you're working onright now?
Speaker 3 (06:29):
well, the world was very different when we began in
2017, you know, when wepublished a paper in 2017 me
coming from the silicon valleybackground the situation was so
weirdly different when it comesto policy and when it comes to
what you can and what you cannotdo legally.
My first shock in terms of howto build this company came when
(06:51):
talking to lawyers think it wascooley which was the innovator.
When it comes to marco centauri, I think he created this like
thing called, say, saft right,the, the saft document based
from safe and we were like, okay, we're just going to build a
product and you know it's goingto be a web-based product, and
people are how do we get thesetokens for two people to use?
(07:14):
Uh, we're just going to chargethem with a credit card.
I was told that I'll go to jailif I do that.
So first shock in terms becauseI'm grew up, you know, raised
the whole notion that creditcards are the default ways you
purchase something on theinternet.
When I was told that you cannotuse credit cards to sell, I was
(07:34):
like, okay, can we throw an ad?
How do you get users?
Can we do google ads?
Like, no, crypto is banned fromgoogle.
You cannot use crypto, cannotdo anything.
So the whole notion of userengagement and demand generation
and payments that I knew waswrong in crypto.
So a lot of things, things arevery different now, right, I
(07:57):
mean you can launch a meme coinand go to billions of dollars of
value and you still are okay.
In fact, you didn't getpresidential immunity.
So lessons are different, right, and we have to go and do quite
a lot of work to even get tousing credit cards on Akash.
Now we have credit cards onAkash legally, without going
through all the money services,license and all that stuff.
(08:18):
And also what's the definitionof a security and what's a
commodity.
I mean, these days it's a lotmore relaxed, right, I mean it
got to a point where trulydecentralized companies are
getting like investigated, weregetting investigated at least,
but current administration hastaken a very different policy
position when it comes to crypto.
So, like I almost have to goback to 2017 and rethink what I
(08:43):
could have done.
You know that I was not allowedto from a legal standpoint, you
know, and that's why wesurvived eight years, right,
without getting in trouble inany way because we went by the
book.
So some of the things we didright, that's across the board
was building a large communitybased on users and not based on
(09:06):
speculators.
So even before we had a tokenin the market, we had a series
of events we call them testnetswhere I mean remember, alex, I
think your company was involvedtoo, so you can.
Actually we ran, we ran it outfor years yeah, it was 2018.
We started testings from 2018,2019, 2020 and we had, um, all
(09:29):
kinds of cool things.
We had like challenges andwhatnot.
Everyone's.
To complete a challenge byusing the product, you get some,
some, some credits.
You can exchange the creditsfor tokens that do not exist yet
.
So we were able to bootstrapour community pretty selective
community, because they have todo be technical and they have to
do a deployment, they use acommand line, they have to be a
provider, they have to be avalidator, they have to do
(09:51):
things on the network and youget some rewards right.
And when we launched, we had acommunity that you know, we
airdropped tokens to thatbootstrap our community and that
I think, in my definition, oneof the best things.
We kickstarted this wholenotion of Airdrops in a very
different way.
Things are different now interms of how people are
launching Airdrops, but I thinkincentivizing your early users,
(10:14):
not speculators, is a very, verygood way of launching things.
We went open source first, likeday one, even before we
launched.
When you see my background, I'mvery open source.
I open source from license toand readmemd.
That's how I open source.
Would I do it from launch orwould I not?
(10:34):
If you ask me, I would still doopen source, but it comes down
to the comfort level of the team.
Right, because I've been doingopen source for a long time.
I'm comfortable doing opensource.
I'm not embarrassed of my codeand no, it is code because I
understand there's quite a lotof like restrictions.
When you go to someone else inmy team being like, hey, we want
to open source, they get very,very, very, you know,
(10:58):
uncomfortable.
And that I think you got tomake sure you have an entire
team buy-in, not a founderbuy-in when it comes to open
source.
And it's very, very importantbecause that might impede your
progress because people will beafraid to ship code to open
because they feel judgmental.
So there's all kinds ofemotional aspects that you've
got to deal with open source.
And third, I think like if Iwere to redo things right, we
(11:22):
obviously launched the sovereignstate chain and cosmos and
akash is the first cosmos chainbecause there were there was a
little option.
The other option was ethereum,which is unusable till date.
It's still unusable.
Like you can't expect people topay 30 gas fees to make a
deployment that's, you knowthat's way more expensive than
actual deployment.
(11:43):
You're paying with gpus.
So from from a shade statechain ecosystem, it was
non-existent beyond Ethereum.
If I were to do today, I wouldmost likely do it on Solana or
one of these shade state chainsystems.
When you do sovereign, yes, youhave the benefits of control,
but if you ask yourself deeply,deeply, do you really need that
(12:07):
control in the early stages orcan that control come later?
I would most likely wouldchoose a more modular system
where a mechanism that lets mestart off with a shared state
but I can transition to asovereign state if I need to.
Classic example is SQL rightand the reason why you want to
do SQL, because SQL is astandard that can be used in
(12:30):
MySQL or Postgres.
All you need to do is SQL dumpif you're using a shared
database, and SQL import to asovereign, a completely
controlled database.
If you need to right Similaranalogies.
I mean blockchain, I thinkshould be.
It's a key value pair system.
You should be able tointeroperate technically from
one key value pair to anotherkey value pair.
(12:50):
I mean there are nuances interms of transactions and block
space and whatnot, butultimately, if you remove all
the wrappers, it's just a keyvalue pair state system.
That's your.
You know it's immutable keyvalue pair.
It's it's immutable key valuebased.
Yes, it says right.
So some of the lessons like yes, that would save you quite a
lot of security budget that youcan repurpose for incentives.
And one more thing we didabsolutely right, absolutely
(13:13):
right.
I think a lot of them don't getit get, get these incentives.
So people think of usingincentives to bootstrap a
network.
We did the opposite.
We're using incentives to growthe network.
Now, is it the right way, thewrong way?
That really depends on whatyou're trying to do.
Right, it's.
There's no silver bullet.
You know that answer is notthat helpful.
(13:34):
But in a scenario like, uh,let's say helium, helium, you
need the network even before youget product market fit.
So there's no way in hell youhave to bootstrap the network
because it doesn't exist theresources, but something like
Akash, where there's abundanceof compute everywhere.
They've got 7 million datacenters.
11 million of them are over 1megawatt data centers.
(13:56):
There's abundance of computeeverywhere, so you don't need to
bootstrap.
People don't need money to gobuy compute units, they already
have.
So bootstrap people don't needmoney to go buy compute units,
they already have.
So now the question is do youwant to incentivize early for
that computer to come on boardor do you want to experiment
with the understand the behaviorbefore you incentivize?
I think we chose the latter andthat's working out really well
(14:16):
because you know like mostcompute networks today, straight
up pay, you know, for talkingin tokens to have their compute
on the network.
Like I would literally give youblock rewards, every block for
you to go put your compute onthe network.
The problem is you know youdon't know the quality of the
(14:37):
compute, how good enough it isto your users.
Does that fit your that compute, fit your market right?
How would you know?
By measuring utilization rates.
So if you have high utilizationrates for a certain type of
compute that obviously is inhigh demand, on akash, for
example, h200s, which are whichare best gpus to run deep seek,
(15:01):
are 100 utilized.
That tells you that h200s arehigh demand.
H100s were pretty highutilization.
Now they're at 70 a100, similarright.
So we know that h100s, a100s,h200s, 49s have high demand.
We know that we be 100s.
The older chips have low demand.
Now the question is how do youincentivize each chip?
(15:22):
So our incentive model now is,instead of straight up paying
for the compute, we guaranteeutilization because we know that
if you have h200s you don'tneed incentives.
I mean, utilization is 100, youdon't need to be incentivized.
But we do know that if you haveh100s, where you, there's
(15:43):
volatility in utilization.
Sometimes it's high, sometimesit's lower.
If you go to a provider and belike, I can guarantee you 80%
utilization, like if you getunder.
I mean, provided that you havehigh quality compute and the
provider you pass all thesemarks.
If you're underutilized, we'llmake sure we increase
utilization.
How Well?
Just lower the price so we cansubsidize the cost to the tenant
(16:06):
.
Because we know for sure fromour data that if we lower the
price for H100s to like 99 cents, they're gone.
So we have the product marketfit.
Now can we throw in a discount?
Everybody has discounts, right?
Cloud providers do providediscounts.
So if we throw an extradiscount they're gone.
Know that for sure.
(16:26):
So incentivizing afterunderstanding we came to this
conclusion after we saw how ourcash works.
There's no way in hell we would.
I mean we could draw math,model all that stuff in a in a
room, but ultimately thecustomer's behavior is what's
going to be the most valuable,most accurate inputs.
(16:47):
You need to design incentives.
So designing incentivespost-launch, in growth phase,
post-pmf, I believe, is workingout really well for Akash.
The numbers are very clear.
Our growth is very clear.
Our utilization Not onlyrevenue growth, but we also
measure earnings per gpu and wealso measure utilization per gpu
(17:08):
right.
So across the board utilizationright now is 70 and that was
what 10 when we began.
Over the last 12 months it camereally high.
On per gpu right now is about20 per day compared to 10
roughly.
That was about 12 months ago.
So we clearly see an uppertrend right.
So some of the ways we look athow we design incentives,
(17:31):
instead of blindly following andI think like Filecoin, I mean I
can comment on incentivestructures based on the outcomes
of each deep end.
I can tell you with a degree ofconfidence what went wrong,
what came right, what went wrong.
A lot of opinions on likedifferent incentive models at
this point now.
Speaker 2 (17:47):
Like Akash, was far and away one of the best
investments seed investments,angel investments I've ever made
you and I haven't actually hadthis conversation before, but
I've, like, I've really thoughtabout those early days when I
first met you and the feelingthat I got about you as a person
and the team and like thevision you guys had for a cash.
(18:08):
For me, investing is all aboutpattern recognition.
Right, it's like seeing thesepatterns and feeling that
feeling and knowing to haveconviction and take action on
that conviction.
It's a very like it's as muchas a science and an art.
I get that same feeling withbid on.
What strikes me of what you justsaid is you guys both approach
these problems from firstprinciples and with certain
(18:29):
constraints in mind.
Right, and it goes back to theearlier conversation we were
having about like being verycapital efficient.
Right, like super capitalefficient than open source from
day one.
It's like these three thingsare very common between both of
you.
Like the last conversation thatDon and I had last week was I
asked him like what was theinsight that led you guys to
(18:50):
uncover this zero knowledgeproof, innovation and the whole
market?
All these researchers that werelooking at how to execute this,
were focused on burning compute.
Measuring was compute used, butreally their insight was,
instead of looking at the burn,like at the burning of the
compute we're going to look at,did the model actually get
improved?
And when they switched to likemeasuring that that's what the
(19:14):
zero knowledge proof is provingis did the model improve by a
certain amount.
It changed everything.
Right, that was the big, thebig difference.
Speaker 3 (19:21):
So you measure that some work has been done by
actually seeing the loss ratereduce.
How do you measure modelimproved.
Speaker 4 (19:32):
Yeah, I can take that up.
So first of all, before goinginto that, like Greg, totally
agree with the incentivestructure you described.
Mark Andreessen had a quote, Ithink recently or famously.
Is that like if a system is notworking and if you put more
money into it, that actuallymakes it worse, all right.
(19:54):
So if a network is not workingand you just put more money into
it, that actually makes itworse.
So if a network is not workingand you just put more incentive
into it, that makes it worse.
Speaker 3 (20:04):
It's a rule of thumb, of course, like there are
exceptions all the time.
No, there are no exceptions,actually, it's a first principle
at this point.
Yeah, it is a first principle.
Speaker 4 (20:14):
So if a network does not have users and you give some
incentive and some of them showup, but they will leave after
the incentives dry up anyway,you cannot keep that tap on
perpetually and does not work.
You have to do it.
You have to figure out whatworks and what doesn't without
incentives.
That's when you know what'suseful in terms of the for the
customer customer and then youcan supercharge that with the
(20:35):
incentives.
That's how I see it as well.
So totally, I see eye to eyewith you on that.
And now going back to theprotocol verification protocol
that alex was mentioning.
So it's basically what we havedone.
We figured out this like amodular structure for models
where each contributorcontributing they're building a
(20:57):
model together, but they are nottraining this monolithic dense
transformer together.
Instead, like they're providingthis modular contributions,
which are called like adapterlayers, like lauras, they're
providing that and they stack ontop of each other like Lego
pieces and all those Lego piecescome together and build a model
(21:20):
.
So it's a fundamentallydifferent approach of seeing
this thing.
Like a lot of people, a lot ofteams, very talented teams, are
trying to reduce thecommunication overhead over data
centers to be able to train adense transformer, monolithic
model.
We looked at it a different way.
We looked at it in a way likewhy can't we just make the
(21:41):
architecture itself modular,which is very much in line with
the industry trend at the moment, because MOEs are a rage right
now?
Deepseq is a mixture of expertand this is a modular
architecture, so they keep thetransformer core and modularize
it for efficiency.
We did the same and then whatthat enabled us is that the
(22:01):
contributions are, first of all,they can come from one data
center each contribution, so youdon't need that much of a
communication to begin with.
And second, the contributionsare small enough that you can do
zk verification of that.
So when the person or developeror the data center developing
this modular adapter, they run azk verification on top and it's
(22:24):
way less overhead.
But then even that was likehigher than what would be
acceptable in a productionenvironment, because in
production environment you wantseconds, not minutes, not hours
or not days, and the previousexample of verified training
were weeks.
So what we did?
We looked into how this worksand we saw that the previous
(22:48):
attempts of ZK verifying thesecontributions were tied to
compute.
They were trying to verify.
If compute was burned Then, likewe, I personally have been in
machine learning for more than10 years and my team has like
more than 40 papers published inmachine learning together.
So we have like extensiveexperience of training models in
(23:09):
the traditional ML and we knowthat's not how it works In
traditional machine learning.
You do these massive trainingruns, you get the result and you
look at the evals and if theydon't match up to your standard
you just throw them out.
So they don't really count howmuch compute they burnt.
They count how muchcontribution, how much
(23:30):
improvement to the model thatwas done for this specific
training graph.
So that was kind of themismatch between the traditional
ML and the Web3 AI side.
In Web3 AI we were justincentivizing compute, which
makes sense to some extent.
But if we don't actually lookat the quality of those compute
(23:52):
usage then we get lower qualitysupply on our network and that
does not incentivize theconsumer set to come in and put
their capital in and use theresources.
That's there, yeah.
Speaker 3 (24:03):
I mean incentivizing compute.
Your level of abstraction issignificantly higher than
compute and you shouldn't carewhere it comes from.
I mean, I can train the modelby hand if I have to right.
It doesn't matter what GPUs Ido, as long as your evals I mean
they're good enough for youknow, whatever your standards
are right.
I think Moose is doing asimilar approach.
(24:24):
If I'm not wrong, in the distrothey did present a mechanism
where they're reducing thecommunications overhead for
distributed training, but alsofor the verification.
I believe they took a similarapproach.
But it's fascinating, very cool.
Speaker 4 (24:39):
Yeah, thank you.
So I want to close the thought.
So what we did with this likewe actually are doing ZK
verification of evals, ofimprovements, so it doesn't
matter, like you might just typethe weights by your hand, we
don't care if you increase theevals, that's okay.
(25:00):
And that's how we were able toreduce the verification overhead
.
And we believe this is thefirst and we open sourced it the
both the paper and and the code, and it has been peer-reviewed
by people at Stanford, U of Tand whatnot.
It works and we believe we havesolved the ZK verifiable
(25:20):
training.
We are using that frameworkinitially for fine-tuning only,
but this is the stepping stoneof going towards actually having
fully ZK verified training in adistributed setup.
So that's what we are up towith this.
When token, we have launchedthe first version of our product
and we are planning on going tomainnet very soon as well, and
(25:44):
token, as we already discussed,like a kind of similar thought
process as you, we believe thatwe want to do the product market
fit and feed in the token intothat.
Speaker 2 (25:53):
I mean it sounds like you're getting there because
you've got like 14,000 peoplethat just rushed to use the
bakery.
That's what they call.
The product is the bakerybasically overnight.
I mean it's been live for what,two months.
Speaker 4 (26:04):
Yeah, yeah, it has been live for a month, a month
and a couple of weeks and wewere trending in the top 10 of
product trend when we launchedit.
So I don't think that has everhappened in crypto Top 10 of
product trend worldwide when welaunched.
So first ever crypto product todo so.
But anyway, I'll stop shillingmy own thing.
But going back to what I wassaying, is that, like that's how
(26:27):
we see it, like incentivizingactual equality, not
incentivizing, you know, justwhatever contribution, and
that's that was the gap in allof these machine learning
related resource networks.
Right, and this is so versatile, the framework we have built.
It can snap on top of anycompute aggregated network, like
Akash, and instantly convert itto a verifiable training
(26:49):
network.
So the compute can just feedinto that, and that has a lot of
value as well, because itincreases the value of this like
compute aggregated networks atthe same time.
And the last point I want tomake about open source we love
open source.
I believe in open source aswell.
There are lots of upsides tothat, but there are some
(27:10):
downsides.
We have noticed as well I'msure you have as well.
I was discussing that with Alexa couple of episodes ago that
we open sourced our research ona very fast model verification
algorithm watermarking andfingerprinting early last year
and that has been adopted byactually quite a few well-known
(27:35):
Web3 AI projects in the spaceand they're using that, which
I'm happy about.
But the downside of this isthat we were not credited when
that algorithm became the corepart of the protocol.
So sometimes that happens rightLike you open source it, people
like it, but they just use itfor their own use.
Speaker 3 (27:54):
What kind of licensing did you have on the
open source model, on the opensource code?
Speaker 4 (27:58):
On that, we had MIT.
So again, like I'm happy, Verypermissive.
Speaker 3 (28:04):
It's one of the most probably permissive license
right.
So an advice I would give toavoid such scenarios in the
future is to have Apache 2.0license, where if a protocol
includes your code, they have tocredit your or GPL like GPL V3,
I think.
Speaker 2 (28:24):
Is it like AGPL or GPL V3 has the same thing?
I mean, because they have toopen source it, not just credit
you.
I think with Apache they stillhave to credit you, but it
doesn't have to be open source.
Again, I'm kind of biasedbecause of Brian who helped
write those licenses, but for me, if it's my company and my team
making it, I like it to be aGPL or GPL.
(28:45):
Just because open sourcesoftware for me is my legacy, I
want it to live forever.
I want it to go to Mars, likebash Bash is literally powering
quadcopter on Mars, and I wantit to be there for everybody,
for the world.
Right, and I feel superstrongly about that.
It's a personal choice forother teams, like I don't judge
other people for choosingdifferent licenses and whatnot,
(29:05):
I get it, but for me that's it's.
I don't know again, it's mylegacy etched in stone, etched
in bare metal on the internet.
Speaker 3 (29:13):
Make me think.
I mean, I think I probablyshould consider a GPL license,
because there are similarscenarios with Akash too, Like a
lot of people fork a codeblindly, copy our code.
Just, I mean it's so funny.
Even some of our plugin codepeople just copy, paste and just
control F, change namesliterally, Like we saw that
(29:35):
happen so many times.
A lot of the deep end networksthat are launching just with
Akash straight up code base, youknow, change the names.
There was even instances wherethe founders were like doing
demos and I could clearly seeAKT in there as a payment
mechanism.
They haven't even bothered tochange the currency.
But yeah, but it is a problem.
(29:56):
I mean, crypto is sadly.
I think a more restrictivelicense is definitely an answer,
which I hate to be saying that,but I bet the ones that copied
you have enormous funding.
Speaker 4 (30:12):
Yes, yeah, close to nine figures.
Yes, yeah, close to ninefigures, actually.
Speaker 2 (30:17):
Yeah, and we have the receipts and there's been some
drama on Twitter about thisparticular company, with some
unethical things in the past,from one of the founders
especially, and that areunrelated to open source, right,
but it's kind of like the Apple.
I don't know what the rightphrase here is, but the Apple
doesn't fall from far from thetree or whatever.
I'd be curious to hear what youthink about this, Greg, but I
was like ready to go to war.
I was like dude, let's callthis out Like it's, it's
(30:37):
bullshit, you know, and I wouldgo to war.
You know what I mean?
Yeah, Like right, Like it'sit's, it's complete BS.
Maybe there's still time for usto do that bit.
Back to where we started thisconversation.
It's about execution, thinkingabout things from first,
principles and the peopleinvolved, and even if it was
(30:59):
closed source software, you knowit's easy to copy people and
we've seen that happen in SaaSall the time.
Smaller team comes by, buildthe exact same product out,
executes the well-funded Goliath, and then they've got a huge
company and the other companyends up dying at some point.
That's just been a repeatedstory that we see, and I think
this same thing applies fromopen source and the thing that
(31:19):
happens when you don't opensource is you lose all those
network effects, you lose allthat community.
And you know, one of thereasons why I joined Solana is
because of that patternrecognition we're talking about.
I saw the same thing that I sawin you, Greg, with Tully and Raj
, and they were shipping day oneopen source.
And I even asked Tully as kindof a test, like, hey, what
happens if someone rightwhatever left click copy or
(31:41):
right whatever the right click?
You got to do copies, your codebase?
And this is back when they hadthat office on Howard street.
They were just starting out,they hadn't even launched main
yet, yet, I don't think.
And they and they and they wereshipping a really good idea in
public and he was like pleasefork it, go for it, Cause the
other part of it about theexecution, especially in crypto,
is getting to where you are,Greg, with a real community, a
(32:03):
real customer base, actualpeople providing node services.
You know, leasing, leasing andregistering like providers, dude
, that is a whole ball of wax,it's a it's.
It's hard to replicate that, Um, and you're only stronger
because of the open sourcecommunity.
I just, I just believe that youknow to my core and ultimately,
you know, karma is a bitch forthe people who don't respect
(32:25):
that.
Speaker 3 (32:26):
Um yeah, so, yeah, on that point right.
So I mean open source has somany benefits that outweigh the
risks, right, I mean risk being.
You know you're getting copied,but I think on the long run you
eventually end up winning,because you can fork the code
but you cannot fork thecommunity.
I remember Solana launching.
You know, I was there literallytwo weeks ago.
(32:47):
I did the first video and theyneeded testimonials.
If you go to Solana's YouTube,I'm like one of the first videos
they did In that office.
I had a photo shoot too.
I took like black and whitephotos of them, literally like
one week from shipping.
So funny.
That was like right before covidhit and they hit.
Right when covid hit theylaunched and I was concerned
(33:07):
because we were like planning tolaunch.
A few months from there theycrashed and it was crazy.
Unfortunately, I was able tolike rake up some solana tokens
at like one dollar, two dollarrange or less than that.
I think it's significantly lessthan that, but they were.
So I mean totally.
I remember totally was like thesoftware was not ready to get
shipped, but totally was likewe're gonna ship and that's
(33:30):
that's execution right, becauseyou're never ready.
It's like being a dad you'renever ready.
It's like being a dad You'renever ready.
You just got to go for it.
It's never going to be perfect,but I think it takes the
pragmatism that tells you quitea lot on execution Out-execute,
you know, people are going tocopy you.
People are going to copy youbecause they're good at it, but
they cannot execute you right.
Out-execute you right.
So I would execute ship asaggressively as possible.
(33:53):
Build that community based ontruth.
People like drama, you know,people like the fact that you
know somebody copied you.
That actually builds a strongercommunity in my opinion.
It did for me, for us, right,and you get more sticky
community.
But focus on the community veryimportant, I think, one of the
best community builders I see inthe space and, of course, like
(34:21):
and toli and they're very, verygood at it.
But I think in jacob constreborn jacob from bit tensor or
even shaw walters from from ai16z.
I mean, it's more modern times,but but I I admire community
builders and I studied them.
Um, I think Jacob has quite alot of admiration.
I mean really think about howhe builds BitTensor.
Go to the discourse andunderstand how the discourse
(34:41):
works and how the communicationsare working there.
Spend a day in BitTensordiscourse.
Spend a day in AI16 discourse.
I was so impressed by Shawaw'sgtm in terms of like they had
this eliza framework open sourcefrom day one of the top
trending repo in in the women,it's like for a month or
(35:05):
something.
Like it's crazy, and I wastraining on github like 10 years
ago, so I understand what itmeans to be trending and what
that brings you in terms of likeinbounds and it's just a crazy
phase you go through.
And the reason why I think hewas so successful in shipping
software is you know his hewould do like this, like massive
(35:28):
three hour long videos onexplaining the software and you
would think like this video hasto be succinct and like to the
point because nobody has time,like no.
He proved that if you explainthings well, people will sit and
listen to you and that's whatJoe Rogan did and a lot of the
new podcasters long-formpodcasts are about.
(35:51):
If the content is interestingenough, people will listen to
you and think you know.
Shaw definitely understood thatand his videos are like I watch
a three-hour video on fridaynight to understand eliza and I
sat through it, right, you know,and I was surprised that I
could spend.
I mean, I don't watch tv thatmuch, I mean especially youtube
(36:11):
and whatnot, not for that longbut but there's something to be
understood about these foundersand how they build communities
around open source systems.
Same thing with Jacob.
Jacob would spend hours on out.
Jacob's weapon is Discord right, and everybody has their own
choice of platforms, right.
That is something that I thinkcannot be forked, no matter how
hard you try, and I think that'sone of the reasons why I think
(36:33):
I'm not endorsing ai16z as atoken or any of that sort.
I'm just observing communitybuilders and seeing their
patterns.
Uh, like you say, it's patternrecognition.
Right now, I can recognizethese patterns among successful
founders, successful builders.
They all have one thing incommon is cut the bullshit, be
pragmatic.
Don't worry about gettingjudged.
If you have to go on athree-hour podcast, three-hour
(36:56):
video, where are you going tomake mistakes?
It's not going to be perfectbecause it's life, right.
I've noticed so much, so much,I guess, like, like engagement.
When I do live raw videoswithout filters, you know people
love it.
I mean, I just randomly go ontwitter and be like hey, I'm
(37:19):
going to live stream my codingsession.
People love to see me code,right because, like people, love
to see other people play games.
You know things of that naturewill give you quite a lot of
like true community.
That well-funded company withbig PR teams will be prohibited
from going live because theyhave they're too afraid to lose.
Speaker 2 (37:41):
Let's, let's build some AI agents on live stream.
Next time, bidhan, I'll watchyou, I'll do commentary, I'll be
.
I'll be like the sportscommentator I like.
Speaker 3 (37:50):
Sahil from Gumroad, another like Sahil.
Sahil from Gumroad, anotherfavorite Sahil from Gumroad.
All these phenomenal communitybuilders, right, I don't know if
you know Sahil, but he's acharacter.
I follow him on Twitter.
Have you seen his YouTubevideos where he does this PMing,
where he does this amazing wayof how to write a PRD?
He's a great product manager.
(38:11):
The way he would approach a PRDusing AI these days is just
amazing because I need to checkthat out it's amazing.
So he starts with, like blackdocument why and what, why they
want to do something and whatthey're going to do Like a few
lines and sentences, plainEnglish, and he develops this
PRD using different tools, usingDeepSea, using, you know,
(38:36):
chaigpt, you know O3, andactually he does the V0, the
whole, like you know, userinterface thing, and he does it
live.
And that's awesome Because yousee the mistakes people make and
you see someone work.
A great way of improving yourwork is to imitate somebody's
work and I had that in my careerfor a long time.
I imitated when I was learningGo.
A lot of my imitation came fromMish Hashimoto, the founder of
(38:56):
HashiCorp, developing Terraformand whatnot.
I learned from how these peoplecode.
Looking at the code.
I love learning by other,watching experts or people that
I consider good to work, and Ithink that speaks quite a lot in
how these founders buildcommunities, and we need open
source.
I think is a great way toimitate them.
Speaker 2 (39:19):
So we only have a couple of minutes left and I
couldn't agree more witheverything you said.
But there's something I reallywanted to ask you, greg.
The conversations I've had withBen on the last couple of
months have really helped medevelop this thesis around open
intelligence, and the term openintelligence is actually came
from a brainstorming sessionthat Markeisha and I had.
She leads marketing over atBagel, but basically it's
(39:41):
artificial general intelligenceor artificial super intelligence
.
On crypto rails.
It's open source, right Opensource or AGI on crypto rails.
I have come to believe that thisis really really important for
humanity and for our industryand that it really should be the
North Star of every founder.
That's, building criticalinfrastructure should be to help
(40:04):
make this happen, to accelerateopen intelligence.
Bagel is critical for it and Ithink a cash is critical for it.
I think for obvious reasons,but I also am asking myself,
like, am I crazy?
Right, like you know, can we dothis?
Can we get a bunch ofentrepreneurs who are building
these systems to work togetherto create open intelligence for
(40:24):
the sake of humanity?
Am I being like overly dramatic?
Is it even possible?
Like, and then, if it ispossible, like, what are the
components?
It's like I look at it like afactory system and there's like
inputs and outputs and there's abunch of tooling that needs to
be built in between to make ithappen, and I I'm like trying to
develop like a vision for whatthat is.
So that's, I know that's a ton,but generally speaking, like
(40:47):
how am I doing?
Here are we, are we going tomake it to open intelligence?
Speaker 3 (40:51):
I mean it's not something we, you know,
something that it'd be nice tobe open source.
It's something that weabsolutely need.
It has to happen because,historically speaking, any
sufficiently importanttechnology that reached mass
adoption is open source.
Linux, world Wide Web, you know, name it.
(41:15):
In modern era, even, as amatter of fact, phones, I mean,
without Android, you wouldn'thave the same level of adoption
if it just were, yeah,transformers for open source,
right?
So open source is so criticalfor mass adoption because the
network effects would not comeotherwise.
And we've seen very clearlywhat happened with DeepSeek,
(41:38):
right, very, very obvious.
Like, until DeepSeek, thestate-of-the-art model was a
closed model and DeepSeek cameand changed things, and this is
just going to be a steppingstone in reaching ultimate
intelligence.
Now, have we done that before?
Yes, we have World Wide Web.
Like, without the companies thatinnovated early failures,
(42:01):
successes, whatnot we wouldn'thave the open internet that we
so much use.
Right, 20 years from now, youlook back and be like we'll be
asking the question.
So, looking back to internetitself, right, it did not begin
as open.
Internet began very close, aolwas literally shipping cds.
You needed aol to get on theinternet, right, look where aol
(42:24):
is today.
We're still around.
But it's not the internetcompany, right, it is the
companies that actually operatethe internet.
So, similarly like open ai, Ithink would be the new aol of
the companies that actuallyoperate the internet.
So, similarly like OpenAI, Ithink would be the new AOL of
the world.
Yes, they were the first toinnovate with ChatGPT.
I mean not first to innovate,but at least first to get to
market with ChatGPT anddefinitely like proved product
market fit with ChatGPT.
(42:45):
That kicked off an entireindustry.
But I think the trueintelligence through AGI will be
open source.
And if you look at how do youmonetize and how do you sustain
open source software,decentralization is the way.
Crypto is the way.
There is no better mechanismthat I can think of and I have
(43:06):
factual evidence for this,looking at successful open
source companies that open thisproducts that could not create a
successful mechanisms tosustain themselves.
So, looking back, there's nodoubt in my mind that crypto is
the only mechanism, onlyframework to have monetizable.
You know machine intelligenceright.
(43:28):
So that's really comes down toit.
I'm giving a Bindan's pitchright here.
But open source is good as longas it can be monetizable and
sustainable, and I think cryptois a great way.
So, yeah, it's going to bemultiple people, multiple
organizations, multipleprotocols, rather interoperating
(43:48):
with each otherpermissionlessly.
If Bagel wants to use Akash, noone can stop Bagel from using
Akash.
No one can stop you fromgetting the H200s for $0.99.
Right, and that's what it comesdown to.
If I want to use Bagel, no onecan stop me to use Bagel, and I
(44:09):
can build protocols on top of.
Speaker 2 (44:11):
Akash.
Speaker 3 (44:11):
Unstoppable.
Open intelligence Correct.
It has to be unstoppable, ithas to be.
Imagine you're a developer andyou're coding at 2 o'clock in
the night and you want to use aproduct.
Last thing you want is anapproval from someone that's
going to wake up on a weekdayand give you an approval to get
your account, increase yourlimits, all sorts of things.
I mean.
You want to review code, youhave a question?
(44:32):
You should be able to go readthe code yourself and you have
AI.
Now they explain a lot ofthings on how things work.
I mean, back in the day it'slike actually read the code,
figure things out.
That takes a while, but I canfeed that into an AI and be like
explain what's going on.
The logic here, right.
There are so much tools nowthat open source adoption is
(44:56):
going to be accelerated withtools as well as creation as
well.
Speaker 2 (44:57):
Yeah, I couldn't agree more with you, man, and I
appreciate you coming on.
We can wrap now.
Thanks so much, dude, awesome.
Speaker 1 (45:10):
You just listened to the Index Podcast with your host
, alex Cahaya.
If you enjoyed this episode,please subscribe to the show on
Apple, spotify or your favoritestreaming platform.
New episodes are availableevery Friday.
Thanks for tuning in.
I feel like.
Welcome to the index podcast hosted by Alex Cahaya.
Plug in as we explore newfrontiers with entrepreneurs,
builders and investors shapingthe future of the internet hey
(00:33):
everybody and welcome toEverything.
Speaker 2 (00:35):
Bagel, I'm your co-host, alex Cahaya, and I'm
joined by Ben Enroy, founder ofBagel Network.
Today, we're going to explorethe future of open source AI
with our special guest, gregOsuri, who's the founder of
Akash Network.
Today, we're going to explorethe future of open source AI
with our special guest, gregOsuri, who's the founder of
Akash, which is a decentralizedcompute marketplace.
Greg, thanks so much for beinghere today.
I appreciate you joining us.
Great to be here, alex.
Thank you so much for having me.
(00:57):
So for some context, I've knownGreg for a while since, I think,
before Akash actually launched,because we were pretty early
investors back in the day, andone of the things I love about
both Greg and Bidon is they'relike open source purists, which
is, you know, I got red pilledby open source by my partner,
brian Fox, who's the, you know,the author of the Baschel Rola
(01:19):
GPL licenses, and both Greg andBidon have been massive
advocates of open source and arean inspiration to me in that
realm.
Greg, I know you've beenfollowing Bagel, but I'm sure
one of the things you love aboutthem is like they ship really
innovative software technologyaround AI open to everybody and,
like most recently they figuredout how to use ZKP zero
(01:40):
knowledge proofs to prove whocontributed what to an AI model,
and that is what enables themto power monetizable open source
AI, which is huge, and theyopen source the white paper and
all the code that shows itactually working and they like
just to give you like an idea ofthe order of magnitude here,
they reduce the time that ittakes researchers to do the same
(02:03):
thing from like 1000 thousandsof hours to like two seconds
with a team of like threeresearchers Like there are nine
people on their team total.
So I've been like, yeah, reallyinspired about that.
We've been talking a lot aboutthat on the show, Just like our
last episode that we justrecorded was a decent.
Part of the episode was likewhat's the insight?
Like how do you get to thoseinsights that lead to these
(02:23):
kinds of big innovations whenthere's teams that have, you
know, 10 times the amount offunding and 10 times the
headcount?
But then this little team, thislike underdog, the Dave and
Goliath kind of story, shipsopen source, kind of like the
deep seek thing that justhappened.
I really feel like that whatthey did is very similar to that
where it's way cheaper, wayfaster and came from this like
way less funded, smaller team.
Speaker 3 (02:43):
It's funny how that works.
I wrote something in 2019 or Ithink 2018, a post that went a
little viral, that talked abouthow the amount of funding is
inversely correlated to progressin early stage, not later stage
, especially seed stage, and Iwrote this saying I mean, there
(03:06):
has historically hasn't been asingle success story that raised
or over fund, that's overfunded, that actually innovated and I
don't want to pick names incrypto, because if I say
something it's going to cause alot of scare, but classic.
Do you remember?
There used to be an app calledColor.
Speaker 2 (03:25):
I vaguely remember this, like vaguely 2011?
It was like a social mediathing.
Speaker 3 (03:31):
Yes, it was all social media back then and we
had companies that raised $50million.
I think they raised $15, $20,$30 million Back then.
It was a lot of money.
There were companies thatraised $100 million.
I even don't remember some ofthe names.
There hasn't been a singlesuccess story that came out with
companies that raised millionsand millions of dollars In
(03:52):
crypto very classic.
Now.
Solana is a classic story.
Ethereum is a classic story.
Pretty much any teams I meanthat are considered top in terms
of usage.
Forget the market cap, becausemarket cap can be may not be the
actual way to estimate progress, but in terms of actual usage,
solana raised what?
30 million before they launched.
(04:13):
I mean it's significantly lowerto actually build.
I remember because I was therethe first, like the first round,
ethereum raised what?
18 million in their icosignificantly lower to actually
get started.
Atom similar right and 15million or so just enough to get
you started.
And most of the successfulteams always have significantly
lower funding.
(04:34):
Because my thesis is once youhave money, you have
distractions.
Now you're under the radar tospend the money because people
are not going to give you moneywithout you know either, having
conviction that you're going tospend the money, because people
are not going to give you moneywithout you know either, having
conviction, that you're going todeploy the capital because you
know the opportunity.
Cost of money is fairly high,right?
So you can't just like sit andhave money waiting in the bank.
(04:56):
You get a lot of pressure frominvestors and usually they have
controlling.
You know authority to a degreeand then you're screwed.
So screwed.
But teams that have less funding, smaller sizes, your
coordination cost is much lower.
Similar to how I believe JeffBezos has a saying that
(05:16):
apparently in Amazon there is nosingle team that is too big for
a pizza to share.
If you cannot share a pizzawith a team, that team is too
big for a pizza to share, and ifyou cannot share a pizza with a
team, that team is too big towork.
So the pizza team size is anoptimal way of thinking like
small teams actually make a lotof progress versus large,
humongous teams with overfundingusually typically don't make
(05:39):
progress because they have toomuch distraction.
That's a thesis.
But here's a very classicexample here of how Bagel is
innovating with less funding.
Because they're moreresourceful, they're more
ruthless when it comes to focuson the user and the market In a
cash ride.
We raised $1.8 million tolaunch what is a billion-dollar
chain.
Right, but a lot of that comesfrom our hyper-market focus and
(06:04):
hyper-ruthless execution, andthat comes with less money.
Speaker 4 (06:08):
So, greg, as we were talking about this already,
about Akash, some insight.
I'd be interested in hearingwhat worked in the early days of
Akash, what kind of problemspace you were exploring or
usually that's called likeproblem maze, idea maze you were
exploring, and how did youstumble upon this exact problem
space that you're working onright now?
Speaker 3 (06:29):
well, the world was very different when we began in
2017, you know, when wepublished a paper in 2017 me
coming from the silicon valleybackground the situation was so
weirdly different when it comesto policy and when it comes to
what you can and what you cannotdo legally.
My first shock in terms of howto build this company came when
(06:51):
talking to lawyers think it wascooley which was the innovator.
When it comes to marco centauri, I think he created this like
thing called, say, saft right,the, the saft document based
from safe and we were like, okay, we're just going to build a
product and you know it's goingto be a web-based product, and
people are how do we get thesetokens for two people to use?
(07:14):
Uh, we're just going to chargethem with a credit card.
I was told that I'll go to jailif I do that.
So first shock in terms becauseI'm grew up, you know, raised
the whole notion that creditcards are the default ways you
purchase something on theinternet.
When I was told that you cannotuse credit cards to sell, I was
(07:34):
like, okay, can we throw an ad?
How do you get users?
Can we do google ads?
Like, no, crypto is banned fromgoogle.
You cannot use crypto, cannotdo anything.
So the whole notion of userengagement and demand generation
and payments that I knew waswrong in crypto.
So a lot of things, things arevery different now, right, I
(07:57):
mean you can launch a meme coinand go to billions of dollars of
value and you still are okay.
In fact, you didn't getpresidential immunity.
So lessons are different, right, and we have to go and do quite
a lot of work to even get tousing credit cards on Akash.
Now we have credit cards onAkash legally, without going
through all the money services,license and all that stuff.
(08:18):
And also what's the definitionof a security and what's a
commodity.
I mean, these days it's a lotmore relaxed, right, I mean it
got to a point where trulydecentralized companies are
getting like investigated, weregetting investigated at least,
but current administration hastaken a very different policy
position when it comes to crypto.
So, like I almost have to goback to 2017 and rethink what I
(08:43):
could have done.
You know that I was not allowedto from a legal standpoint, you
know, and that's why wesurvived eight years, right,
without getting in trouble inany way because we went by the
book.
So some of the things we didright, that's across the board
was building a large communitybased on users and not based on
(09:06):
speculators.
So even before we had a tokenin the market, we had a series
of events we call them testnetswhere I mean remember, alex, I
think your company was involvedtoo, so you can.
Actually we ran, we ran it outfor years yeah, it was 2018.
We started testings from 2018,2019, 2020 and we had, um, all
(09:29):
kinds of cool things.
We had like challenges andwhatnot.
Everyone's.
To complete a challenge byusing the product, you get some,
some, some credits.
You can exchange the creditsfor tokens that do not exist yet
.
So we were able to bootstrapour community pretty selective
community, because they have todo be technical and they have to
do a deployment, they use acommand line, they have to be a
provider, they have to be avalidator, they have to do
(09:51):
things on the network and youget some rewards right.
And when we launched, we had acommunity that you know, we
airdropped tokens to thatbootstrap our community and that
I think, in my definition, oneof the best things.
We kickstarted this wholenotion of Airdrops in a very
different way.
Things are different now interms of how people are
launching Airdrops, but I thinkincentivizing your early users,
(10:14):
not speculators, is a very, verygood way of launching things.
We went open source first, likeday one, even before we
launched.
When you see my background, I'mvery open source.
I open source from license toand readmemd.
That's how I open source.
Would I do it from launch orwould I not?
(10:34):
If you ask me, I would still doopen source, but it comes down
to the comfort level of the team.
Right, because I've been doingopen source for a long time.
I'm comfortable doing opensource.
I'm not embarrassed of my codeand no, it is code because I
understand there's quite a lotof like restrictions.
When you go to someone else inmy team being like, hey, we want
to open source, they get very,very, very, you know,
(10:58):
uncomfortable.
And that I think you got tomake sure you have an entire
team buy-in, not a founderbuy-in when it comes to open
source.
And it's very, very importantbecause that might impede your
progress because people will beafraid to ship code to open
because they feel judgmental.
So there's all kinds ofemotional aspects that you've
got to deal with open source.
And third, I think like if Iwere to redo things right, we
(11:22):
obviously launched the sovereignstate chain and cosmos and
akash is the first cosmos chainbecause there were there was a
little option.
The other option was ethereum,which is unusable till date.
It's still unusable.
Like you can't expect people topay 30 gas fees to make a
deployment that's, you knowthat's way more expensive than
actual deployment.
(11:43):
You're paying with gpus.
So from from a shade statechain ecosystem, it was
non-existent beyond Ethereum.
If I were to do today, I wouldmost likely do it on Solana or
one of these shade state chainsystems.
When you do sovereign, yes, youhave the benefits of control,
but if you ask yourself deeply,deeply, do you really need that
(12:07):
control in the early stages orcan that control come later?
I would most likely wouldchoose a more modular system
where a mechanism that lets mestart off with a shared state
but I can transition to asovereign state if I need to.
Classic example is SQL rightand the reason why you want to
do SQL, because SQL is astandard that can be used in
(12:30):
MySQL or Postgres.
All you need to do is SQL dumpif you're using a shared
database, and SQL import to asovereign, a completely
controlled database.
If you need to right Similaranalogies.
I mean blockchain, I thinkshould be.
It's a key value pair system.
You should be able tointeroperate technically from
one key value pair to anotherkey value pair.
(12:50):
I mean there are nuances interms of transactions and block
space and whatnot, butultimately, if you remove all
the wrappers, it's just a keyvalue pair state system.
That's your.
You know it's immutable keyvalue pair.
It's it's immutable key valuebased.
Yes, it says right.
So some of the lessons like yes, that would save you quite a
lot of security budget that youcan repurpose for incentives.
And one more thing we didabsolutely right, absolutely
(13:13):
right.
I think a lot of them don't getit get, get these incentives.
So people think of usingincentives to bootstrap a
network.
We did the opposite.
We're using incentives to growthe network.
Now, is it the right way, thewrong way?
That really depends on whatyou're trying to do.
Right, it's.
There's no silver bullet.
You know that answer is notthat helpful.
(13:34):
But in a scenario like, uh,let's say helium, helium, you
need the network even before youget product market fit.
So there's no way in hell youhave to bootstrap the network
because it doesn't exist theresources, but something like
Akash, where there's abundanceof compute everywhere.
They've got 7 million datacenters.
11 million of them are over 1megawatt data centers.
(13:56):
There's abundance of computeeverywhere, so you don't need to
bootstrap.
People don't need money to gobuy compute units, they already
have.
So bootstrap people don't needmoney to go buy compute units,
they already have.
So now the question is do youwant to incentivize early for
that computer to come on boardor do you want to experiment
with the understand the behaviorbefore you incentivize?
I think we chose the latter andthat's working out really well
(14:16):
because you know like mostcompute networks today, straight
up pay, you know, for talkingin tokens to have their compute
on the network.
Like I would literally give youblock rewards, every block for
you to go put your compute onthe network.
The problem is you know youdon't know the quality of the
(14:37):
compute, how good enough it isto your users.
Does that fit your that compute, fit your market right?
How would you know?
By measuring utilization rates.
So if you have high utilizationrates for a certain type of
compute that obviously is inhigh demand, on akash, for
example, h200s, which are whichare best gpus to run deep seek,
(15:01):
are 100 utilized.
That tells you that h200s arehigh demand.
H100s were pretty highutilization.
Now they're at 70 a100, similarright.
So we know that h100s, a100s,h200s, 49s have high demand.
We know that we be 100s.
The older chips have low demand.
Now the question is how do youincentivize each chip?
(15:22):
So our incentive model now is,instead of straight up paying
for the compute, we guaranteeutilization because we know that
if you have h200s you don'tneed incentives.
I mean, utilization is 100, youdon't need to be incentivized.
But we do know that if you haveh100s, where you, there's
(15:43):
volatility in utilization.
Sometimes it's high, sometimesit's lower.
If you go to a provider and belike, I can guarantee you 80%
utilization, like if you getunder.
I mean, provided that you havehigh quality compute and the
provider you pass all thesemarks.
If you're underutilized, we'llmake sure we increase
utilization.
How Well?
Just lower the price so we cansubsidize the cost to the tenant
(16:06):
.
Because we know for sure fromour data that if we lower the
price for H100s to like 99 cents, they're gone.
So we have the product marketfit.
Now can we throw in a discount?
Everybody has discounts, right?
Cloud providers do providediscounts.
So if we throw an extradiscount they're gone.
Know that for sure.
(16:26):
So incentivizing afterunderstanding we came to this
conclusion after we saw how ourcash works.
There's no way in hell we would.
I mean we could draw math,model all that stuff in a in a
room, but ultimately thecustomer's behavior is what's
going to be the most valuable,most accurate inputs.
(16:47):
You need to design incentives.
So designing incentivespost-launch, in growth phase,
post-pmf, I believe, is workingout really well for Akash.
The numbers are very clear.
Our growth is very clear.
Our utilization Not onlyrevenue growth, but we also
measure earnings per gpu and wealso measure utilization per gpu
(17:08):
right.
So across the board utilizationright now is 70 and that was
what 10 when we began.
Over the last 12 months it camereally high.
On per gpu right now is about20 per day compared to 10
roughly.
That was about 12 months ago.
So we clearly see an uppertrend right.
So some of the ways we look athow we design incentives,
(17:31):
instead of blindly following andI think like Filecoin, I mean I
can comment on incentivestructures based on the outcomes
of each deep end.
I can tell you with a degree ofconfidence what went wrong,
what came right, what went wrong.
A lot of opinions on likedifferent incentive models at
this point now.
Speaker 2 (17:47):
Like Akash, was far and away one of the best
investments seed investments,angel investments I've ever made
you and I haven't actually hadthis conversation before, but
I've, like, I've really thoughtabout those early days when I
first met you and the feelingthat I got about you as a person
and the team and like thevision you guys had for a cash.
(18:08):
For me, investing is all aboutpattern recognition.
Right, it's like seeing thesepatterns and feeling that
feeling and knowing to haveconviction and take action on
that conviction.
It's a very like it's as muchas a science and an art.
I get that same feeling withbid on.
What strikes me of what you justsaid is you guys both approach
these problems from firstprinciples and with certain
(18:29):
constraints in mind.
Right, and it goes back to theearlier conversation we were
having about like being verycapital efficient.
Right, like super capitalefficient than open source from
day one.
It's like these three thingsare very common between both of
you.
Like the last conversation thatDon and I had last week was I
asked him like what was theinsight that led you guys to
(18:50):
uncover this zero knowledgeproof, innovation and the whole
market?
All these researchers that werelooking at how to execute this,
were focused on burning compute.
Measuring was compute used, butreally their insight was,
instead of looking at the burn,like at the burning of the
compute we're going to look at,did the model actually get
improved?
And when they switched to likemeasuring that that's what the
(19:14):
zero knowledge proof is provingis did the model improve by a
certain amount.
It changed everything.
Right, that was the big, thebig difference.
Speaker 3 (19:21):
So you measure that some work has been done by
actually seeing the loss ratereduce.
How do you measure modelimproved.
Speaker 4 (19:32):
Yeah, I can take that up.
So first of all, before goinginto that, like Greg, totally
agree with the incentivestructure you described.
Mark Andreessen had a quote, Ithink recently or famously.
Is that like if a system is notworking and if you put more
money into it, that actuallymakes it worse, all right.
(19:54):
So if a network is not workingand you just put more money into
it, that actually makes itworse.
So if a network is not workingand you just put more incentive
into it, that makes it worse.
Speaker 3 (20:04):
It's a rule of thumb, of course, like there are
exceptions all the time.
No, there are no exceptions,actually, it's a first principle
at this point.
Yeah, it is a first principle.
Speaker 4 (20:14):
So if a network does not have users and you give some
incentive and some of them showup, but they will leave after
the incentives dry up anyway,you cannot keep that tap on
perpetually and does not work.
You have to do it.
You have to figure out whatworks and what doesn't without
incentives.
That's when you know what'suseful in terms of the for the
customer customer and then youcan supercharge that with the
(20:35):
incentives.
That's how I see it as well.
So totally, I see eye to eyewith you on that.
And now going back to theprotocol verification protocol
that alex was mentioning.
So it's basically what we havedone.
We figured out this like amodular structure for models
where each contributorcontributing they're building a
(20:57):
model together, but they are nottraining this monolithic dense
transformer together.
Instead, like they're providingthis modular contributions,
which are called like adapterlayers, like lauras, they're
providing that and they stack ontop of each other like Lego
pieces and all those Lego piecescome together and build a model
(21:20):
.
So it's a fundamentallydifferent approach of seeing
this thing.
Like a lot of people, a lot ofteams, very talented teams, are
trying to reduce thecommunication overhead over data
centers to be able to train adense transformer, monolithic
model.
We looked at it a different way.
We looked at it in a way likewhy can't we just make the
(21:41):
architecture itself modular,which is very much in line with
the industry trend at the moment, because MOEs are a rage right
now?
Deepseq is a mixture of expertand this is a modular
architecture, so they keep thetransformer core and modularize
it for efficiency.
We did the same and then whatthat enabled us is that the
(22:01):
contributions are, first of all,they can come from one data
center each contribution, so youdon't need that much of a
communication to begin with.
And second, the contributionsare small enough that you can do
zk verification of that.
So when the person or developeror the data center developing
this modular adapter, they run azk verification on top and it's
(22:24):
way less overhead.
But then even that was likehigher than what would be
acceptable in a productionenvironment, because in
production environment you wantseconds, not minutes, not hours
or not days, and the previousexample of verified training
were weeks.
So what we did?
We looked into how this worksand we saw that the previous
(22:48):
attempts of ZK verifying thesecontributions were tied to
compute.
They were trying to verify.
If compute was burned Then, likewe, I personally have been in
machine learning for more than10 years and my team has like
more than 40 papers published inmachine learning together.
So we have like extensiveexperience of training models in
(23:09):
the traditional ML and we knowthat's not how it works In
traditional machine learning.
You do these massive trainingruns, you get the result and you
look at the evals and if theydon't match up to your standard
you just throw them out.
So they don't really count howmuch compute they burnt.
They count how muchcontribution, how much
(23:30):
improvement to the model thatwas done for this specific
training graph.
So that was kind of themismatch between the traditional
ML and the Web3 AI side.
In Web3 AI we were justincentivizing compute, which
makes sense to some extent.
But if we don't actually lookat the quality of those compute
(23:52):
usage then we get lower qualitysupply on our network and that
does not incentivize theconsumer set to come in and put
their capital in and use theresources.
That's there, yeah.
Speaker 3 (24:03):
I mean incentivizing compute.
Your level of abstraction issignificantly higher than
compute and you shouldn't carewhere it comes from.
I mean, I can train the modelby hand if I have to right.
It doesn't matter what GPUs Ido, as long as your evals I mean
they're good enough for youknow, whatever your standards
are right.
I think Moose is doing asimilar approach.
(24:24):
If I'm not wrong, in the distrothey did present a mechanism
where they're reducing thecommunications overhead for
distributed training, but alsofor the verification.
I believe they took a similarapproach.
But it's fascinating, very cool.
Speaker 4 (24:39):
Yeah, thank you.
So I want to close the thought.
So what we did with this likewe actually are doing ZK
verification of evals, ofimprovements, so it doesn't
matter, like you might just typethe weights by your hand, we
don't care if you increase theevals, that's okay.
(25:00):
And that's how we were able toreduce the verification overhead
.
And we believe this is thefirst and we open sourced it the
both the paper and and the code, and it has been peer-reviewed
by people at Stanford, U of Tand whatnot.
It works and we believe we havesolved the ZK verifiable
(25:20):
training.
We are using that frameworkinitially for fine-tuning only,
but this is the stepping stoneof going towards actually having
fully ZK verified training in adistributed setup.
So that's what we are up towith this.
When token, we have launchedthe first version of our product
and we are planning on going tomainnet very soon as well, and
(25:44):
token, as we already discussed,like a kind of similar thought
process as you, we believe thatwe want to do the product market
fit and feed in the token intothat.
Speaker 2 (25:53):
I mean it sounds like you're getting there because
you've got like 14,000 peoplethat just rushed to use the
bakery.
That's what they call.
The product is the bakerybasically overnight.
I mean it's been live for what,two months.
Speaker 4 (26:04):
Yeah, yeah, it has been live for a month, a month
and a couple of weeks and wewere trending in the top 10 of
product trend when we launchedit.
So I don't think that has everhappened in crypto Top 10 of
product trend worldwide when welaunched.
So first ever crypto product todo so.
But anyway, I'll stop shillingmy own thing.
But going back to what I wassaying, is that, like that's how
(26:27):
we see it, like incentivizingactual equality, not
incentivizing, you know, justwhatever contribution, and
that's that was the gap in allof these machine learning
related resource networks.
Right, and this is so versatile, the framework we have built.
It can snap on top of anycompute aggregated network, like
Akash, and instantly convert itto a verifiable training
(26:49):
network.
So the compute can just feedinto that, and that has a lot of
value as well, because itincreases the value of this like
compute aggregated networks atthe same time.
And the last point I want tomake about open source we love
open source.
I believe in open source aswell.
There are lots of upsides tothat, but there are some
(27:10):
downsides.
We have noticed as well I'msure you have as well.
I was discussing that with Alexa couple of episodes ago that
we open sourced our research ona very fast model verification
algorithm watermarking andfingerprinting early last year
and that has been adopted byactually quite a few well-known
(27:35):
Web3 AI projects in the spaceand they're using that, which
I'm happy about.
But the downside of this isthat we were not credited when
that algorithm became the corepart of the protocol.
So sometimes that happens rightLike you open source it, people
like it, but they just use itfor their own use.
Speaker 3 (27:54):
What kind of licensing did you have on the
open source model, on the opensource code?
Speaker 4 (27:58):
On that, we had MIT.
So again, like I'm happy, Verypermissive.
Speaker 3 (28:04):
It's one of the most probably permissive license
right.
So an advice I would give toavoid such scenarios in the
future is to have Apache 2.0license, where if a protocol
includes your code, they have tocredit your or GPL like GPL V3,
I think.
Speaker 2 (28:24):
Is it like AGPL or GPL V3 has the same thing?
I mean, because they have toopen source it, not just credit
you.
I think with Apache they stillhave to credit you, but it
doesn't have to be open source.
Again, I'm kind of biasedbecause of Brian who helped
write those licenses, but for me, if it's my company and my team
making it, I like it to be aGPL or GPL.
(28:45):
Just because open sourcesoftware for me is my legacy, I
want it to live forever.
I want it to go to Mars, likebash Bash is literally powering
quadcopter on Mars, and I wantit to be there for everybody,
for the world.
Right, and I feel superstrongly about that.
It's a personal choice forother teams, like I don't judge
other people for choosingdifferent licenses and whatnot,
(29:05):
I get it, but for me that's it's.
I don't know again, it's mylegacy etched in stone, etched
in bare metal on the internet.
Speaker 3 (29:13):
Make me think.
I mean, I think I probablyshould consider a GPL license,
because there are similarscenarios with Akash too, Like a
lot of people fork a codeblindly, copy our code.
Just, I mean it's so funny.
Even some of our plugin codepeople just copy, paste and just
control F, change namesliterally, Like we saw that
(29:35):
happen so many times.
A lot of the deep end networksthat are launching just with
Akash straight up code base, youknow, change the names.
There was even instances wherethe founders were like doing
demos and I could clearly seeAKT in there as a payment
mechanism.
They haven't even bothered tochange the currency.
But yeah, but it is a problem.
(29:56):
I mean, crypto is sadly.
I think a more restrictivelicense is definitely an answer,
which I hate to be saying that,but I bet the ones that copied
you have enormous funding.
Speaker 4 (30:12):
Yes, yeah, close to nine figures.
Yes, yeah, close to ninefigures, actually.
Speaker 2 (30:17):
Yeah, and we have the receipts and there's been some
drama on Twitter about thisparticular company, with some
unethical things in the past,from one of the founders
especially, and that areunrelated to open source, right,
but it's kind of like the Apple.
I don't know what the rightphrase here is, but the Apple
doesn't fall from far from thetree or whatever.
I'd be curious to hear what youthink about this, Greg, but I
was like ready to go to war.
I was like dude, let's callthis out Like it's, it's
(30:37):
bullshit, you know, and I wouldgo to war.
You know what I mean?
Yeah, Like right, Like it'sit's, it's complete BS.
Maybe there's still time for usto do that bit.
Back to where we started thisconversation.
It's about execution, thinkingabout things from first,
principles and the peopleinvolved, and even if it was
(30:59):
closed source software, you knowit's easy to copy people and
we've seen that happen in SaaSall the time.
Smaller team comes by, buildthe exact same product out,
executes the well-funded Goliath, and then they've got a huge
company and the other companyends up dying at some point.
That's just been a repeatedstory that we see, and I think
this same thing applies fromopen source and the thing that
(31:19):
happens when you don't opensource is you lose all those
network effects, you lose allthat community.
And you know, one of thereasons why I joined Solana is
because of that patternrecognition we're talking about.
I saw the same thing that I sawin you, Greg, with Tully and Raj
, and they were shipping day oneopen source.
And I even asked Tully as kindof a test, like, hey, what
happens if someone rightwhatever left click copy or
(31:41):
right whatever the right click?
You got to do copies, your codebase?
And this is back when they hadthat office on Howard street.
They were just starting out,they hadn't even launched main
yet, yet, I don't think.
And they and they and they wereshipping a really good idea in
public and he was like pleasefork it, go for it, Cause the
other part of it about theexecution, especially in crypto,
is getting to where you are,Greg, with a real community, a
(32:03):
real customer base, actualpeople providing node services.
You know, leasing, leasing andregistering like providers, dude
, that is a whole ball of wax,it's a it's.
It's hard to replicate that, Um, and you're only stronger
because of the open sourcecommunity.
I just, I just believe that youknow to my core and ultimately,
you know, karma is a bitch forthe people who don't respect
(32:25):
that.
Speaker 3 (32:26):
Um yeah, so, yeah, on that point right.
So I mean open source has somany benefits that outweigh the
risks, right, I mean risk being.
You know you're getting copied,but I think on the long run you
eventually end up winning,because you can fork the code
but you cannot fork thecommunity.
I remember Solana launching.
You know, I was there literallytwo weeks ago.
(32:47):
I did the first video and theyneeded testimonials.
If you go to Solana's YouTube,I'm like one of the first videos
they did In that office.
I had a photo shoot too.
I took like black and whitephotos of them, literally like
one week from shipping.
So funny.
That was like right before covidhit and they hit.
Right when covid hit theylaunched and I was concerned
(33:07):
because we were like planning tolaunch.
A few months from there theycrashed and it was crazy.
Unfortunately, I was able tolike rake up some solana tokens
at like one dollar, two dollarrange or less than that.
I think it's significantly lessthan that, but they were.
So I mean totally.
I remember totally was like thesoftware was not ready to get
shipped, but totally was likewe're gonna ship and that's
(33:30):
that's execution right, becauseyou're never ready.
It's like being a dad you'renever ready.
It's like being a dad You'renever ready.
You just got to go for it.
It's never going to be perfect,but I think it takes the
pragmatism that tells you quitea lot on execution Out-execute,
you know, people are going tocopy you.
People are going to copy youbecause they're good at it, but
they cannot execute you right.
Out-execute you right.
So I would execute ship asaggressively as possible.
(33:53):
Build that community based ontruth.
People like drama, you know,people like the fact that you
know somebody copied you.
That actually builds a strongercommunity in my opinion.
It did for me, for us, right,and you get more sticky
community.
But focus on the community veryimportant, I think, one of the
best community builders I see inthe space and, of course, like
(34:21):
and toli and they're very, verygood at it.
But I think in jacob constreborn jacob from bit tensor or
even shaw walters from from ai16z.
I mean, it's more modern times,but but I I admire community
builders and I studied them.
Um, I think Jacob has quite alot of admiration.
I mean really think about howhe builds BitTensor.
Go to the discourse andunderstand how the discourse
(34:41):
works and how the communicationsare working there.
Spend a day in BitTensordiscourse.
Spend a day in AI16 discourse.
I was so impressed by Shawaw'sgtm in terms of like they had
this eliza framework open sourcefrom day one of the top
trending repo in in the women,it's like for a month or
(35:05):
something.
Like it's crazy, and I wastraining on github like 10 years
ago, so I understand what itmeans to be trending and what
that brings you in terms of likeinbounds and it's just a crazy
phase you go through.
And the reason why I think hewas so successful in shipping
software is you know his hewould do like this, like massive
(35:28):
three hour long videos onexplaining the software and you
would think like this video hasto be succinct and like to the
point because nobody has time,like no.
He proved that if you explainthings well, people will sit and
listen to you and that's whatJoe Rogan did and a lot of the
new podcasters long-formpodcasts are about.
(35:51):
If the content is interestingenough, people will listen to
you and think you know.
Shaw definitely understood thatand his videos are like I watch
a three-hour video on fridaynight to understand eliza and I
sat through it, right, you know,and I was surprised that I
could spend.
I mean, I don't watch tv thatmuch, I mean especially youtube
(36:11):
and whatnot, not for that longbut but there's something to be
understood about these foundersand how they build communities
around open source systems.
Same thing with Jacob.
Jacob would spend hours on out.
Jacob's weapon is Discord right, and everybody has their own
choice of platforms, right.
That is something that I thinkcannot be forked, no matter how
hard you try, and I think that'sone of the reasons why I think
(36:33):
I'm not endorsing ai16z as atoken or any of that sort.
I'm just observing communitybuilders and seeing their
patterns.
Uh, like you say, it's patternrecognition.
Right now, I can recognizethese patterns among successful
founders, successful builders.
They all have one thing incommon is cut the bullshit, be
pragmatic.
Don't worry about gettingjudged.
If you have to go on athree-hour podcast, three-hour
(36:56):
video, where are you going tomake mistakes?
It's not going to be perfectbecause it's life, right.
I've noticed so much, so much,I guess, like, like engagement.
When I do live raw videoswithout filters, you know people
love it.
I mean, I just randomly go ontwitter and be like hey, I'm
(37:19):
going to live stream my codingsession.
People love to see me code,right because, like people, love
to see other people play games.
You know things of that naturewill give you quite a lot of
like true community.
That well-funded company withbig PR teams will be prohibited
from going live because theyhave they're too afraid to lose.
Speaker 2 (37:41):
Let's, let's build some AI agents on live stream.
Next time, bidhan, I'll watchyou, I'll do commentary, I'll be
.
I'll be like the sportscommentator I like.
Speaker 3 (37:50):
Sahil from Gumroad, another like Sahil.
Sahil from Gumroad, anotherfavorite Sahil from Gumroad.
All these phenomenal communitybuilders, right, I don't know if
you know Sahil, but he's acharacter.
I follow him on Twitter.
Have you seen his YouTubevideos where he does this PMing,
where he does this amazing wayof how to write a PRD?
He's a great product manager.
(38:11):
The way he would approach a PRDusing AI these days is just
amazing because I need to checkthat out it's amazing.
So he starts with, like blackdocument why and what, why they
want to do something and whatthey're going to do Like a few
lines and sentences, plainEnglish, and he develops this
PRD using different tools, usingDeepSea, using, you know,
(38:36):
chaigpt, you know O3, andactually he does the V0, the
whole, like you know, userinterface thing, and he does it
live.
And that's awesome Because yousee the mistakes people make and
you see someone work.
A great way of improving yourwork is to imitate somebody's
work and I had that in my careerfor a long time.
I imitated when I was learningGo.
A lot of my imitation came fromMish Hashimoto, the founder of
(38:56):
HashiCorp, developing Terraformand whatnot.
I learned from how these peoplecode.
Looking at the code.
I love learning by other,watching experts or people that
I consider good to work, and Ithink that speaks quite a lot in
how these founders buildcommunities, and we need open
source.
I think is a great way toimitate them.
Speaker 2 (39:19):
So we only have a couple of minutes left and I
couldn't agree more witheverything you said.
But there's something I reallywanted to ask you, greg.
The conversations I've had withBen on the last couple of
months have really helped medevelop this thesis around open
intelligence, and the term openintelligence is actually came
from a brainstorming sessionthat Markeisha and I had.
She leads marketing over atBagel, but basically it's
(39:41):
artificial general intelligenceor artificial super intelligence
.
On crypto rails.
It's open source, right Opensource or AGI on crypto rails.
I have come to believe that thisis really really important for
humanity and for our industryand that it really should be the
North Star of every founder.
That's, building criticalinfrastructure should be to help
(40:04):
make this happen, to accelerateopen intelligence.
Bagel is critical for it and Ithink a cash is critical for it.
I think for obvious reasons,but I also am asking myself,
like, am I crazy?
Right, like you know, can we dothis?
Can we get a bunch ofentrepreneurs who are building
these systems to work togetherto create open intelligence for
(40:24):
the sake of humanity?
Am I being like overly dramatic?
Is it even possible?
Like, and then, if it ispossible, like, what are the
components?
It's like I look at it like afactory system and there's like
inputs and outputs and there's abunch of tooling that needs to
be built in between to make ithappen, and I I'm like trying to
develop like a vision for whatthat is.
So that's, I know that's a ton,but generally speaking, like
(40:47):
how am I doing?
Here are we, are we going tomake it to open intelligence?
Speaker 3 (40:51):
I mean it's not something we, you know,
something that it'd be nice tobe open source.
It's something that weabsolutely need.
It has to happen because,historically speaking, any
sufficiently importanttechnology that reached mass
adoption is open source.
Linux, world Wide Web, you know, name it.
(41:15):
In modern era, even, as amatter of fact, phones, I mean,
without Android, you wouldn'thave the same level of adoption
if it just were, yeah,transformers for open source,
right?
So open source is so criticalfor mass adoption because the
network effects would not comeotherwise.
And we've seen very clearlywhat happened with DeepSeek,
(41:38):
right, very, very obvious.
Like, until DeepSeek, thestate-of-the-art model was a
closed model and DeepSeek cameand changed things, and this is
just going to be a steppingstone in reaching ultimate
intelligence.
Now, have we done that before?
Yes, we have World Wide Web.
Like, without the companies thatinnovated early failures,
(42:01):
successes, whatnot we wouldn'thave the open internet that we
so much use.
Right, 20 years from now, youlook back and be like we'll be
asking the question.
So, looking back to internetitself, right, it did not begin
as open.
Internet began very close, aolwas literally shipping cds.
You needed aol to get on theinternet, right, look where aol
(42:24):
is today.
We're still around.
But it's not the internetcompany, right, it is the
companies that actually operatethe internet.
So, similarly like open ai, Ithink would be the new aol of
the companies that actuallyoperate the internet.
So, similarly like OpenAI, Ithink would be the new AOL of
the world.
Yes, they were the first toinnovate with ChatGPT.
I mean not first to innovate,but at least first to get to
market with ChatGPT anddefinitely like proved product
market fit with ChatGPT.
(42:45):
That kicked off an entireindustry.
But I think the trueintelligence through AGI will be
open source.
And if you look at how do youmonetize and how do you sustain
open source software,decentralization is the way.
Crypto is the way.
There is no better mechanismthat I can think of and I have
(43:06):
factual evidence for this,looking at successful open
source companies that open thisproducts that could not create a
successful mechanisms tosustain themselves.
So, looking back, there's nodoubt in my mind that crypto is
the only mechanism, onlyframework to have monetizable.
You know machine intelligenceright.
(43:28):
So that's really comes down toit.
I'm giving a Bindan's pitchright here.
But open source is good as longas it can be monetizable and
sustainable, and I think cryptois a great way.
So, yeah, it's going to bemultiple people, multiple
organizations, multipleprotocols, rather interoperating
(43:48):
with each otherpermissionlessly.
If Bagel wants to use Akash, noone can stop Bagel from using
Akash.
No one can stop you fromgetting the H200s for $0.99.
Right, and that's what it comesdown to.
If I want to use Bagel, no onecan stop me to use Bagel, and I
(44:09):
can build protocols on top of.
Speaker 2 (44:11):
Akash.
Speaker 3 (44:11):
Unstoppable.
Open intelligence Correct.
It has to be unstoppable, ithas to be.
Imagine you're a developer andyou're coding at 2 o'clock in
the night and you want to use aproduct.
Last thing you want is anapproval from someone that's
going to wake up on a weekdayand give you an approval to get
your account, increase yourlimits, all sorts of things.
I mean.
You want to review code, youhave a question?
(44:32):
You should be able to go readthe code yourself and you have
AI.
Now they explain a lot ofthings on how things work.
I mean, back in the day it'slike actually read the code,
figure things out.
That takes a while, but I canfeed that into an AI and be like
explain what's going on.
The logic here, right.
There are so much tools nowthat open source adoption is
(44:56):
going to be accelerated withtools as well as creation as
well.
Speaker 2 (44:57):
Yeah, I couldn't agree more with you, man, and I
appreciate you coming on.
We can wrap now.
Thanks so much, dude, awesome.
Speaker 1 (45:10):
You just listened to the Index Podcast with your host
, alex Cahaya.
If you enjoyed this episode,please subscribe to the show on
Apple, spotify or your favoritestreaming platform.
New episodes are availableevery Friday.
Thanks for tuning in.
I feel like.
Popular Podcasts
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com