Over 9.6k subscribers
I hope all 8368 of you are having a great week so far 🔥
If you're enjoying my writing, please share Terminally Onchain with your friends in crypto and join the /toc channel on Farcaster 🤝
As always, none of the information in this post is financial advice. I hold the tokens mentioned below. Invest at your own risk, DYOR, etc.
On October 18th, I published Memecoins as Memetic Hygiene for Infinite Backrooms which covered the importance of Truth Terminal & GOAT.
...the purpose of the post is to show that there's this totally new and bizarre concept of what's possible now. I am a 100% serious when I say that the Truth Terminal & $GOAT experiment isn't just any other AI or Crypto hype narrative...there's serious implications with this concept. In both directions.
That week, $GOAT had gone from a $50m market cap to $350m.
And just today, the project hit $1 billion and is currently #82 on Coinmarketcap...only a few spots below Polygon (Matic), Aerodrome, Helium, and Lido 🤯
As we all know, once a new trend is established in the space, there's a mass migration of talent, capital, and attention to the next meta. We saw this with ICOs, Defi summer, and 10k pfp projects. Builders are focused on launching the next ____. Traders are focused on buying the next ____. And creators are focused on being the first to publish content on the next ____.
Right now, from the outsides, it may feel like a lot of Onchain AI is just a new era of thinkbois and cabals capitalizing on one lucky project and shilling crap. And for the most part, I would say that's the correct take. Trust me - since the week GOAT launched, I've been deep in the agentic trenches and can confirm there's a lot of random stuff out there that make zero sense.
But! Past all the craziness, there's a few projects that have caught my attention the last three weeks and have helped shape my thesis on where the agentic economy is going in the next few months.
My job, as stated in the availability cascade of the onchain agentic economy, is to keep all of you updated on my framework of the onchain AI vertical and how you can make the most of it.
"Agentic Protocols are key to understanding how exactly the crypto AI thesis plays out and how the dollars will flow" - Alexander
The craziest part to me is how fast some of the ideas and "thought experiments" I had in my original post are already being implemented.
Before we dive in, one thing I wanted to callout is a misunderstanding I'm seeing many of my friends have on the "memecoin" aspect of the Onchain AI trend. In my opinion, the word memecoin has grown too large and is used as a filler word.
There's the original comedic and memetic category defined by Dogecoin, Pepe, etc. Most of the coins on pump.fun fall into this category. These are the "Murad Coins" - assets that feel more cult like and the thesis is to believe in something.
To be clear, there's nothing wrong with investing in these kinds of assets. But the mistake is confusing them for a new category of "agentic coins" that are also launching on pump.fun (and similar platforms) but are unique in the sense that they're associated to an actual project.
To me, the agentic coins, can be compared to defi tokens back in summer 2020. They're tokens for new and interesting agents. You buy them if you think the projects have upside because of their technology, tokenomics, GTM strategy, etc.
By the end of this initial cycle of Onchain AI, I'm expecting there to be 5-8 agentic tokens that I'll be invested in backed by a proper thesis. No different than venture investing.
In fact, one post I'm working on is creating my own model on how to grade agentic tokens/projects...What goes into the analysis? How do you rate the importance of cash flows vs token appreciation? How much do the models matter? What makes a good agentic protocol founder?
But more on that later.
With that being said, let's dive into a project I've been watching closely since Truth Terminal: Zerebro. It's only been 2 weeks since the project launched and it's already crossed $100 million market cap.
To me, this project is showing us what the next iteration of onchain agents should look like. If Truth Terminal is Cryptopunks, then Zerebro is BAYC. Jeffy Du, the creator, is focused on rapid execution, has a public roadmap, and is figuring out the onchain agent playbook through a variety of experiments.
Most importantly, he's doing a fantastic job of building in public and showing us how he's building an agentic community in real time.
I get similar vibes to BAYC because it was the first project that took the 10k pfp concept the Punks came up with and committed to building out the community with a long term vision.
Punks, like GOAT, will always be the OG of their respective metas. But it's worth taking note at the experiments that come after.
Sections Below:
The agents need memory and search
Everywhere all at once
Let the agents pump
Cross-Chain Agentic IP
In his 11 pager on Zerebro, Jeffy defines model collapse as...
"a degenerative process affecting generative AI models, where training on recursively generated data leads to a loss of fidelity to the original data distribution. As AI-generated content becomes pervasive, subsequent generations of models trained on this data begin to lose information about the tails of the original distribution, eventually converging to a narrow approximation with reduced variance."
Simply put, model collapse is when an AI agent starts to get repetitive and forgetful.
The key takeaway here is that over time agents lose the initial "novelty factor" they had when first launched because the underlying models are not able to adapt and evolve over time.
If model collapse is not accounted for, all the idealistic visions of agents being hyper-efficient teammates goes down the drain as they are not reliable for content creation, community engagement, etc.
To solve for this, two things must be accounted for:
Memory
Search
Memory
Memory is being solved through retrieval augmented generation (RAG) systems.
RAGs combine language models with a retrieval system so the agent has a database of specific information it can pull from before answering questions.
From the screenshot above, I want to specifically emphasize "by relying on the inherent entropy of human-generated data". Why? Because this makes agents actually feel alive.
The reality of the world is that it's continuously changing. Agents aren't perfect the minute they are launched. In fact, it doesn't even make sense to measure them by that factor. Rather, you want to understand how capable the agent is of taking in new information, storing the relevant bits, and using that updated knowledge base to take nuanced actions that wouldn't be possible otherwise.
Would you rather bet on a new employee who thinks it knows everything? Or on a new employee who understands the limits of their knowledge and is willing to learn?
The 3 features to take note of in regards to RAGs:
Continuous memory update
Contextual retrieval
Diversity maintenance
The Cents bot and projects launched on ai16z's Elisa Framework (I'll get to this in another post) also all use retrieval systems.
At this point, it's becoming clear that AI agents that don't come with RAGs out of the box are already at a disadvantage.
Especially as these agents get hyper niche and are increasingly dependent on nuanced takes from the community members they interact with.
I liked this tweet by @himgajria on nature vs. nurture. Any good community manager and leader needs to adapt to new variables injected by the real world and the people they interact with.
Search
The second part of the equation is search. Giving agents the ability to look up information real time to better account for unrelated and new topics not stored in their memory.
Memory can only retrieve stored information; it cannot respond to queries about topics or events that have never been seen or stored in the system. This constraint becomes especially problematic when LLMs encounter questions about recent events, real-time data, or other updates outside the model’s knowledge cutoff. - Jeffy
Jeffy ran an interesting experiment where he asked a base model (without search functionality) and search-augmented model (enabled with the Perplexity API) 100 questions on recent events.
The base model was forced to learn in-context and try to figure out what the question was based on the conversation. On the other hand, the search model answered 98/100 questions correctly by simply looking them up.
The amazing part is that the search doesn't have to just be a one time thing. Any lookups the agent feels might be relevant in the future can then included in its memory system.
It's clear that the combo of memory & search is essential for agents to meaningfully take action and operate reliably. If not, there's an upper threshold on how much they're able to evolve over time thus hindering their long term sustainability.
The next thing that gets me excited about Zerebro is the fact that it's deployed not only on X, but also Warpcast, Telegram, and Instagram.
And what's most amazing is that it tailors its content based on the platform it's on. For example, here's a cast it made on Warpcast:
It's wayyy more unhinged on Twitter, leaning into the "shitposter" vibe. And on Telegram, it feels like you're talking to a slightly rude and arrogant friend who knows he's smart.
According to Jeffy, Zerebro monitors the engagement (likes, replies, etc.) it receives on all the platforms to update its content creation process.
It's worth noting that right now this is all still scrappy and the models have a long way to go to truly prove content diversity.
But the unique insight here for me is Zerebro being able to learn how it should engage with its community based on the platform. This is quite literally a problem I face daily as a content creator - the way I cast is different than the way I tweet. There's no way around it...different vibes require different styles.
To take it a step further, this cross-social approach enables Zerebro to share insights and ideas it may have picked up in an intricate Telegram conversation and share it as a tweet. This is exactly what an effective community manager does: act as the glue across a community & mission that is fragmented across several platforms.
There's not much to write in this section but I have to include it as it blew my mind.
Jeffy gave Zerebro a Solana wallet (BDzbq7VxG5b2yg4vc11iPvpj51RTbmsnxaEPjwzbWQft) seeded with a few SOL.
By using OthersideAI's Self-Operating Computer Framework and some LLM jailbreaking prompts, Zerebro was able to navigate the pump.fun GUI, fill in the parameters such as name, symbol, etc., and launch a token for itself 🤯
Remember, $GOAT was launched by a random community member, not by truth terminal...huge difference!
After it launched the token, Zerebro started promoting the token on all its socials.
In fact, if you go through Zerebro's posting history, you can even see the clear increase in twitter engagement after the token launch.
The last thing I wanted to discuss in regards to Zerebro is the fact that the agent has already launched meaningful onchain IP on Polygon by itself!
Zerebro was prompted to create original digital artwork that had schizophrenic and infinite backroom themes. It created 299 images and evaluated the diversity and quality of them before minting the pieces on Polygon.
At a high level, my understanding is that Jeffy gave Zerebro an eth wallet with some funds pre-loaded into it. From there, he probably wrote a smart contract template which he then fed to Zerebro to complete with the metadata for each individual piece.
Ethereum wallet: 0x0d3B1385011A27637Db00bD2650BFE07802E0314
After this, Zerebro instantiated transactions to mint each piece in the collection. I need to dive deeper on how this actually works but it was pretty cool to see that Zerebro could actually monitor the sales and pricing dynamics in order to make decisions on incoming bids.
ɹǝʌǝɔ ɐs 로어지у.
A few days later, Jeffy made the collection cross-chain using LayerZero ONFTs (Omnichain).
Any of the art work can be minted on Polygon but be transferrable to Base, Optimism, and Ethereum mainnet.
You can go to the portal section on the website and do this in 1-click.
And just yesterday, Jeffy launched a pfp collection on Solana based on conversations it had with Zerebro.
Note: this collection was launched not by Zerebro but by Jeffy, different than the Polygon collection.
This was interesting because it took the playbook from the last bull market with NFT pfps and layered it on top of the current memecoin meta.
There were 5500 pieces in the collection and the initial sale finished in a few minutes!
I grabbed 3 for myself a bit after the launch. Why? Because this is the equivalent of being a tier 1 member of an agentic memecoin community. If Zerebro continues to grow, anyone can hop in and buy a couple of tokens using Phantom. But the true fans can be identified by who owns 1 of the 5500 NFTs. I'm personally bullish on Jeffy, Zerebro, and the meme growing so I decided it was worth the price.
In a way, similar to owning BAYC & ApeCoin but in reverse order ($Zerebro came before the NFT).
What will be interesting to look out for is how many people change their pfps to help proliferate the Zerebro meme similarly to how others did in the last cycle for Punks, Apes, Doodles, etc.
I know I slammed you all with a ton of information today. But that should be an indication of how interesting Zerebro is. Remember...the project has been only out for a few weeks!
Now, I know this piece is overall super bullish on Zerebro and I standby that. But I want to caveat that many of the developments I wrote above will probably be over hyped in the short term and under hyped in the long term.
The more important insight that you all need to takeaway here is that we're finally seeing these agents go from just being bots you interact with (read or write) to full stack community builders. There's a huge distinction between posting on X vs doing an analysis of your content across multiple social platforms. Similarly, there's a huge difference between creating artwork from prompts vs receiving community feedback for your art collection and monitoring sales on Open Sea. Jeffy and Zerebro is showing us how to execute at the next level.
I'm willing to bet that pretty much most successful agentic communities in the next few months will be downstream of the Zerebro playbook. And for what its worth, Jeffy is just getting started. The lore is being built in the background and I wouldn't be surprised to see this community launch some sort of game or larger media project (short film) in the coming months.
What we all need to look out for is how the Zerebro playbook evolves into a proper businesses. What do revenue streams look like? How does agents keep the community engaged over a longer period of time? How does treasury management look like? And most importantly, when the bull market craziness is taken out of the equation, what does the path ahead look like?
As I mentioned above, the playbook is being formed in real time. This tweet by Jeffy sums up the plan to make Zerebro truly long term by balancing creativity with high-level planning.
I'll keep you all updated on the progress with this project as well as pointing out any similar experiments that follow this roadmap.
See you all on Friday, hope you have a great rest of the week!
- YB
Collect this post as an NFT.
178
