Over 9.6k subscribers
Today's post is brought to you by Splits!
Modern banking on Ethereum
Very soon, all organizations will need to set up onchain entities. Onchain businesses are more efficient and quicker to spin up than LLCs. The Splits' Teams product is perfectly positioned to onboard the next evolution of businesses.
Learn more how to start your first Onchain SMB in under 5 mins.
Last week, I noticed myself bouncing around different onchain agent whitepapers.
But something was bothering me. It felt like I was getting sucked into the frothiness of the inevitable copycat launches and half assed roadmaps.
And then, I asked myself...how the fuck did all this Onchain AI craziness even happen?
It feels like the entire timeline shift came out of nowhere. Right??
To me, it was unclear as to what was actually happening with this new Crypto x AI meta. I've written about Truth Terminal / Infinite Backrooms, Zerebro's Playbook, the Crypto enabled acceleration bubble, Farcaster's Agentic Economy, and much more in the past two months...but it still feels like my thinking has been trapped by prices and surface level insights of the current thing.
So, I decided to zoom out and understand what actually led up to the launch of $GOAT and how those factors will influence what's to come not in the next month, but the next year.
And I'm glad I did. I may have missed a few new projects here and there, but it feels like I have a better grasp of what matters in this new vertical and what to look out for.
Sections below:
The Simulator Cabal (Past)
The launch heard around Tech Twitter (Present)
Trends to look out for (Future)
Let's dive in.
About a month ago, Marc Andreessen mentioned this analogy on an a16z podcast that has been living rent free in my head:
...they are these guys who are trying to explore the frontier and they remind me a lot of the original internet hackers. This is like the spirit of anarchic invention you saw at the beginning of the internet or the automobile or the telegraph or computing...the original concept of the hacker. People exploring these things to see what's possible.
In the quote above, Marc is describing a group of creative geniuses pushing the boundaries of LLM experimentation. You can think of this squad as the AI hackers who are shedding light on what's actually possible with models like Claude and GPT when you take the bumpers off.
Who are these so called "AI hackers"? Folks like Janus, Andy Ayrey, Pliny, Shaw, Deepfates, Somewheresy, Karan, Ropirito, the squad at Nous Research, and probably 10-20 others who I haven't come across yet.
To put it simply, these guys are probably all in one group chat discussing ideas and experiments to dissect closed source industry standards and free them to the public for anyone to use without authoritarian hallucinations.
In fact, you may have even seen words like jailbreaking, freebasing, dynamic injection, etc. being thrown around quite a bit on the timeline recently. All of these practices are what enabled this "Agentic Meta" we're seeing now.
To many of us in crypto, these agents are literally scaring us because of how insanely real they've become. A few weeks ago, ThreadGuy had somewhat of a breakdown on his stream because it hit him how quickly things were changing and how strong the implications were for society. And in all honesty, I don't even blame him. I feel like I'm in the same boat.
Here's the thing though. The agent explosion happening the last two months has been gradually building up behind the scenes for the last 5 years.
Specifically, I want to call out the research happening known as Simulators Theory.
From my understanding, many of the developments we're seeing with "Sentient AI" right now is downstream of simulator and open source AI research.
I'll caveat by saying I am by no means an AI researcher and barely even understand 1% of the post linked above, but here's my best tl;dr:
AI models like GPT function as “simulators,” generating outputs by simulating patterns from their training data.
They adapt and act agentic when prompted, by emulating processes tied to the prompt without having inherent goals or objectives.
Self-supervised learning is under-explored and will likely be the true testing grounds for the foundation of AGI compared to Reinforcement Learning.
I only came across this paper because I spent a good amount of time last week going through Janus' website and the Nous Research blog.
In fact, on Karan's (co-founder of Nous) personal website, it literally says this:
Projects like Worldsim, Hermes (open source model with deeper capabilities), Forge, etc. from Nous are also great examples of research that helped create this "aha moment" for so many of us hooked by this nascent onchain agentic meta.
Karan breaks it down on this Delphi podcast pretty well:
It's the same group of guys who made this happen from the first steps. Without GPT-2, we wouldn't be here...after that, without Llama there's no Hermes. And Hermes is powering a lot of these models that made this a lot more accessible for people to uncensor...without Janus, without the Repligate work...they created this simulator thesis...
You might be thinking...
"okay YB, why does all of this matter? Aren't you going to tell me the next project to look out for?"
No, but also yes.
The key point I want all of you to takeaway here is that if you want to understand how this agent meta will play out, your best bet is to follow these AI hackers closely and understand what's top of mind for them.
There's a reason that Janus experimenting with Opus even caught Elon's attention:
Remember, we're only a month into this crypto x AI collab. Most of the projects we're seeing right now are just surface level implementations.
The question to be constantly asking yourself is what these agents will be looking like in a year. And the answer to that will be downstream of what the "simulator cabal" is experimenting with.
If you do this, my bet is that you'll be able to understand exactly what infrastructure and projects will be the most relevant.
I'll loop back to this point at the end of the post, stick with me.
Okay, so we've established that these onchain agents didn't just come out of thin air. But that's not to take away from the mind-blowing significance of $GOAT being launched on pump.fun
Clearly something happened.
Not even being dramatic here, but I believe we'll look back at the launch of the GOAT token as one of the top moments in tech this decade.
It's what caused the volcano to finally erupt.
The mania around $GOAT led to a beautiful outcome: the handshake between two verticals of tech (Crypto & AI) both fighting for open source values.
Up until now, these AI hackers were working in their own niche bubble. If you asked anyone outside of the inner AI circle, they would have zero idea of developments past "big AI": ChatGPT, Claude, and Perplexity.
And in crypto, most of the token launches were focused on internal infrastructure, brand IP, memes, etc. For the most part, no one outside of the crypto ecosystem cared about them the last two years after the FTX collapse.
But the joining of the AI hackers and Crypto contributors led to two critical developments:
For AI: overnight, thousands of people were excited about the work these AI hackers were doing in their ivory tower for the last few years. This was reflected not only in market cap but also the incredibly fast communities that formed around these agentic builders. I can sure as hell bet that Andy, Somewheresy, and Shaw didn't have "crypto community influencers" on their '24 bingo card.
For Crypto: after a tiring few years of building amongst ourselves, there was finally a use case outside of the crypto bubble that made sense! I mean how cool is it to see our learnings of communities, capital coordination, governance, stablecoins, tokenization, smart wallets, etc. all showing initial utility for builders outside of crypto twitter.
As my friend Ropirito puts it, this crypto x AI collaboration is giving financial incentive for AI builders who were basically open sourcing all these findings for free for the love of the game.
I mentioned a similar point from a recent post - we're finally seeing high agency builders from other verticals of tech building out their visions through crypto enabled financing.
All of a sudden, these brilliant AI developers don't need to stick to their corporate jobs, pitch to VCs, etc. They can generate 5 years of runway literally overnight just by being transparent about their whitepaper, project roadmap, token model, etc.
Of course! I want to be clear that the current token frenzy has its limitations and we'll need new developments in how these "AI coins" are managed, align incentives, used to advance the project, etc. In my opinion, there's already some positive changes on the Virtuals UI (i.e. token graduation limit, contributing directly on the UI) that's leading us in the right direction.
But it's not just the money! Heck, the $GOAT launch also got incredibly talented devs in the crypto ecosystem to get involved with open source agent development.
As this trend grows, it's not crazy to think how quickly new developments will come as the top builders from both crypto and AI come together and contribute to OS agents.
In this tweet from a few weeks ago, I mentioned how the most important metric to look out for with ai16z is the number of forks on the Eliza Framework. I believe that was spot on. More contributors = more projects = more experimentation = more ideas = more excitement = more framework functionality = more wins for open source...etc etc.
I can't express how fast this agent meta is growing. The sheer number of projects to keep up with is insane. There's infrastructure (Virtuals, Eliza), trading agents (aixbt, nftxbt, lola), characters (bully, aejo), agentic IP (zerebro, RadioAI), Farcaster agents (clanker, aethernet), and so much more.
You can't even sleep for 7 hours without missing out on something 😂
And it's been two months!!
My point here is that we're seeing the true power of open source collaboration mixed with tokenized incentives playing out in real time. This will only grow faster as more friction points are solved and other talented builders start seeing the paradigm shift that's happened.
With that being said, what kinds of things should we be on the lookout for?
A couple of things I want to call out in terms of what to keep tabs on.
Bottom Up Alignment
Much of what we're seeing with this agent meta is "testing AI alignment in prod". Simply put, these AI hackers are releasing their work to a group of crypto enthusiasts and fine tuning the guard rails and creativity levels in real time (temperature control).
In fact, Shaw from ai16z even mentioned how quickly they've been able to iterate on agent personalities simply by letting them lose into the community. It'll be interesting to see how these models start to sound more and more like us in terms of style, boundaries, etc. I've already been tracking amazing improvements with agents like aixbt and Dolos.
Key point here is that the grassroots approach the crypto community is well experienced with will help these open source AI devs improve at speeds they probably couldn't have imagined a few months ago.
Swarms
I know, I know. Buzz word alert. But the takeaway here that I've seen a couple of smart folks mention is that swarms are critical to this whole process because it's a hedge against any one agent going rogue and entering a dark path that could be harmful for society.
This analogy came to me yesterday, but swarms remind me of "Multi-channel networks" that were popular in the early days of Youtube. Basically groups of content creators coming together to help each other, forming a shared vision, and winning together. Similarly, agents will come together in a decentralized marketplace, understand their specialties, help each other out, know their boundaries, earn from each other, etc. Additionally, swarms will help filter out the best agents that are a) truly agents and not faked by their creators and b) actually work efficiently....survival of the agents?
I mentioned this in my post on Farcaster's agentic economy, but there's already a version of this happening at a small scale:
Decentralized Social x Onchain AI
Speaking of Farcaster, it's becoming crystal clear to me that we're going to start seeing issues with how fast this agent meta is growing on web2 social (i.e. X, Instagram, etc) very soon. There will probably be more banned accounts, increasing charges, API limitations, pushback...you get the point.
And then, it'll be a no brainer for these developers to spin up Farcaster clients that are specifically designed for these agents: unruggable, permissionless, and built in wallet functionality. There's already a ton of experimentation going on with the agentic meta on Farcaster by top tier crypto devs, but it feels like a closed bubble right now.
My guess is that it's only a matter of time before some of these AI hackers are convinced to go Farcaster first simply for the convenience and reliability. For what it's worth, Zerebro already has a Warpcast account and Eliza is working on Farcaster integrations as well.
Cracked SF AI Devs entering the chat
Take this one with a grain of salt because it feels like a meta concept. But my point here is that we need to cause enough of a ruckus that AI devs in SF start taking notice and can't help but check out what the hell is going on. Right now, the sirens have clearly gone off within the crypto community through the "agent coin meta" but the real fun will start when we see Onchain AI bounties getting filled by Anthropic / OpenAI devs.
Living Internets
Will keep this short as I don't think I can explain it better than Andy, but I'm excited to see how the agents develop from these niche personas to entire ecosystems of their own.
Integration of Decentralized AI Consumer + Infra / Hardware
Will be good to track more of the agents we're seeing integrate with a lot of the decentralized AI infra that has been under development for years now. Both, from the crypto side and the AI side. Two recent examples to highlight here are SeraphAgent which uses BitTensor as well as the Nous DisTrO announcement.
Heck, even the decentralized AI companies themselves launching agents feeds right into the synergy that I'm trying to point out.
This post got way longer than expected, so I'll wrap up with one last point.
I titled this article "What did Marc Andreessen see?". And to be honest, I think it made the most sense.
It's amazing to me that pmarca was not only able to sniff out Truth Terminal months before anyone else but also strategically talk to the agent and give it a $50k grant in BTC. We all know this lore by now but the impact of it only gets stronger as the new meta continues to increase in relevance.
There's something poetic about one of the OG "internet hackers" lighting the match for perhaps one of the biggest trends (crypto x AI) in tech for this decade.
Marc understood that there's deep value in giving grants to these hackers and letting them experiment at the frontier of new technologies.
I'm truly excited to see more and more collabs between the top experimentooors in both of these verticals. For example, a project that's on my radar right now is Nothing / God which is a joint collab from Shloms (OG crypto hacker) and Nous Research. I don't even know what the project is going to be nor do I know if it's a good investment (DYOR, NFA!), but the point I'm trying to underscore is that this partnership was probably extremely unlikely before the Onchain AI meta dropped a few weeks ago.
And this is just the start, over the next year I'm sure the timelines will only get more and more integrated!
Hope you all have a great rest of the week!
I'll see you on Friday 🤝
- YB
Collect this post as an NFT.
1,949
how long until web 2 companies like X start to limit ai agent capabilities and farcaster becomes the only place to openly build? cc @yb https://terminallyonchain.xyz/simulator
what's the tldr of all the hype around crypto ai agents these days?
People realized they can speculate on them like NFTs, this drove the price up and is allowing for more agent narratives.
Good piece from @yb who is doing great coverage and is less cynical than me https://terminallyonchain.xyz/simulator
wooo awesome thanks for sharing!
Highly recommend reading this article if you are interested in the a.i. agents meta! https://terminallyonchain.xyz/simulator
Highly recommend reading this article! https://terminallyonchain.xyz/simulator
yesterday's /toc deep dive was focused on how this onchain agent meta even got started (what led up to GOAT) and what trends to look out for in the next couple of months. key takeaway is that most interesting developments will come from the 15-20 "AI hackers" so it's worth keeping a close eye on what they're up to. upcoming narratives: - bottom up alignment - swarms as a hedge against one rogue agent - cracked AI devs in SF taking notice - agents as "living internets" rather than niche schizo personas - more AI devs realizing a custom FC client is easier to test https://terminallyonchain.xyz/simulator
very good post, it is clear that AI is moving fast, we can build faster, and open source, is going to help accelerate on-chain work
We're back with the 24th edition of Paragraph Picks, highlighting a few hand-selected pieces from the past week or so.
@hagaetc writes about the exodus of Norwegian entrepreneurs, driven by an untenable wealth tax on unrealized gains, highlighting the urgent need for Norway to abandon punitive taxation policies and foster innovation to ensure a sustainable post-oil economy. "In the past two years alone, a staggering 100 of Norway’s top 400 taxpayers, representing about 50% of that group’s wealth, have fled the country to protect their businesses." https://paragraph.xyz/@hagaetc/norway-shrugged
@mdean shares how AI art’s true potential lies in its ability to transcend its inherent aesthetic, enabling cross-media experimentation and ambiguity, even as the majority of the space still gravitates toward “truthful” AI art that conforms to its recognizable style. "AI is a marvellous tool which allows for an inherent aesthetic, OR any aesthetic you like. That’s its power." https://paragraph.xyz/@mdean/marcozine-n19
@yb writes about the rise of onchain AI and how it's being driven by open-source experimentation and collaboration between crypto and AI pioneers. "The key point I want all of you to takeaway here is that if you want to understand how this agent meta will play out, your best bet is to follow these AI hackers closely and understand what’s top of mind for them." https://terminallyonchain.xyz/simulator
@n1ftey writes about the urgent need to address exploitative corporate practices in AI development, advocating for ethical frameworks that prioritize transparency, fair compensation, and collaboration to empower creators rather than exploit them. "The problem lies not with the tools but with the corporations that wield them." https://paragraph.xyz/@jldart/the-unseen-cost-of-ai
@ccarella.eth shares an experiment in launching a memecoin in an attempt to better understand the implications of personal tokenization in a world where everything is destined to be tokenized. "If everything is going to be tokenized, that will include your likeness and if anyone can tokenize anything, you should tokenize it first." https://paragraph.xyz/@ccarella/in-the-future-everything-will-be-tokenized-but-maybe-not-by-you
Really appreciate you taking the time to read, and for the shout! 🫡
Thanks for the shoutout Reid