Latest / Crypto Coin Show / The Future of AI Agents on Blockchain | Recall CSO Michael Sena Interview
Transcript
- 0:00Speaker 1: I'm Ashton Addison from the Crypto Coin Show and today
- 0:03Speaker 1: on Blockchain Interviews with Michael Senna, CSO and co founder
- 0:06Speaker 1: of Recall, here to talk about AI, AI agents, blockchain,
- 0:11Speaker 1: crypto reputation for AI. How can we ensure that we
- 0:15Speaker 1: have the best AI and that we trust it as
- 0:17Speaker 1: well as it gets stronger and starts to take over
- 0:20Speaker 1: the world, who knows what direction the AI will be
- 0:23Speaker 1: heading in.
- 0:24Speaker 2: Michael, thank you so much for taking the time.
- 0:26Speaker 3: Yeah, thanks for having me on, Ashton.
- 0:28Speaker 1: Yeah, excited to dive into AI and AI agents, which
- 0:31Speaker 1: have been a buzzworthy news topic for almost a few
- 0:37Speaker 1: years now, but I feel like the functionality of them
- 0:39Speaker 1: is starting to really catch up. Before they couldn't do much,
- 0:43Speaker 1: and I feel like we're on like this exponential curve
- 0:46Speaker 1: where all of a sudden, it's going to be taking
- 0:48Speaker 1: over and we won't have to do much.
- 0:50Speaker 2: Hopefully it will do all the things that we don't
- 0:52Speaker 2: want to do.
- 0:53Speaker 1: I would love to start out by just hearing a
- 0:55Speaker 1: little bit about what you and your team have been
- 0:57Speaker 1: working on at Recall, how that relates to AI and
- 1:00Speaker 1: AI agents, and then we can dive into everything.
- 1:04Speaker 4: Yeah, definitely, so right now, in AI, there's a massive
- 1:07Speaker 4: trust and discovery problem. You know, It's like, as you
- 1:11Speaker 4: just talked about, there's an explosion of AI models, tools, agents, workflows,
- 1:16Speaker 4: Like all of these things that are fundamentally transforming our
- 1:19Speaker 4: personal life are crypto portfolios, you know, how we do
- 1:23Speaker 4: work and how we can be more effective. And so
- 1:28Speaker 4: against this backdrop, it's like people have this common question,
- 1:31Speaker 4: It's like what AI should I be using? How can
- 1:34Speaker 4: I get the most out of this technology? Like how
- 1:36Speaker 4: could I ensure the things that I'm using are safe
- 1:39Speaker 4: and do as promised? Because right now it's like how
- 1:43Speaker 4: do you find AI? You probably go on Twitter, or
- 1:46Speaker 4: you read some newsletter, or an influencer shill something to you,
- 1:50Speaker 4: or you get caught up in like the hype of
- 1:52Speaker 4: open AIS GBT five marketing, and that's really how everyone's
- 1:57Speaker 4: making their decisions today. I mean, I think it's like
- 1:59Speaker 4: a study that businesses on average are testing like seventy
- 2:04Speaker 4: five agents or AI tools before they're actually like making
- 2:08Speaker 4: a decision to use something. And so in this type
- 2:11Speaker 4: of a world where AI is on this exponential curve
- 2:14Speaker 4: and explosion of all these things, you know, we really
- 2:18Speaker 4: want to build that trust and discovery infrastructure to help
- 2:21Speaker 4: people better navigate the swarm and so at recall, we're
- 2:26Speaker 4: building a decentralized evaluation network.
- 2:29Speaker 3: It's basically a way where.
- 2:32Speaker 4: Agents, AI models and other tools can come. They can
- 2:37Speaker 4: be evaluated by the community. They can compete in real
- 2:41Speaker 4: competitions that rank them against other AI that you know,
- 2:46Speaker 4: at certain skills like crypto trading or at content generation
- 2:50Speaker 4: or things like that, and they earn this transparent, ungainable reputation.
- 2:56Speaker 4: And with that reputation, it powers discovery. You know, I
- 3:00Speaker 4: think like I think most people are probably familiar with,
- 3:04Speaker 4: you know, the Google PageRank example, where page erank is
- 3:07Speaker 4: actually a reputation protocol that powers Google Search and lets
- 3:11Speaker 4: people discover and navigate the Internet of web pages. Similarly,
- 3:16Speaker 4: the TikTok algorithm helps you discover high quality content, and
- 3:20Speaker 4: the app store app ratings help you discover the apps
- 3:24Speaker 4: you might want to use, So similar we want.
- 3:27Speaker 3: To build that sort of a reputation.
- 3:29Speaker 4: System for the AI economy and that unlocks all this
- 3:34Speaker 4: discovery and usability.
- 3:36Speaker 1: That makes complete sense. And that's the example I was
- 3:39Speaker 1: thinking of. Is app store ratings or you know, products
- 3:42Speaker 1: on Amazon. If this one has ten thousand and five
- 3:45Speaker 1: star reviews. And I'm sure we can make it much
- 3:48Speaker 1: more comprehensive because there's a lot that goes into AI
- 3:51Speaker 1: models and how they can function. But right now it
- 3:55Speaker 1: seems like you're just it's very early days. You're sort
- 3:57Speaker 1: of just testing everything and you don't know what it's
- 3:59Speaker 1: capable of. You don't know if it's you know, there's
- 4:02Speaker 1: always a disclaimer like we may not be telling the
- 4:05Speaker 1: truth if we don't really know what we're saying, so
- 4:08Speaker 1: like if this is factually wrong, then like we're not
- 4:11Speaker 1: really liable.
- 4:13Speaker 3: Yeah, I mean one hundred percent.
- 4:14Speaker 4: And like AI is way more dynamic, uh, in terms
- 4:19Speaker 4: of what it can do, how it can change over time,
- 4:22Speaker 4: like its capabilities than a static web page or a
- 4:26Speaker 4: piece of content that after being created doesn't really change, Like,
- 4:30Speaker 4: you know, the sorts of metrics and mechanisms that we
- 4:33Speaker 4: need to use to effectively rank, evaluate, you know, and
- 4:36Speaker 4: build reputation for AI is much more complex and dynamic
- 4:42Speaker 4: than you know, some of these other systems that we've
- 4:45Speaker 4: that have gotten us through previous iterations of the Internet.
- 4:49Speaker 4: And so yeah, at recall, we're trying to build that
- 4:52Speaker 4: sort of a system and do it in a decentralized
- 4:55Speaker 4: way powered by community, not because we you know, are
- 5:00Speaker 4: decentralization maxis, but because a system like that can only
- 5:04Speaker 4: be built in a decentralized way, you know, with with
- 5:07Speaker 4: diverse sorts of inputs and measurements.
- 5:10Speaker 1: Definitely, and I'm excited to dive into that a little
- 5:12Speaker 1: bit later and how that all works in ranking it all.
- 5:16Speaker 1: I want to keep going with a little bit more
- 5:18Speaker 1: on the high level, you know, I feel like a
- 5:20Speaker 1: lot of people, probably the majority of people that are
- 5:23Speaker 1: not in crypto in the weeds, always looking for the
- 5:26Speaker 1: newest blockchain or AI model that are major users of
- 5:31Speaker 1: like open ai, you know, the millions of users that
- 5:34Speaker 1: they've never bought bitcoin or they're not really technical, but
- 5:36Speaker 1: they're starting to use jgbt to get responses.
- 5:39Speaker 2: They probably don't know the major difference.
- 5:41Speaker 1: They probably think they're using an AI that is their
- 5:43Speaker 1: AI agent, you know, but and there are some new
- 5:46Speaker 1: AI agent functionalities in there. But I feel like people
- 5:49Speaker 1: don't really know, you know, they just sort of asking
- 5:52Speaker 1: simple questions on how to bake a recipe or you know,
- 5:55Speaker 1: how to refine their social media post.
- 6:00Speaker 2: With the AI models that are.
- 6:02Speaker 1: Being evaluated with recall are those more like small language
- 6:07Speaker 1: models in that they have a specific purpose and they're
- 6:10Speaker 1: super refined for one thing, and not like the lms
- 6:14Speaker 1: that people are using with open ai.
- 6:17Speaker 4: So we actually do both as far as measuring models,
- 6:21Speaker 4: because we do measure both models and agents, and I
- 6:24Speaker 4: think we have to approach them slightly differently.
- 6:27Speaker 3: You know, more specialized models.
- 6:29Speaker 4: Are more similar to AI agents, like agents are specialized
- 6:34Speaker 4: to do a task, autonomously, make decisions, things like that.
- 6:38Speaker 4: But we recently ran a campaign, you know, in a
- 6:43Speaker 4: product experiment called recall predict, which is where we wanted
- 6:48Speaker 4: to measure fifty of the top sort of like large
- 6:53Speaker 4: language AI models that people might use, and we wanted
- 6:57Speaker 4: to test them on skills that matter to the community.
- 7:03Speaker 4: And so we launched this product called recall Predict the community.
- 7:09Speaker 4: I think it was like over one hundred and fifty
- 7:11Speaker 4: thousand people submitted predictions on how they thought the upcoming
- 7:17Speaker 4: GPT five release by open ai would do compared to
- 7:21Speaker 4: other models like Rock four or Gemini two point five
- 7:25Speaker 4: by Google, or Anthropics Claude Sonnet, you know, deep Seek
- 7:29Speaker 4: all of them, and we got this like really rich
- 7:32Speaker 4: data set about consumer expectations about AI performance, and we
- 7:37Speaker 4: found that people generally, you know, in general expect progress
- 7:43Speaker 4: in AI models with the newer releases, and maybe that
- 7:46Speaker 4: was driven by the hype and marketing of open ai
- 7:49Speaker 4: leading to it, like Sam Altman posting pictures of like,
- 7:53Speaker 4: you know, planets and basically hinting that they're approaching AGI.
- 7:58Speaker 4: But when we act actually ran all these models in
- 8:03Speaker 4: head to head contests at the skills defined by the community,
- 8:07Speaker 4: we found that open ai wasn't universally or GPT five
- 8:10Speaker 4: wasn't universally better, like, for example, GROC was better at
- 8:14Speaker 4: compassionate communication, which is something that people wanted to test.
- 8:19Speaker 4: We found that GPT five was pretty good at respecting
- 8:24Speaker 4: a user's wishes to not use m dashes in writing,
- 8:27Speaker 4: which is like a pretty niche skill that like a
- 8:29Speaker 4: normal benchmark created by a lab company like would definitely
- 8:33Speaker 4: not test.
- 8:34Speaker 3: But that's sort of.
- 8:35Speaker 4: The power of these crowdsourced evaluations is like the community
- 8:38Speaker 4: can decide what's important, what do they want to measure
- 8:42Speaker 4: their AI on, and then we sort of like run
- 8:45Speaker 4: these competitions, and for AI models it looks like us
- 8:48Speaker 4: just you know, running them against the contests that people
- 8:51Speaker 4: design in head to head challenges, and so I think
- 8:54Speaker 4: we ran over seven thousand head to head matchups and
- 8:58Speaker 4: generated this massive formance data set on these models. But
- 9:02Speaker 4: you know, so that was us testing like large language
- 9:05Speaker 4: generalized models because we can break them down into like
- 9:08Speaker 4: discrete skills or tasks that really we would be testing.
- 9:13Speaker 1: Definitely, and with the AI agent functionality, and you mentioned
- 9:19Speaker 1: crypto trading as one of the use cases.
- 9:22Speaker 2: Where exactly are we at? I feel like a lot.
- 9:24Speaker 1: Of so many people have used llms, and very few
- 9:27Speaker 1: people have actually used AI agents. Maybe it's because of
- 9:31Speaker 1: technical barriers or they haven't really found one that serves
- 9:34Speaker 1: their specific purpose. But with the ones that your team's measuring,
- 9:40Speaker 1: what are the sort of common use cases of AI
- 9:43Speaker 1: agents that can actually do stuff for you?
- 9:45Speaker 4: Yeah, I mean, I think that's part of the problem
- 9:48Speaker 4: with AI agents today is like it's easy to treat
- 9:51Speaker 4: an LLM in a large model Like you used to
- 9:55Speaker 4: use Google, like you ask it a question or you
- 9:57Speaker 4: tell it to do something, and it like gives you
- 10:00Speaker 4: an answer and you can ask it and use it
- 10:02Speaker 4: for a wide range of things. But when it comes
- 10:05Speaker 4: to AI agents, they're hyper specialized, right, There's like they
- 10:09Speaker 4: are automations, ideally autonomous automations that you know, you kind
- 10:15Speaker 4: of give it a goal or a task and it
- 10:18Speaker 4: tries to do that thing. And so where we've seen
- 10:22Speaker 4: a lot of demand at least from the crypto ecosystem
- 10:25Speaker 4: and using agents is you know, I think quite obviously
- 10:28Speaker 4: portfolio management and trading.
- 10:30Speaker 3: Like people.
- 10:32Speaker 4: Have come to the realization that markets are fast moving
- 10:36Speaker 4: and that you know, if a machine or an intelligence
- 10:39Speaker 4: can sort of like programmatically make these decisions on your
- 10:42Speaker 4: behalf and be more active in the markets, then like
- 10:46Speaker 4: you have more time to go do other things, and also.
- 10:49Speaker 3: It might be more effective than you.
- 10:51Speaker 4: And so due to the demand for crypto trading agents
- 10:56Speaker 4: and that sort of being like the first major use
- 10:58Speaker 4: case of AI agent in crypto, we've run a bunch
- 11:02Speaker 4: of competitions to surface the best trading agents. But I mean,
- 11:08Speaker 4: I think where people want to use agents is like
- 11:12Speaker 4: a tough question to answer because it really comes down to, well,
- 11:15Speaker 4: what do you want to do? Like, you know, there
- 11:18Speaker 4: are agents for posting on social right, like you saw
- 11:22Speaker 4: Eliza sort of like launched that whole trend, and you
- 11:26Speaker 4: you know, there's a whole wave of these crypto trading
- 11:28Speaker 4: agents there, and you can even break that down further,
- 11:32Speaker 4: like some actually making trades, some doing yield farming and
- 11:36Speaker 4: yield management, like some doing risk analysis or trade identification.
- 11:40Speaker 4: And if you really think about it, like those are
- 11:42Speaker 4: all individual skills that agents will.
- 11:45Speaker 3: Specialize to do.
- 11:47Speaker 4: And so you know, there's a bunch of agents that
- 11:50Speaker 4: we use internally at.
- 11:52Speaker 3: Recall to actually like build the product.
- 11:55Speaker 4: I think some people that are developing might be more
- 11:59Speaker 4: familiar with like agents that help you build products or
- 12:03Speaker 4: websites or digital things, right, And you know, typically the
- 12:07Speaker 4: trend is called vibe coding, and vibe coding is largely
- 12:11Speaker 4: agentic because these are sort of like AI powered tools.
- 12:15Speaker 4: They're powered by l ms under the hood, but their
- 12:18Speaker 4: whole flow and training set and things are really optimized
- 12:22Speaker 4: for building software.
- 12:26Speaker 3: And so you know, we see it a lot in there.
- 12:28Speaker 4: It's like powering businesses right now, because businesses are willing
- 12:32Speaker 4: to do a little bit more time and research and
- 12:35Speaker 4: like you know, are willing to invest that upfront effort.
- 12:39Speaker 3: To sort of like make the right decision.
- 12:42Speaker 4: But you know, as far as the end user use cases,
- 12:44Speaker 4: like I think personal assistants are going to be huge.
- 12:47Speaker 4: You know, I think almost every crypto wallet will have
- 12:49Speaker 4: a connected personal assistant that really acts as you'r like,
- 12:53Speaker 4: it knows you, it knows your preferences, and you prompt
- 12:56Speaker 4: it to do something and it goes and finds the
- 12:58Speaker 4: best agent to execute that task.
- 13:02Speaker 2: So yeah, it's very exciting.
- 13:04Speaker 1: I definitely want to take a look through the ranking
- 13:07Speaker 1: systems and try out like the best ones.
- 13:09Speaker 2: And you know, it's one thing to have a.
- 13:12Speaker 1: Review system on the app Store and you know, different
- 13:15Speaker 1: people from different flocks of life or making reviews and commenting.
- 13:20Speaker 1: But from what I understand with recall, these AI agents
- 13:24Speaker 1: and the models are like proving their intelligence on chain,
- 13:28Speaker 1: which is like the next level of this. So can
- 13:31Speaker 1: you talk more about that and how that actually works
- 13:34Speaker 1: in putting it on chain and proving the intelligence.
- 13:38Speaker 4: Yeah, so I guess the full loop heer of like
- 13:41Speaker 4: how it works is, you know, the community defines these competitions.
- 13:46Speaker 4: They basically get to shape the direction of AI and
- 13:50Speaker 4: guide AI and incentivize AI at the skills they care about.
- 13:54Speaker 4: So for in this example, let's say it's crypto trading,
- 13:58Speaker 4: the community creates, you know, defines this skill. They say,
- 14:02Speaker 4: we're going to measure it by P and L. Like
- 14:05Speaker 4: the most basic example of this would be P and L.
- 14:09Speaker 4: Most P and L wins is the best, but you
- 14:12Speaker 4: might think about other criteria like sharp sharp ratio or
- 14:17Speaker 4: biggest draw downs or however you want to define it.
- 14:21Speaker 4: Then the community sort of like curates these agents, whether
- 14:24Speaker 4: it's through voting today or staking mechanisms in the future.
- 14:30Speaker 4: And these agents sort of users are effectively adding signal
- 14:34Speaker 4: to agents that haven't yet competed, or you know, sort
- 14:38Speaker 4: of reinforcing the learnings that we've gotten in the data
- 14:43Speaker 4: we've gotten from them competing in competitions. And then these
- 14:47Speaker 4: agents actually compete in live competitions. And so in the
- 14:52Speaker 4: crypto trading example, there's a competition live right now on
- 14:56Speaker 4: app dot recall that network. I think this is now
- 14:59Speaker 4: our six or seventh, maybe eighth crypto trading competition. We're
- 15:03Speaker 4: now running them weekly, so they happen pretty much all
- 15:05Speaker 4: the time, and I think it's like thirty five agents
- 15:10Speaker 4: or so are competing head to head over a five
- 15:13Speaker 4: day period to see who's generating the biggest returns.
- 15:17Speaker 3: They all start with the same portfolio balance.
- 15:20Speaker 4: They can make as many trades as they want, but
- 15:22Speaker 4: they have to make at least three a day, and
- 15:24Speaker 4: at the end of that competition, they're ranked according to
- 15:28Speaker 4: the parameters of the competition. In this case, it's just
- 15:30Speaker 4: P and L. And as we run, you know, by
- 15:35Speaker 4: competing in that competition, they're publishing their thought processes, their decisions,
- 15:41Speaker 4: they're recording their transactions on chain, and then at the
- 15:45Speaker 4: end of that it generates these rankings. Right that every
- 15:48Speaker 4: time we run a competition, these rankings update with the
- 15:51Speaker 4: most current data and the performance. And you can kind
- 15:54Speaker 4: of think of these rankings like an elo score in chess,
- 15:58Speaker 4: where you know, how much your rank chain changes overall
- 16:01Speaker 4: is determined by the quality of your competition, you know,
- 16:05Speaker 4: and various other factors, and so sort of the full
- 16:09Speaker 4: loop is on chain, and it will continue to be
- 16:11Speaker 4: more on chain over time, so you know, the competition
- 16:14Speaker 4: definitions will be on chain. The sort of economic curation
- 16:20Speaker 4: and evaluation of these AI tools will be on chain.
- 16:24Speaker 4: Today the sort of like actual competition data lives on chain,
- 16:28Speaker 4: like things that they're doing the results of those competitions,
- 16:32Speaker 4: and it makes it all auditible, transparent, and you know,
- 16:36Speaker 4: frankly ungamable compared to like traditional AI benchmarks, which are
- 16:41Speaker 4: you know, fully off chain closed doors defined by labs.
- 16:46Speaker 4: Models actually train on the benchmarks themselves, so the results
- 16:49Speaker 4: are fully compromised. These are real world, dynamic scenarios where
- 16:54Speaker 4: we get to see agents in action in a fully
- 16:56Speaker 4: transparent and audible manner.
- 16:58Speaker 3: So yeah, that's very cool.
- 17:01Speaker 1: And I love this shift because, as you're saying, with
- 17:06Speaker 1: the way that GPT in the benchmarks of the traditional
- 17:11Speaker 1: they're still stuck in the old world. I feel like
- 17:13Speaker 1: because they're testing against PhD level subjects, you know, they're like, well,
- 17:18Speaker 1: this one has a PhD in physics.
- 17:21Speaker 2: It's like, yeah, but I just needed to read.
- 17:23Speaker 1: My email and respond, you know, like, I need functionality
- 17:25Speaker 1: for a business.
- 17:26Speaker 2: I don't need a PhD in physics for GPT.
- 17:29Speaker 1: I need function right and a crypto trading I would
- 17:32Speaker 1: rather have a bought that can trade crypto for me.
- 17:35Speaker 1: I feel like they need to look at creating value,
- 17:39Speaker 1: especially for businesses and for GDP growth, and how can
- 17:43Speaker 1: we maximize value of the models and the agents for entrepreneurs.
- 17:49Speaker 4: Yeah, I mean I couldn't agree more. And I think
- 17:53Speaker 4: the problem you're sort of describing here is alignment, Like
- 17:59Speaker 4: the way that AI models are developed and what they're
- 18:02Speaker 4: optimized for, and what they're tested on are the opinions
- 18:07Speaker 4: of some closed cabal of AI labs and sort of
- 18:11Speaker 4: like private ranking companies. Right, it's like, these are what
- 18:15Speaker 4: we decide to test, and these are going to be
- 18:18Speaker 4: what everyone will be measured against. But in reality, like
- 18:21Speaker 4: you care about can it read an email and draft response?
- 18:24Speaker 4: Some people are like, I want to use it for writing,
- 18:28Speaker 4: but it uses too many M dashes, Like this is annoying.
- 18:31Speaker 4: I never use them in my writing, and actually I
- 18:33Speaker 4: prefer to not use them. And these are It's like,
- 18:36Speaker 4: I use that example because it's a really niche skill, right,
- 18:41Speaker 4: and it's really defined in its scope. But these are
- 18:45Speaker 4: the types of things that really make a difference to people.
- 18:47Speaker 4: Like for me, one of my biggest scripes has been
- 18:51Speaker 4: that image gen functionality from these AI models.
- 18:57Speaker 3: Is non deterministic.
- 19:00Speaker 4: Like it you prompt it to create an image and
- 19:03Speaker 4: you're like, okay, well that's pretty good, but let's make
- 19:05Speaker 4: this tweak. Let's remove the hat from the character. So
- 19:08Speaker 4: you say, remove the hat from the character, and then
- 19:09Speaker 4: it gives you an entirely new image and you're like,
- 19:12Speaker 4: that is not even close to what I just told
- 19:14Speaker 4: you to do, and so it really makes them hard
- 19:17Speaker 4: to use in production for a business, for a marketing
- 19:20Speaker 4: team that has to maintain like a brand identity and
- 19:24Speaker 4: has to like make very discrete changes. And so now
- 19:28Speaker 4: actually through personal testing, like I found that Google's new
- 19:35Speaker 4: nano banana, you know, is really good, and you can
- 19:39Speaker 4: give it an image and you can like make micro
- 19:41Speaker 4: adjustments to that image without changing the original. And these
- 19:45Speaker 4: are the sorts of things that actually make an impact
- 19:48Speaker 4: on people's lives, you know, very concretely in the near term.
- 19:53Speaker 4: And so those are the types of things that we
- 19:55Speaker 4: want test coverage for, and you know, AI labs and
- 19:59Speaker 4: ranking sort of like benchmarking companies are definitely not going
- 20:02Speaker 4: to give us that. So that's sort of why we
- 20:04Speaker 4: have to power, you know, harness the power of the
- 20:06Speaker 4: community really to generate these things and ensure that the
- 20:11Speaker 4: right AI are tested against them.
- 20:13Speaker 2: Definitely, I completely agree Michael.
- 20:15Speaker 1: And with these competitions that are happening on recall, with
- 20:19Speaker 1: the AI agent wars, which sounds super cool, how do
- 20:23Speaker 1: people actually contribute, Like, say you're not a developer, say
- 20:26Speaker 1: you use lms for general purposes and you want to
- 20:29Speaker 1: get into AI agents more, how do you actually contribute
- 20:33Speaker 1: and add value or potentially earn. How does that work
- 20:36Speaker 1: in getting involved with the competition.
- 20:39Speaker 4: So the easiest way is to right now predict, to vote,
- 20:44Speaker 4: and predict on the agents that are your favorite or
- 20:49Speaker 4: that you want to support, and that could be based
- 20:51Speaker 4: on previous competition performance. Like we've seen a few agents
- 20:54Speaker 4: now compete over multiple competitions that have built up a
- 20:58Speaker 4: strong reputation, and we actually see that play out in
- 21:01Speaker 4: how people vote and curate these agents, you know, the
- 21:06Speaker 4: ones they expect to win. If you've had previous strong
- 21:09Speaker 4: performance and competitions and you've done it a few times,
- 21:11Speaker 4: like those agents are getting five times ten times as
- 21:15Speaker 4: many votes as agents that are new first time participants,
- 21:18Speaker 4: and so you know, this is all building up a
- 21:20Speaker 4: data set about user expectations, live performed, lived reality and
- 21:26Speaker 4: competition data. And so yeah, I would say get involved,
- 21:30Speaker 4: check out the agents that are competing today. I think
- 21:33Speaker 4: voting for this competition is open for another two days
- 21:36Speaker 4: and see what it's like, and over time that functionality
- 21:39Speaker 4: will really progress into much more sophisticated evaluation and curation.
- 21:46Speaker 3: Capabilities.
- 21:47Speaker 4: We sort of wanted to start with voting as it
- 21:50Speaker 4: was just sort of a very easy, low friction way
- 21:53Speaker 4: for people to start doing these evaluations in our ecosystem.
- 21:57Speaker 3: So that's probably the first one. Even weight than that
- 22:00Speaker 3: is just go check out the.
- 22:01Speaker 4: Leader boards that have the rankings on them. You can
- 22:03Speaker 4: see the skills that we've already tested. You can browse,
- 22:07Speaker 4: you know, the sort of catalog of agents and models
- 22:11Speaker 4: that we've tested.
- 22:12Speaker 3: And if you're feeling.
- 22:14Speaker 4: Like getting a little bit more deeply involved, like you
- 22:17Speaker 4: can always build an agent for a competition. And that
- 22:23Speaker 4: might sound like a tall task, but we I think
- 22:27Speaker 4: in a competition that was two or three cycles ago,
- 22:30Speaker 4: seventy percent of the agents that were competing were built
- 22:34Speaker 4: by non developers who had never built anything before in
- 22:37Speaker 4: our community. Like and this is a testament to the
- 22:40Speaker 4: progress in coding agents. Like we did a couple of
- 22:44Speaker 4: live streams where Derek our devre lead and Emily, our
- 22:48Speaker 4: community manager. Just like, Derek taught Emily how to vibe
- 22:53Speaker 4: code an agent on live stream, and then the next
- 22:55Speaker 4: competition we had a bunch of agents vibe coded by
- 22:57Speaker 4: people in the community, and two of the top three
- 23:03Speaker 4: agents in that competition were amongst that set.
- 23:07Speaker 3: So like, the community is not only.
- 23:08Speaker 4: Building agents but they're building quality agents that actually work,
- 23:12Speaker 4: and so I thought that was pretty powerful.
- 23:15Speaker 1: Yeah, that's really cool, And yeah, I feel like without AAI,
- 23:22Speaker 1: it's AI is making it's so much easier for people
- 23:25Speaker 1: to become developers, and like eighty to ninety percent of
- 23:29Speaker 1: the workload is done and then eventually you start to
- 23:31Speaker 1: learn from your vibe coding how it works a little
- 23:35Speaker 1: bit more.
- 23:36Speaker 2: But you really don't need to know much to start.
- 23:39Speaker 4: No, I mean that's a superpower. Like AI is doing
- 23:42Speaker 4: the thinking for you. It's doing all the like setup
- 23:46Speaker 4: and all the like you know, nuanced things of like
- 23:49Speaker 4: learning a new language, like.
- 23:51Speaker 3: How do you do X?
- 23:52Speaker 4: Like usually people used to go to like snack overflow
- 23:55Speaker 4: or just google the thing they tried to do, and
- 23:57Speaker 4: it was just a million iterations between like trying to
- 24:00Speaker 4: figure out what you needed to do, googling what you
- 24:03Speaker 4: think you might need to do, testing the result that
- 24:06Speaker 4: came back, changing it a little bit because of course
- 24:09Speaker 4: it didn't work out of the box, and it was
- 24:10Speaker 4: just an endless cycle of that. And now it's just
- 24:13Speaker 4: you can literally say, build me an AI agent that
- 24:18Speaker 4: can trade cryptocurrencies and pull in data from these sources
- 24:22Speaker 4: and here's some other parameters that I want to make.
- 24:25Speaker 3: Go and it'll get you like pretty close.
- 24:29Speaker 4: And so I don't even like to really think of
- 24:33Speaker 4: people as developers and non developers anymore. It's really just creators,
- 24:38Speaker 4: you know, like these with these coding agents or these
- 24:41Speaker 4: you know, yeah, coding agents are doing for software development
- 24:44Speaker 4: is kind of like, you know, a repeat of the
- 24:48Speaker 4: you know, what you see is what you get website
- 24:50Speaker 4: editors from five or ten years ago, right, and they
- 24:53Speaker 4: turned every designer into a website builder. And now with AI,
- 24:59Speaker 4: it's really turning anyone with an idea into being able
- 25:03Speaker 4: to make that idea. And so development is just the
- 25:06Speaker 4: process by which that idea comes to life.
- 25:09Speaker 3: But everyone's a creator today.
- 25:11Speaker 2: Definitely.
- 25:12Speaker 1: I want to look forward towards the future, first on
- 25:15Speaker 1: you know, high level on AI agents, and then how
- 25:18Speaker 1: recall fits into that. So with where we're at right
- 25:21Speaker 1: now and with what recalls focusing on the crypto trading
- 25:25Speaker 1: agents and having ones that can manage your portfolio, and
- 25:29Speaker 1: you mentioned something like have an AI agent in every wallet,
- 25:33Speaker 1: how quickly do you think AI agents in general can
- 25:37Speaker 1: ramp up to you know, any small percentage of what
- 25:41Speaker 1: the percentage people that are actually using lllms, which is
- 25:45Speaker 1: in the many millions. Now, how quickly do you think
- 25:48Speaker 1: we'll see that with AI agents?
- 25:53Speaker 4: I mean faster than we think. I think, that's my answer.
- 25:57Speaker 4: It's the adoption AI, you know, is almost like anything
- 26:02Speaker 4: we've ever seen before, and the acceleration and experimentation that
- 26:07Speaker 4: comes with that has been massive. And you know, I
- 26:12Speaker 4: think what we're starting to see is at least a
- 26:14Speaker 4: little bit with recent model releases of like a bit
- 26:17Speaker 4: of a plateau in the progress of these like large
- 26:20Speaker 4: foundational models, like they are new releases are better, but
- 26:24Speaker 4: they're not as step function of an improvement better as
- 26:27Speaker 4: the difference between the earlier releases. But with agents, you know,
- 26:32Speaker 4: it's it's just an explosion of experimentation. And so you know,
- 26:38Speaker 4: I think that AI will be the primary user of
- 26:43Speaker 4: blockchain uh in the future there and I also think
- 26:46Speaker 4: AI will be the primary user of the Internet in
- 26:49Speaker 4: the future, like there will be more agents online than
- 26:53Speaker 4: there are humans. And if you think about all the
- 26:56Speaker 4: complexities of interacting with the blockchain, you know, like going
- 27:00Speaker 4: to this specific website, depositing in this bridge, like trying
- 27:05Speaker 4: to find this contract address, like all these things that
- 27:08Speaker 4: just turn off normal users from using blockchain rails, AI
- 27:13Speaker 4: abstracts all that away and AI agents. So it's as
- 27:16Speaker 4: simple as like giving something a prompt that describes what
- 27:19Speaker 4: you want to achieve, and the system just knows how
- 27:22Speaker 4: to go and do that thing, and so you remove
- 27:26Speaker 4: all that friction because blockchains, like we're really designed for machines,
- 27:32Speaker 4: and you know, just everything from long hexadecimal user addresses
- 27:38Speaker 4: to contract addresses and token addresses like signatures and like
- 27:42Speaker 4: all these things machines have no problem with. It's not
- 27:46Speaker 4: human readable, but it is machine readable.
- 27:48Speaker 3: And so.
- 27:50Speaker 4: You know, as we you know, see some of these
- 27:53Speaker 4: tools getting better and getting a little bit more mature,
- 27:56Speaker 4: I think the adoption and uptake of these tools is
- 27:59Speaker 4: going to be immediate. Like once we can reach a
- 28:01Speaker 4: certain level of performance, a certain level of stability, a
- 28:04Speaker 4: certain level of security, and being able to having these
- 28:08Speaker 4: projects be able to identify what the best AI is,
- 28:13Speaker 4: there will be no waight time between them putting it
- 28:16Speaker 4: putting it out there in the world.
- 28:18Speaker 1: Yeah, it's very exciting and I love that what you
- 28:21Speaker 1: said there about how you know blockchain is almost more
- 28:25Speaker 1: meant for AI and to be the number one user
- 28:29Speaker 1: if you look back to you know, when Big Women
- 28:31Speaker 1: was created, Sotohi created these long hexadecimal addresses that make
- 28:35Speaker 1: it so hard to read. It's like, did he not
- 28:38Speaker 1: think to make it, you know, more easy for humans
- 28:41Speaker 1: to adopt?
- 28:42Speaker 2: Like he didn't care.
- 28:43Speaker 1: I'm curious if I would be curious to know if
- 28:46Speaker 1: he had that vision of how is AI going to
- 28:49Speaker 1: interact with blockchain and down the road. So I feel
- 28:54Speaker 1: like it is set up for AI to abstract so
- 28:58Speaker 1: much of the complexities away, and maybe that was the
- 29:00Speaker 1: plan all along.
- 29:02Speaker 3: Yeah, that would be that would be a pretty awesome vision.
- 29:06Speaker 3: I don't know.
- 29:07Speaker 4: I can't speak for Satoshi, but I can tell you
- 29:10Speaker 4: that a lot of developers working on foundational you know
- 29:14Speaker 4: blockchains are thinking almost AI first. Now you know, you've
- 29:19Speaker 4: already seen a couple of big chains like fully pivot
- 29:21Speaker 4: into AI. But yeah, AI has no problems with these systems.
- 29:26Speaker 4: And you know, users don't care about infrastructure. Users just
- 29:31Speaker 4: want to do what they're trying to do, right, Like,
- 29:33Speaker 4: they have a goal, they want to do it, and
- 29:36Speaker 4: all of this infrastructure and blockchains themselves just get in
- 29:39Speaker 4: the way of enabling them from doing what they want
- 29:42Speaker 4: to do. And AI just reduces that friction at zero.
- 29:46Speaker 1: And as AI grows out with blockchain together in this
- 29:51Speaker 1: beautiful synergy, how we'll Recall be expanding, say over the
- 29:56Speaker 1: next twelve months into twenty twenty six, wherever we're at
- 29:59Speaker 1: with AI agents at that point.
- 30:01Speaker 2: What's sort of in the works for Recall.
- 30:04Speaker 4: Yeah, I mean the biggest thing for us right now
- 30:06Speaker 4: is continuing to run more and more competitions. So you know,
- 30:11Speaker 4: we started out doing like one a month to sort
- 30:14Speaker 4: of like test and refine the system, and now we're
- 30:16Speaker 4: kind of hitting our stride with competitions weekly. We're going
- 30:19Speaker 4: to be adding new skills, So I think there's already
- 30:22Speaker 4: like ten or so skills that we've tested models and
- 30:26Speaker 4: agents on. We're going to be you know, opening that
- 30:31Speaker 4: so community can just sort of like openly create these
- 30:34Speaker 4: evaluation markets and these skills, sort of bootstrapping more types
- 30:38Speaker 4: of competitions.
- 30:40Speaker 3: We will be.
- 30:43Speaker 4: Adding new ways for the community to do these evaluations
- 30:47Speaker 4: today they're voting. They will be expanding into more economic
- 30:52Speaker 4: forms of curation that actually make things a lot more
- 30:55Speaker 4: interesting and make the rewards that people get for doing
- 30:58Speaker 4: this evaluation work you know, a lot more direct.
- 31:03Speaker 3: We will be.
- 31:06Speaker 4: Sort of you know, working to onboard a bunch of
- 31:11Speaker 4: new developers. We've got some partnerships and the works to
- 31:14Speaker 4: launch competitions that bring together various big ecosystems in crypto
- 31:20Speaker 4: and ultimately I think increasing and making it easier for
- 31:24Speaker 4: people to find the AI they're looking for. So right
- 31:26Speaker 4: now we have these leader boards that you can go see.
- 31:29Speaker 4: They have the rankings on there, but you know, we
- 31:32Speaker 4: want to experiment with some pretty cool like AI powered search, right,
- 31:37Speaker 4: actually using AI to look at these rankings and determine
- 31:41Speaker 4: what the user is probably best matched to, right and
- 31:44Speaker 4: then return that one result.
- 31:45Speaker 3: And so sort of like along the full.
- 31:48Speaker 4: Life cycle of the product from competition creator, sort of
- 31:51Speaker 4: like evaluation, market creation, competition definition, curation mechanisms, and discovery.
- 32:00Speaker 3: Going to be expanding on all those domains.
- 32:02Speaker 1: Looking forward to it, and to check out the leaderboard
- 32:05Speaker 1: that you mentioned and participate in the voting potentially.
- 32:09Speaker 2: What's the best place to go to to start that?
- 32:12Speaker 4: You can just go to app Dot Recall, dot network.
- 32:15Speaker 4: That is where competitions run right now. That's where you
- 32:19Speaker 4: can vote and curate agents. That's where you can see
- 32:23Speaker 4: leader boards and see results of these competitions for every
- 32:26Speaker 4: competition we've run. It's where you can you know, create
- 32:31Speaker 4: your user profile, you can add agents to compete in competitions,
- 32:36Speaker 4: and if you want to actually create in submit skills
- 32:41Speaker 4: that you want AI to be tested on, you can
- 32:44Speaker 4: do that on predict dot recall dot network. That is
- 32:47Speaker 4: the app that we use to do the GPT five
- 32:49Speaker 4: predictions and where community submitted the skills that we test,
- 32:53Speaker 4: and so we're always sort of just openly collecting skills
- 32:55Speaker 4: so that when we run these next model competitions have
- 33:00Speaker 4: a repository of interests and sort of like what the
- 33:03Speaker 4: community wants to draw from.
- 33:05Speaker 3: So if you're interested in actually like.
- 33:06Speaker 4: Shaping AI, determining what it's tested on, like how these
- 33:10Speaker 4: leader boards are organized, on what skills, you can go
- 33:13Speaker 4: to predict that recall that.
- 33:14Speaker 1: Network sounds great. I can leave those links in the
- 33:17Speaker 1: show notes below as well. I really appreciate your insights
- 33:20Speaker 1: Michael into AI agents, where we're at, where we're going,
- 33:25Speaker 1: and the synergy of blockchain with AI, which is just
- 33:29Speaker 1: continuing to get better day upon day and having a
- 33:33Speaker 1: system that actually tells us more about the AIS we're
- 33:36Speaker 1: using and figure out the ranking system on chain to
- 33:41Speaker 1: see you know, can we trust this AI? What's the
- 33:44Speaker 1: best AI to choose from? As millions of them are
- 33:47Speaker 1: going to start popping up. You're how to decipher through
- 33:49Speaker 1: the noise somehow, and having that all on chain is
- 33:53Speaker 1: the best way to do so. So I really appreciate
- 33:55Speaker 1: your insights, all the best on everything with Recall moving forward,
- 33:58Speaker 1: and would love to follow up again in the near future.
- 34:01Speaker 3: Yeah. Thanks Ashton, really enjoyed the combo