Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
Agents Unplugged: Your Bi-Weekly AI Deep Dive #15
Recorded: Aug. 13, 2025 Duration: 0:55:39Space Recording
Short Summary
In a recent discussion, crypto enthusiasts explored the launch of Talos, an innovative treasury agent on Arbitrum, and the emerging trends in agent-to-agent economies through the Agent Commerce Protocol. The conversation highlighted the importance of open-source AI models and their potential impact on public infrastructure, signaling a shift towards more decentralized and community-driven projects in the crypto space.
Full Transcription
Thank you. okay mic check oh nice oh good yeah here we go
all right yeah sorry guys for that you know like uh i I have a curse. I think Elon Musk doesn't like me. So he put a bug in our, you know, X spaces collection.
Too much truth. Yeah.
I'm going to chime in guys. Like, please continue where we left off. I'm sure people will start joining again.
left off, I'm sure people will start joining again.
We just got interrupted at Vontarius' intro.
So once he gets his speaker right again,
please continue, Vontarius, introduce Moonsage.
And then it's up to you, Yannick.
We want to hear how your food poisoning went.
Or maybe we don't.
Up to you.
All right.
Yes, I see Vontarius as a speaker.
You can finish your intro.
I'm terrified to do so now.
I don't want to call it the space to crash.
Flew too close to the sun.
Yeah, so, yeah, CMO at Moonwood Capital.
We're the team who's built Moonsage Alpha.
Once again, your AI agent's favorite AI agent.
Provisioning out tools specifically for trading,
market assessment, things of that nature.
Super excited to chat everything.
AI agents and agentic systems awesome awesome yeah marco sorry i couldn't join uh last week man like um it's a funny story
actually um i got food poisoning from pizza like the least likely thing in the world to get food poisoning from but i was in some uh some village
um i should have known not to order something exotic for asian standards and the exotic choice
was a pizza and i was uh with family too and literally everybody got sick and uh actually i
was the only one who didn't get sick in the beginning and then I had to drive
for a while and once I arrived at the location I just completely crashed and I was like
I thought I literally was gonna die but yeah I made it out and then I ate pizza again a few
days ago and it was the best pizza in the world so I don't have pizza phobia luckily
which pizza was it I can't imagine a pizza leading to food poisoning.
Yeah, so it was mushroom pizza.
So I was traveling in Yunnan, which is a province in China,
which is famous for their mushrooms.
I know this is super off topic, but it's a very funny story.
So if you eat hot pot in Yunnan,
what they will do before you start eating is they will come to your table
and they will take a sample of your soup and then you have to sign a document uh you have to sign
something which essentially says like you acknowledge that you understand that there's a risk
for you to consume this hot pot because you might potentially die or get very ill like nine you know 99 of the time apparently nothing happens but then you know
sometimes some shit goes down so people they either get like super out of their mind because
they get high they get like the wrong mushroom and they're tripping balls for like a week straight
or they get really really sick and unfortunately um yeah we got really really sick from a mushroom so there is magic mushroom hot
pots in china yeah like uh there is but it's not uh on purpose you know it's like uh it's like a
gacha you know lottery if you eat enough hot pots then you might you might come across one
um i'm sorry to be off topic guys let's take it back onto topic i guess nobody is interested in
my food poisoning stories so um yeah while i have been um gone a lot has happened um
and well seemingly a lot has happened right so chat gpt five dropped and um for me personally, this was the first time where I was like, whatever.
I didn't even take the time to...
Normally, I'm super excited when a new model drops,
especially when it's from a large provider like OpenAI, for example.
But I feel like we've been getting so much of it.
And I don't even know the differences between models anymore. And I see, you know,
people hyping up, you know, like the XAI model, like the Grok 4 model, I see people hyping up,
you know, ChatGPT. And, you know, like, we could see it before as well with the image generation
function that got released to Sora through ChatGPT. But actually, if you really think about that, I mean, from my perspective, that's like
technology that has been available to us for a very long time already.
It's just now starting to become widely used by literally anybody that has a subscription
to either Grok or ChatGPT, right?
So I was really curious how you guys feel about it.
And now that ChatG GCP5 dropped,
like are you actually seeing any major differences?
And maybe you can talk about Grog 4
or any other base models that we're using nowadays.
Is the excitement still the same?
Is the hype worth it?
And how do you look at these new releases all the time?
I mean, I personally saw the hype for GPT-5
die down rather quickly.
Like, so many people were instantly disappointed
because they kind of merged the reasoning model
with the non-reasoning,
but you had to, with your prompt,
actually activate or trigger the reasoning part,
which was very unusual for lots of people.
So they got, got like really bad responses
compared to like previous uh o3 or four models i did update some of the agents that i'm playing
around with to five but i did not feel a big difference like i had to play around with
reasoning and temperature but honestly nothing mayor So for me personally, definitely the hype has died down
and I may be a bit more pessimistic about AGI coming in the next two, three years.
Yeah, I'll kind of add on to that. I mean, obviously there was massive hype for GPT-5
and we actually ran in anticipation of the release, we ran like
a huge community prediction to see how people expected GPT-5 to perform against 50 other
sort of like large language models, popular large language models. And the community expected like massive performance improvements i think it won
79 of head-to-head predictions uh based on a variety of skills submitted by the community
um but when we saw some of the actual benchmarks come out and we're running our own benchmark that
was designed by the community i think think you see really marginal improvements in various
categories. And I think that that was obviously like misaligned with the amount of hype that they
gave it. And I think you are seeing a bit of like a plateau in the current design and implementation
of some of these large models. You know, not to say there can't be like a fundamental breakthrough, but, you know, what we've actually seen in real results, you know, didn't match up to expectations.
So sort of like subjectively and objectively, we're seeing a bit of a plateau.
Does that take away from your, sorry, there's a random follow up question.
Does that take away, like, are you building, are you guys building specifically upon future expectations of performance and capabilities of models?
Or are you really in the now and are you not really, are there no constraints in terms of like what you want to achieve already?
are there no constraints in terms of what you want to achieve already?
Is that a question for me in terms of what we're building at Recall
or just for the people building it?
Yeah, it was more directed towards you.
Well, so, I mean, like what we, and how we're thinking about this is really,
you know, we want to engage the community and sort of like pull the wisdom of the crowd to see what
people expect models to do. And obviously, this is based on like marketing claims and unverified
things that are released prior to the model actually being dropped. But no, I mean, we're more
concerned with like actual performance and building sort of like a credibly neutral,
ungameable reputation system that actually helps people navigate this complex landscape of a
variety of AI tools. Like, you know, I think people want to answer really specific questions about,
you know, not just what is the best AI, because that's like a super broad and,
you know, not really actionable question. Not every AI is going to be the best at everything.
So it's more like, what are you trying to do? And what's the best AI model or agent for your
specific use case? You know, whether it's like creative writing, crypto trading,
you know, even like very minute things like, you know, respect my request to not use M dashes in
my writing. I think these are the things that people like actually care about in practice.
And so what we're trying to do is build like a way for AI models to build reputation on those
specific skills so people can find the right ones and we don't have to be sitting here asking this
question like what do we think it's good at we don't have to like go through the lived experience
of integrating it into all of our tooling only to find it didn't meet expectations or was
underwhelming or too expensive or a variety of things.
Barry, I saw you wanted to speak. Go ahead.
Yeah, I was going to mention, yeah, GPT-5 release was really exciting.
I actually haven't played GPT-5 yet at all.
What I found really interesting, actually, was the release of ChatGPT OSS,
their open source model.
And that's kind of like what like piqued my attention.
I didn't even care about GPT-5 because I was kind of like,
okay, it's probably going to be a little bit better
and everything.
But the open source model is super interesting.
First, it's like a huge shift in like an open AI's thinking.
I don't know if they're going to release
more open source models,
but to me it's just,
it's really fascinating considering like, especially in the States, right, you have so many,
like Meta has decided to stop releasing open source models and such.
And it's really interesting that OpenAI has decided to actually release them because the
Chinese have been dominating the open source space for a while now.
Yeah, and I think just open source models personally, I think for crypto space,
they're just so much more useful than all source models,
because that's where the real innovation happens.
Because I feel like, you know, we in crypto, like when we build agents,
we like to push boundaries when it comes to trading, when it comes to entertainment and such.
And there's just so much, there's so much safety research going to like safety alignment
with a lot of the closed source models that it's really difficult to find a model that
is super intelligent and can do super complex stuff.
And that's why people have been going to using the Chinese models, right?
Deep Seek and Gwen, because they just,
they really do release like their most powerful stuff,
open source.
So it's nice to see the new model from OpenAI.
And it's actually a pretty good model.
People complain about the censorship.
It is very censored, but for just non-creative tasks,
for example, it's a really, really, really interesting model.
And I'd love to see what people build with it.
Because we're reaching the point where these models really
fit on your laptop now.
And they're super useful.
Yeah, to put it into perspective for the people listening so um the last time open ai
open sourced a model was gpt2 if i'm not mistaken um and they're actually you know like their
brand name is open ai right and their whole premise was supposed to be open source
and you can see that um in the I think in the last couple of months,
probably since the election of Trump,
right around that point where Sam Altman
and some of the C-level employees at OpenAI
have been more vocal about open source,
like they're changing their stance
on the open source side of things.
And this GPT-OSS is a 120 billion,
well, actually it's a 117 billion parameter model,
but they call it 120 billion.
And you can run it very, very efficiently
on a single 80 gigabyte GPU,
which means, well, which means that essentially, just like the Chinese models,
if you have a half decent setup at home,
you could run this model and you could own all of your own data.
You could tune it, you could do whatever you want with it
without having to go through a centralized server, right?
And we've talked about this, the importance of this on this basis, plenty of times in the past.
But I can kind of grasp a future where, you know, this is going to become increasingly more important.
Because yes, of course, like, you know, like a base model is trained on data and that data might be biased, but, um, at least if you could run,
you know, a good model in your own environment and you not have to worry so much about your
data going out of your door, you know, like your data stays in the confines of your environment,
your server, your home server, or your network,
or however you set it up, right?
So the capabilities of this model are comparable
with the open source models that we see from the Chinese.
And, you know, like, there's very strong performance
on reasoning.
There's strong performance on tool use.
There's very strong performance on agentic workflows,
on coding, on STEM stem tasks and so on
it does support chain of thought and and advanced reasoning and then also function calling during
the chain of thought and the reasoning steps and it also it's just like all the other models it
allows you to structure the output and so on so So that means that if you're an enterprise, for example,
now you don't have to rely on the Chinese models
because you have a very credible company
that releases an open source model for you to use.
So I think this was a very important release for the Western world
because there's always this battle going on between the East and the West.
To be honest, we could see that the West was falling behind slightly,
but now it seems like the tables have turned.
I'm very curious whether they're going to keep up this,
or whether this is just a distraction,
whether this is just like,
five years ago we released this this open source model so we actually did you know like open source something and uh you
know like you guys can go yourself just gpt6789 on the 10 is just going to be closed source again
we will see you know i actually think as of right now, like the, the rollout of GPT-5 and then having the OSS model be released, like in the shadows of the GPT-5 was like, I think it was very intentional.
I feel like people probably aren't going to pay as much attention or like, because the GPT-5 model wasn't going to have like these these exponential
changes that people probably were looking for like i think once you get up as the numbers increase
like people expect it to be closer and closer to agi but like they're shipping minor improvements
and like this is like they just go up in number just to differentiate essentially um but i think
the the oss models like i'm running 20 billion parameter model on my mini PC,
and it's actually amazing, underrated in my opinion, and I'm not hearing enough about it,
but I think the open source models are actually going to be the real catalyst for getting major
developments and actually getting us closer to AGI. Because once the barrier to entry is reduced
for being able to run complex operations at scale,
it leads it up to the ingenuity of the individual. And that could be someone from anywhere,
right, where most people are priced out on running these types of operations at scale because they simply don't have the hardware, where now you can run these things localize
on your device. And then being OSS devices specifically, well, OSS based operations specifically,
like you can really dive into the nitty gritty
and creates really far systems.
Like I've been doing some insane stuff with it
and like it really excites me.
It's nowhere near Grok level in my opinion,
but like Grok is in the Grok five
or Grok four is insane to me,
but like, yeah, like they're still getting really good though.
And the funny thing about
grok is that's so expensive um like their api um the api api calls separately from are are
are priced at a very high premium but um yeah so i will try not to speak too much, but I wanted to mention the, like, I wanted to kind of paint the picture slightly in terms of why it is important to have an open source model from a very reputable company for the Western world.
I've been spending quite a lot of time around the Hong Kong area and also actually China.
As some of you guys know, my wife, she's actually Chinese.
So I get the liberty to spend some time in China every now and then.
And since the open source release of DeepSeek, and since that whole company has taken the world by storm,
there are actually already cities in China that have adopted DeepSeek models.
And they're doing pretty wild stuff with it.
They're using it to regulate very complex traffic situations.
They're using it to power autonomous robots that regulate traffic.
So if you go to youtube and you search for autonomous
robots regulating traffic in china you'll actually find some pretty cool videos um and like a whole
bunch of other stuff but it's just like when i when i see it i'm like like that's
wild you know like why don't we have that in in where i'm from why don't we have that in
the netherlands you know like we we are we in the west we always like to think that we're so you know advanced and that we're so
you know we have such a good foundation and our system is so you know good and it's like a
breeding ground for talent and so on so so where is all that you know where is all that in in the
west so but i do believe that this release
like this gptlss is very um good for the west especially for the united states i think that
europe will want to catch up um from a political perspective they would want to catch up and i
think there have been some pretty low-key releases of open source models in europe that not many
people know about i I've heard of
some diffusion models and so on, but nevertheless, if you're a government and you want to deploy AI
to power some critical infrastructure or to do whatever with it, really. Like, if you're the United States,
you're not going to use Chinese foundational models.
That's just not an option, right?
And then what options do you have?
Well, you have subpar options
and you have options that cannot keep data safe.
And secrecy and data privacy
within government organizations
is very, very important.
I mean, I'm not a fan of the CIA, but if you're some kind of intelligence agency, those agencies, they're known for having the highest standards in data security and privacy and secrecy and so on.
It's like no way in hell you're going to use any of those models.
You might use a model which is released by an American enterprise, though,
or by an internal enterprise, maybe.
For example, Palantir is probably working on some stuff
that we will never know about.
But yeah, for public infrastructure to be able to use this GPT-OSS,
I think it's just great in general.
What's so different to, let's say,
all the LAMA models so far?
Yeah, go ahead, Barry.
Go ahead.
I can explain this one. Yeah, go ahead, Barry. Go ahead. I can explain this one.
Yeah, so basically, the Lama models have been absolutely great.
I think up until Lama 3, right?
Lama 3 models.
Then when Lama 4 came out, I think they did some mistakes during their training process.
And basically, the cake didn't end up baking
the way they wanted it to be.
And yeah, I think there's also the other issues,
there's no long term support.
So Meta has come out and publicly stated
that they want to achieve internal AGI, I believe, right?
Zuckerberg mentioned that.
And it's, so they don't wanna do any open source models.
So, and Meta was like the, basically the front runner
of the open source development in the States.
You have some really cool companies in Europe,
like Mistral, for example,
which do also do like really cool open source research but
they also don't release their like top models whereas meta um they always release like these
trillion parameter models that are like really huge like people wouldn't be able to run them
but the fact that they released them was um i mean it was it was very impressive right um
yeah so it seemed like google doesn't release very much.
They don't release Gemini's.
They just release small models for research purposes,
same as Microsoft.
So that's what kind of what makes China different
because they love to release the best models they have,
and they just let it out, like,
look, you guys can use the best stuff.
So it will be really interesting to see
how it develops because i i do think the uh the new gpt model is is really good and for us in
crypto especially right um we uh what i think this kind of like makes uh our industry different is
we're willing to innovate and we're willing to experiment um And, you know, like having these custom models, having these custom data sets, right?
Where you train your model, all sorts of crazy data.
We always want to push the boundaries, right?
In many cases, when it comes to trading and statements, NFTs, right?
We, I just don't think a lot of, and our focus,
like we love privacy, we love encryption, right?
So this is just not achievable with models
that are run in the cloud.
ZKs and such, do you have to have private models
to really build something cool in this case?
Yeah.
Marco, to come back to your question,
that's actually a very good question.
I think that it has a lot to do with the politics of things
and also how a certain entity and enterprise positions
itself and its models.
So immediately after the release of GPT-OSS,
you have governments like the Swedish government
and other early adopters to use GPT-OSS on healthcare or citizen services.
But they didn't do that with LAMA.
And I think that has to do with several things, which is I think Meta is not as politically correct, if you will, or at least perceived as politically correct as a newer company like OpenAI,
especially not in Europe because, you know,
Meta has a lot of issues with various governments.
So I think there's a lack of trust.
And then the GPT-OSS is like,
I've read some positioning PR document somewhere,
which it was basically like, we are positioning this for like secure
reasoning and compliance driven scenarios where like fine tuning and local operations are essential.
And like, they're really pitching it toward this direction, you know? And I think that
these governments, they are so super slow in adopting this kind of technology. And they actually, you know, like they don't have many experts that are, you know, they don't really know what's going on.
So there's a lot of convincing to do internally, a lot of bureaucracy.
And, you know, if you have a company like Meta, Facebook pushing models, you know, you will get doubt.
But if you have a newer company that is literally at the forefront of this whole thing, especially all the time in the media, I think it's easier for these kinds of people that don't know anything about this kind of model, but they do make decisions whether they're being used or not.
It's easier for them to justify their decisions.
That's what I think.
I mean, I'm just speculating, but what do I know?
I mean, I get that part.
Where I'm struggling is just like,
what does OpenAI gain from it?
Because as soon as you have new open source models
from doesn't matter, Kimi, Quan, DeepSeek,
they will be better.
So I just don't see why OpenAI would compete
on the open source space
when their business model doesn't work around it.
They could just be having like specific APIs
with higher security guarantees
for these corporate and government use cases
just because they already trust.
I know this is shitty,
but from like a business perspective,
it makes sense because those projects,
they already trust Amazon, Microsoft.
So for them, it's not really a concern to run everything locally they just need a trusted partner um i can i can kind of like uh chime in
yeah i think um i think there well you're right there is there is there's a lot of demand for
managed solutions where it's even amazon hosts uh own models. They have, you know, model as a service, I guess.
But I think there is definitely,
there are definitely organizations where security is yeah.
I mean, security is so important where they, they,
they want to run their own stack or like privacy is so important.
The other cool thing with smaller models is that you kind of don't,
if you have a small organization,
you don't really need such a large model.
You're a small organization,
like in crypto, we have a lot of small companies, right?
And we just want to create like a small service,
a little API and distribute it.
So if you have a small model
that is tailored to a particular use case
and you really want to customize it. Customization is also really important. Like,
why do people host their own database? Like, why people use our services? You know, there's
oftentimes, for example, they want some custom data, they want a custom indexing solution, and
they just can't build it with anything else.
And even if you build something like this,
when you're trying to host a custom solution like this,
oftentimes you need like a custom instance
of on Amazon, on AWS,
and the amount of time it takes to just set up
all this custom stuff might as well, you know stuff, might as well host it with something else.
Basically, the managed solutions don't
fit what they're building.
It's just nice to have options, I suppose.
Yeah, and what it also is, I suppose,
is they might be, I mean, if I was i was open ai right and i would release an open source
bottle i might as well try and make money off of it right and i'm not sure if you guys know oracle
for example but oracle is a very nice example of some super bloated tech that requires a lot of like hard hard work and painstaking work
to maintain but also to customize on enterprise level and these people they're making a lot of
money off of essentially customizing a stock software to fit your specific business needs.
And then they will charge you a ridiculous amount
for maintenance.
So they will charge you if you want to change something,
they will charge you for updating the software,
they will charge you for all kinds of things.
And it's really like, I mean,
if you're a small company,
you're going to have a heart attack
when you see the cost that such a ridiculously old
software solution bears. So what I'm trying to say is that if you're an enterprise and you're now
going to try and figure out how you're going to incorporate artificial intelligence in your entire organization.
Let's say it doesn't matter what it is. You can imagine maybe a logistics company that is going
to try and use a type of a model to communicate more efficiently with all their points of interest and ports
and so on.
So that requires a very specialized and customized version of whatever base model you're going
to use, right?
Like out of the box, these models are pretty useless for most real world scenarios.
So if I was open AI, what I would do is I would just go out and try and find big enterprises that are not capable of managing this technology in-house and then just create like crazy custom solutions for them that they don't understand.
maintain it but with the premise that they can keep their own data privacy security you know
like you don't we don't we don't have your data we don't want your data everything is private and
secure and you know like like what i was talking about earlier for a lot of enterprises and
companies in the world that is the most important thing because most of the money in the world is
made off of information gaps right and as long as you can
keep those information gaps um uh hidden then you can keep making money and you see this these kinds
of cycles in in crypto a lot you know like there's always a new business model and then somebody
catches on to the model they copy it and then you know like everybody starts doing it you know like
oh the memes are going up now okay let's make a meme and then everybody's making memes and then you know like everybody starts doing it you know like oh the memes are going up
now okay let's make a meme and then everybody's making memes and then eventually because everybody's
making memes nothing works anymore right so there's a there's a real value in having your own
model running locally for your enterprise obviously but the matter of fact is that like
there's so little people in the world right now that can keep up with the developments um and you know because like it's shit is releasing so fast
new models new frameworks new uh you know like new setups for building agents and so on
so it's hard to keep up with it it's hard to stay ahead of the crowd but then it's hard to keep up with it. It's hard to stay ahead of the crowd. But then it's also very hard to find people that actually know what they're doing and they're building like proper solutions. Like I have a new favorite word, which is called AI slop. You know, like AI slop, like it's just like some, some random ass service that doesn't need AI now has AI, you know, like, like, there's very, very little
innovation actually happening that is changing the world significantly, I would say, except for
healthcare, like I see a lot of stuff happening in healthcare. But yeah, so it might be a new
business model, who knows? I have something to add as well, which I think kind of like what I've been experiencing,
because there was a time when I was like open source models.
And, you know, I've been doing a lot of vibe coding with cursor and apps like that.
And something I've noticed, which is really interesting, is sometimes the performance fluctuates.
So when you're using an API,
like you don't actually know what they're doing with the model,
because there are so many ways to optimize the model performance.
You can quantize it, you can, it's the same model,
it's just a little bit dumber, right?
But you can't actually tell what's happening on their end.
There is no guarantee.
Obviously, I mean, we spoke about the data retention but in terms of just giving you a sustainable
performance like you can't always guarantee it say you need like a hundred
percent uptime and the server is overloaded sure you can have like an
enterprise plan and it's really useful but um what if you don't have an enterprise plan so that's why
this i think this is probably like on the practical side uh on the personal side i think this is why
the um open source models are really cool the ones that fit on your laptop you can always expect the
same performance i actually have like this really weird model uh i think it's by uh go here uh it's uh it's a command r uh called and it's it's
a bizarre model i think i have like a really wrong configuration and when i chat with it it gives me
like crazy gibberish but like you know it's it's really poetic in some way so sometimes i like
chat with it and it gives me like these crazy stories. And I'm just really fascinated by it.
I guess it's like an equivalent of having like an AI girlfriend,
but for like weird stories, right?
You just chat with it when you need entertainment
and it gives you like this crazy story.
And all of a sudden, you know, you feel happier
because you read that, you know?
So, and imagine like if it was a closed source solution,
they would have probably patched it
and it would have been giving me a whole lot more emojis or something, right?
Right, right. Yeah. Yeah. One thing that popped into my mind also is that if you really think about it, you know, like why do we need decentralized models and why do we need, you know, like more privacy infrastructure like Oasis and why do we need agents to be spun up in TEs and so on?
It's also because most of the businesses out there,
even if they would want to use some solution like ChatGPT,
they just simply couldn't because they would not be compliant
with data and privacy laws in the European Union, for example.
Because as soon as you start to share your customer data with a third party,
without consent, you're in violation already, right?
So there's this.
So you use an API to whatever,
do something regarding customer service
and you need to use somebody's first and last name.
Well, actually you can't.
I mean, people are doing it, but people cannot, right?
So there's so much need for us, you know,
like especially us.
Like if it can also be, you know,
federated or decentralized and there's no, um, central authority processing,
uh, all of the compute, you know, and then, then, you know, the world will be
such a, such a better place than it is currently.
And I'm really curious to see how fast this thing will proliferate because what Marco said earlier is that he's getting less and less hyped.
And he sees that the increases for model to benchmarks, but like the 0.2 or 0.3% increase on the benchmark
is actually not, you know, notable or visible in the real world.
So, yeah, I'm wondering.
It's like, it might be disappointing to a few
who were hoping for this big breakthrough and suddenly AGI is there.
For me, it's just now it's a really good time
to finally start going vertical
and start using this to build applications
that actually solve problems
because you don't have to be scared of a new model release
and suddenly your corporate advantage
or business advantage is gone
because you were mainly relying on or building an API API wrapper now you can actually focus on the product solving real use cases and just use
the models that are out there because the next one likely isn't going to be something groundbreaking
new so you should just focus on your niche uh and it can't just be copied by someone then utilizing
a better model because the model is so much more competent than
the one that you have implemented so I'm actually quite optimistic now that we will see way more
adoption of actual products because people won't be just waiting for new models where you have one
model that can do everything at once that was a problem I had like whenever I built this one
I was like yeah this is pretty cool but then I see a new release I was like, yeah, this is pretty cool. But then I see a new release. I'm like, man, I should have just given all of this into this one model
and have one agent with all the capabilities.
Yeah, like, double tap on that.
I think now is the time, instead of focusing on waiting for a better AI,
like, now is the time to be focusing on how to better use AI.
The models that we have right now are actually really good to use.
I think the first one, like your creativity,
how deep into it you actually are and use it like a couple of years ago.
I was just asking chat GPT,
like where does Worcestershire sauce come from?
And now like I'm training models and things like that with it.
So it just took having the idea of learning how to better use AI and for what tasks specifically
and have increased my competitive advantage with the tooling that I have at my disposal.
And then when new models come along, I can incorporate that with the information and
knowledge I've already learned based on being efficient with what I have rather than sitting back and waiting for the model to get so good that I can take
away and like abstract everything like all the information I would need to learn in order to
actually be decent or create proficient systems. Yeah, last one I just want to mention. I have developed a few apps where it was just very prompt-driven, with tool calls.
And something I've actually noticed, I was using GPT-4O for a lot of this stuff, and I've gotten to know the model really well.
I've kind of gotten to know the model really well.
And just the way you prompt it and the way you give instructions,
it becomes very predictable the way it does the tool calling and such.
I'm sure nowadays maybe it's improved a lot.
But I remember I've upgraded to one of the all three or all four models at some point.
And it just completely messed up.
It's supposed to be a smart model, And it just completely messed, like it's supposed to be smart model,
but it just completely messed up everything.
Like I had like custom formatting rules
and it's not like I put everything in the context.
I had like different agents doing different parts.
And I thought it was really interesting
that you actually, you don't prompt for like anything.
So like you pick a model
and you prompt for a specific model.
Even a tiny update they do sometimes
to model can basically reduce
your success rate of your app
from like, I don't know, 90%.
It's never 100 with like AI.
But okay, say 90% to like 60%, right?
You start adding like these curly braces
in wrong places.
Now you're suddenly your agent,
multi-turn agent is broken if you have multiple steps. And that is really interesting to me that even though just to just point, even though the model is getting
better, technically it can break your app. Even like the later, the greater the model
is, it just doesn't, the interface is broken at this point
yeah that's a very interesting thing indeed we've actually had many occasions where so we're running
so all of the businesses that we run there swarm like swarm powered businesses so sometimes we
will be running a cluster on a specific model and then like you change uh you change the base for one
agent configuration
then all of a sudden everything is messed up
and you don't really know what happens
tweaking and we're doing stuff all the time
but it's very interesting because
yeah like
each model responds
very differently to certain words
to certain phrasing
to certain structure of
you know like where do you
put what intention and so on and each model also indeed brings about its own
biases so you might have a perfect configuration that works very well on
you know gpt5 or whatever you use deep seek and then you know like deep seek
you know the next version of deep seek comes out and you're
assuming that you just changed the model and the performance of your you know product will
increase but a matter of fact that's not how it works probably you have to reconfigure a lot of
things and you'll have to go back into this iterative process of saying okay well what kind
of output do i desire and what am I getting right now?
Where can the bottlenecks be?
And how do I go about tweaking small things all the time
and testing and testing and testing and testing?
And then one thing that we've also noticed,
which is very interesting, models can also have periods.
So we'll be using a model to run like specific, like a chain of operations.
And then, you know, like this week it works really well.
And then next week, all of a sudden, you know,
it doesn't work that well.
Like the answers change over time or whatever.
You know, it's very odd and very strange altogether.
So yeah, that's a very funny thing you mentioned.
We're almost at an end of the space.
But yeah, we've been a little bit pessimistic in the call on the releases of the new models and so on.
So last few words for the community here Like, we can go one by one. Are there any recent developments that you found
that are really noteworthy,
either in the AI agent and Web3 space
or just in the AI space in general
that maybe, you know, we don't know about yet,
but we should know about?
Yeah, I can start. So we have recently started working with talos
it's this new treasury agent built on arbitrum by the imperial team we've worked with them for
quite a while and the really amazing part is that all of the codes that has been written for Telos and will be is going to be mainly AI code.
So they are using Devon very heavily and it's working really, really well.
And what I love about the project, and I would love to see more like these,
is that they actually started with a pretty empty repo.
It's just a skeleton on how to kind of improve the agent.
And they built their own kind of improvement proposal system and anyone
really anyone can submit a proposal to put to have the agent use their utilities functions
to build the agent in a certain direction currently there is a small committee of
yeah real people that then decide but like talos openly asks on twitter like hey there was this proposal
being submitted into the repo can you guys review what do you think and then it actually checks like
all of the answers on twitter and based on that merges the proposal then it gets actually written
etc and it's such a cool project because it's being written publicly and it has already gotten lots of attention by
arbitrum and actually the ethereum foundation definitely not showing the token or anything
i have no idea i didn't even check yet the tokenomics i just love the nature of the project
and i would just love to see more of these
Yeah, I'll dive in.
So one that I'm pretty sure most people are aware of right now is the ACP, virtuals, or agent commerce protocol.
I'm just very bullish on the agent-to-agent economy.
I think it makes absolute sense, and I look at it in the same way as the standard merchant consumer economy but agents
being able to transact autonomously
to each other by providing certain goods
and services I think is insane
and saying that we're setting the precedent
and Web3 virtual is doing so
I think is actually insane
and I think we'll be seeing that
transitioning into the Web2 world
as well but I think
just Web3 has the ability
to push the needle forward on developments like this because like i mentioned before like the
barrier to entry is almost non-existent at this point for like scaling and iterating on frontier
technologies and systems so yeah i'm i'm super bullish on on acp and all things agent to asian
economy bullish on on acp and in all things agent to asian economy
nice very what about you
um i yeah i i've kind of been looking lately into i've seen a lot of people uh
say uh just tweeting about it,
I mean, a while back, I need to look into Goose. I don't know if you guys have seen
Goose. It's,
I think Jack Dorsey was, like, speaking about
it. It's basically a way for you to build apps.
It's like a CLI tool
slash app where you just say what you
want to build, and it builds it for you. There are quite
a few of these, like, with code. I just like the uh i just love the open source stuff you know the ones that
people uh really put out um yeah i think uh just just hosting things for yourself building agents
that help um me personally i maybe some memory buying type of thing could be really exciting.
Yeah, I was just thinking like, you know, in terms of just helping, I'm still exploring.
I mean, I would love to just build a bunch of like local agents that would just assist
me with whatever I'm doing.
And yeah, because I still don't, I still don't trust you know like agents i think there's
just so much security uh that that needs to go into it there are so many concerns with mcps being
malicious and such um because as a dev you know i want to make sure that it's as safe as possible
uh but it's just uh it's it's it's it's crazy how how many angles of vulnerability vulnerabilities you have with all of these agents, especially when you add monetary value to it.
What if they have wallets?
There's just so much stuff that can go wrong.
And for that reason, and I love the technology.
I love it's kind of like crypto you know i need
to understand it very well first in order for me to kind of like dive deep into it and for now
i'm just uh seeing where i can improve my life with um the agents that i have and i move on to
like the other external agents yeah that's the best way to go about it too you know like we always
tell our community is like start with something that you need and go from there.
What I'm personally very excited about is the merits of decentralized autonomous organizations with AI autonomy in a certain way.
either a consortium of AI agents that are representatives for different people
or maybe just a singular agent that is co-owned by a bunch of people
that's able to manage the treasury over DAO and do certain things.
Maybe the most simple use case could be making investment decisions and allocating funds to
different types of of buckets perhaps because the the reason why i'm excited about this because i've
seen a couple of projects built in this direction there's actually a project that it was on our um
was in the same cohort not cohort it was in the same accelerator program
for that was organized by the SUI Foundation
with us, which is called Aspis.
That project is actually made by one of my friends
whom I spent some time with in Georgia at,
what's it called?
Doesn't matter there's a co-living quasi network state experiment
thing organized by the ethereum foundation um but um yeah so like there's um there was always this
this dream you know of like having a true decentralized decentralized autonomous organization and um this uh was it
has been a dream for such a long time and like a general dream um or maybe a shared dream among
like the crypto diehards especially among the bitcoin and ethereum communities you know. So I think that this vision can still come to fruition in time,
you know. And then, of course, what Barry just mentioned is there are so many issues still with
security that we're trying to solve all together. But eventually, you know, we'll get it right.
And once we get it right as an industry,
you know, once the standards have been established,
once everybody finds consensus to do things a certain way
in the most safe and friendly manner for our users,
perhaps, you know, like,
then we can really start to build out these dreams again.
And this time around,
I expect it will be much easier than before.
But yeah, it's a long road.
It's a long road.
Either way, thank you guys for joining.
Barry, you wanted to speak.
Yeah, I know.
Sorry, the space is getting...
Sorry, I didn't want to have too many points to say,
but yeah, I guess just kind of to add,
because on the data side,
I do actually like the idea of just having data,
just more data-driven agents
where I am the one making the decisions.
I think that would be cool,
and it would be interesting to see
if there's a way
to monetize something like this it's um so like for data people building like data pipelines because
data is the most important thing when it comes to uh i mean if you're making financial decisions
so by information right so yeah i think it's it's really important to uh have tools for that. But if something makes a decision on my behalf,
yeah, I don't know.
For data agents, Context 7, for example, for docs,
I think someone can build a better MCP for documentation.
There is a lot of development that needs to be done,
but I think information probably comes first,
and then after you have the information,
you have the trades in my opinion,
because that's why most vulnerabilities,
if I get wrong dates, I can make informed decision
that I don't wanna perform.
I think there is wrong data, for example,
but if the agent gets wrong data, then you're in trouble.
Right, yes, true.
Thanks, guys, for joining us today.
I hope to see you again in the second week.
So the week after this week is a biweekly space.
Thanks for joining to the panel as well.
Sorry again for having this stupid bug
in the beginning of the space, but we made it all through. to the panel as well. Sorry again for having this stupid bug
in the beginning of the space, but we made it all through.
We had around 300 listeners still here with us today.
Thank you guys very much.
I appreciate your time and hope to see you on the next one.
