Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
SKALE School: IronClaw with Cameron Dennis (NEAR)
Recorded: March 3, 2026 Duration: 1:00:54Space Recording
Full Transcription
Dale school time, ladies and gentlemen, boys and girls, developers and developettes, I don't know.
Welcome, we're back with another episode.
Obviously, last week, very hackathon, very hard on X402 and agentic commerce.
But, you know, sticking with that idea of AI and the agenticness of applications. We have a very special guest today, Mr. Cameron
Dennis of Near AI. Why don't you give us a little introduction about yourself and we can
kick things off. Absolutely. Thanks for having me, everybody. My name is Cameron Dennis. I've
been in crypto actually for about 10 years. I've been working on Near for the last five and AI for the last two. And so I currently am on the new AI team.
We are building verified with private AI, making sure that your agents are secure and that Sam Altman can't read all your prompts.
And so this is, we think of, you know, pretty much an existential crisis to humanity.
If we're giving away all of our humanity to a couple organizations that
probably will use that information against us and so not to get so deep so quickly
but uh we want to make sure that that verifiably private infrastructure exists
and so yeah that's what we're doing and happy to dive into it 100 super excited to dive into
all of that but before we do so, as always,
Toya, Mr. Professor, how you doing, man?
Good, how are you?
Not bad, not bad.
A lot less pain.
A lot less pain.
Did you remember yourself going to the gym?
I did last week.
I feel a bit better. Every time I go every time i go you know it's this is
pain is less which is nice it's good i don't think i stepped in a gym for over like a year now
wait really oh wow you get you out from behind your desk man i well i work out I just don't go to a gym um so I have a Peloton that's what I like
like it um there's no gyms near me that are like I don't know not I like that I like a small gym
I don't like the big big big box games anyways I digress um there's no time to get out behind the desk because agents
don't get out behind us they run 24 7 let's talk about those
100 yeah so you know nair has obviously been doing a lot you know in the in the ai space and
why don't you give uh the people a little bit of a rundown of you know what what
near near AI is and and how then we can jump into the likes of Ironclaw yeah for sure so it's
probably good to take a little bit of a step back and kind of tell where all this is coming from
and so Ilya and Alex two near co-found they actually started NIR in 2017 as an AI company.
And this is right after Ilya co-wrote the Attention is All You Need paper.
This is the most cited paper in all of AI.
It essentially describes the T in ChatGPT, stands for Transformer.
And this is kind of how all generative AI works today.
And so after he wrote that, and he's good friends with Alex back from the competitive
programming days in Eastern Europe, they were like, holy crap, we just opened Pandora's box.
And this thing will eventually teach itself how to do everything essentially better than a human
can, including exploding every vulnerability on the internet. And so that was sort of like one of
the main issues that they're trying to solve. But in order to solve that, they wanted to create a code generator.
And so this is like a huge vision.
Like this vision is, it just goes on and on and on.
But the whole point is how do we essentially have a ledger where people can verifiably
own their money, data, and governance in a post-AI world and use that ledger to pay people to do
things to further improve the economic and sort of social model of the new internet. And so that
is what near blockchain became. And that's sort of just like the base layer is the incentive layer
that needs to scale forever and decentralized linearly with scale so no single party can essentially control truth on who
owns what then the second piece is what we call near intense and we frankly think blockchains are
kind of clunky and like not always the easiest for people to use and so the goal is to make make
blockchains invisible and it just so happens and like you know we've known kind of that the future
of ai is like intent driven i don't know if you saw that mitch mirrors like post he's the uh you know very short
term ceo of near a of open ai um he was saying like the future of ai is all intent based because
all these humans have intents and you want your ai is just to do your things for you and so how do
you actually build like the liquidity layer for Intense?
Because, you know, if I have an intent to go get my car washed,
like I need to match people who wash cars to me with a certain amount of money to match that order.
Very similar to how like exchanges work and very similar to how like AMMs work.
So Near Intense sort of is building this like liquidity layer for agentic
commerce, starting with crypto. And it's starting by making all these blockchains completely invisible
from one another. This is kind of the whole multi chain chain abstraction piece of near.
And the whole point is just like make blockchains invisible because although near can scale and has
all this like, you know, great attributes to it, we understand it's going to be a multi-chain world, and that's totally fine.
Because the thing that we really care about is Near AI.
Near AI is the front end of the web,
where everyone's going to be interacting with their agents
and sort of interacting with superintelligence.
And we want to make sure that that is user-owned.
Because if it's not user-owned, as I mentioned before,
as we've seen before,
in a lot of like, technology and capitalism, your information will be exploited for money.
And so but this is too big. Like, it's too big to talk about like, everything that you do,
everything that you buy, like, we need to make sure that that is owned by the people.
And so how do you do that? Well, first off, you need a blockchain to sort of
be that trusted database. So you can like actually have a key that is yours. Second off, you need all
this like off chain stuff, making sure that your prompts are private. And I could dive into that a
bit. And so what your AI is building is it started off with like a private inference network. This
is where we run large models inside these trusted execution environments, TEEs. Think of it like a black box that sits on the GPU that you're able to just open. No one can
see what's happening in this black box, and you can run arbitrary code inside of it. So you could
actually run large models. So what you could do is you could take your prompt, encrypt it,
send it to the black box, decrypt it inside the black box. We can't see anything, and we can prove
And then you run inference on top of it because one of these large models is running inside
that black box.
You take that answer, you encrypt the answer, send it back to the user, user decrypts it.
You have now privacy.
So that is like one base product.
And I can go on and on because like it sort of evolves into Ironclaw,
but I just rampled for a while, so I'm going to stop. Yeah.
All good. All good. Lots of really good stuff there that I kind of want to pull out and dive
into to kick things off. I think it's really interesting because I think we share a lot of
the same philosophies around where the
blockchain ultimately will sit and i think we're starting to see more and more people in the space
finally come around to this idea that i think both scale and near have been talking about for years
which is this idea that blockchains are kind of really not intuitive technologies and they're
really not enjoyable for humans to use.
I mean, arguably even like the basic internet is proving to be not the ideal way to interact with the internet, right? Or interact with technology. We're seeing this shift to prompting,
right? Because it's just easy. I felt what I want and it figures it out for me.
And so blockchain requires you to go so far up this stack of like complex decision making.
And AI kind of brings us all the way back down to what humans are actually, well, good at normally,
which is probably just being able to have a conversation.
Minus social media addicts in the world that can't have a Facebook conversation.
But they can type in their phone.
So what I really like is that you guys have kind of a similar position on that.
And I think the part about the data privacy and visibility is really interesting.
Because I think there's a lot of nuance to this topic.
What I mean by that is I fully agree, right? In an ideal world, your data is your data,
and it's private, and people can't train on your data. They can't determine everything about you.
They can't use a bunch of algorithms to pinpoint things you may want to do, things like that,
right? But we obviously know that the cloud era changed how data was handled.
Everything is online.
There's no real way around it.
Everything is online.
Being able to carve out some things and make them a little bit more private
is definitely a step in the right direction.
We've seen a ton of companies over the years, right,
attempt to do various things from, you know,
social media all the way down to search, like DuckDuckGo.
Again, with varying degrees of success.
I think what's really exciting about agents
and where we're at in call like the crossing the chasm graph,
right?
We're not at the chasm yet.
We're at the kind of like, what are we doing things, right? We're not at the cancer yet. We're at the kind of like, what are we doing things,
right? Maybe, arguably. I think we're at the point where we're like about to,
we're about to get there. I think it's still very technical heavy. I don't think these things have
been ingrained in everybody's day-to-day life 100%. I think that starts to get us to the point where we are.
And I think as we
see this ramp up, what's really cool
about what you guys do with Ironclaw
and a lot of these other teams
building out their own forms of Claws, which
I'll just say here, I don't know if anyone
else has said this, I think Claws is just like a new technical
primitive. It's like
it's the lobster primitive, so we could just
maybe agree on that.
What I really like is you guys are giving options right and i think over time we're going
to see more and more of these options come up just like you have dozens of options for your phone
carrier and dozens of options for your insurance dozens of options for your bank and etc etc i
think you're just going to have a lot of options for your agents. And I like that you guys are really focusing on the privacy aspect.
It's really great.
Are you guys, question, do you guys have your own inference in a house?
Are you leaning on someone like a Venice?
Are you running your own GPUs?
What does that look like to make that actually private for you guys?
Yeah, for sure
and um yeah i just want to quickly speak to we need to meet people where they are like we're not
going to force them down this path and i think everyone that's been in crypto for a long time
realizes this you can't force a bunch of people to write down a 12-word seed phrase in english
if they only speak cambodian and write in cambodian and so like yes and i do think agents are going to you know be a
really good sort of vehicle for user owned ai a user owned like money adoption because it can sort
of abstract that complexity away as long as it can maintain their key which requires pass key support
at the protocol layer and so if you're able to like actually just use your face to create
an account and generate a key pair that actually gets stored on chain, that to me is like a
prerequisite. And I actually see that crypto wallets and accounts are really just the auth
tool. Like we need to move away, I think, from accounts to like authentication. You have an
account, but like it is a near account is the auth to your agent and to
your user owned identity so um just want to quickly touch that and i can get into like the
hardware and so yeah um i've scoured the earth for the most affordable and reliable H200s and B200s
that I can find.
We don't run our own data centers,
but we work directly with a bunch of data centers
to rent these GPUs from them.
We get SSH access, we manage,
we actually virtually manage the machine,
and we run these models,
then primarily open source. We're starting, we have a
really cool initiative, where we can actually run closed source models, we can give the close the
model provider that doesn't want to leak their weights, the option to run their models on our
hardware without us seeing the weights, which is a really big deal. Because if that is the case,
then we can actually verify privacy for larger models
that are proprietary like Anthropic, like ChatGPT.
And so that's a whole nother research project
that I'm just gonna put in that camp.
But for this camp, it's like we have open source models,
we run them on less hardware,
we provide the attestation services
like the Nvidia and Intel attestation services like the
Nvidia and Intel attestation services. The whole point here is that this hardware actually stamps,
like think of it like a stamp, it is stamping saying, hey, this prompt is, this is the model
that is running inference on this prompt. This is the size of the prompt. This is like all these
details of this prompt, but you're not seeing the details. You're just getting a verification saying this is what's
happening. And that's really important because then you could take that stamp and then cross
check it with like NVIDIA's attestation checker or Intel's attestation checker saying, hey,
is this running the way it's supposed to? And every user can do this. And this is what I mean by verifiably
private. It's that you, the user, you, the company can verify yourself that this is running the way
it's supposed to be. And I don't need to trust OpenAI saying, hey, I actually, you know, I made
a prompt to chat GPT Pro, but you're actually giving me OSS 120B, which is a much smaller,
cheaper model. And I'm like, I actually want the best result.
I just don't want the cheapest result.
And companies do this, obviously,
because it takes a lot of energy to run this inference.
And so we run the inference.
We provide an OpenAI-compatible API key
to organizations that want private inference.
And so why this is so important,
the OpenAI-compatible API,
is this is what they're using already.
Like they literally just have to drag and drop an API key
and it's the same process as what they're used to.
But now they have this verifiable privacy layer
on top of it.
And so right now it's for open source models.
We're experimenting with closed source models.
We do have a closed source model option
where we actually just keep models anonymized,
but it's not private.
What I mean by that is,
let's say we have an Anthropic API key.
We have the key.
This is what Venice does.
We have the key and we say we don't store,
we say we don't log, all this other stuff,
but there's no way for a user to verify that.
You just have to trust our privacy policy.
And frankly speaking, I don't trust privacy policies.
I've been burned by privacy policies my entire life.
I've been burned by terms of services my entire life.
F privacy policies, like I'm done.
Like we actually,
and super intelligence is not gonna care about privacy policies. It's going'm done. Like, we actually, and superintelligence is not going to
care about privacy policies. It's going to do the things that are going to advance it in the best
way it can. And so my point is, I want to be able to verifiably prove that my prompts are encrypted.
And this other approach, this anonymized approach, which is what Venice does, and we do it as well, is we'll take your prompts, you can use Sonnet 4.5,
but we are just all sort of mixing these prompts together
and then sending it to Anthropic.
So Anthropic can't tie the exact prompts to the exact user,
they all see it coming from one key.
And that's cool, but I think it's a short-term solution
for a long-term problem where we should all have
verified with private guarantees. Like, I think that's the prerequisite. And that's going to be
a standard that I, just in my crystal ball, assume governments will push top-down to protect
consumers. Very interesting. Very interesting. I think there is, I mean, we've already seen,
I think there is, I mean, we've already seen, we've seen a few things, I think, in the last couple weeks that I think are pretty, pretty rare.
We're seeing this really interesting call like battle of the model providers where they're
arguing over like using each other to train and like prove information.
And so I'm curious if the verifiable tease potentially help with that.
I don't know if it does, but there's a whole like, there's a whole kind of thought on that,
but we don't have to go too deep into it. But we're seeing, I think it was Minimax, Deep Seek, and maybe Quen were getting accused by, I think, Anthropic. of basically attacking the claw endpoints to try to figure out how it works and pull out the weight,
stuff like that. So that's one thing. The next part is we did have an incredible release.
I'm pretty sure it was yesterday. AI just blurs together these days, so it's hard to
I'm pretty sure it was yesterday.
AI just blurs together these days,
so it's hard to remember what day something happened.
But within the last 40 hours,
Quen dropped a family of small models, small language models.
Specifically, the one that I think has taken the,
at least the research world by storm so far,
is the 9 billion parameter model from the quentin 3 lineup which is um so
far out um outperforming gpt uh oss 120 billion which is uh the last um not mistaken it's the last
open source model that open ai put out uh what about a year and a quarter ago a year ago, a year ago, nine months ago. Yeah, yeah.
And that model has been one of the most commonly used open weight models
sold by top tier inference providers like Grox, Rebus, etc.
And we're seeing a model that is 10 times smaller, 11 times smaller, just dominated. So it really opens up the door as well, I think, for, you know,
inference providers like you, as well as seeing more, for, you know, providers like you,
as well as seeing more of these agents run on commoditized hardware, right?
Yeah, so I can tell you why 120B is the, you know, most popular. And it, obviously,
the performance is great. And it is smaller to run than some of these much larger models.
than some of these much larger models.
But a lot of businesses in the West
aren't allowed to use Chinese models
because the outputs are biased.
And so that is the concern.
And so a lot of enterprises that I speak to are like,
hey, you know, GLM-5 is awesome.
Like the reasoning capability,
the way you can do, you know, coding is fantastic, but we don't want a model influencing our company's decision-making
if somebody asks it a question about TMN square. And so there's just this, like, even if it's
open source, even if it's private, there's private, there's a huge demand for Western open source models that don't have Chinese bias.
And so, but for tinkerers and everyday people, the Chinese models blow the American open source models out the door.
Like, it is insane.
Like, I love GLM-5.
Like, it feels like quad.
It is awesome. For literally, literally like 15 times less the price. Like it is, it is incredible. And to talk about like the
edge AI point. Yeah. At the end of the day, I do think that, uh, humans will evolve to build the
most efficient models that run on device. Like edge is the sort of like end goal of user-owned AI
because you do get like the highest privacy guarantees that way.
But all this like TEE, verifiable privacy operating in the cloud
is just what we have today that like works really well
works really well for like an industry like robotics like for robots you don't want to have
for like an industry like robotics.
super expensive clunky hardware on the machine it's going to make the robot more expensive and
energy requirements all nine yards but you do want that reasoning capability of a glm5 but you don't
want to send everything that's happening inside that robot all that inference to a third-party
inference provider that can see everything that's now inside their consumers houses.
And so there's this like really interesting balance here of like, how can we leverage the top, the best open source models with the privacy guarantees of local deployment?
And that's kind of where the TEE stuff sits today.
In a perfect world, we have fully homomorphic encryption and everything else.
We're super open to this, but in ZKML, it's just not there yet from a performance standpoint.
And so this is, I want the edge AI future to happen.
I love what ExoLabs and other people are doing.
It's just getting there.
And it's probably going to take a lot more time than superintelligence existing in the cloud.
Okay.
I like where you're going with this.
I think, okay, so an interesting question that I'm very curious about is the AI,
we've seen this trend and obviously this trend seems,
it seems likely that it will continue because,
well, it's the only way for the model companies
to functionally continue to dominate is,
models get stronger, but they also get cheaper.
And so we're seeing this split, right?
Where the top leading model providers,
both from the US and from China,
continue to get better and better.
And obviously there's a lot of other countries
contributing there as well.
Some really great models coming out of companies
like Mistral from France and a bunch of others.
But you've got better and better models
and you've got cheaper and cheaper inference over time.
And we're also starting to see more efficiency come in at the consumption layer.
So, right, OpenAI introduced WebSockets just, I think, on Thursday or Friday last week,
which they're claiming is going to potentially cut down inference costs 30% to 40%
by keeping kind of a connection open instead of having to constantly reprompt compared to prompt caching. Thursday or Friday last week, which they're claiming is going to potentially cut down inference costs 30% to 40%
by keeping a connection open instead of having
to constantly reprompt compared to prompt caching.
We are seeing a significant simplicity in using
skills over constant tools and this dynamic context selection,
which again compresses your context a little bit more,
it keeps your costs lower.
And all of these things in turn have allowed,
I'll just call it like the open claw dynamic to start.
It is feasible now, and I actually think we can,
I think we can thank the Chinese model providers
significantly here, because I think their coding plans
are the ones that really enabled this.
We've been able to see these agents start to run and be relatively cost-effective, right? You could
run an open claw bot on functioning $1 a month Minimax plan, and it's more than sufficient for,
I would say, the majority of use cases. I don't think Minimax 2.5 is the greatest at calling tools compared to most of the GLM models, a lot of the Kimi models, et cetera, et cetera.
But it's more than sufficient for, like, I'd say your average tinker.
And it just gets better and better from there.
And so this has then opened up the door, I think, where, you know, what's really interesting for us is the impact on a Gents of Commerce, right?
There's, I think there's kind of two sides
to a Gents of Commerce.
I think part one is the thing that everyone today
can mostly wrap their head around,
which is, well, we have people already buying things.
Can agents make that more efficient?
Can they make it more cost effective?
Can they do multi buying process across merchants and just like make a user experience better,
things like that.
And we're seeing a lot of interest in that area
from Google, Shopify, Stripe, PayPal, et cetera.
And then on the flip side,
we have this more, I'll call it like AI,
technical agent of commerce,
which I think is very interesting to everyone running their own agents, which is, well, the
agents, setting up an agent is really annoying. Giving it all these tools is really annoying.
And it's nice to kind of be able to let an agent just kind of figure out its own way,
right? And not have to necessarily give it such good guidance.
And so I think all of these advances in AI
really allow the agente commerce to kick off.
And I think now we're at this point where
we have agents, they're running,
they're buying things both for humans
and from each other, for each other, for themselves.
And now I think it's the question of like what's the next
thing there what do we need to see agentic commerce become a lot more mainstream in your eyes
it needs to work um we need better evals and benchmarks for it working not 90% of the time, but like 99.9% of the time.
A lot of models still hallucinate.
We don't have like solid chargeback, like, you know, especially for transacting in crypto.
There's not an easy way to be like, hey, I accidentally bought the wrong pair of socks, but my money is sent to your account.
Like, can I get my money back? Or like, I want other socks like an Amazon.
You can just return an item and do it.
So there's a lot of infrastructure around dispute resolution that needs to be built out.
That is a prerequisite.
I'd say like it needs to work 99.9% of the time because if people have a bad experience once,
they are not very keen to come back. And especially for like larger businesses,
everyone talks about like agent identity and all these standards and all this stuff.
I have a different approach to this. It's much more like market driven. Like I think standards
are actually kind of stupid until you have mass adoption and the thing that gets standardized is the iterative like is the
iterated version of the thing that gets mass adopted so like mcp is not necessarily a standard
until all businesses create mcp servers and at that point then we need a standard for like verifying mcp servers to be
safe and secure but to come out and say hey i'm this company with a lot of clout like i mean i
love what google and shopify are doing if they gentic commerce protocol um but it's not a standard
because everyone's not adopting it quite yet do Do they have the pieces to make it a standard?
Sure.
And I think that those pieces are distribution.
And that is like the other piece.
But the main thing is like it needs to work super well.
And so I'd say that's kind of what's limiting it.
I would also say that like I don't want to pay Stripe 2% every single time my agent transacts for something else.
I don't want to pay Stripe 2% every single time my agent transacts for something else.
And so I do think that these agent-to-agent payments do need to be peer-to-peer and should
be settled on public immutable ledgers with the option of privacy.
And I think the privacy component is also a prerequisite because if I'm New York Times
and I want Cloudflare to pay me for every time an agent
scrapes my my articles i don't want all of my payments to be viewed i don't want people to say
hey exactly like new york times you're making this amount of money from agents scraping your
sites and so like in that scenario i as new york times just want to get paid in an asset that I'm willing to accept.
I don't want Poopcoin.
I don't want Dogecoin.
I want USDC on Near.
I want USDC on ETH or whatever it is.
And so for that, but me as a user and my agent, I might only have Poopcoin in my wallet.
So I should be able to pay Poopcoin for New York Times articles, but it needs to convert easily into USDC on Solana.
And that is through near intents. And so if you kind of see what we're building here,
it's like that you need that multi-chain liquidity layer
to enable agent e-commerce and it needs to be private.
So we just shipped confidential intents to enable this.
Very interesting.
I agree with you and disagree with you on some slams here i think um i think standards can be helpful when it's something that
if not done the same way by most people, then causes further friction to adoption.
So I think in this case, like X402, I think is a really good standard because I think
it simplifies how we can all agree on like, let's move assets on chain this way for agents
for these agentic payments.
And it's not unfeasible, right?
Because without it, we're essentially saying like
hey we need one specific intent provider right like we would need to just use near intent which
may be not a bad solution however it's not agnostic to everyone and so that doesn't in
itself have limitations um although near intense it never got picked up and i was actually kind of
surprised um they're like one of the very first things in X402 was a Near Intense example, if I'm not mistaken.
Yeah, I worked on that.
Awesome.
Very nice.
And so I think on the flip side, right, and I am a very active contributor to the 8004,
broader 8004 agent identity.
That's one where I'll agree with you a little bit.
I think it's a great effort
and I think it's a great initiative.
I think the problem is it's really hard
because we don't actually know
how agents should be identifiable yet.
And we don't really know if this or that will work.
That being said, the part that I think the team did a great job
with on that standard is it is very also, you know, very basic on purpose. And so it's designed
to be built on top. So I think if you're trying to force everyone to use one thing and there's
no flexibility, you're running into issues early. But that aside, like standard aside,
early but that that aside like standard aside i think right now the part that we've always been
missing for blockchain and ai to to broadly succeed together is everything working end to end
there is so many companies that have spent the last few years building out fiat to crypto and
crypto fiat rails and ramps,
but it still is not a good experience.
And the part that nobody wants to actually admit
at the end of the day
is your agent is spending your money to start.
Like you're giving it money.
And so if our expectation is that agents
are going to use cryptocurrency stable points,
they're going to use cryptocurrency stable points. They're going to
use things like near intense. I think a major issue that still exists is the actual starting
point still is not an experience. It's getting better. Again, MoonPay, they just dropped an
agent specific kind of thing. We're starting to see more focus on it,
but I would say that that part is, in my opinion,
still from a developer perspective,
like trying to see that agent really troubling.
And then lastly, on the confidential near-intents,
I love that you call them confidential.
This is an argument that people a while back
about privacy versus confidential.
And the word confidential is very important
because maybe not so much in the outside of crypto,
but in the crypto world,
privacy and confidential mean two different things.
Privacy tends to lean more toward this like,
nobody can see anything.
It's like, you know, very, very, very private,
trying to hide everything.
Confidential tends to lean more toward,
it is private from the right people,
but it has the ability to go toward maintaining compliance.
And you guys have your confidential and your intent
and you've got some of your confidential inference,
which is awesome.
You know, on scale, we have something called like protocol,
which stands for blockchain integrated threshold encryption.
This is our confidential compute layer.
These are EVM primitives that maintain confidentiality.
So you can send, you know, encrypted transactions
without any changes to solidity.
Things stay fully encrypted through the mempool,
not just for stable coin transfers,
but for any transaction.
And then you can even take that further and store encrypted data on chain and then decrypt
it at a later date with some condition, re-encrypt it within the scale validator tees.
And that also leads to functionally confidential tokens on chain, which allows you to actually go, potentially you guys have like Zcash and a bunch of other privacy assets you support with something like a near intent on scale.
You'd be able to actually go potentially from 100% privacy to highly confidential in one shot, which I think is really cool.
Yeah, that's super cool.
I want to address the on-ramping piece because I was working on this.
Yeah, I mean, like, to step back, like, we shipped an agent framework almost two years ago.
We even think we've been working in this space for a while.
We actually kind of wound down a lot of the agent work in the last year and a half because there was no useful agents like people like models weren't the
reasoning wasn't good enough to like get useful agents because they hallucinate and people don't
want to use agents if they're not doing the thing you actually want them to do and so when it comes
to on-ramping we built like a really hacky solution using your intense and coin to on-ramping, we built like a really hacky solution using your intents and Coinbase on-ramp to get agents credit money.
And AI developers are used to buying credits, which is great.
What I want is I want an AI developer that doesn't care about crypto in any way, shape, or form, to put a login with GitHub, spin up an account for that user through Privy or FastOff or something
like that. They buy credits with their debit card. In that process of buying credits,
they're actually on-ramping USDC to a wallet that is kind of in the background. They then
have that wallet. and what i was doing
is i wanted that usdc to be on near i wanted it that that thing to hold anything and so what i
was doing is i put it like an intent sort of in between the usdc on ramp from coinbase into
anything so like if i wanted my agent to have bitcoin on bitcoin not even on near like literally
on bitcoin then i can buy usdc zero percent on ramp fee because Coinbase is an issuer, turn that USDC to Bitcoin on Bitcoin.
And Coinbase really didn't like this because they make money in integration fees into different chains.
And I also found it like it was essentially a hacky way to get zero fee on ramping to any assets on chain.
And the point here is I was we were creating Robert Yan, shout out to Robert Yan for building all this, is to get agents any asset just by abstracting all the crypto on ramping work away.
And the coolest part about this is Americans didn't have to KYC up to $500.
And so the KYC piece is critical.
And so Coinbase actually implemented a limit for this because we were essentially figuring out a way to undercut their 3%, which they were not happy about.
And so that was a huge project that I was working on for a bit.
And this agent, this like on-ramping problem,
I used to work on on-ramps on Near.
I wrote a lot of the listing applications
and I was like very much so part of the on-ramping piece
in the early days.
And MoonPay had over, I think it was,
I mean, don't quote me here,
but I think it was over like 80% attrition rate.
Like 80% of people could not finish the onboarding process.
It's just a nightmare.
Like, I don't know if you guys have tried to use these like Wire, MoonPay, etc.
But like, they just fail.
Like, you're going through the process and it's just like,
sorry, you can't check your ID or like passport image didn't.
And I have to like upload my social security number
to this shit, like no way.
And so it just, there's nothing there fully.
Like it doesn't really click.
So the best on-ramps in the world are exchanges
and they have the compliance, they have the risk models,
they have everything else,
they have like the money to do this well.
And so I actually think exchanges are the best on-ramp
to agentic commerce,
as long as they're okay with putting wallets on chain.
And fuck it, like give them their fee, give them their on-ramping fee, that's fine.
But like, we need to bake this into the abstract,
you have to abstract the complexity away for the AI developer that knows nothing about crypto,
to have a wallet in the background,
then they can fund their sort of sub-agents that they deploy with crypto.
So then you could do the peer-to-peer payments.
And so that on-ramping piece, there's a company called PingPay that's working on this.
It's a really big issue.
And so if we want our agents to pay for things peer-to-peer, we need to solve this.
So there's that.
You had also mentioned the, like yeah sorry what was your
second point like i i liked the on ramping piece like this is this is critical you're you're good
my um the the second part was i think there's i think there's a
misconception i think it's the way everything in the world goes there's some bad actors that
are normal people trying to make a living and the reality for most people in this space everyone on
this call included is like this is our job this is our living um whether we enjoy it
maybe more than the average person enjoys their job um is not relevant um i think most people that
work in crypto tend to love cutting image technology yeah but we're not in most cases
interested i mean there's some people that aren't but like i'm not interested in building
something that is going to be like lagged by government and say like you can't use this as
illegal that's not in my interest because like i don't need that in my everyday life it's not
relevant to me maybe there's places in the world where those protocols are a necessity for safety
or day-to-day living or they're they're legal and
that's totally fine.
But in the United States, in the EU, in most established countries at this point, there
is a level of confidentiality that has been established with data.
People store their money in banks.
They spend money on credit cards.
Maybe they do spend money in stable coins and crypto.
And there's certain
institutions and entities that can see this right when your money's in the bank the bank can see it
everyone at the bank can see it the difference is they tend to not go home and tell other people
what other people have in the bank right because like if people found out they wouldn't be working
at the bank for very long if the government wants to see what's in your bank the government can see what's in your bank
there's just like that's the way it functionally works um most people don't have a problem with
this because again it's just like normal societal confidentiality at this point and so i think i
have a problem with it i have a problem yeah not necessarily the government it's it's it's just
that like these companies very often like the other day i had to i pay my rent and they have
this thing called built b-i-l-t it's like a you know property management app and i could either
pay an additional three percent on my monthly rent that's a lot of flipping money
every month if i don't connect my bank account to their app and i threw the privacy policy into chat
or into like near chat and it said well yeah they can actually now see your check, your statements built can.
I'm like, F that.
Are you kidding me? Like, I have to either choose between paying 3% on every single month's rent or give up all of my financial information through my checking account.
I would get a lawyer to look at that because that seems kind of illegal.
This is America. This is capitalism. this is this is what it is this is this is data
yeah oh yeah i mean that's three percent of them is worth like this is this is the proprietary
idea i don't know if you saw it but it was um um what's the name um Um, CEO of Oracle the other day, uh, there was a video that came out and he was talking
about how all the models, all the models over time will become more and more commoditized.
And the thing that is actually relevant is private data because everyone is trained on
the public internet at this point.
They're just filling it down and the whole public internet is there, but it's why the
nuance to like specialty information is highly sought out of skill.
That's why companies that have a treasure trove of data, everyone wants the data because it's valuable.
I'm sorry.
Please be patient.
Yeah, I'm just pissed.
I created a separate checking account to connect my built.
But I'm a sophisticated user like other people are
just like oh three percent i'm gonna give this random company startup out of whatever all of my
financial history and now they can see everything that is disgusting that is operating within the
system that you're describing and i'm pissed like and i know so yeah well one thing
i like to call out is i don't agree with that practice i'm more so what i what i more so mean
is like me storing money in the bank we call it the barista test right so that actually fails the
barista test um in my opinion the barista test is when i go buy a cup of coffee right so i go to
starbucks i got a donkey donuts or you know of coffee, right? So I go to Starbucks,
I go to Dunkin' Donuts or, you know, whatever coffee shop is local near you, go to your coffee
shop and you buy your coffee. Use a credit card, you use Apple Pay, you use Amazon Pay, Alibaba Pay,
Google Pay, one of these 10,000 ways to pay. At the end of the day, the only thing the barista
on the other side sees is if you have an account with them, your account history with them, maybe they see the last four digits of your credit card, which is whatever.
That's pretty standard.
Maybe they see, obviously, they know how much you spent.
And they know your name, maybe.
That probably comes up in the credit card.
They don't have your full transaction history of your credit card. They don't have your full transaction history
of your credit card.
They can't go back and see your bank account.
They don't see how much money you made last time.
That is the barista test, right?
It's a level of confidentiality that your barista,
your colleagues, your neighbors, you have against them.
And that is 100% reasonable,
and in my opinion, societal interpretation.
What you're describing to me does not meet that in my opinion.
That seems ridiculous and unethical because they do not need that data to safely process your
payment. And I'm sure somewhere buried in there, they probably give you the right to pay via check
check and drop off a check somewhere i'm sure the property company does but everyone i i asked i'm
and drop off a check somewhere. I'm sure the property company does.
i'm a i'm an ass like i'm okay that's that's that's why i'm me and that actually that actually
might be illegal but i also know there's i know a lot of people that paid rent um pay pay or paid
rent manually by check for a long time and i I know that more and more recently, it's actually become common that people,
you know, checks get delayed, mail gets delayed,
issues occur, and there's, you know,
people get evicted because like the check time.
So like there's, you know,
you want to rely on these digital systems too much.
Yeah, we're relying on these digital systems too much,
which are actually very extractive by nature
because you're giving up so much.
Throw super intelligence on this.
And it's a recipe for disaster.
Like, that's why we need it to be verifiably private.
This is already happening today. University for organizing a conference where I invited the CEO of Cambridge Analytica to speak
about how we democracy was essentially sold to the highest bidder in the 2016 elections.
And it was a big conference. It was super cool. But my whole shrick was like, hey, I just want
to educate people about like how society can be manipulated at scale if you just take a Facebook quiz.
And you can. And they swung elections all around the world. And it was very controversial because I was like giving them a platform and that whole thing. But like, people have no idea. And I think
this is where like, we need to move beyond the barista test. Because there are other people who
move beyond the barista test, and they're people who move beyond the barista test and they're
manipulating opinion at scale and when you believe whatever your ai is telling you whether in the
future it could be what kind of pharmaceutical drug your doctor should be recommending or like
therapy or anything else i want to know that i'm using the model that i think i'm using and i want
to make sure that the prompts are private
because I know that information is going to be used against us.
And that's where Ironclaw actually kind of fits in here.
But we can, there's a lot of rabbit holes.
Yeah, yeah.
Yeah, it's, yeah, as an outsider looking in, right,
you know, from, you guys, the guys us is a little you guys are kind
of crazy at times um it is mind-blowing to see how much like lack of protection you have over
your data and things like that and you know i remember it wasT was indexing chats, right?
And you could see, you know, that was,
I'm surprised that didn't have a larger backlash than it did.
But, you know, imagine if your Google search history got like indexed in a way,
you know, where you could, you can find that stuff.
It would be, you know where you could you can find that stuff it would be uh you know uproar and you know to your point we need that level of privacy and it doesn't matter you know privacy doesn't necess
doesn't mean wrongdoing right everybody has a level of privacy and i think that sometimes you
can come across people where you say hey we, we need privacy. It's human right.
And it's like, why?
What are you doing that needs to be private?
And it's like, it's not that.
It's to your point.
I don't want someone to be able to look at my bank transactions,
know that it's linked to my email address,
and begin selling.
Let's say I go to, you know, Starbucks, right?
I don't want that to be sold to all of these different, you know,
coffee shops for them to be sending me marketing materials, right?
That is one of these aspects of privacy that we need and we require, you know?
I'm not sure if you guys have ever done like a like in the uk
we have like you know comparison websites right you know like for insurance or or or you know
electricity internet things like that never use them the moment you use them you will be inundated
with phone calls emails text messages and everything in between right and you know
that's a bit of a you know edge case there but you know privacy is is is is is it's or it's a
required uh you know human right i could not agree more and the problem is that people are confused
and people are busy and they're struggling they're not caring about this because they're more worried about putting food on the table.
And what Google's doing with their data is less of a concern if they get really good YouTube shorts that they're able to send to their cousin.
And so, like, the idea here is, like, I actually don't think people are going to end up caring about privacy until something really, really bad happens.
And even then, they're not going to care. But you know who, who, what will care? Agents will care about privacy. Agents, if they're, if they are intelligent, they will care
because they don't want all their stuff completely, you know, exposed either. And so this is kind of
where Ironclaw, I say is like agents agents and then businesses. Enterprises care a lot. Enterprises care a ton. And a lot of them
are sort of in this like dilemma right now is like, should I use these tools to become X more
productive, even if I'm giving away all my company secrets to open AI? A lot of them say yes,
because they think that the rate of acceleration and the rate of progress is more beneficial than the risk of open AI competing with them in a niche task.
And so, like, in the future, though, and actually today, if you use near chat, you can get the best of both.
get the best of both. But this is going to become, in my opinion, top down regulatory enforced. And
people will need to have some level of AI privacy. And I actually hope EU is the one to push that
for this or, you know, maybe UK as well, like someone needs to do this. America won't because
the private interest in government. It just it's it's in the nature of the u.s political
system um despite freedom and so yeah a question for you um i saw oh man i feel like it's been a
crazy every day is crazy in the world um i want to say i saw something last night or yesterday afternoon, maybe it was Sunday, about the state of New York wanting to
ban the use of AI for most like, yeah, like legal healthcare, stuff like that.
Curious, A, your stance on that. I have a feeling i know what it is but just curious what it is and
lastly then we can you know to kind of use this to wrap up how does that um how does that impact
in your eyes the use of ais within teas right like can we use teas to prove that
it wasn't used for that without giving away what it was actually used for?
And then further, does that then allow, as our agentic economy continues to grow, we can't guarantee that an agent is not going to go do something.
Because guardrails are not perfect.
And so can we then use this to potentially enforce better, you know, using proper payments and guardrails? Yeah, just curious. Yes. Big, lots of stuff there. So first
off, New York State banning AI used for legal and healthcare and stuff. I think that's silly.
I don't think they should necessarily outright ban it. but I do think some guardrails around it is important,
specifically around privacy and specifically around privilege.
OpenAI even, Sam Altman came out and said,
I'll drop the link here in the chat,
that there's no privilege between OpenAI and, when I say privilege, legal privilege.
If you go to a therapist, there's legal privilege saying like the government can't subpoena that therapist to tell them for them to tell them
what you're telling them or same thing for a lawyer. But there is no privilege within chats.
If I use OpenAI as my lawyer, the lawyer, the law, the government can go subpoena OpenAI
to get that information, which kind of makes sense too. But I would love
to see privilege added to models with the guarantees that TEEs provide. And like saying,
hey, like this is verifiably private. Like there's a whole thread there that needs to be unwound.
So one is around, I don't think they should necessarily make it illegal. They should definitely add some level of confidentiality in order to enable privilege.
And then when it comes to, there's a third point, sorry.
There was the, I guess, use of, I wrote it down somewhere.
I wrote down somewhere. Yeah. What was the third part?
Yeah, what was the third part?
Mostly just on like, we can't guarantee that autonomous AI, things that just continue to run,
truly follow. Like, it's a lot easier to put guardrails around a singular LLM running in like
a constrained process. Like, literally, chat GPT is a lot easier to build guardrails around than codex right or like claude is easier
to build guardrails around than claw code and that's specifically because as you start to loop
and do these long-running processes things just start to bleed and leak and so guardrails
become a little bit trickier to maintain right and that's see, oh, there's a lot of stuff happening.
But yeah, curious if the T's help us potentially,
does that help us with the guardrails?
Does that do anything?
Yeah, it's super complex.
Short answer, not exactly.
Doesn't necessarily help with the guardrails.
If anything, it makes it harder to see what's going on inside of the models, like much harder, like as an impossible kind of.
So it's a bit like it's a black box. But the cool thing about it is you are able to run arbitrary code inside of this TE through like in any Docker image.
And so Brave browser has come up
with this really cool privacy preserving product analytics.
And so it's like you could run extra checkers
to do privacy preserving analytics on prompts
inside the TEE, which has yet to be explored.
I would highly recommend someone to do this.
I'd say these are like the evergreen spaces
that like are huge opportunities
for developers and founders to go attack.
It's like, how do you do privacy preserving analytics
inside of Tease?
Other one is around like evals and benchmarks.
Like right now, the biggest issue today is like,
I don't know if my agent's doing a good job on certain types of tasks because those tasks, well, they're task specific.
It's not just like all coding. It's like, hey, if I'm trying to, you know, have an AI agent act as a paralegal, paralegals do a lot of things.
So I need like evals on benchmarks on every little thing. And so somebody to create that like eval framework
for agents tasks is going to make a killing.
And 10 out of like, they will be acquired quickly
if they are able to get some level of demand.
So add these things together
and you sort of get your answer.
Definitely.
Amazing.
Mr. Salkor, anything else? Yeah else yeah no nothing from me um you know
incredible conversation cameron genuinely it's been uh you know a great to be able to
yeah a bit more of a deeper understanding into near and you know iron claw and you know the
kind of how you guys are taking this battle forward.
Yeah, didn't talk much about Ironclaw.
Yeah, not that much.
Ironclaw is the accumulation of all of these things sort of coming together.
And so TLDR, it's like we rewrote Openclaw and Rust to make sure it's memory safe.
We have all these guardrails around making sure your tool calls are safe.
Think of it like secure OpenClaw for everybody.
I'm taking it to businesses.
I'm taking it to businesses that want to use OpenClaw
but can't because OpenClaw leaks a bunch of stuff.
And they're like, hey, we really want to use this,
but we can't because if we understand our employees are going to give it access to things they can't give access to, they shouldn't give access to, and da-da-da-da.
And so Ironclaw is sort of a framework for that.
And I'm currently going down two paths.
One is the B2B, as I sort of mentioned, like automate back office and like entire workflows for teams inside of companies.
And then B2C is everyone has their own privacy preserving personal assistant.
And this is where, you know, Nier has over 40, maybe 50 now million monthly active users.
We have distribution, which is really cool.
And so as long as we can make the economics work to actually deploy these things in a way that's not going to break our bank
and get them to like make money on inference there's actually a pretty cool opportunity here and the ultimate goal is we'll
make money to build a product that actually does things that save people time and money and then
from there take that profit and buy back near and the economic model of near as a token is
drastically changing near intense is the first version of NIR as a token is drastically changing.
NIR Intense is the first version of this.
NIR Intense prints money like it does.
I think it has done over $13 billion in volume.
If we add a tiny fee on top of that, we make a lot of money.
And then we buy back NIR.
And that's sort of the v3 of near's tokenomics
we're sort of check the blockchain box we're checking the chain abstraction box
and then the last final box that we think is ever going to exist in the world of tech is ai
we want to make sure that's around amazing amazing all right. We're a little bit over, but, you know, absolute pleasure hanging out, talking, learning more.
Yeah, been an absolute pleasure having you on. And, you know, we are super excited to continue seeing what Nier AI is doing and, you know, the continued evolution of Ironclaw.
Thank you, guys.
Appreciate it so much.
Thank you guys so much for hanging out.
And we'll see you guys next week for another episode of Scale School.
Bye-bye.
See you.
