I hope all 9486 of you had a great week 🔥
Welcome to the 6 new members of the TOC community.
If you're enjoying my writing, please share Terminally Onchain with your friends in Crypto!
Hi everyone, happy Friday!
Imagine you've had your eye on this awesome puzzle for a while, and you finally decide to get it. It's on the table in front of you. All the pieces are there. You have your coffee ready and some background music playing. No one's in the house. Just you and this puzzle.
There's only one problem. IT'S SO OVERWHELMING. Where do you start? Am I stupid...why aren't any of these fitting? Ugh, should I have just stuck to watching a movie?
But then, you calm down and think logically. Okay, first let me do an initial sort of all the pieces based on which quadrant they fall in. And then I'll start connecting the low hanging fruit so I can have some small wins. And so on.
This is basically how my past week has been.
The puzzle that's been frustratingly delightful is called "Decentralized & open source AI".
The true TOC fans may have noticed that I didn't send my usual Tuesday post. Well, that's because I hadn't fit together any of the pieces by then and decided it would be dumb to draft a half ass piece and waste all of your time.
But, fortunately, on Wednesday and Thursday I had some luck and I've finally given this puzzle some shape and life. However, to be totally transparent...I'm not done with the puzzle and it might take a bit longer than I thought which I'm okay with as long as I'm making progress.
My mission in these upcoming months is to become the crypto x AI guy.
I want to be the subject matter expert that glues both of these sectors together by understanding what's happening at the frontier of open source AI and decentralized AI.
This means depth and breadth - I don't want there to be a single thing that is past me in this vertical.
For today's post, I want to share my framework on how I'm approaching this goal. Below is the "puzzle quadrants" that make the most sense to me and brief thoughts + a todo list for each section.
AI fundamentals: It's not that hard to catch up
Open source AI community: How far behind is open source?
Decentralized AI: What's the golden basket?
Open source history: What are the key learning lessons?
Let's dive in.
AI Fundamentals: It's not that hard to catch up
I started this week by attempting to draft a post on Nous Psyche. Obviously, I started by going through the docs and even managed to understand most of it at a high level.
But I didn't truly get it.
So then I decided to read the Nous Distro paper thinking it might help me. And even that - with the help of ChatGPT - I was able to understand at a high level.
But once again, I didn't truly get it.
It was clear to me there was an AI knowledge gap. There's no way I could have novel insights without understanding fundamental AI concepts.
So, I decided it was time to go off into the deep end.
I started with Karpathy's famous talk.
After taking notes on this, I already felt a huge boost in confidence. A lot of the terms that I had read and heard on Twitter / podcasts started making more sense.
And then, as a math person at heart, I decided to understand how neural networks and transformers work at a more deeper level. So, I binged the entire 3blue1brown series on LLMs.
By this point, I was ecstatic. So many of the AI concepts started clicking. If you're anything like me, you'll understand how annoying it is when you don't understand the entire stack of a new concept (not there yet but finally in the rabbithole).
It’s actually insane as to how quickly you can get up to speed on basic AI fundamentals by simply watching the @3blue1brown neural network series and @karpathy intro to LLM video (he published a new one today that I need to get to)
25
And then! It's as if the tech gods were listening to me. Right after I finished those videos, Andrej Karpathy drops this 3.5 hour video on everything you'd possibly want to know about LLMs at an intermediate level. His first video in several months came literally at the perfect time.
Believe me when I say this...if you are feeling AI imposter syndrome, I promise this video will single handedly cure the problem. I'm only 2 hours in (going really slow + taking notes) and I can already say this is one of the most amazing tutorial videos I've ever watched.
going through it rn and can’t believe this is just free content for everyone
It was super helpful to see the tools/resources he was using throughout the video as well:
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
GPT-2 from scratch: Karpathy's repo to do everything from ground up
Hyperbolic: access pre-hosted models without any setup
UltraChat: synthetic data improvements repo for post-training process
Nielsen's textbook: follow 3blue1brown examples and get hands on practice
After finishing the last hour of the LLM tutorial above, here's the upcoming agenda:
Familiarize myself with Hugging Face
Follow the ChatGPT-2 tutorial and run it from scratch
Briefly go through both the Llama & Deepseek papers to understand open source SOTA models
Read through Leopold's situational awareness essays
Just to clarify, the AI fundamentals section is one that requires a larger ramp up at the beginning which is where I'm at now. Once my base knowledge clears up, then it's more of just keeping up with the latest papers and releases. No different than the rabbithole you have to go down when first learning about crypto.
Open source AI community: how far behind are the closed competitors?
I know there's been a lot of recent hype around Deepseek and open source, but I'm still trying to figure out where exactly we are with open source AI developments.
Fortunately, Lex Fridman's podcast episode this week introduced me to Nathan Lambert and his work with Ai2. Nathan has been an open source AI advocate the past few years and writes a fantastic substack called Interconnects that covers many of the happenings in the OS AI world.
And just yesterday, he published a post on why he thinks the recent Deepseek news should be a huge wake up call for Americans to invest even more in open source.
Making the U.S. the home for open-source AI
Open-source AI is here to stay, but it is not a given that it will be American.
The core point was that in the last 30 years, China has just been copying western tech and improving the margins. But in this current AI race, it's clear that they are trying to be the ones spreading the innovations. So if U.S. companies don't make the effort to open source their models, the rest of the world (including Americans) will be quick to adopt Chinese tech.
The section that was a reality check in the essay was this:
Building strong AI models is far, far easier than building a sustainable open-source ecosystem around AI.
Building a better, truly open ecosystem for AI has been my life’s work in the last years and I obviously want it to flourish further, but the closer you are to the core of the current open-source ecosystem, the more you know that is not a given with costs of doing relevant AI training skyrocketing (look, I know DeepSeek had a very low compute cost, but these organizations don’t just fall out of the tree) and many regulatory bodies moving fast to get ahead of AI in a way that could unintentionally hamper the open.
Yes, efficiency is getting better and costs will come down, as shown with DeepSeek V3, but training truly open models at the frontier isn’t much easier.
Specifically, "the closer you are to the core of the current open-source ecosystem, the more you know that it is not a given with costs to doing relevant AI training".
Look, I'll be the first to admit that I have no idea what's actually going on in the open source AI world. But! I have to say that I was surprised that there was no mention of crypto whatsoever from my readings this week by OS AI researchers like Nathan and Tim Dettmers.
And I know I know. They don't take the crypto industry seriously. Scams, grifts, blah blah blah.
But fuck that.
As I was reading Nathan's essay, I couldn't help but notice the core issue he was presenting was a cost one. And if that's the case, then to me it's a no brainer that crypto incentives could enhance open source research big time.
So, is it that these guys have already considered crypto and came to the conclusion that it's not worth the efforts? Or is it more of an avoidance to the crypto space in general?
If the former, then I'd like to see examples. And if the latter, then that's a blind spot they have that I'd like to double down on.
Anyways, I've barely scratched the surface in terms of understanding the open source AI landscape but Lambert's work was a great start.
One thing that was painfully clear to me is that these OS AI guys are actually preaching the exact same thing as folks in the crypto world. I mentioned it in last Friday's post, but the open source AI challenge is truly so similar to Bitcoin that it's a no brainer these two communities (crypto and open source AI) need to be collaborating more and more. Hopefully, I can help bridge this gap.
Some checklist items for this quadrant:
Watch Lambert's talk on open source AI that he gave to Harvard Law
Understand what Ai2 does and their progress so far
Find companies and researchers similar to Nathan/Ai2 and get a big picture understanding of the state of open source AI (create Twitter list as well)
Understand the different kinds of licenses and their impacts (i.e Amazon has to pay Meta for Llama but Deepseek is MIT license and more "free")
Form a thesis on the current state of OS AI from the perspective of Lambert type researchers. Try to understand the biggest bottlenecks and what solutions they think will work (keeping crypto bias aside)
Decentralized AI: what's the golden basket?
To start this section off, I recommend everyone take 10 minutes to read this post on why AI needs crypto specifically. I've already covered the core thesis in previous posts, but I think Daniel does a fantastic job of really laying out the past, present, and future.
Why (Good) AI Needs Crypto - Variant
Open source AI combined with crypto leads to better AI by enabling larger models and driving more innovation.
He nails the point about how open source AI has a resource problem and why crypto solves for that.
good read by @dbarabander on crypto x ai:
- It's better to compare open source AI to Bitcoin rather than to Linux. Both are software projects that require massive amounts of resources & hardware to function
- OSS contributors need to give time + compute & data
- almost…
- It's better to compare open source AI to Bitcoin rather than to Linux. Both are software projects that require massive amounts of resources & hardware to function
- OSS contributors need to give time + compute & data
- almost…
One thing that caught my eye was the folks that helped review and edit the essay. The thanks section gave me a solid list of companies to start off with in terms of breaking down the decentralizedAI sector once I brush up on AI fundamentals. Note: I'm aware there may be some bias considering Daniel works at Variant but helpful regardless.
A question that has been top of mind for me recently is how aware are the folks in AI of decentralizedAI (what we would consider crypto) companies?
I don't have an answer for you just yet, but I'll admit that it was pretty damn cool to see Andrej Karpathy use Hyperbolic in his new LLM tutorial video I mentioned above. A few years ago, I remember having a call with Jasper (founder of Hyperbolic) when seemingly no one cared about crypto x AI, and it's so cool to see the team's hard work pay off.
We are honored that Andrej Karpathy @karpathy has recognized Hyperbolic as his favorite platform for interacting with LLM base models. In his latest deep-dive video on Large Language Models (LLMs), he also commented that “Hyperbolic is the only place that hosts my <3 Llama 3 405B…
You may be wondering...YB why is it taking you so long to understand these DeAI companies? Well, like I mentioned above with Nous Psyche, I'm bottlenecked by my AI knowledge. For example, 48 hours ago, I wouldn't be able to understand what this tweet below even meant.
afaik Hyperbolic is the only place that hosts my <3 Llama 3 405B Base, and in bf16 precision. So thank you :)
But! I can confidently say that as I wrap up my AI fundamentals crash course, I'll be able to speed through all of these companies and form a thesis on them over the next 2 weeks.
My goal will be to create a golden basket of DeAI companies. If I could angel invest into just 5 of these startups, which ones would they be and why? And then my plan is to become the biggest hype man for these teams. Two of them that I already feel like may be on the list are Nous and Prime as I've mentioned them several times in my posts the last few months.
Teng, a friend and fellow crypto x ai researcher, is doing a fantastic job of covering these companies in his newsletter and I'm excited to start going through his primer as well. Check out Chain of Thought if you haven't already, its a must read!
At the end of the day, the thesis for DeAI is as follows. Our job is to make this a reality.
The next AI X Crypto wave is going to be driven by real AI builders leveraging crypto rails and not crypto native grifters slapping AI on their token slop.
With that being said, here's the checklist for this section:
Go through Nous, Prime, & Hyperbolic to get started next week & publish Tweet summaries
Read through Ronan's piece Can decentralized AI compete?
See if other DeAI companies are talked about and/or used by AI researchers (i.e. Hyperbolic-Karpathy type examples. I know Nous is well discussed, but checking for others as well)
Start reading through the resources in CoT's primer
Dive deeper into what Teleport and Nous are working on with TEEs
Draft my own thesis on why crypto needs AI and create the golden basket
Open source history: what are the key learning lessons?
And finally, open source.
I'll keep this section short. Technically, it's not a requirement but I think by spending time learning about the history of open source, I'll have a much better grasp of why the fight for DeAI even matters in the first place.
The truth is that I'm not old enough to remember the Linux days or early 2000s era of computing when open source was the norm. I grew up in the FB / Twitter / Uber era and centralized platforms have always been the default for me.
That's why I want to better understand what happened in the past with open source and what I can learn from those examples.
The other day, I was listening to the BG2 podcast, and Bill Gurley was comparing Meta's Llama strategy to what MongoDB and others did back in the day. I thought that was fascinating and I really want to understand the nuances there. Many OG Terminally Onchain readers know that I'm obsessed with tech history, so this quadrant is really scratching my itch to make obscure comparisons from the past to what's happening now.
I think this will help me not only discover unique insights in this vertical but will also help me feel philosophically more aligned with becoming a subject matter expert on crypto x AI.
Read Working in Public
Read Cathedral and Bazaar
If you made it this far, I'm guessing you are just as passionate and excited about the DeAI vertical as I am. I purposely hid this sentence at the very end, but reply to this e-mail with your TG handle if you'd like to be added to a TOC chat where we can discuss all the topics / resources mentioned above. I haven't made the group yet, but if I get interest from even a few excited folks, then I'd love to make the chat so all of us crazies can help each other out and have fun.
That's all I have for today's post!
I hope all you have a great weekend and I'll see you next week 🤝
- YB
Yesterdays /toc deep dive https://terminallyonchain.xyz/horsemen