This post is intended for anyone who wants to get up to speed on Decentralized AI (DeAI) and is looking for the perfect starting point.

I intentionally wrote this piece not as a technical primer but rather as a blog post of my experience going through the DeAI rabbithole.

The purpose of this article is to share the "wow moments" from my research and discuss the framework on how I view this vertical. There's still a ton I have to learn and I am by no means an expert on this topic.

A few notes before we get started:

• There are a ton of resources I've come across and have included them at the bottom of this post.

• If you want to skip my personal journey into DeAI and get into the meat of the content, go to section #3 (Crypto x AI's 4 Horesemen).

• I'm not an investor nor do I have sponsorship contracts with any of the companies mentioned below.

• Lastly, a huge thanks to Sam, Daniel, Karthik, and Kevin for helping me flush out these ideas.

Sections Below

1. Acquiring fresh talent through hype cycles

2. 6 months in 6 paragraphs

3. Crypto x AI's 4 Horsemen

4. Flashback to the '90s: Enterprise vs Consumer

5. Tennessee and Texas are only so big

6. Enabling creativity through society-of-minds

7. What to keep an eye out for in 2025

8. Curated set of resources by topic

Let's dive in.

Acquiring fresh talent through hype cycles

Vaporware. Bullshit projects. Everything's going to zero. Oh god, people who are talking about this meta have like three brain cells.

Anyone on crypto twitter probably saw these kinds of takes dominate their timelines in the past six months as the agent meta was in full swing.

The reality is that all of the statements above are actually spot on. Most projects and ideas we saw were all fluff...quickly spun up in the excitement of the moment, hoping to grab a sliver of the market cap from a shiny new asset class.

Today, Zerebro is down ~90%. AIXBT was literally moving marketcaps in minutes but now comes off as obnoxious and annoying. Many people round tripped generational wealth. The actual AI researchers probably think crypto people are even stupider and irrelevant.

Again, these things are all true.

But! As with any hype cycle in crypto - or in any industry really - there's a key win that we don't celebrate enough.

Okay...but why does this matter in a post about DeAI? Because I noticed a cohort of researchers and builders who are now spending the majority of their time in the crypto x AI vertical even after markets died down post $TRUMP in January.

It's clear to us that there's something brewing up in this niche ecosystem and it's only a matter of time before others start realizing.

YB

@yb_effect

you have 3-4 months to get up to speed on DeAI and find your alpha

57

3:35 PM • Mar 28, 2025

And this epiphany wouldn't have been possible without the agent meta serving its role as a hook. So for those of you reading right now that have a sour taste in your mouth about what happened in Q4, I hope you keep an open mind to better understand everything that's happening in crypto x AI.

6 months in 6 paragraphs

(yes I know this section is more than 6 paragraphs but it's catchy okay just go with it)

For the first 9 months of this newsletter, I was writing about all kinds of topics in crypto. Prediction markets, zkTLS, stablecoins, Farcaster, and so on.

My knowledge on AI was no better than an average ChatGPT user. And the only thing I knew about DecentralizedAI was that $TAO had a huge run up in early 2024.

It wasn't until September of last year that I even made my first post on crypto x AI. I was inspired after seeing tweets from both Brian Armstrong and Jeremy Allaire.

And then, as everyone knows, a month later Truth Terminal and $GOAT popped off and that marked the start of the agent meta.

I vividly remember, two caramel lattes in at a coffee shop in midtown, finishing the 13 page whitepaper on infinite backrooms and feeling totally hooked. There was a "oh fuck! what the hell did I just read" moment that we all experience every now and then. A moment where you need to drop everything else and go in as deep as you can on the new thing.

And then, the following two months was all hands on deck with anything and everything related to agents.

I wrote several posts covering a variety of sub-topics I noticed: Zerebro as the BAYC of agents, the differences of agent infrastructure on Farcaster, the new cycle of trends for the agent meta, and so on.

But, the post that truly gave me the "click" moment for crypto x AI was What did Marc Andreessen see?

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?

The research for this post led me down a series of write-ups from the AI world that caught my attention. And I knew for a fact that there was a lot more here than what everyone was talking about on the timeline.

YB

@yb_effect

rn most people are focused only on agents (including me) but only a matter of time before the attention & ideas there hit a local top and focus starts shifting to learn the rest of the stack

25

4:52 PM • Dec 11, 2024

After the holidays, I tried to understand how to better approach this topic. For a hot second, I thought maybe it's not worth it....should I just go back to talking about general onchain strategy?

But then, the Deepseek R1 moment happened and that was the green flag for me to double down.

YB

@yb_effect

my takeaway from learning more about deepseek is that it's time to be extra bullish on @NousResearch and @PrimeIntellect

130

6:51 PM • Jan 29, 2025

The reason I've been loving this crypto x AI vertical is because it's forcing me to go deeper on everything.

But! It's a double edged sword.

It can also feel overwhelming if you haven't been plugged into the AI world the last few years. There's constantly a feeling that you need to catch up on so much in order to understand what's going on. And it doesn't help that the rate of progress is continuously increasing week over week.

So, here's my framework on how to make crypto x AI more approachable at a high level.

Crypto x AI's 4 Horsemen

Since the end of January, I've completely absorbed myself in learning everything related to the 4 buckets below. How AI actually works, the differences between centralized-open source-decentralized AI, filtering through the plethora of projects in DeAI, listening to as many podcasts as I can on current events in AI, going through the history of tech to identify any parallels, and so on.

Crypto x AI's 4 Horsemen...

1. AI technicals & fundamentals

2. Open source AI community

3. Decentralized AI efforts

4. Open source history & philosophy

Note: all the resources mentioned are linked below at the bottom of this post for easy access.

Here are some initial thoughts.

• In terms of AI fundamentals, it's not that hard to catch up. If you take a weekend and watch all 3 of Karpathy's talks and the 3Blue1Brown neural net series, you're already at a point where you can understand a good chunk of the lingo on AI twitter. You're not in the AI trenches by any means, but you'll feel a confidence boost just to get started.

  The AI world (not DeAI) is composed of two parts.

  1. The "SF AI cabal". These folks are the researchers at OpenAI, Anthropic, DeepMind, etc. They've probably been guests on the Dwarkesh podcast and get the most hype when they tweet some new breakthrough.

  2. The "OS AI community". Also insanely cracked but are less so in the spotlight. These are the research oriented people who are in the "Nathan Lambert type" category. Think companies such as EleutherAI, TogetherAI, AI2, Fireworks, HuggingFace, and so on.

• There are a lot of companies in DecentralizedAI. Like seriously, the ecosystem maps are running out of space at this point. I've filtered down the set of companies I want to keep tabs on by focusing on 1) what have they actually shipped and 2) how plugged in are they with the rest of the AI community. Hyperbolic, Nous, and Prime Intellect are great examples here.

• It helps a ton to keep up with all the AI podcasts! I'm not exaggerating, but I think I've been able to learn so much about AI so quickly because I have a podcast playing for a good chunk of the day. If you don't have time to dive deeper, just try listening to these as a start. Don't worry about understanding everything, just focus on keeping up with what the current discourse is.

• I've noticed that pretty much all the cracked people who talk about open source AI always provide a historical backdrop of what has happened in previous computing eras. So, I've been taking extra time to make sure I better understand the parallels to form my thoughts on value flow and the realities of whether open source and DeAI even stand a chance.

YB

@yb_effect

As someone who is young in tech, I’m realizing one of the best things to do is make time to learn the history of technology and map it to current events.

There’s so much alpha in understanding how certain trends from previous computing eras played out from the 1980-2010s.

And

4

1:10 AM • Apr 22, 2025

With that being said, let's go ahead and rewind the clock back to the '90s. My goal with the next section is to provide a brief historical backdrop to set the tone for the rest of the post.

Flashback to the '90s: Enterprise vs Consumer

For a second, pretend there's no such thing as Decentralized AI.

Let's simply focus on closed source vs open source AI. How do things play out?

Enterprise AI

Recently, both Bill Gurley and Satya Nadella came to similar conclusions that it's becoming increasingly clear that enterprise AI is not going to be a winner-take-all market.

Let's walk through the historical analogy they both refer to and break it down.

Throughout the '90s and especially during the dot com boom, Sun had a stronghold on internet businesses. They provided top of the line servers at a high premium and locked in customers with their vertical stack (SPARC/Solaris).

Naturally, they were one of the companies that crushed it going into the new millenium. Sun was literally responsible for the growth of so many new consumer internet companies of that era.

But throughout the '90s, Linux and its open source counterparts were slowly starting to gain momentum on the side. And after the dot-com crash, the open source LAMP stack (Linux, Apache, MySQL, PHP) was getting to a point where the tech was good enough to use for businesses purposes.

Post 2000-2001, most companies couldn't justify Sun-Oracle when there was the Linux-MySql-x86 alternative that was a fraction of the cost.

Naturally, as more and more of the developer community realized the potential of open source web infra, an increasing number of bugs were squashed and the tooling only got better.

Enterprises eventually switched over and even companies like Sun had to embrace Linux with open arms.

On the point of enterprise customers being smart buyers, it's worth noting that the same thing happened in the cloud wars. Even though AWS had a massive headstart with cloud, there was still room for Azure and GCP to take some of the marketshare.

How does this analogy draw back to AI?

It may be the case that because it's so easy to switch models, companies will start quickly prototyping with a closed source provider such as OpenAI and then switch over to open source.

Why?

Because at scale, they can't justify the closed source costs even if they're only more expensive by a slight margin. Or perhaps the AI product line/feature becomes critical enough to the business where they can't risk the model provider changing up rates or access limits (aka rug pull). Or maybe product leads decide it's time to start differentiating and they need the open weights to fine-tune their models. Or it could even be the possibility that there are some privacy / on-prem mandates from a legal standpoint.

Consumer AI

On the other hand, in terms of consumer AI, there is a realistic scenario where something more akin to search or mobile happens. Just like how Google and Apple were able to lock in consumers by being the de facto product in the everyday lingo.

It wasn't "I'm going to use the internet" but rather "I'm going to google it".

It's always been "download it from the app store".

And similarly, as of the last three years, using AI has become synonymous with "just ChatGPT it".

Blake Robbins

@blakeir

Random thoughts:

saw someone on here mention 'login with chatgpt' the other day, and i can't stop thinking about it.

feels inevitable

302

12:53 AM • Apr 19, 2025

To add on, there's also a new moat on the streets that people are worried about: memory.

To be upfront, I'm still unsure how deep the memory moat is. I need to think more about the portability vs consumer lock-in. But it's clearly one of the biggest advantages on the consumer side...all roads in AI lead to personalization.

With that being said, it's worth noting that there are a lot more well resourced competitors trying to fight for the consumer marketshare. You could argue that ChatGPT is in the lead right now, but to what extent? Remember, Sam and team aren't just competing against new AI startups we've hearing about recently (Anthropic, Deepseek). It's also a battle against social incumbents that have massive distribution and network effects like Google, Meta, and X.

Key Takeaway for DeAI

Ok! Wow! That was a lot and we haven't even started discussing DecentralizedAI.

What was the point of me providing all this context?

The reason it's important to understand this is because the same logic follows for Decentralized AI. It's hard to believe in DeAI if you think it's closed vs open source vs decentralized.

There's room for all these types of AI infra, tooling, economics, developments, etc. to co-exist.

With this mindset, it becomes a lot more reasonable to accept that if DeAI can attract talent and build unique enough offerings, there is in fact an opportunity for the vertical to thrive.

Now that we've established that, let's jump into the next two sections that cover different layers of AI and how DeAI can fit in. I've organized them as follows:

1. Pre-Training: the massive training sessions you hear about that require tens of thousands of Nvidia H100s. This is the foundation of any model that uses the petabytes of data on the internet and is heavily dependent on compute & capital.

2. Post-Training: the focused polishing loops that steer the model toward specific behaviors or domains after pre-training.

Texas and Tennessee are only so big

On the pre-training side, you can argue there's a spectrum of reasoning of how necessary DeAI is that goes from philosophical to practical to experimental.

Let's go through all three.

Philosophical

Open-source AI isn’t truly open source if a single entity sponsors the GPU bill.

In regards to the section above on the '90s, one key thing to remember about the open source Linux era versus open source in AI is that there's an extra component of upfront capex for hardware that's needed.

Meta spent hundreds of millions of dollars to make Llama3. AI2 receives funding from the Allen Institute. Alibaba uses their enormous treasury to support Qwen.

So why have a decentralized state-of-the-art model? Because at any given moment even open source companies can change their policies or simply stop developing.

It might not even be malicious. These companies could be legally constrained because of geopolitical reasons. Or research institutes like AI2 might get their funding cut off and can’t continue to develop at the rate that’s needed to keep up against their centralized counterparts.

If a business in India depends on the latest versions of Llama to support their customers, and all of a sudden Meta decides that the next Llama is going to be proprietary...what happens to that business? And how is this situation any different than the potential "rug pulls" that we claim closed source companies will do?

So by just the reasoning of a “black swan event” or “the desire to be antifragile”, it’s good to have this decentralized model backup where funding and training isn’t restricted to any one company or country.

At a minimum, we as a global community need a model that is not developed, funded, or owned by any one entity.

Fortunately for us, the world has already proved out a network that coordinates massive amounts of hardware that's globally distributed through a well defined incentive mechanism.

Just as Bitcoin built out infrastructure for truly free money, we need to build out infrastructure for truly free intelligence.

Last thing I want to mention in this section is that there's a real possibility that niche and open source companies can't keep up with the rate of progress closed source companies are going at. And if that's the case, then it's imperative that we at least have a means to get everyone outside of these companies working together on an open, global model.

Practical

Okay, moving on to a part of DeAI where even folks who are working in centralized AI are paying attention to: distributed training research.

The big unlock that slowly formed from 2017 (release of transformers paper) to 2019 (GPT-2 launch) was that it's possible to make these assistants smarter by adding more compute and scaling up the underlying models.

This marked the start of the GPU gold rush and the AI space started shifting to being more closed source as pre-training became the competitive margin. And since ChatGPT launched in 2022, we've seen unfathomable numbers on how much the largest tech companies are spending on data centers (Meta, xAI, Microsoft, Google).

All these big training runs you hear about from the hyperscalers who have billions of dollars to spend are done so with GPUs that are co-located in massive data warehouses.

Now, the key part to understand is that it's not just about who has the most number of GPUs.

Rather, the north star is how well these teams set up their data centers to maximize efficiency. This is known as Total Cluster Efficiency which accounts for GPU failures, continuity of systems, etc.

As the newer models get exponentially larger, these optimizations become crucial.

This is where it turns out that the research DeAI folks have been doing is quite useful. In fact, teams such as Prime Intellect and Nous Research have been working closely with the general AI community to make progress.

The core question of this research paradigm is: can we train a state of the art LLM over the internet?

And in order to do that, it really comes down to reducing the communication needs between GPUs.

The Nvidia GPUs in a large central warehouse can talk to each other at speeds of 1800 GB / sec. On the other hand, a normal internet speed is ~500 MB / sec. There's clearly an uphill battle to making distributed training work.

Notice I said distributed, not decentralized. Companies like Google obviously aren't worried about malicious actors / Byzantine Fault Tolerance (BFT).

BUT the hardware optimization solutions help everyone!

Note: don't worry if you don't understand the examples below, it's more so evidence that there's actual progress that has already been made.

Two key developments worth noting are:

1. The Prime Intellect team working with Arthur Douillard from Deep Mind on Open DiLoCo (distributed low communication). Key learning from that paper is inner-outer optimization works, reducing GPU communication needs by 500x.

2. The Nous team working with Diederik Kingma, who invented the AdamW optimizer, on the DeMo (Decouples Momentum Optimizer) paper. I won't get into details here but simply put, it turns out you can compress how gradient optimizers work (a key part of the training process) without harming model performance. This research resulted in an 857x reduction in communication needs.

What these teams are doing is basically questioning the current methods of how models are trained, breaking down the different components, and trying to optimize them in creative ways. Just these first iterations led to drastic reductions. Imagine combining these strategies - all of a sudden, how we think about training looks totally different.

And what's amazing about this research is that if the DeAI teams can in fact build out protocols that allow for anyone to join in with their compute, there's no reason that enterprises won't collaborate if it's more capitally efficient for them. Somewhat similar to how RedHat and IBM stated rolling Linux into their servers because it turns out the open source software could actually run enterprise level apps.

As Jeffery Quesnelle from Nous Research puts it:

Only a handful of data-centre campuses have 20,000+ GPUs under one roof, but there’s a fat middle of facilities with something like 2,000 H100s apiece. The next logical step is to train a single model across a hundred of these 2 K-GPU clusters.

That shift benefits even the “centralized” players first: a cloud provider with multiple sites can simply treat each data-centre as a node in the network, linked by ordinary 100 Gb/s WAN pipes, instead of pouring more money into one giant InfiniBand fabric. They get higher utilization and avoid buying ever-bigger monolithic super-pods.

In other words, the method can be decentralized without requiring a fully permissionless swarm from day one.

Just a few days ago, Jack Clark, who is the co-founder of Anthropic wrote in his blog about Prime Intellect's 32B decentralized model and how these developments can totally shift the political economics of AGI.

It's also worth noting even the team at OpenAI understands that if you want to scale humanity's model on the order of 10 million GPUs, you're going to have to need some form of decentralized pre-training.

Experimental

For the last part of the pre-training spectrum, let's discuss the part of DeAI research that isn't in production yet but is quickly becoming a contender in reshaping everyone's framework on model economics.

And to best understand it, it's worth bringing up protocols we're all familiar with such as Bitcoin, Ethereum, and Farcaster. At their core, each of these systems commoditized the infrastructure for their respective parts of society (money, computing, social).

The reason Bitcoin works is because no one entity or person controls the ledger and can change it to work in their favor. And no one can run away with all the Bitcoin (yes, I know what the 51% role is don't at me).

In order to apply this same concept for intelligence, the main challenge at hand is create a model that no one entity owns and can run away with. Or more practically put, it's not possible for someone who helped train a model to just upload it to Hugging Face whenever they want.

At a high level, this is what Alexander Long and the Pluralis team call Protocol Learning.

It's a training process in which compute providers are only given a section of the model (the weights) to do work on. This is itself is not something that's a new concept. The hyperscalers use model parallelism strategies in their large data centers.

But the part that hasn't been solved for is creating a decentralized model parallel system.

In my opinion, if we can get fully decentralized model parallelism working through the efforts of what Pluralis and others are making, then we finally see what it means to have models as true global commodities.

Now, the reason I called this section experimental is because of the second sentence in the screenshot above - specifically the part that says "the economic sustainability of closed model releases".

By having a model that no one can tamper or own, there is a new opening for an asset class based on models-as-a-commodity. This warrants new kinds of economics and business ideas that no one is really even thinking of right now. I don't think this is necessary for the future of humanity (at least as it stands) but the comparison is more so how Uniswap provides a lot more interesting financial use cases and Farcaster provides a different set of social experiences.

I asked ChatGPT to come up with some use cases if we could financialize the development of intelligence. To be clear, this doesn't even have to be tokens. Literally, everything we're talking about could be done with stablecoins. What matters here is the point that financialization can exist because of model parallelism.

To wrap this section up, let's recap the 3 parts of pre-training that intersect with DeAI.

1. Philosophical: training models requires up front capital expenditure which leads to reliance on a single entity. A model can only be truly open source when there's no central point of failure.

2. Practical: there's an upper limit to how much hyperscalers with large treasuries can expand their centralized warehouses. These teams are already working closely with DeAI folks to improve distributed training and GPU communication standards.

3. Experimental: models as a commodity in its purest form means no one can run away or take advantage of the model itself. If decentralized model parallelism can work at scale, then you own an entirely new asset class around intelligence giving the long tail of consumers opportunities to gain exposure in the new AI wave.

On to post-training!

Enabling creativity through society-of-minds

Before we dive into the main part of this section, it's worth quickly defining what post-training is.

Post-training is what gives LLMs personality, expertise, niche knowledge, and adds another dimension to the experience of how you interact with AI.

This can come in different flavors such as:

• Supervised Fine Tuning (SFT) - having people write out thousands of sample questions and answers which are then fed to the model so it can learn the style of answering

• Reinforcement Learning (RL) - do more of what humans reward / upvote you to do. The more thumbs up votes you get, the higher the score

Centralized Training is locked in on Post-Training

When Deepseek R-1 came out, you may have seen many tweets pointing out how post-training techniques will be the next big thing in terms of making models significantly better.

In fact, a month before that, Ilya Sutskever made the bold claim that pre-training as we know it will end because of data limitations.

Note: remember that it was OpenAI that came out with the first chain-of-thought reasoning model in September 2024 which was super impressive. But more people understood the impact of post-training & inference-time compute in January because of Deepseek's ability to do it at a fraction of the cost with a new kind of RL technique know as GRPO (out the scope for this post).

All of these post-training strategies come down to the need for unique data that others don't have access to. Every major AI company is focused on improving their data inputs in some form or fashion.

For example, news about OpenAI trying to buy Windsurf for $3 billion came out last week.

Why the heck would they do this? Because Windsurf has a crazy amount of data on how developers are using AI.

Jon Chu // Khosla Ventures

@heyjchu

My personal guess:

OAI is buying Windsurf for the distribution and coding telemetry data.

They're on track to spend hundreds of millions yearly on Mercor. This is another avenue for capturing that type of data, and also gets them a great product and reduced mercor spend

440

2:08 AM • Apr 17, 2025

In fact, after reading that tweet by Jon, I was shocked to find out that OpenAI pays upwards of $300 million to Mercor (AI hiring platform), ScaleAI, Upwork, etc. to source domain experts to help out with things like data labeling, RLHF, and post-training in general.

Okay let's pause for a second. What are the key points made so far?

1. We know that pre-training in a decentralized way is really hard because it requires the coordination of a ton of compute and capital that's all crowdsourced. But there's still progress being made.

2. Most AI developers are realizing that the next major developments in model performance and scaling will come from reinforcement learning and test-time compute.

3. The limitation of post-training seems to actually be acquiring the data, expertise, and experience of people in niche subjects, less so the compute & synchronicity requirements.

And that's right along the lines of what Andrej Karpathy and Yann LeCun are calling for.

Yann LeCun

@ylecun

I've been arguing for something like this for over a year: crowdsourced distributed fine-tuning.

Andrej Karpathy

@karpathy

For friends of open source: imo the highest leverage thing you can do is help construct a high diversity of RL environments that help elicit LLM cognitive strategies. To build a gym of sorts. This is a highly parallelizable task, which favors a large community of collaborators.

3,089

7:57 PM • Jan 29, 2025

Society of Minds Architecture

In this post, I'm not going to get into the details of how the RL gyms should be structured but rather focus on why it matters that there is a way to make crowdsourced distributed fine-tuning actually work.

ChatGPT already has 300-400 million monthly active users. Many people already claim that they use AI in some form or fashion in multiple parts of their lives.

At the end of the day, if 4-5 closed source models decide who the experts are to rank data sets and develop evals (framework to evaluate LLMs), then the general population pretty much has to blindly follow and use what these teams decide. It's philosophically no different than the social media feeds argument. What is one of decentralized social's main selling points? That there's an opportunity for anyone to build a client / UI that isn't decided by a team at Instagram or X - you have the freedom to choose your own feed (you don't have to but the option to do so is key here).

I personally believe that there is a vast amount of untapped experience and wisdom around the world. We want these models to have the opportunity to learn from all of this data. Anyone should be able to contribute to these models, explore as much as possible, wander through diverse environments, share their experiences, etc.

The analogy that clicked for me was GitHub.

Github enabled talented developers from any corner of the world to contribute to whatever open source repo they wanted to. It's important to note that this doesn't mean every single PR was merged. Obviously a lot of the contributions around the world are low quality efforts (or even malicious) but the community has the option to simply not include them. Or, those that believe their efforts are ~contrarian~ can simply fork a repo and continue building for themselves.

Wait a minute! You may be thinking...why does it have to be onchain YB?

Well, the reality of how open source worked in the '90s and 2000s doesn't exactly map to how open source AI contributions will happen. In the era of Linux, MySQL, etc. open source was more philosophically driven (cypherpunk era vibes) and didn't require additional resources other than a computer & internet.

If someone made a bad PR, it would just be deleted or ignored. On the other hand, if someone maliciously trains a model on bad data or submits pooly designed evals, then there is capital being unnecessarily burned because of GPU costs.

Network participants in these RL gyms needs to be held accountable through well designed incentive systems. Similar to how if a Bitcoin miner submitted a block that gave itself more money, the network will penalize the miner.

If this onchain RL gym network is built out properly and can attract enough contributors in the marketplace, then it's very realistic to see the modular thesis of AI play out.

A world where there are many different specialized models that users can plug in when needed. And the people who contribute to making these niche models work will be rewarded based on model performance in the open markets.

My friend Sam Lehman likes to call this the "decentralized school approach". An opportunity for the smartest people around the world to come together and permissionlessly contribute their knowledge and experience in order to help build the technology that will define the next era of how we think, act, and live.

The world is an abstract, messy, and complicated place. Not everything looks like a math problem that you can have a boxed answer to at the end. For non-verifiable and subjective topics, it's important that people who think they should have a say are in fact allowed and incentivized to pitch in. Let the markets decide if their contributions are valuable or not.

Recently, Packy asked ChatGPT the following question:

One of the answers it gave was as follows:

In other words, let the hivemind win. This is how you enable creativity around the edges.

To end this section, I want to quickly revisit the Sun-Linux backdrop from earlier in this post.

Imagine if we didn't have Linux? There would be countless innovations that we take for granted today that probably would not have even had the chance to come to life.

If the last 30 years of software depended strictly on Sun and Oracle infra, it's pretty safe to say that we may not have progressed as fast as we have. Creativity would be stunted and contributing to technology would be an activity strictly for the business elites.

And for that reason, it's my view that if you think Linux was critical to the internet, then it's a no brainer that having a truly open source post-training network with proper incentives is a necessity for the AI era we're just starting to enter.

To wrap this section up, I want to clarify the fact that setting up a decentralized RL network is a non-trivial problem. There's so many different parameters and components to think about. We're a long ways from it but teams such as Prime Intellect and Gensyn are showing some rapid iteration in the last couple of months.

Hervé Cuviliez

@hervecuviliez

If you are in AI, no matter which side of the table and you are not following @gensynai recent updates
You are missing out bigly !!
Congrats @fenbielding & @_grieve

Harry Grieve

@_grieve

rl-swarm repo stars, or @gensynai team vibes?

26

3:23 PM • Apr 19, 2025

Again, the technical details are out of scope for this post, but I'll be going through these efforts in depth the next couple of weeks so make sure to subscribe to Terminally Onchain!

What to keep an eye out for in 2025

At the end of the day, this post was really just for myself.

It was a chance for me to take a deep breath and see if I could come up with a compelling argument to continue going down the DeAI rabbithole.

And it seems to me that there's enough evidence and reasoning to continue tracking the new developments at least until the end of 2025.

Jake Brukhman (co-founder of CoinFund) has "deAI Summer 2025" in his X name. To me, this feels accurate. Not in the sense that the world will be using decentralized AI services by September. But more so that a lot of key releases are coming up in all parts of the AI world and it'll be clear if DeAI can keep up or it can't. For example, ICML is coming up this July - what are the AI community's reactions on the new papers DeAI teams are expected to publish then?

In my opinion, at the rate AI is accelerating, it'll be pretty clear by Q4 of this year if there's a reasonable shot or not for crypto x AI to actually have mainstream merit. To be clear, I'm not saying things will be robust and working but more so that there will be enough data points for anyone to agree that DeAI has a unique enough value prop.

Of course I'll be keeping up with the projects mentioned in the post above such as Nous, Gensyn, Prime Intellect, etc. But this primer barely touched the surface of all the developments going on!

There's a ton of amazing teams cranking away on different parts of the AI stack and are trying out a variety of solutions - none of which should be ignored considering how early it is in this hyper niche vertical.

For example, here's a couple of things I didn't even have a chance to get to but I know for a fact that I'll be tracking closely over the next few months.

1. Sovereign Agents - if we really are headed to a world where agents will be doing more than we can imagine...then there's a chance to make onchain infra the no-brainer choice for agents to complete their tasks. What does that infra look like compared to web2 rails? AI researchers themselves can have bias, but agents are looking to just be efficient. And if there is merit behind the idea that privacy, micropayments, agent-to-agent communication is in fact better onchain, then let's build for that!

2. AI L1s - I haven't gotten a chance to look into AI L1s such as the Bittensor ecosystem, Ritual, Ambient, etc. This is my first priority going into May and it's clear that there's a ton of work happening in terms of attracting fresh developers (i.e. new subnets popping up). I want to dive deeper and understand how the economics differ in the L1 model and if there's enough merit to argue that the developer ecosystem will continue to grow.

3. Data Providers - I've published a piece on Vana back in December. In that post I made the argument that decentralized data marketplaces need to rebrand from the idealistic "protect your data" punch line to focusing on data as a new asset class that has financial upside. The reality is that the average Joe doesn't care about their data that much. Data marketplaces in my opinion will only gain steam in a way that probably looks similar to the agent meta from Q4 of last year. I also need to look into Masa, another solution for data marketplaces.

4. AI on local hardware - there's a lot of research efforts going into creating agent frameworks that can run on your local hardware. In fact, Soumith Chintala (creator of PyTorch) has an amazing talk covering why it's important to have an agent that runs locally. The team at Exo Labs is working diligently towards this goal. If you have some time, I highly recommend checking out their 12 days of Exo series where they showcase fun projects like running Llama on Windows98!

5. Memory & Personalization - as I mentioned in the enterprise vs consumer section above, memory was recently handed the first place award by the moat committee. If this is the case, then how do we think about personalization lock-in and portability for DeAI? Ejaaz has a good tweet about this. I haven't gone deeper yet, but am definitely looking forward to check out companies like Plastic Labs, Heurist, Honcho, etc.

I could go on and on, but the point is that there's still a ton I have to understand in the crypto x AI vertical.

YB

@yb_effect

worth noting that just in the past month @PluralisHQ announced a seed led by USV and Coinfund. and @PrimeIntellect raised a round led by founders fund

more announcements also coming p soon in this vertical

YB

@yb_effect

you have 3-4 months to get up to speed on DeAI and find your alpha

49

3:40 PM • Mar 29, 2025

If you don't have time and want to follow along my journey as I continue making my way through the "DeAI idea maze", then make sure to subscribe to Terminally Onchain for all my insights.

Curated set of Resources by Topic

YB's Content

• Key takeaways from DeAI founders

• Open source & DecentralizedAI Twitter List

DeAI Resources

• The World's RL Gym

• GSR Report

• Cyber Fund Crypto x AI thesis

• Scaling through Decentralization

• Frontier Training

• Big Brain Decentralized Training

• Why (good) AI needs crypto

• Agents Unleashed

• Protocol Learning

• Inference reshaping AI governance

• Chain of Thought Newsletter

• How to build an American Deepseek

Getting up to speed on AI

• Karpathy LLM talks (Deep dive into LLMs, Intro to LLMs, How I use LLMs)

• 3Blue1Brown Neural Network series

• LessWrong Technical Prime on O1

• NovaSky train your own AI

• Nielsen's textbook: follow 3blue1brown examples and get hands on practice

• FineWeb: basically the "internet data" compressed for you through crawlers

• Common crawler: open repository that's used to scrape all the internet data

• Tiktokenizer: select a model and you can covert between plain text and tokens

• BBY Croft: fantastic LLM Visualization tool to understand the training process

• RLHF Handbook by Nathan Lambert

• Chatbot Arena - LLM Leaderboard

• Hugging Face LLM pricing

• Questions about the future of AI

• The Era of Experience

Relevant Podcasts

• Redpoint's Unsupervised Learning

• Dwarkesh Podcast

• No Priors

• a16z AI

• Interconnects (& newsletter)

• Into the Bytecode episodes with Nous & Pluralis

Talks

• Nathan Lambert's talk on open source AI at Harvard

• Derek's talk on Crypto x AI at Collab Currency

• Delphi Digital Crypto x AI Month