home

Transcript

Happy Independence Day!

I’d like to go through this article “The American DeepSeek Project” that just came out today.

A little bit of background. This is an idea that’s been floating around for a while. An article came out in April called “How to Build an American DeepSeek and Why We Need It Now.” It’s a good read.

I agreed with all of this. I didn’t feel any particular urgency to do anything specific for this in part because it was like, oh, you know, Meta is kind of this big institutional anchor. They’re going to be pouring all their money into building these open models and the community will kind of take that Llama release and spin it and fine-tune it and we’ll be able to collectively catch up to the big labs.

Look at this prediction:

Llama 4 will likely deliver amazing results.

It didn’t. It didn’t. It was so bad that Meta is now ‘reconsidering their open approach to AI’. This came out last week:

What other signs do you have? Their spokesperson was like, no, no, we’re committed to it. Yeah, they don’t want a huge flight of people away from Llama.

But is the writing on the wall? Well, they’re in the process of trying to hire these two gentlemen who are the largest funders of closed AI labs outside of the big labs.

We mentioned them in our episode 124. One of the companies they funded raised $465 million, shipped nothing but a blog post a year ago, and have done absolutely nothing except for call for regulation.

As well as SSI, who has no customers other than the government.

And that ClosedAI guy who was talking about ‘nationalizing the AI labs’.

So that’s who Zuck is trying to poach. Gross just stepped down from SSI. He’s probably going to be joining Meta.

Okay. This is hugely problematic.

It’s safe to assume that Meta is now just a bad actor. They’re leaving Team Open.

I’m happy to be surprised if they come out with some super strong statement or new model, but it looks like they’re ditching Team Open.

So I’m now feeling the urgency here.

Let’s go through this article from Nathan Lambert.

What I think the next goal for the open-source AI community is.

While America has the best AI models in Gemini, Claude, o3, etc. and the best infrastructure with Nvidia it’s rapidly losing its influence over the future directions of AI that unfold in the open-source and academic communities. Chinese organizations are releasing the most notable open models and datasets across all modalities, from text to robotics or video, and at the same time it’s common for researchers worldwide to read far more new research papers from Chinese organizations rather than their Western counterparts.

I’m not going to read all this. I do want to go down to some stuff at the bottom.

Can I show you this? I have been enjoying DeepSeek R1. I love DeepSeek. I love this Chinese lab DeepSeek. Their model is awesome.

Say hello DeepSeek. This is a work in progress. Anyways, we are integrating DeepSeek alongside US models and going to be doing sort of a best of all worlds.

So we need an open source version of DeepSeek. And I want to kind of jump down to some stuff here about why I’m really bullish about this approach generally. In part because open agents, I think are going to help be that sort of like magic X factor that helps to make up for the lack of the insane amount of resources and training base models.

The efficiencies of open-source software style development are dramatically stronger for agentic systems than models. Models are singular entities built with expensive resources and incredible focus. Agents are systems that can use many models off the shelf and route requests depending on what’s needed.

Speaking of routing, I’m literally had my agents today build a router using Effect TS routing between different models and agents all in Effect will open source this pretty soon. But anyways, they’re speaking my language here.

This agentic era is the opportunity open models have needed, but we need to clear much stronger performance thresholds before the open counterparts are viable.

I don’t even know how much of this requires model level innovation. What’s out there is already like quite good. But hey, I’ll appreciate further compute at the base layer.

We have companies like OpenAI and Google launching Claude Code competitors that pretty much flop.

Yep. Yep. Yeah, we’re doing our own. You have to be neutral between models. You can’t just like push your own models. Maybe Anthropic can. But anyway…

For this reason, we have finite time to get there. … The best American models are plagued by the Llama license (and rumors that future versions will be discontinued).

Yeah, this is done. The Llama ecosystem is done. It doesn’t matter if they put out a Llama 4.3. They’re leaking now. They’re going closed.

In many ways, obtaining this goal is a quintessentially American volition. In the face of a technology that is poised to bring such extreme financial, and by proxy literal, power to a few companies, opening AI is one of the only things we can do to reduce it. Technology proceeds in a one-way direction — for a variety of geopolitical and capitalistic reasons it is impractical to pause AI development to “do AI another way” — the best we can do is chart a path that makes this future better.

Our take is that AI needs to be decentralized with multiple power centers. There’s not going to be zero power centers. But we can at least make sure there’s not one or two. We think there should be like a thousand. Dude, dude.

We need new systems to mitigate misuse, but it shouldn’t be solely up to corporations to control this. Safety by isolating technology to a select few is something we’re in the later stages of with nuclear weapons5, and AI progress is far harder to monitor. Robustness to AI can only come from designing systems that expect it to be pervasive — not that it is an easy task.

Realistically, all of this is fighting gravity. The corporations will win, but we can control to what extent. We can control how good the other options are. The open options.

The call to action here is simple — consider how you can slightly shift your decision making to make The American DeepSeek more likely. This approach succeeds just as much by having one model at the end of it, as it does by having the community form better habits and norms around the way AI models are conceived, built, shared, and used.

Okay, so for my part as a CEO of a tiny American lab, we will support this to the extent that we can. We’re more focused on agentic stuff than the base layer, but it sounds like even they acknowledged that the agents are going to be increasingly important.

We’re basically trying to build the front door of the interface that allows you to very easily use any number of models and agents to pull in different datasets and have it connected to this sort of like decentralized open marketplace that people can get financial contributions for contributing to.

We’re gearing up for a larger product release later this month, so we’ll have more on that specifically.

I’ll push back a little bit on the term American. I get the point that there’s sort of the different poles of west and east, America and China. But it more seems like it’s the underlying ideal that matters, not the country. And so I hope for this to be a global effort that can engage happily many people other than just Americans.

And if we need a better name than American DeepSeek, we can always go with Atlantean DeepSeek. More on that soon.

Stay tuned!