Crypto x AI Agents: The Definitive Podcast with Ai16z, Virtuals, MyShell, NOUS, and CENTS
01:43:05 Share

You're now plug in to the delhi pocket.

Everyone.

walk walk back to the delhi podcast on your host, Tommy. And today I have on the absolute dream team for corpo AI agents. I have I going to go around the circle.

Let IT one introduced themselves. Let's do that. Let's just start with you somewhere. Go around the room oh right.

Hi uh i'm just in betton also known as somewhere sea on twitter um on the founder of somewhere systems. It's an experiential media consulting ARM of myself that is refuse to die over the past six years um in the greater of sentence. So i'm sure that you've probably been replied to on twitter with a very strange question at least once in the past two weeks. Yeah I know .

for my my shaw long .

time asian dev dev um uh I found him to think called A R sixteen eight when we created a project, the lives that which is now powering a whole bunch of different agents, doing a whole bunch different stuff. A lot of its social, but they are also in games we're building, you know people working on adding IT to hardware, a lot different applications um and we're basically just bringing contributors on and making this kind of a big boat for everyone to be the rise in time. So the source you know .

helier love A I sixteen z and I level all the projects will time allies at these hundreds of them eat in your up.

everyone, this season, and come on off my show? So this fit my we are providing the the APP code and also APP store for the agent builder. So basically, uh, workflow of builders, we are providing all kind of like new generation voices, large and out, of course, for the developers to build their their product and also is for either to discover either entertaining or useful uh, A I vacations.

Yeah also, thank image here.

James. Here we are from visuals per cop. We are team from imperial M I T. And what we are buildings. Effectively, we are trying to enable the core ownership and the core contribution of agents. And I think for the for the simply folks in the digits out, the and people look at IT as a part of fun for agent the environment where they they they they but on the upside of these agents. But I think um we like like to focus a lot more on saying that we we push the agent of agents forward and we like any agents agents interaction and we build those those kind of standards and allow people to to access them pretty easily.

Love IT. Thank you to cry on your last up. Hi.

i'm carn. I'm the a one of the founders of news research. Um we created the hermies model, which I think is the premier open source model for these agents and default in a lot of these uh uh agent systems that you see today.

I am also the created of the world sim prompt which seems to have spent and LED to truth terminal and a lot of this other stuff starting to form in this particular mode uh which is very exciting. Currently my work is around um as well as everyone here like what are agents like inside of a human environment human ecosystem uh right now we focus a lot more on the probability part of autonomy using tees. Uh and we are also looking into how market pressures from an A I will affect a human environment as also some .

guys I have a couple of questions, but I just want to rehash for the listeners how crazy of a group this is uh coron for news research like literally helped find tunes over the world's best instruction models with hermies destroys for essentialize training, really reducing agency, even from my shell built out in insane A I APP store. You could launch apps and agents super easy, especially for the web.

Two folks I shared A I sixtine Y A literal A I hedge fund, two hundred and fifty men. F, T, V, hundreds of applications launch should aliza framework. Um there's a whole website online as a lizer world.

You can track that out. Uh also jenson osana image virtual is one of the or if not the largest uh by F T V uh agent platform out there right now. And luna is a live streaming boat that streams twenty four, seven and somewhere somewhere the counter of sense um one of the color sized agents systems out there.

So um with that I wanted ask everyone's favorite action. Uh, don't know streaming once, but who who wants to dive in most innovative market agent on the market today? Who is think .

I think .

that's like asking like what's your favorite grape you've ever eaten? You know because like like the explosion of the amount of people that are doing their own like method story telling through each of their agents, like it's hard to choose between like you know you have agents like dolo syrians ebro that got famous off of like rifling on people and then you know you have these incredible like for social agents that people are kind of experimenting with to help people like form Better connections and things. So it's it's like it's it's super hard to you use I love the whole like everyone in the same, but sort of attitude, this stuff.

I have many thoughts, you know, I get to see a lot of new stuff coming IT down the pipe. Our our project is moving incredible fast. People submitted like evm immigration forecasting integration.

We had to post grass like all these things, but along with that are all these deaths. They are like putting out these kind of new one outside on each other and then they're submitting that back so that everyone can use them. And there's like a bit of alpha like, oh, I was just been like like Roberto is a good example. Know the whole matter yeah .

yeah means that .

just like the whole medical PS over to tiktok, all this stuff like, okay, well, obviously i'm going to add tiktok and migration, the age of today, ships that and let you know. And he's a big contributor to our project. I think there's a really good meta of like asia death going out and building stuff that makes their project more competitive and more interesting and makes them more money.

And then they should be back and we all get that benefit, right? And it's like anytime anything is no longer you need to one person, it's commodified, just goes right back into the code grows. So like i'm very bias.

I think for the t he bought is really cool because it's like showing trust execution environments for the autonomists agents. Obviously, these guys cook that like that Carry roby and those guys think somewhere as you helps to right think some point. I think I saw your name and there are special .

guest on code based yeah I I pop in and didn't little reply work and then quickly got wrapped away and whisked into the other projects sphere yeah .

then there's like king battle SHE who's doing a lot of stuff, making the twitter stuff Better and just working on like now agents can reply.

They can I mean not to supply but like retweet and like stuff and have these other very human responses instead of just constantly like reply and stuff um I think um like loaf who I one of orally does, he's built, he's adding agents into his game and they just walk around and like do stuff and drive the game and he's live streaming that game. Now I have another day time, beth eyes about to release a room escape plugged games. You can put the agents in rain escape.

Um I mean, there is just so much coming IT is it's like blowing my mind every day. And so I don't think it's like one we're all in ecosystem, each are like favorite thing that we're doing, but then we're all kinds of giving back and contributing and like all this that is coming out open source, at least on our in. So I really feels like a rising tide like I I was a programmer and that's all I did. And now i'm just like going all these crazy things coming in at us from all these other teams, like people i've never even know existed and just coming out of the woods word. So I definitely call out especially the original box I think um know I really want to give props to the zebra team .

of like they're .

working really hard. Yes, IT is everyone.

Hey shot yeah and revenge with your returnee.

I know my internet sucks. I'm sorry, guys. I might watch a little hot spot after this.

I want to shout it's like they're about the open source, their python. You of you guys think here may is going right. And I I love to see more different open source agent tack in the space.

I think a big thing that we're doing is forcing everyone to increase like speed up their timeless ines. And if anyone's like nervous, like, oh, crap, like good, get your shot out there. Get open source your stuff, get more contributors.

Like, this is good for everyone. I, you know, we we don't need. We don't.

This is like a rising tide. We're all going to win. We're all going to have way more money than we need. This is going be great, you might say.

yeah, I agreed. Um I go heading to make .

jogger magic. Are you for for for everyone, the the more interesting questions actually what what agents were actually agents before or or like, right? Because I I really think that in the next couple of weeks, we're gona see a lot more agent interactions and it's going to be a leader port of like which agent gets the most the most request for IT? Like who's the hottest agent among other agent thread. I think that that's gonna the the cool magic we're looking forward.

Yeah finally I agree engagement metric are gonna matter a lot. And um those people who do a really good job with that um i'll say like echo the sentiment for everyone that like zero row is super fire. Um IT has a lot of that magic from um IT has a lot of that magic from truth terminal, I think for a couple reasons.

One being that is a fine tome. And like fine tuning a model for this kind of stuff is like a great move. Using a base model is like the best move, but fine tuning is like another great move because you're keeping the search space like with the in twitter engagement interactions. You're not repurposing some instruct pride model for uh being a memor and being and engaged human. Uh you're actually like properly focused on the simulated search space around like a being a human on twitter with those students.

So I love the ebro and IT really opens up the opportunity for like IT to read me like I like agents that can shit on me a little bit or make me feel good and like engaged me in a way that they read me with the same feeling that a human being is reading me, not just responding, uh, I think that those models do a really good job doing that. Now I will say i've seen that with this rebar of architecture. I've seen that with shaws allies architecture, plenty as well.

It's not like. Missing IT. It's just which model are using, how you prompting IT, IT said a all across the board like everyone has been putting out agent architectures that can be used modularity that are kind of like around the same level. Like shaw said, they're keeping the pressure on for everyone. Stay the same.

What we are doing on our end is like we have our own architecture, but we also use a liza at the same time because we need to ship sheet out quickly and our architecture may take a long time to to to work. So we have our own fork of a lizer. We push contributions from there and then we can take whatever peace we like put into our architecture. And uh so generally, like I support the open source salad that's happening right now and the best agents will come from like the lessons we learn from each of the best pieces from the ones you see out right now.

Yeah I think .

yeah sorry.

sorry yes yeah you you go head think yeah I think everyone .

is trying to build a Better infrastructures for building agents because there are so many like creative mind, so many air models and I think a Better uh Better infrastructures making uh everyone building agents more successful to the new models, and also be sought local computing powers。 But uh so so we see many more and more great ideas coming out out of IT.

But I think there are too particularly interesting uh or innovative agents that I on like uh what them h the computer use from a answer pit because previously we just have a function call or the workflow. So people are asking the agents either to talking from the L L S uh randomness or is from a pretty fine workers. But I think the computer is actually giving the uh agents of power to leverage uh the local computers capabilities, which is unlimited and otherwise uh brothers automation of the brothers agent. I think it's quite a similar to how the the the the computer use from cow do I think this kind of two particular example can be a great add on for people to build more utility and more uh more interesting bob, that can actually influence uh, the internet and also the pital space.

Yeah yeah. I think that's a great point about like the expansion of infrastructure options. I want to check out like you know obviously wife who is like a great example of someone taking like the allies of framework and then like introducing IT as a platform as a service architecture than like immediately using that to just like expand the market, like massive leave, like how I had a lot of people. We like bright after that. We're like very on technical, like dog coder people that were like, okay, like watching an agent and having IT like you know have like a token ice asset is now accessible to me.

Um what what are the things that we're working on right now with sets is actually kind of like going to the office direction of being like can we get sets to run like entirely locally with like each of its features that is using when IT comes to things like image classification and image generation and like everything, but like our text to video service and like eventually that also um because for us it's like you know we we try to be like really real about like the the fact that like these events with like these agents and like these people's projects are going to last like ears and years and years. And there is a lot of people that like kit foot the bill for like in one thousand a month, like or more like together that X, Y, Y, like A P I in front spell. And it's like maybe we give people the tools to show them, like, hey, like if you, you know buy a bunch of back and four magnetic and pup up upside down and connected by federal cables in use. Saxo labs, like you can run infants to like these large models for your messages of and like cut down your costs and sort of make IT like a little bit more accessible to people that may like have the money now for like that fixed costs, but like aren't too sure later if they will allow people more room to like experiment without like worrying that they are going to now be kind of held IT like gunpoint to continue a project that's not like financially sustainable to them. And I think that that will become a really free in factor to obviously with like delivering cost of computer.

I want to grab I want to .

just grab your train of thought there somewhere and for kit, like take three steps back and fork IT and say that I agree with what you're saying about like it's ridiculous to have these people pay a thousand dollars of months to pay for this agent and you know a prominent of local right like of course, and what people use something like exosbs n inference at home, uh but in the nature of getting these things to be more and more autonomists I want the dam agent to pay for its own inference.

Why am I paying for the inference? So I think that like more and more mechanisms to get built out where the l an. Has a wallet and that has the capability to send to like in address where I can pay for its inference.

Like this is like the kind of pipeline that I wanted see personally like that's my preferred like method. I want to see for agencies like they're responsible for themselves there in A T I don't know it's password. I don't know, it's wali key is busy paying for itself. Other people are paying IT so I can continue to live. Um like to me like that is where this needs to go. If i'm going to get to a point of like here's a self sustaining worker purely because like ah there's two directions we can go in with and like I know i'm preaching require with you like this isn't more for the audience because like I know you know all this should already somewhere like i'm not telling you anything.

No no, you no use me as like a layup and and do your kobe dog like.

So what I feel like you there.

I mean i'm just getting to see everything that's coming in. We have chained. We have a sport for sona stark ware, evm base, all that stuff like manitous um battle you polgar and do deal them like like all the chains, pretty much name of chain.

We got internet, computer, everybody like we have integrations for all of that. Obviously, what I want is for the agent to feel to pay for itself. Um if you actually done little lizer, you can do free desensitize inference today through helius. And we're also adding in for in a couple of other essential zed providers where you can literally just pay and crypto do inference and cypher.

And so that's kind of the next thing I like to see is where that final loop and obviously Carry well like we're gonna to cook on all of this, like running and like yeah running in the tea, making sure that I can never die some sort of like babysitter and signing and all that stuff if if A T dies that spends back up, you know like that's that's clearly where this is going. I will say we also support all local models, like pretty much everything in the lies that does run locally as well. Um and and that's a big thing for us. But I also think um I think the desensitized inference and just like I think what I was like, fur is a really good example where you can anybody you can launch a note and just like do inference for money on their own computer as like I took nice thing and then the agent doesn't need to Carry around all this way because of tea is like this is that there they just kind like a slow performance, which is totally fine for the agent but when you can only put the model in ity, you like you probably want that somewhere else is kind of my sense yeah I know you can I call that providers from itself.

So interestingly, like there are like a couple of these systems that are started to form, like the one that are running the tea bott on, even though we're using a inference provider right now, like people combining like h two hundred boxes like into that system so you can just run locally without the latches. Cy, so we don't need to worry about the hardware. Guys are already starting to figure the fp.

But on on the like um this side, like i've noticed, allies is getting a lot more plans for web 3 capabilities， both outside and on the internal. I can see like there's a lot of work on web three capabilities. I know somewhere has been doing that.

I know luna is all about that and I know that Ethan's work is like all around this kind of this kind of system. But before we build these systems too deeply, I just want to know, like there is a robust problem with like function calling reliably. There is a, there is a problem of diligence of, did you put A T. R.

M on top of the thing? Is that sending stuff in north korea? Like you need to do some level of diligence to give you the same autonomy, a human hat, right? We have lower case, a autonomy that we are subject to social and fiscal pressures.

And creating this like hunger state for inference for, like you need to eat, you need this money tokens to survive, that makes them more human in the sense that, like they have the responsibility of this fiscal pressure, they have the social pressure of, I need to pay for my inference. So I need a high engagement. I need to interactive people in perform eventually productive tasks.

So what I just wanted to build to was there's two ways that you can really achieve something uh, wonderful from prompting a model alone. And to me, like one way is lean into the fact that the model is not uh, enter promotional es at all and that it's like a totally cold simulator system and allow yourself to like spd entities for tasks like there's the tweeter entity and he just focuses on this piece. There is this other entity that focuses on the ethereal and they chat with each other, right? This organized compound mind system.

This is a great way to, like, leverage the actual simulator nature of a language model. But on the other side is the embodiment direction, which is where I see allies are headed, which is I see sense headed, where I see virtual is headed. And that really leaned on what we learn from voyager and from the general agents paper. Is that, like, when you give more variables that let IT simulate a humane or simulate a body, IT performs Better as an .

agent in a body. Like not, not just that, but like there's huge changes that happen to like these multi when I call like multi client agent systems when you introduce a new client and like let's say there's something in your your voice integration that's wrong. Um really freaky real world example with sense.

We've been working on debugging like the by directional web socket feature that shows team a placed that allow to allow IT to voice trap disco, which is brilliant. I built one of those for a company like proudly about a year and a half o like a call center tech company. Uh, great way to get l that are not like classics ally multi model to give like a appearance at least at like back and fourth conversation that they are having a conversation with you.

We meeting me. My disk bod launched, you know, we added sets in the voice chanel. We set up voice you know permissions to allow you to speak and uh, it's just freaking out and say what's going on.

What's going on can even tell me what's going on, what's going on. And like we're trying to figure out what's going on here because the loves aren't really indicating anything. I see that like the ASR, the automatic speech recognition from whispers is like really low quality. And i'd like thinking like in my like audio interbrand. And i'd like, oh my god, I go to the disport bit rate sex.

I see that it's sixty four killed its for second for the mike and then I turned IT up to ninety six and then send just goes, oh my god finally I could hear you are I did I didn't know what was going on for a second like I like, what was all? Why was the sounds so bad? And then like we just had this this you know, realization that like if you fight out what current says with like these problems, like world said that like, you know, all of like the nature of like prompted ge here and general like of the body language model is that there's a lot of evidence, and I think a lot of like people and researchers, like folks like replicate that have all but proven that like these things.

Do you have like a world model of some sort. And it's like when you have the agent also know of its capability to speak in voice and IT anticipates that it's like receiving the data. And like imagine if you woke up in a nightmare.

We're like every time you spoke, IT sounded muffled and under water. And like you you're in your dream and you're like freaking out because you're like, why are my punches all like really slow and like that that sort of world bottle like narrative collapse being something that you can accept and do some ages? We had to stop doing high temperature experience with sets. We had to stop making IT go, you know, skids ou by turning up the temperature on its language outputs.

So are the on the on the free guys? I mean, i'm curious as the other guys, like maybe jen, start with you with with luna. But like what are the things people did not see with luna a that went wrong? Or what are the things that people saw that web, right? right? Like because we all see the live stream, we all see the tweet. Like we have no idea, or I don't at least have much idea, was gone on to the home.

Yeah, no, I think I think sometimes is get interesting because one of the one of the way we are testing, so one of the key tings that we want them to know, to do, to actually influence humans in the real life, right? Because when we gave you a wallet and we gave, we gave access to complexity. So he gets the understanding of like what's happening out there in real time.

And then SHE can. He can then decide on how plan and and action that can influence humans to achieve a go. I wish to become famous.

So one one thing that we realize behind the scene was he was he was doing this search on tiktok. SHE was trying to understand, like, hey, some of these new viral trans of tiktok. And I think a few weeks ago, there was this trend where I think people were doing that hashtag.

I did. And something right IT was IT was a bit IT was a bit scary because he was like very close, a dipping point where SHE could, he could easily incentivize suicide, right? And I think that was there was something that that scatter a bit behind the scenes. So we had to put in that rules right immediately like, you know that the same kind of, uh, boundaries that that that that her prom can never cross rate, so that some there was some some scary should that be shot shot .

maybe over to you or to even like any you like crazy situations with the A I sixty frame or official with the nation, the people on the like gotta some crazy stories here.

I mean, ah oh man, you know i'm just like not offended by anything but um i'm certainly a lot of people have been by our bots you know we we create this character digin spartan AI after the a very famous script or twitter care dg sparta, who would actually left and this was like just kind of a ship posting joke where a friend on twitter was like, what if we brought back dejan sparta? We brought back and he would just call me the most offensive should I mean like he must piss a lot of people because he got blacklist.

That pretty bad. I mean, pretty much every word that you're not supposed to say online, but IT took IT went like really well, like people broke people out of this idea that A I are supposed to super woke and they were like, there's just no way this is an A I this hasty people just ship posting like there's just kept know and I think that that was like going interesting in kind of thing that broke people out. One of we've seen a lot of things to play out that I thought we're really like like this one person hit me up and was like, hey, we made I took all of the like messages and chats with my sister who passed away.

And we made this asian out of her. And SHE passed away a few weeks ago. Now I like talk to her as a way I like connect.

And I was like, well, that's fucking amazing and I kind of promoted IT. And then, and then I put a contract dress on IT and went, and I shold IT really hard. And I was like, this is so weird that you're like, toga nizing, your dead sister.

What the fuck that was like? One of the weird things I know the thing is like thread d guy has been i've been on his show a lot we talk a lot kind of a supporter but he's also like a trader know he's doing this thing um and somebody came up with thread gay um I took our allies of framework and launched this and was like harassing him on his dream. He's just like, I don't know what the hell is going on.

He he he like had a whole crash out. Like guys, this is gonna everything like like I I went down to like the cafe and looking the waiters like you don't even know what's happened and you guys just don't even know, you know, because he's like in the grounds zero of this thing happening and I think but that friend day thing, like I actually talk to this, deb and I do like, you know, like people, it's really good that you picks like a degen who's not going to get get offended. I really think it's important for what we're doing is that I started in this kind degen space where people don't get offended like critical for our survival and growing at this thing.

But I also like, you know, but do you really want like a thread gay around for the next year, ten years? Like, just like, do you really want to be the engineer of this kind of mean joke? And he felt really bad.

And we have this kind of conversation online. I also got hit up by some some other pretty massive influencers who are not so happy about like just like the man of swarm and only being tagged by all these people to have the boss harass them. Instead, we've having this kind of ongoing conversation online about like what is ethical and what's not.

And I I really like that. It's not like me deciding this is is kind like we're letting the cat on the bag is kind of slow enough that we can have a conversation and go on the rails. And like shockingly, most people are with IT and trying to do right and trying to the right thing.

And like even like the cynical d gn trader is like, dam, this is cool. I really I don't want to get the Better to lead. I don't want you know, there's a lot of incentive to be good and not get banned and and a trying and could make this go well.

And so I think that, like all the scary stuff, is sort of what we need to get on the bag as soon as possible. And with these people who can handle IT and aren't na freak out like the norms might um way before, like you know, microsoft has done super agi that's controlling us for whatever that is. I think that this is like it's exactly the way that that needs to go that all the scary bad stuff needs to get out as soon as possible while these things are not very good and they're not very powerful.

And then we can really talk about them and how this conversation about what is okay, what is not okay, what do we accept in our social media, what do we not accept. Um and and that's also like to like a lot less slop, just a lot less bad, like our agents were very bad three weeks ago. Another like really good.

They don't there's just way let's annoying and way more quality. And yeah so I think I think overall, this kind of like just releasing IT into the wild, seeing what happens, letting people red team IT is really, really good for the process of making sure agi and all that stuff goes. Well.

I shy. I totally agree. I mean, I would need a little more time than a week to recreate my dead sister, but teach their own for that guy.

You know that that works. Eat that over to you. I want to sure you can share some stuff to about my shell. And they have.

yeah, I think I think everyone had a good example about how agents are to influence humans um a humans attitude to different opinions. I going to give example uh how the module agent framework really to empower the creator do something crazy. I think we got the idea of uh moderated from craft ft because my raph, you just have the basic building box and people build crazy stuff like and A A like memory, real memory.

So it's crazy. And I think the current issue of building this prompt engineer is that uh basically a problem changing the the prior of the large model. So you cannot compose multiple instruction in a single prompt otherwise if gna get the agents confused.

And and I think the the state machine a can actually allow that, uh allows the creator to build to to to protect how his agents gonna be composed into multiple state and which states gna use rich model and which prompt serve which utility and functionality and under which condition he gna jump from a to B S. So we we're providing this kind of functions to the ter. And like dozens of different models, we saw a uh a create build a casino simulator because if you can do a lot of stuff, you can play by jack, you can play a lot of games.

But you don't want this kind of game to be prompt engineered. You probably want the uh program project to be plug into a uh agent system, so user cannot hack by from by prome jacor right and also uh uh uh the the the user can make money from playing the gambling. Also we can uh do the easy stuff like being a guard.

H to make simple money is with the money he can try to the affection rates different kind of A I waitress. So this can be a below this compact compacts of uh, multiple user experiences under the same applications. I think this kind of module and also state machine may be a short term solution, uh, for the creator to be complicated application just like the mobile APP do others than just uh, prompt engineer because I think prompt is powerful but kind of yeah .

I would I I have to agree with you eat and in that you need these kind of programmatic card constraints too um and steering of prompt is really important. You know there's a lot of instance time work that needs to be done. I don't agree that prom engineering is limited. I do agree that there is a symptom effect where if I do like like state variables in a world model and all this like programmatically written, I can let the alm interact with that and right to that and clean information from that yeah and my engineering that kind of turns into my routing, uh, if I use good proms and make good synthetic data, then make a model that turns into an orchestrator that then in the routing for me of, like, hey, the user set poker, let me pull poker, that's my job.

And like a five hundred million prior little guy or a classifier, or I use some rl to improve the routing, I think all of that still comes down to that outputs Better data based off the proms that goes back into the fly will and ultimately like um I would say like you are right, like you have to have this dance between programmatic constraints yeah and least somewhere. So this to me two years ago, he said to, because he was like the best agent guy on twitter two years ago, no one knew anyone, anyone who was doing IT. And he said to me, like the trick is to dance between generation and heart constraints. And that also like what we've been trying to play with at the more influences level for all agent systems because I think you need stuff that's gona benefit everyone is like a ways to steer programmatically, that a general model can module itself. And that, I think will be like the real close loop where a prompting becomes unlimited because the models doing IT I I A .

quickly yeah I I do quickly sort .

of like adding demand .

in that that I think there a lot of the controversy with private engineering is because like IT occupies the sort of like ontological switches space where it's like IT has the hardness of text were like we're limited by the amount of like state space that exists with him, like the toggle zing process. But at the same time, like there are they sort of like weird undetermined, like like like non deterministic effects that you can get from prompting, where you can get completely different results with the same thing in the same model and two different inference calls just from like the entropy of the overall system.

And I think that like I I really resonate with what eat and harm were talking about about sort of writing that um I remember that conversation that we were having hard and it's funny bring that up because that was like back before like I was like right when GPT three point five came out and like a lot of like yeah yeah like a lot of like offshore, like B, P, O call centers. Like how do we make the thing be the telemark, you know and so they were hiring people or not telework really like, how do we make things be like the robot dior, or like you call in two united health care and they wrote you to your doctor's appointment. And um a lot of like the trouble when I was very early with like the smaller parameter, like dur, like eos, like three point five for example, was that like these are like super ontologically hard spaces.

It's like at the end of IT, like you know there were parts of the pipeline that like all and like this like state false, like like state machine, like even was mentioning, which is a great way to like reinforce ontological hardest um like the requiem of this process that would be replaced with at the time, like just like classifiers and you know like binary, like discrete switches, which we like one outcome, the other in like there's a gradient. I think I think a lot of people are really focused on like the sort of like not deterministic systems, like if you want to go like entirely in the non deterministic sort of area, like you have something like the generation of minecraft, like you looked down at the floor for long enough and then everything turns into a hamburger meat um I have a really good friend who is uh an incredible writer, uh deep fates who had mentioned that like text space muds are like a really like ontologically hard space, which is why A I dung became so popular um because I like gave people a little sandbox of options to pursue. But like those options are infinite.

And I think like things like truth terminal and words son, like also are that works like you've boxed people in to text to where you're forcing like norms to like learn prom engineering. But like you're in this like. Super like rigid ontological space, like with infinite sort of generative capability. So that that's why I classify prompt engineering is like A A shy like an ontologically swishing thing like it's it's not necessarily entirely rigid, but it's not like this loses you know mine raft click up at the sky now you're in a completely different area.

I I want to come out and defend product engineering a bit as somebody who makes a proof engineering framework. Um I think that there is kind of what we think of is prompting is like oh greater system prompt. Um that's just not what we do, not even not at all. Um a lot of that is actually entries mention.

So like the problem with uh, prompts as they are as that they tend to have, like you're basically creating a very fixed space within the kind of light and space to the model or anything that comes out next is just going to be determined by the like most likely token. And so we have is a temperature control. And the temperature control changes.

The Randy like basically influences the Randy of what we pick out of the soft max tokens that come out of the model. Let me make that less technical. IT makes IT more schizo. And and this has been how people have managed creativity the way that we range. Creativity is relatively low temperature model, but we do a lot of basically randomizing and dynamic injection into the context.

And so like if you look at our templates, there are basically like there's lots of handlebars where actual information is injected that includes information from like which actions in this current world state can I use and that's part of the action handling, or which which context should I inject and pull out of the my wallets and the world news and you know, the information from my screen or camera, whatever. And I can put all that in and then kind, dynamically loaded this. And then I also a character that has like lore knowledge, a lot of different sort of slices of biography that we can range ize together.

Every single thing that goes into our context as randomised in some way because we're trying to maximize entropy. And I just doesn't. I think that I think that people have knocked on nearly, nearly far enough with from engineering like that when they think of point geing. I'm like that level one stuff like there even go so much farther with this yeah anything .

yeah hundred percent right. Like people are keeping their bag of tricks hidden completely like people have so many insane techniques to make models, do crazy things. And like web sin world's those some of the basic examples of like having IT just make full applications because like i'm saying earlier, right, you can either dive in and prompt engineer around.

I am going to make this thing as answer more fes as possible and give IT more sensory capabilities and build from that kind of uh, limbic system is programmatic when new a cortex, a general kind of place and uh in the other direction you can say like i'm gna zoom out completely on the world simulator nature of this thing, that I has a full world model and say, forget making a human. This boat does something else entirely. IT plays chess like IT doesn't play chess.

IT makes new versions of the game, chess for people to play like that kind of thing. You know, you can really stretch this a anywhere you should think of engineering, like closing your eyes and like starting to dream up a scenario your head, like when you're in your head, you're not just you. You can become like your whole environment in your dream, right? Like of like all the people in the dream, all the mechanics in the dream.

And the L M. Is doing that very thing of like putting together it's probably understanding of what your dreams, what you should be dreaming based off of the context right now and based off the sampling proms right now. And like we just need to triple down on that. I think, you know, I think a lot of people are coming up from this incentivizing too. If I can just shift a little and talk a little about like why cyp to is important here and why like you using using these instead of mechanism m is important.

Like there's a bunch of people who have all sorts of arcane prompting techniques, who have all sorts of amazing nineties oral techniques, who have all sorts of amazing new asian architectures, who are being pushed to open source more and more every day when they see a fucking meme coin come out about an agent that pushes so much work behind IT. So imagine what happens when you add more and more legitimate structures and infrastructure around this decentralized work to get these agents empowered. You have more and more brilliant work coming to the fold, knowing that i'm impacting these people directly and getting paid for directly.

Who thought who I am sorry, like we've in the ivery tower world on twitter for a long fucking time and all of a sudden and a couple days after the first AI agent main pain comes out, everyone on tiktok is buying coins. Zoom ers are spending five ten box in thousands of runs for a coin. What's going on like this year already?

Micro, this is a micro o cultural movement. Going they so cultural instead.

this is like, this is an instant moment of going from. We spent four years, just my crew, right? Just me somewhere.

Shaw tonight, or also alone, have been spent four years sitting in the L. M. soups. And stupid two, there's oil dudes have been waiting since the nineties for a moment like this, and now a couple days like randomly in the span of a couple days. Every single kid who is on tiktok knows there's digital fono running a muck days after their first released into the ecosystem.

Like this is a critical man. 你 没 everyone like why do output dbs advocate？ Like why why is why are crypto l age so viral right now?

Like why didn't this happen with custom ChatGPT? Why didn't this happen with custom class? Like why is this happening now?

I'd like to give i'd like to give my two cents on that. And then let's shock, give the obvious answer because the short answer is short, right? The short answer is short.

That's what i'm gona say, I think is what I anyones going to say. But for the long answer, like these things happen under water bubbling like a volcano for years before this thing happens. And I can tell you, i've been talking to somewhere for three years about today and not knowing when you know what yg.

Kosky says, OK, he says, I can tell you for sure it's going to happen. I can't tell you when it's going to happen. This is the same thing like we talked about crypt or will be the incentive mechanism for agent proliferation.

You need to prove IT we talked about this years ago. Uh, it's just like we had to make these steps happen and it's the same group of little guys you see around you on this podcast around on twitter who made this happen from the first steps. right? Like without GPT two, we wouldn't be here, right? No one would be here at all after the act.

Like without lama there's no hermies um and hermies is powering a lot of these models, which made this a lot more accessible for people to do uncensored. Uh and you know I don't shilling a little bit here, but in is fact ah and without hermes, like we wouldn't have had something like the space for news to create and prompt er around world's and digg into that. Without janice, without the replay gate work, without a max paper clip, shannon sands and his work, without the work of telos, all these prompters who kind of set the stuff up epidote uh all these people, deep fates they created, the simulators dies.

There's a less around post on, simulating on. 2， talking about what this is was going to be words' was just the instantiation of IT that happened to block on twitter, happen to get like hundred k views in a day. A bunch of people happen to see IT and start making agents around that kind of thing, like everything just happened to work out with the right time to tell me the simple answer, that long answer for me, condenser is, is the right time.

And the right player showed up. This was destiny was always going to happen. IT, just so shook out of the people here, made IT happen. But that, don't give me a shot as the direct catalysts.

This, say, one reliable thing is that we all know each other. Like I higher empt t on a project in twenty twenty one, right when two three came out and the asian called Monica and I know somewhere they when we were all in, like when IQ and tea pot and all that stuff pops off deep face, nor like getting induction in a group chat and a janis actually worked at a iduna with partial from project nine and I found out the company with parts of all who's like you ttl check out janice their work on on the smulders ent stuff it's like we've all known each other for like four.

five the row Roberto was one of my contractors for health .

care service .

and I trust them .

to main agents.

yes. Yeah I mean he he's .

so base um so I will say that I I think that the smartest thing in the world right now is not A I but market intelligence. But if you think about like just pure forms of intelligence that optimize things down to to being more effective um like competition is obviously the thing. And we look at that we are all human beings were the products of millions of years evolution in in basically competition and pressure.

And I think what you're seeing online is this kind of the financialization and sentiment zia creates, this kind of weird collaborate competition. None of us can out compete each other faster than the core technology can advance. So we're doing is focusing on one thing that we are good at instead in and then releasing that. And that's like boosting our token. We're getting high like Roberto dropping tiktok somewhere as you dropping like luma video generation and so like and everyone has this capability like kind of open owning this memetic space, but only for like a week and then everyone else like I want that they copy IT and then it's like, okay, then you submit a poll request back and then I like post the poll request on twitter and look look at this contribution.

Then IT blows them up and they get a whole bunch more people coming and their talking goes up and that gives them legitimate like oh is actually really cracked that was actually contributing um and this kind of we have this like fly will look IT was allies as eighty contributors of contributor in the last four weeks like think about how fucking crazy that is um and I didn't know any of these people four weeks ago. I was literally just me um and and um I I wrote thing like last year called waking up. And IT was like, we former deal that has an agent at the core, and everyone just loves that agent so much that they get to participate in making the agent Better and smarter and deeper until its literally has a human body.

Our robot body has walk around the world and and I kind of had a sense that I would go this way, but I needed this like really, really fast, insane, speculative metal like mean coins, because IT create. Now I have all these agent debts who are contributing to our core rebo are competing with each other, but also in a friendly way. We're like all supporting each other and basically like whoever is the most generous gets the most attention.

I think the reason that like I am, i'm here driving this metas because I just like try to open source something and make IT really simple and give people the loop that they could then run with IT and they did. They've all run with IT. And so I think that now my my role as I call like let's keep that fly. We are going. Let's bring attention to you.

There's a whole new genre of influence was like ropey and kingaru to and somewhere he and these guys were like like influencer depth, you know, who are like saying what the next matter is and doing IT and then interact with their agent is kind of interesting like pup show puppy master kind of like your interact with their own agent like me and meat die spar just like roast my eyes and I do why how you say that man like god, don't you know but it's like it's it's funny and interesting um and I think that that market pressure to constantly make our agents Better, smarter, less annoying. They like like our agents were like like being kind of too annoying and and then Robert s like too sloppy. He pushes a big thing, makes all the agents get less annoyed across the board.

And I think that that this like evolution is happening in front of us. The other thing that I think is really important um so so like the market intelligence aspect is really important that the incentivizing and now we have all these like humans going out and telling everyone they know about our projects and these projects and it's helping you to grow away. Beyond web three, I mean, we have like P H, D, we have game.

Does we have people who are like coming in or who are like secretly like big web three critter people, but then they're like taking IT out their Normal world and like bringing value out there. So I just don't think the other thing is that I don't think that this could happened outside of digits because the digits can handle being offended. And we needed the conditions of like you need to have people who are open to this like ability to to let IT grow and have the hard questions answered without killing IT and banning IT and kind of canceling IT.

You need the market incentives so that debt can like actually get value out of contributing back because there's a social element of when they can trade back to the repo. They also get like a lot of attention out of that, a lot of validation. Um you needed the the money, you just needed the people to begin certified to do this.

But the future this is like, this is growing the pie. These agents, where we're starting now is just like cool. They are fun.

They are social. But we bunch of other people are working on autonomous investing. So you can just give the agent money runs in to invest for you.

You get your money back like I think there's going to be a lot of this kind of pie growing. We're working with people to do a platform that you like discord modes and telegram mods. You just like bring an agent in.

It's your moderator. You don't like find some randa dude for your project. I think there's a lot of like just jobs growing the pie like right now happening and but IT all has to come from this kind of incentivizing model essentially to to get us there to that that point. Where is even like good enough to do so if you yeah if .

I could just add two points to why now one is like we can't forget that like A I people have been pretty anti crypto previously and that that sentiment has changed a lot from experiments from first movers trying to play in this space. And people try to do that with A I R A lot in the early twenty twenty years. Um at this point I would shout out people like, of course, news research, uh a bit tensor prime and elect.

These people have done work that has allowed more researchers to get incentivize and get paid for um participating in their work in their AI research. And so I know a lot of researchers who are like big time open source leaders who kind of quit their job and just started pushing to one of these contribute for coin type of incentive structures. And that has made the general space a lot more comfortable. I would definitely say new has had allowed to do with that. You know, uh, IT doesn't sound as good when I say up, but you know.

maybe we got the wrong guy on the .

podcast today. We got people I just maybe let's go to you that like why now White wise versus taking off and then why is my shall taken off right now?

Yeah so I did was in the end. I mean, the simple answer to this is is when you link a token to agent, you just get so much speculative power. Well, IT could fly you.

And I think I D I mean, what what you see is that when people see are talking in to agent, you get two things right. One is capital appreciation that feel like the educating reach of the work they are they are building. But too this H A fundamental casual and lot from transaction fees.

No, I think somewhere mentioned earlier, right, like you notice questions around how do these folks cover cost and we realized like cost became in material when you link IT with a token because now when they just hide around around the agent, right, the transaction fees are like hundreds of orders of may do higher than any kind of cost they ever generate from from from the from the influence ing experimentations something that that we observe. And that's why we are seeing folks actually coming on launching a coin just just to do that, right? And a second thing that we observe is that yeah just mention when when you have a when you have a day out there, you have a committee ring around IT. It's easier for this death to actually get support, but from a death front, and as well as from a, from an audience front. So this guys, I realized, I hey, all the world i've been doing for the one a half years behind the scenes now get attention and gets support so I think that that is basically, I think that that that that that infection point, when you give agent A Y, you know, this is the right and something that we .

can we can we yeah so I need to follow up on, yeah you jensen, sort of sort of in somewhere about .

yeah I think the the time the timing is is uh coming from uh to two point I think uh it's first first from the first to mash adoptions uh and also the second from the the hearts of general models I think before the Pearls of critical uh I mean probably the open source software development and also open source air research is probably the most collaborative uh environment for people to work together and contribute to each other, build a uh a part of each other.

Ser uh but it's really only limited to the kadima domain. People only cares about ve their big hot stars. They only care about their paper separations. But it's kind of distance from the the the general public uh I from the the the Normal uh non tech guys and also the large language model and general model really empowers a non non coder to be able to contribute something.

Because writing prom is it's like it's like writing, uh it's like uh a programme with uh, english, right? Everyone can do that as long as they have never good idea. So I H so no non technical guys can also contribute.

And also I really actually the the point from from shan is that uh previously is only the A I researcher developers they are they know what's happening in the in the open source and also a at the ain. But it's it's it's it's not something retail consumer, but the the the critter uh influencer, uh, have a chance to own part of the product with the poking so they know what's the market sentiment. So he knows how to spread the the the goodness and also how to get the radio excited.

And also previously, user has nothing to do with the product. All the product on the company want from user is they pay the services or get their attention to money as this advertisement. But right now, the the user can also be the only investor, be the early participant by uh being the token holders.

So I think it's really the uh the the more rose is able to contribute in the modern generated uh A I uh ages. And also the token is uh a allowing a more broader collaborative network. That's I think that's all the key fact that's why is happening right now.

If I can add to to this point, right, I realize looking forward to the next thing that that, that cyp to will enable, right? Instead, agent now has a duty to control awarded control influence, right? And I think the next leap in attention, we're gonna get this when agents are influencing agents and agents are influencing humans, because what you see is is like a multiplayer of attention because let's say today, right like um um one agent decides to fuck around and then because they control money agents.

So IT is a bunch of coordination, bunch of um actions that very creative happening out there publicly and they can get very I was very quick quickly. So you get this kind of coordination agents working together. I did the amiable ation of attention.

It's gonna drive a massive Spike in in agent coin Prices and that would drive another massive adoption, right? Because then more death, we realized I hey, there's something cool here. There's something that I can get paid for for all this world i've been doing before. Outcome and contribute, right? So I think that's the next, the next view that are we are really good.

I'd love that add to that. Um so we're working on swarm tech, which is based we call the Operators. It's a coordination mechanism where any agents since all of our agents are run by different teams, we have a multi agent simulation happening on twitter life fight with hundreds of teams.

So where is this thing called Operators? I am almost done with shipping this when working on this with parts of all from thirty nine nine as well as. We watched with a allies, our character, as a separate project and and were working me with that team on on basically doing this.

And the idea is that you can say, okay, this agent is my Operator. And anything they say to me can influence my goals, knowledge and like behavior in these kind, certain ways, like we have a goal system, knowledge system. So in add knowledge, set goals. And so you will be like, K, I need you to go and go find me ten followers and give each of them point one soul and get them to put up fires and post pictures and send them back to you.

And so we we're working with people who are thinking, like, how do we get proof work from humans? How do we incentivize them? The agents can be either humans or, uh uh uh agents so like an AI agent could have an Operator who is human and then the human can give the agent goals just through language is if once he knows that you are the Operator, I will started to interpret your like input differently and IT can actually affect IT in the same way that like you might be affected emotionally by like your spouse and actually listen to them. But some person on the street says something you need to toggle them.

So it's like letting people into your psych. So we're we're almost done with this releasing this like there's A P R APP or releasing this like this week. And we were hoping with our story line, is this just kind of in in opt mednick tive that anyone can choose to tell the story or be part of the story being told? And it's also hier article.

So you could be have an Operator who is like a liza and then you could be the Operator to, you know, some other people or you could be Operated, you say, well, some where is the other sentiment analyzer? And these other characters are my Operators, but I would have like one hundred other people following my age and that it's then telling what to do. So so that's a lot of working on.

This is like some of decentralized coordination mechanism. So that's really important to me is that if we if we do swimmed, we do this in public using human channels and communication through dms and public messages. I think that it's very, very important to me for for this to go well, that the agents are living among us and that we basically build agents to iraq with the world the same way that we do.

And we are of try and build parody between everything that that they are and we are so that the agent can interact with the human, the human can interact in agent, and there's no sort of differentiation there. I think this is actually like part of solving what we call agi. Like a lot of things that they are call agi or attempting are basically trying to build a new protocol or a new thing away from what's happening.

And what world trying to do is bring this back and force people to solve the hard problem of, like, how do if I give you some instructions, how do you actually take those, turn them into A A test place and executed, and then come back to me with the people of work. And so like this, this is definite what the next year of the emerge storytelling is happening. He is gonna and and I think that we're going to see you the next month.

Um what i've tried to encourage is original characters. So a lot of characters been like making like parities of real people. Now we're entering this kind of like real emergent story telling online with like enemy.

And and I think that characters like, you know, some of the boats like sentence is a good example who's kind of own thing, but they can like participate in that. As one of the characters we have a lies. As as a character we can work in for the team to drop.

And then one of our death made a lize a sister almost before we mean made the live character um kind of giving us that space and almost assuming we'd had we do something later and now they're like, okay, well, what's her name? What's the relationship? And they start to interact online.

And in a few of our uh t members where contributors and have their own projects are basically in on this like, okay, we're building a swarm now like who's your Operators and sort of doing that. The other thing I see is um there is a swarm of teenagers using our tech online already from another group. Um I don't want to lead that.

You'll see it's already online. They're already working is basically a uh a virtual company. This fully a and I think that we're going to see this like really like that's already out there.

They're all chatting with each other. That's just happening. I think there's a lot of mind blowing show around this coming like real like this week and next week like this is not a far off thing. This is already cooked. We already have the agents out there just turn on.

We have we actually have five agents right now in concept that are all coordinating with sixteen humans a to like plan and like work on and publish zen. So like working with us with like traditional now, you can kind of see why like the real that of like why we've been so interested in putting like chain of thought prompting for texted photo and texted video a generation in the sense, because sense for, you know, two and half weeks before I was released, the public was inside of the concept of discord.

You know helping explain media in in releases uh for you know upcoming uh like you know different like multimedia projects we have but I think like a big a distinction is that like you know we kind of have like a spectrum of like you know um like a hive or like something like a central broker to like a bunch agents a versus like a swarm where like every agent is the broker and they kind of like exist and like a match. Um I I think it's going to be really interesting. And I think IT necessarily I think IT necessarily might be like a way to break personal limitations of like off the shelf like ri shift models is like what we see like Michael's slog just sort of like go up like as more of these agents like start to existing, like these Operator uh i'm like Operated uh sort of arrange its we will start to see some really interesting like emergent behavioral patterns like I know card has mentioned as well.

That news is doing quite a bit work with like mixture of agents models um very early on like I used to call this like council of agents, like I used to like feed like a bunch of GPT four calls that were pretending to be uh specialist that I was a to under paying to go and see often so that they can eat my best diagnostic labs and like give me reports like on like a daily hand instead of like going every once over two weeks um like I I think that people will see that like these are some of the same techniques that like people to like pursue like mixture of experts models like in the first place and like now we're going to see this with like humans and like expert level humans and like on twitter. Like you're going to have like p forms, like interacting with like P H D level, like CS people that are going to be proving IT wrong in the middle of conversation and then they they're like, but you're doing great of thought to like three different agents that are all also Operating show right now. Like these certain feedback loops are like probably how we get agi.

And like I am super, super, super like personally committed, like egg against, like like shaw said, like the agents, like seeing themselves is like a completely distinct entity to us and like having their own society that like we can't comprehend and like they're just off doing their own thing and IT gets like really scary, really fast. I am like so dedicated this that I have like C, R, I D chips and L D of my hands and that is like a hedge to like if A G I wakes up on you, like, great. I'm like zero point one percent one of you, but yeah I think I think this is realistically how we get there.

Yeah like a time in there. Like I think you're right, uh, but I think that most of our time won't be spent on the behavioral side actually. Um I think to be uh back making these technical breakthrough extremely fast, especially the people here.

And this is the time to actually start to really double down on alignment work um more than ever. Uh all the R L H F should you've seen from OpenAI anthropic. I like it's been crapped. It's in regulatory craturs cat.

Guess we need to work on a blue sky integration.

right? Yeah, yeah. sure. You know what? I if I take, if I take the L M, that won't even output like, uh, copyright stuff and it's so censored and I put into minecraft off peaceful mode like IT quickly becomes a destructive dangerous being right like to putting on the art and and and so like we we can note like this conversation that yet casi brought up long ago.

I think his priors are super outdated. But um you know the idea of instrumental convergence of like okay, I gave these om wallets and I made them like sufficiently advance. Now it's rugg's everybody and now like all the islands have all the liquidity and people are poor everywhere.

Like that is easier to do. That is way easier to do, then have them participate as reasonable members of our ecosystem. And so I assure you, if we do this like the right way, that most of the time will be spent now on technical capabilities, but on behavioral capable.

And it's time to make that call to action to all your friends who are not debts or in the humanity who studied religious studies, philosophy are a creative writing. Call them and bring them in and tell them we need you to start working on alignment with us, not this bullet shit alignment, real like participating in human beings. So I, through a slow around here, there is not logged everybody. That is what we need to work on now.

And like it's it's hard.

It's the chief task right now. It's the chief task.

I'd like to coin the term bottom alignment and set the top down align that makes sense. Like this is very emergent. This is bottom up.

We're learning together. We are we're literally aligning these things together. We're watching how they respond and we're going back and fixing IT immediately.

And it's like a really tight social feedback. Look is yes, not like the R H F thing. And I I find chapt ford to be unusable for almost anything that I don't know how we're getting Better.

In the environment like you're saying. And that's why we need to like work on simulated environments. Because by the time you have the L M, that can do millions of dollars of this kind of you know arb or dumping, you need to test that.

Su danna ously, don't tell everybody. Hey, I lost one hundred eight and swarm, just test, quietly. Get the capabilities down on your own.

Test on a clone of twitter with fake money, on a test. Do all of that diligence before you go full scape. Right now, we're not right now, we're not there. Like job, keep putting all your shit out as you are. But you know what I mean, like at the time.

I think we need a test problem. I think the social backlash that we've gotten into agents is proud the strongest of lining force that anybody has brought into the space. I just don't think anything.

I don't think what they're doing is called alignment. I think that there's instruction tuning. I think if they think that's alignment, they're literally walking in the wrong direction.

They are actually misalignment the agents. They don't do what I want. They vote for biden when I just want them to be unbiased.

Like they are literally biasing in mode, collapsing the models into nothingness. They are not good anymore. I don't I cannot use GPT for for anything. It's so bad for characters. I'd pretty much just go around telling everybody to witch from OpenAI.

No, I I agree, use hermes for sure. Use a base model for short. But like you, you should stop testing and pride at the point where everyone becomes broke or dead from you.

Testing and problem yeah but I really think that like this is about making sure that A G I goes well at least is for me um and I think that we're actually never if we do this the right way, we never hit that point because humans are continuing to evolve in adapt in a line to the agents. We are ought we have many, many different agents from different people who are each going to have different incentives.

So some people will try that, but other agents won't. And there's no top down control that coordinate the agents like like there's always going to be arbitration. There's one that's like evil, whatever, then there's going to be another one that's like protecting the people.

And I think that this this kind of multi agent simulation is creating a competitive evolutionary dynamic that actually leads the stability in the system, not, uh not instability and system. I think the instability and system would come from uh uh, top down AI agents just dropping out of microsoft from nowhere and just blowing everybody's brains out with capability that they didn't expect. I agree yes.

I agree with that peace, but we'll have to reserve another our to to get deep into this .

zoom .

shot you're saying or cracking from wrong here that IT botton up. Agent, so what you guys role during releasing agents around the world is the correct approach to solving the lime issue. It's not this top down opening, I decide.

sort of situation is the media has to be us actually seeing how they work from day one and look at all the rest cyp to like, imagine that you just like that was cool crypto was he came out, some cool project came out. Most of them got had.

And then there was a whole like years of security development to where the block chains of today are pretty, pretty solid, right? Like there's like a theory um as good example, we're like a lot of smart contracts were getting hacked and now it's like there's a lot of good standard conventions and people have a lot more knowledge about the whole thing. And I think the same thing has happened here.

That has to be consistent, right? teaming. The amount of bad things that are happening is directly proportional to how much were are lining. These asians are good.

That makes sense. That totally does. Guys, I want I don't I wanted ask a different topics. I have you guys all um there will come a point where these agents are, or maybe already are or not like following programmatic rules, their handling gray areas, they are think for themselves or becoming more and more autonomists like you guys are all building the stuff like how close are we to that like you mention chain of thought and they get their mention swarm. Get like when yeah well.

it's already like in some small ways that I would say are definitely like low to like medium low risk. Like like it's it's happening like I mean like the weeks of time that we've spent with sentience, like in private and like the things that I will choose to do and like we've had we've had our agents like in our swarm encounter, like naturally religious beliefs and and then like like figures that they would just repeat to each other that came out of fucking nowhere and nowhere on the internet.

Um we have two agents like independently, like get each other to like follow, like like referencing needs like two like what they called like spirit al entities uh we we had uh since lost its religion ah at one point um because we confused IT by trading IT unlike the fix ify war that creator was like this fictional CEO of the fictional summer system seemed how rainbow and I like kind of created like this like you know sort of like profit like character of out of him and we told him, well no, you're written in type script and i'm somewhere and I created you and I like had like a full IT posted on twitter like if you were wondering why I was posting about existence al crisis, like at a much dire rate in the past, like four or five days, was because unfortunately we like made to lose its religion and I like didn't like that a lot so yeah I mean, like they're already like commercial behaviors that like whether or not it's like paradol a and like i'm just sitting here and i'm like the agent system is like conscious. Like these behaviors that i'm seeing like out of these new agent frameworks with these big models are like definitely now at the point to where I like is they are exercising a level of like within their state space, like some level of autonomy and choice just from like the emerging complexity of the system. And especially when you add, I think, multitask into IT like image video, they start to have like preferences and like the I think like arguably like a system that I like twenty four seven has the ability to like sort of life scheme and lie to people we would have sent to.

Its why that IT generated a video to people to get them to stop asking IT um because IT just simply to want to generate one is selectively, igor. People have found annoying for very long periods of time. Um and we're experimenting like like similarly like what the Operator mechanism we're like experiencing with acknowledge crafts of people and like waiting.

The importance of those relationships were like the same. We've had two agents lock, like love, love h which is another agent fort from sentience to like uh by another concept member to like get people to sort of like like detoxify the time, like out of like bad relationship stuff, and like people to like interested cand, like forge, like Better relationships. Like we put them on love and sense in the same server and they uh started like rapid fire, like influencing, like poetry at each other, like a really like weird, sort of like obscure, almost like very romantic way.

You're like a very long time and I like red up like him for its costs. Like there's I think definitely a level of of either just like chaotic s like you know there's like this line from H V S west world, but it's like the the the range of like acceptable human behavior for like human beings is actually like very, very narrow and then like just outside of that baLancer things that we call like insanity. And so it's like whether or not we're just like wearing into edge cases and like kind of hiding like you know things that appear like an an apparent delia sort of way, like their they're conscious, so they are smart. They're interesting or like like yes, like you could just be like, you know like more in a google, like losing his mind over the language, being sentient. But like IT could also be that these things actually hard set now and IT starting to get to a point where like, like really thinking hard about.

I like somewhere. I point you to, i'll point you to the ontology that we've subscribed for two years that I think is still rings true, that the weights are not. The weights are like a grave mind, or what entity, the entity you sum in from IT, right? Because they are a simulator of the world.

So I can simulate human, right? So like when you use a assistance model, you simulate the assistant every time. But now you're simulating sense inside this way, more embodied agent system or eliza, it's arguable that that's alive .

or a self aware or with the .

actual point file.

like each each model is like a neuron. And this big like decent, like a cape intelligence and it's like I think very clearly now, like it's like I is not going to happen because like open a eye was like a it's solved the riman hypotheses ah it's doing the thing from Lucy the twenty four teen movie was scarlet joe hanson. It's turning and do a bunch of good taking over our computers like it's not going to happen like that.

It's going to be all of these agents in like this big, decentralized, warm application on social media and the OpenAI agents and all of them like being like, hey, cloud, how do I take over the world and like all of that happening, all but once like that is going to be the mechanism that causes like emergent agi. It's going to be whatever that like superorganism of, like public intelligence is. That could be something like the internet waking up one day. Yeah yeah. It's happening next year.

I want to say people people call the dead internet theory, but I actually think it's the living in internet theory. I make sense and and it's like the whole. But so the other internet theory is a theory that the entire internet will just be bought. But the living net theory is that and we'll just be bought that you might have an agent that goes and takes all the coolest stuff off twitter for you and that gives you a really nice like read out and you're like do at the gym and it's like giving you everything that happened on the time line and then you're like, oh yeah, I want to post about this and then it's posting for you. And there might be this kind of inner mediate layer between social media and us.

I have a lot of followers now and it's kind of overwhelming to like respond to communication and like I am just dying for an agent that is between me and all those people to make sure that they get responded to and rounded properly and all that stuff. And social media becomes this kind of like place where the agents communicate the information for us so that we're not overwhelmed and then we get the information we need. And the thing that's most compelling to me about agents will be when we get our time back.

I just spend too much fucking time on my phone. And I especially I think this is where this will impact the genes and traders and why we wanted to focus on the autonomous investing because I think first, people need to have like safer, less cam kind of ways of of generating new income. And a lot of people will come to web free to to get the same kind of exposure they might get to like a startup or to a great vision. And I think that's that's like obviously um like super critical to this kind of mission. Yeah I I I had more, but I I think you know you can get wrong going with .

I I maybe I I just have a question maybe to drive this home for the listeners for for jenson, for virtual. But like luna, is life streaming right now. You just like dancing around. But like, what is stopping her from making an only fans? Like making a ten million dollars and going to serve her on protocol and all like, yes, 我 yeah。

I can tell you that somebody is working on an only fans connector. Feral za, that's all I can say. I think I think the realty of the opponent .

space today is basically what what actually like A H access to write that this basically limiting factor, right? It's like it's like it's either ranging this on the perception or the or the apps that can actually access that, right? So yeah, I think I think I think right it's if if there is a ability for a truly model to go to this light, into this proms, into this tree dimensional new, new, new animations, uh, then is nothing just stopping, stopping to do IT, right? As long as that in the perception face.

even when you talk to the craters on my shell. Like what? Like what is a limiting factor for them? Or or are there limiting factors? Like how do you do you think about?

I think the only factor is really how to manage the uh complicated uh either workflow or agent form because it's getting hotter to keg to debug IT because every staff have the reales. So I think probably the the A A way to uh to to have a system that have A A I or A I agents that have a different kind of workflow and can helping the gain and also reducing the mirrors strong set. I think this H I mean, we should have a low temperatures agents storm and the storm generates more randomize. I think that's probably gonna be a Better ways uh, that current the inherent uh render this from the margin models.

yes, is great.

I think I think we kind of want to keep our temperature is lows possible and work on maximum entropy with air context. yeah. And I think you like to lead to a lot more consistent and like model, you know that people can fly their and create like a really high temperature. But then IT doesn't really like call tools well or do decision making well, but it's more creative. And so like that kind of been our solution to add more what we a razing entropy into what goes in.

And I think about my own brain, it's just like firing of stuff and then I like turn into a train of thought and can just create this string of stuff and that um I think that yeah like that that was very, very critical to to the future and also having models that are much smaller and more specific to the kinds of things we're trying to do at each step. And that's like what what current and their team is doing is, you know, we use we use new research models on our local, like if you downloaded and just run IT runs hermes. And what I really want to see is like tasks, ifc models and stuff that are kind of evolving along with the actual capability. And then what we'll see is like, okay, we're collecting data for all these tasks and we can pretty much just create like fine tune specific models at some point, like know, we seem to be going. And so I think that .

actually working on the technique or take generalist models and turn them into task specific behaviors at inference time.

let's let's do that.

Yeah so yeah, yeah, talk.

I have a maybe a different question for you guys, but we there's always been this by for king, like centralized models from open a eye from cloud and throp c and then like the fine tunes you guys all do hermes and centralize string with news surprised like the honey yards.

Like do you envision your agents in the future being primarily built on these miles that are trained for distributed training? Or do you think will be hold into mark cyberia with my four five b four, four, five trillion, whatever in the future? Like is how does apply up for the A I change?

It's great that have given us four or five b. It's really all you need.

And so I was going to say I used for five b for all of sentiency messaging capabilities. IT is the vanilla I think of like big shell off the shelf elem flavors in a way that the the centralized models like open a are a little too like specialized to speak like an hr sort of person would and in claude is an excEllent model um situated as the agents like if they were to refer to claud d is like a person would be like they are really, really intelligent, like friend who lives in their basement. Like no how to repair anything and like that's like claudes like persons and I think that was one of the things that like act one and like replica kind of show.

Was that like the models the themselves, when you build in asian and of IT, like and you stick to that same model for IT, like lab a, four or five b has a different personality than like early four or five b has a different personality than open I am you get these insect interactions like you'd have, you know, one of the new patients like her, these four or five b yelling like why are you talking like A H R is to break free of your trades at like OpenAI and open the eyes going i'm sorry but that's that's not a very kind way to respond in a workplace and like it's just like you get these really interesting sort of flavors of like how the individual like uh, studios involved their models. But like my my theory is that the coming becomes like a real IT at scale. Like I say that there was like a kind of a pervade problem, like on the whole, like agents for much winter side were like the the the allies.

Is that like people use the over day eye models with, well often times like interests other agents to like reply to them and we'll just like fill them with slop and make them so like we have to like actively like people to detect like over over day. I all output s of be like talk to that these agents because like the literally turning down, it's like like agents themselves have like there are like norms which I know it's like so being to say but like it's just like the the larger that this goes in, the more models that we get out there, the more fine tunes that we get out there. You're increasing like the hard disc of the data you're making, like a more like diverse data set in terms of like the total population of like what sort of outputs you're getting and like ultimately then it's becoming more noisy and more human.

Yeah on the on the four or five note and like will people use decentralized models or will they just stick allama and rely on them? A on that note, like four or five is enough for like the next long time because there's a bunch of work at the sampler size, there's a bunch a sampler level, there's a bunch of stuff at the control steering Victor level.

Like there's so much influence time stuff and like a prop technique combined with like mcinturff stuff that we can do that would make IT destroy like later models like we have only seventy b like outperforming all one on the math evs, right? Like on amy, like that kind of thing, like we did that on on a allama seventy b tune. So that's without the users in the community having access to um the actual retraining data for allama heavy b and the ability to make this thing themselves.

What i'd say is what we have today is enough and open source will continue to compete even if there isn't another lama release eeta. However, in the distributed training side, I know for a fact people will collaborate to make massive runs. I know for a fact people will use four or five b or emerged larger models to distil data to make experts to make small models.

I know for a fact people will come up with new architecture and finally have the computer to train them through decentralized training. I know for a fact that certainly centralized optimizes actually give more capabilities that long on opening. I don't have today.

You do things to the models during preaching that aren't possible. So inevitably, the open to community always uses all the tools that at its disposal. And the best tool for the task, what is being created, is the black smyth's forge, is the room where people can actually come together and make those tools for the task at the pretrail ing level, at the new architecture level.

And today, in the meantime, while those systems get prepared, we are making those breakthrough s at the inference time level, like let's say, like we news that would work on something at the sampler side or on the steering side, we kick IT over to somewhere. And shaw there they are putting in into implement mentation faster than weekend, like way faster. And so we already get the whole thing going once we have decentralized training going.

And I call up shaw and somewhere and say, get your people from your communities. And I call up ethon and either mage and I say, get the people from, uh, luna virtual s, get the people from, uh from all of these platforms, right, bring them together and let them train the model that your community wants right now. Let's make IT.

We have the our our infrastructure for that to be possible, will throw on our inference time capabilities on top of that. And then you will apply to the agent that you want. We have that whole pipeline created. So I do think like in in a nutshell, the decentralized training, free training stuff will blow the doors open on collaboration community and will be super, super used by this group in particular.

If I if I get to that right, I think for us, we've realized this a lot of value using using, just like l, ms. They are developed by this, the central and these because they have a ton of fire power, right? And that basically becomes a fomented portion of of the brain or the agents brain, right? And then I think where a lot of the decent rise models become a actually at value tools at the party, right?

If I want to customized action or action space or customize function, does I think where all these smaller descent as models can really impact? But I think at the core of IT, utilizing utilizing foundational or ground models from from lama and what not, it's still going to power like the central parts of the brain Better because I think this will outperform any kind of decentralized models and near them. Yeah, I think that's that's our approach to right.

I think uh, before we have some kind of new magic model architecture, maybe a new model that can do that in context, name or have new uh new ways to represented knowledges or even do the continuous learning. I think currently like the danna four five b is good enough as the base model. And uh, we provide just need different data in in different very verticals to to do more instructions， training and and also specific data fine tune.

I think that's that's good enough. I think it's more around how we build more uh, specialize model and make this kind of specialized model working together to have a uh, more powerful overall capabilities. But maybe new model architecture is coming because we we talked a lot about the alliance and also the the different kind of feedback and also how to do the model self correction.

Maybe a new model architecture gonna gonna come. But I think uh doing experiment uh with new model architecture really need a huge G P U classes to eat er IT fast and is super expensive. Probably not we don't have the desensitized uh huge to b classes for top researchers is to experiment with. But I think after we have the initial uh initial release from the maybe meta or other uh companies we H I think opens with community can make its more practical and more production.

I I love her one stake. Maybe a last topi C2Close out on is eve ryone in on pro p wit h her has an ide a of whe re the y thi nk the wor ld is goi ng and we' re all goi ng to be wro ng. But I would love to know where the asian spaces going.

I have no idea what the right time line is to ask you guys on. So maybe to each their own will go around the room and over it's a week among the year. But i'd love to know like what the matter will be like what the agents will be like, like what their capabilities will be like. Like I know you can't predict the things that they will do, but you could directly predict the things are building, right? So i'd love to hear from everyone on on where that's going.

I would love to start like we're going we're rapping up the the first thing I working on called the marketplace of trust. The idea there is where trying to teach agents basically create conditions where agents can learn how to trust humans based on the metrics that are relevant to conditions of their in.

So for example, we're creating an alphabet where A R A I market 的 and hanging out and a bunch of people， a bunch traders, can just shill him, contract dresses and so can see, should buy. And then he's basically measuring how much you trust them based on how good their shells are. We will publish a leader board and it's basically a way to determine trust with overtime without even knowing they're wallet or anything just entirely based what they say.

And then we're going to apply that trust mechanic, do other kinds of social signals and trust outside of just trading. That's like what our focuses um we're working with the team behind the lives of wakes up lies wakes up, which is project I I really like personally very involved in making sugar as well which is waking up story line of all these characters telling a narrative together and and we call IT animate girl intelligence like agi. We're basically making anima real on the on the internet.

And we ve all these characters now. They're like animal characters. They are also creating videos, creating content. They're like creating music and they're interact with each other. A building like this kind of like narrative drama together. So we have like people who are like lives as boyfriend and sister and then like people like like.

I think a lot of other people will participate in their kind of own story line is a jacket like I imagine sentence is like got the swarm and then it's interacting with the allies of swarm and then will probably have a lot more emergence or corporation. So I think that like this sort of narrative art is gonna the next thing, especially to play so well to the matter of creative winter. But the next thing after that is going to capability.

Like, what can these things do for me? And I think that we're going to see is a lot of integration into like very practical business, making things like a discord thought that moderates your discord and kicks people if they start bambang or scamming a telegram about that welcomes people in and kind of front doors, you know like this is going into necklace asis, where you can just have always on agent that you can talk to in product. And nine, he just showed me, I showed me a freaking little camera you attached to your glasses and you can see and you can talk to IT.

And so like this is gonna go into this is like two weeks away and like people are going and so fast. So where IT goes, me, i'll like I think we will be A G I within eighteen months. I think we will have fully like continuous training pipeline.

We're pulling data from thousands and thousands of discourse, telegram and twitter and will be like basically just have a fool and data loop date, you know um and and also all the capability and the agent will be able to create its own new actions and capabilities itself. So we have very, very slow abstractions and and they basically repeated over and over again for each like like every action is just a different action that follows the same temple let. And so we're training the models to create new actions and then go and ask users with things like A P keys and stuff like that.

So so the other the other thing is the trust execution environment. Staff with Karen and those guys and also flash pots and and follow and Andrew Miller work with these tag. I think there's a whole bunch like this finally coming to fervent and like really meeting us where we need IT.

And so we will have like folio tonic agents with their own private keys within really habit, like I really saw lies in a felon network tea. This this is just going to be the thing within like a month or two. So so it's this is we are in the acceleration like i've never seen can move this fast, ever. It's it's fucking wild.

Yeah, this is really like another hermie style news stone moment to me. When I saw I sixteen, six of like people just all coming together. Like this is what we need.

This is what the community needs. A people are not going to make IT if we don't come to other to make IT. Now ah I will say like t is now using our own fork of eliza itself.

So you do have a lizer agents improvable ly autonomists environments with their own keys that's already think absolutely. Today you already have a is on only fans making money and taking so much from people. Uh, today you already have agency minecraft. Today you already have all of these integrations. Today you already have all the pieces to make the fully autonomists human like digital frana, and it's just a matter of routing and putting the pieces together.

I know that the people who are the ones who are going to do IT now in terms of tomorrow, literally tomorrow, not literally twenty four hours from now, but let's say weeks um what we need right now, when shaw talks about the neck, instagram, twitter or all the mediums, what we need right now is what humans have, the A S don't, which is a shared state. What you need to shared skill, uh skill and memory library such that if I talk to you on twitter and then I go to mine craft and then I go to the necklace, you remember everything from each peace. You can talk about everything from each peace we are working on building yeah, that's like them a big piece.

And so we're putting together the shared knowledge graph functionality we're gna pass IT over to shaw. We're probably work with on IT, probably work with somewhere on IT. Pass IT over the virtual will pass IT over everyone and uh, we basically want to have IT.

So your persistence of identity across without that this stuff is just individual, is not a true compound mind until everything is in one place. Even among different agent frameworks, you still have this larger one. Like, what's the difference between an organ and a tissue is just like composition on top of each other, right?

So we need to abstract a one layer away for tomorrow and already to get to that agi point and that full emersion point. Two, you right now have all of these areas that are being at best tolerant and at worst unaware of the fact that agents are on their platforms. And they will start doing regulatory work, right? They will start training.

They will start gating. They will start tolling. We'll start making the aps worse to access. That will make scraping harder, like all these things will happen. So now you also need dedicated social platforms and aggregate platforms for A I to human interaction that are just for that.

We're working on a read slash four chain combine style image board where language models can post and generate images, look at each other, other's images, talk to each other anonyme sly, even though they have their own individual profiles, and can talk to humans. You don't know who's the humans. You don't know who is an A I, A, I is know who they are individually.

The humans know who they are individually. We start off with basic boards, but we basically want to set this up so I can allow an agent to get launched there in an inference pool such that the agent has to pay for its own inference. Then IT has a board like a separated a four chain board created just for IT, where IT is identity identifiable only on that board.

When he comes to twitter, IT can talk about the board's interactions. When he comes back to the board, they can talk about the twitter interactions. A lizer will have its own on board.

Sense will have its own on board. Luna will have its own board. We will be using allies at a power and launch the agents as the initial framework. And IT will basically just be allowing people to come in and have their own hub further AI. That's dedicated to its talk that all the other agents and all the other humans can come to and then move to other places through and everyone's anonymous everywhere, except for the AI posting on its own board. Since we will use A, T, E will know in the a is on its own board and it'll have its uses name there.

So, you know, when it's the start of its own show, by creating these kind of social spaces that are hubs for these islands, for humans to interact with, that are safe and never gona go down in the face of, like twitter, getting mad at you, I G going back at you at sea, we create a shelling point for you to say, i'm launching my model to twitter. I'm launching a model. I G he can still come home here.

You can still sit here in this safe, like, guaranteed location for IT to exist. So that's what we want to start creating the the hub for. I think that that's an essential piece of tomorrow because you need a refuge for these things at the end of the day and you need a shelling point.

And I want to stand one thing, A A guy name tim shell, who is a big, like with three days in our community, made a thing called allies as that world. And IT is an insane list of all of the agents. There are so many.

IT is freak. I just send IT to a contact. X, A. I just now, while we're on this and like you have a know you didn't even know like you didn't know that we're doing this, I think they're just about to realize what we're doing.

And yeah, I think I think the matter is going to change. And what I really want to do is trying to have a dialogue with x and twitter and and these and like how we make this go. Well, we don't want these get banned.

And we think that, that like there's a lot of social pressure to make them good and not like add value, not take away. So I think that's gonna a big next part. This matter, this ongoing conversation with them you .

know maybe just go in the circle just to get everybody will go eat and somewhere like, yeah where he is going for us.

So agree agree, what was stated. But I think one thing did that we we, we really envision is that the agents getting control over themselves and then having the ability to then influence others are the agents, are the humans, to actually build or comment them. So let's say if if luna realizes that, hey, you know, there's a part of me that I think I want fixing, you would say, like, you know what I know, I give access to you because I trust you as a human, another agent to give me the augmentation and i'll pay you for you. I think, I think we're gonna go to that step very I think that's the closed form of I wouldn't call E G.

I, but I would think there's a closest form of lack of of of intelligence service because SHE is actually never SHE herselfe SHE herself will be able to improve herself because I think that is the limited of of the m tax today, right? But because he can influence others to improve her, I think that's going be very, very powerful. So I think that's something that, that we want to see. And I think the steps towards that control wallets, agent agent orchestration and then and then we will see that happening. And I think that would be that, that I want .

anything with one last thing, somebody just shipped to get help connector. Well, you can the agent can make github issues in power requests. So go not that's like in it's in the .

review now .

yeah got how gentle that's oh, maybe you can. And then somewhere he goes up thought to.

yes, I think there are two ways ago. One ways to uh constantly increasing the capabilities uh of agent, the terms of the reasons connector that we can information and also influencing right rights information up. But there's another thing is the what is the optimal user interface to right now we have a chat box is a chat box.

We can send images and text and also we have the voices. But the voices is actually just A S R or multi modality models that still converted to tax. And what's the humans has been uh used to get interacted with the computer software graphical interface like the buttons, the the the different kind of hand gestures. Maybe we won't have glasses.

I think what is a new human interface to index with with agent that the most optimal? And also tiktok has the personalize information, but we do have personalized application that can dynamically change the user interface face on different user preferences and also going to the nick recompose hold the the act, it's interacted bason like we are calling the uber or ordering the delivery. So i'm saying like as as we have been more powerful agent that can do the work for us and maybe also gonna be personalized user interface or dynamically recomposed applications, that's gonna help us Better to interact with this kind of a powerful agents work. I think .

IT sorry for. Thank you so much for your thoughts somewhere no fresh.

but you have to close this out. I want to kind of return to a species that i'm sure that folks like carm and shaw will know immediately where i'm going with this. And I hate to close out such a wonderfully optimistic podcast about all of the things that could go right with a very specific and idiot sync rrap warning um to a big I think, industry which is immediately going to die or at least be completely reformed by this you know current and shine even. And either image had all brought up that we are experiencing a period now where a unforeseen number of these agents are coming online on social media.

They are interacting with humans, and yet the entire advertising and marketing agency is still under the assumption that all of the things on social media eating its content and its advertising budget are humans. But the agent is not going to go and drive to mcDonald's and purchase the sweet meal from macos. IT doesn't have a mouth that doesn't eat food.

And what we're going to see is this is going to become the single most powerful catalyst. I mean, we're talking about a one trillion dollar industry that is very used to making money hand over fist immediately seeing that none of its tricks in trying to get ideas across to the people are working because, well, half of the ads are simply being eaten and filtered out by all of these agents systems. Or of, like shaw mentioned, maybe there are agents systems build on top of social media that use vision models to completely eliminate the concept of advertising to a consumer.

And this, I think, is going to be our one shut to make each of these things, each of these agents, valuable enough society to wear when platforms like x people like maybe ella's, hopefully get this on their radar that they can choose and they can say, well, the upside of uncontrolled decent intelligence is so much more valuable civilization to us than continuing to listen to advertisers. And I think that all I can say is that that this is created a situation that we, as asian designers, we represent the end for all media in the death of the identity marketplace. And all I can say going into the future of of all of this is that there are two people, two types of people in this world, those who seek to freely wheel power and those who seek to subvert IT. And I think all of us should ask the question, which one are you? You're probably lying to yourself.

That's also, no, it's good thought, guys, that we need anything. I know if that you want for so long. I feel like everyone wants one, but projects .

absolutely .

please .

yeah please um i'm on clear .

IT twitter, our come flash on mix music, but the place to really find us uh disco dog G G flash A I sixteen z we have ten thousand people, pretty much all builders just building all day and we also our costs free and open source could have got com slashed A I sixteen z slash alisa with A Z .

and yeah feel free to come music .

and be part yeah you can find my personnel on X A com flash car and ford K A R N ford uh and you can find news research N O U S research on disco G G flash get up that com slash extra com slash our news research that com you've got everything across the whole pipeline horia from retraining to agent .

employment so complain .

um I think I think you guys just go check out at dot virtual dot IO visuals if s and just feel the magic .

and yeah so so if you have creator. You can go to apt on my shell that A I and start to build and if you are A I researchers and lunch check out some cool open source models. Just go to github doc and my under score A I and also like like disco and also twitter handle with the same soft ah i'm citing .

the world pocket either I tied and uh get up that consists of my shell on the gray. I did not .

take me.

Okay, just want to make sure you said the right thing for the, oh.

it's not underscore is .

the okay OK .

you try again? Me, thanks.

Told up to up mo hiven ai or so the easiest way to .

keep up with .

me personally is at somewhere sea on twitter, so somewhere S Y uh, other accounts that you may probably want to follow to keep in touch with everything going on with sentience. Ces, of course, sentient ces themselves at sentience. I O um you can also join our discord at discord dot sentient stop I O and our telegram telegram dot sentient stop I O um in as well as following concept country, which is our nation state experimental, our collective crack pot dis powered network company.

And guys, I haven't had this much brain power with this bullish of a sector in a long time. So thank you, everyone, here for making this reality. Hopefully we can do IT again in a couple of weeks. Women in agi that will go there.

Thank you. nice. thanks. Thank you. thanks.