How Agents are Shaping AI, Part 1, Software

How Agents are Shaping AI, Part 1: Software

Right now, AI is experiencing two large changes that are better explained together.

Software is trying to look more like hardware

Hardware is trying to look more like what software needs

This may read as obvious, but it’s not. For years, software ignored hardware issues and worked around them, and hardware is, still to this day, largely suboptimal for AI inference.

But with AI already representing the largest investment relative to GDP in US history, we need to make sure this all makes sense, and we all make some money. And agents are our best bet.

In this two-part series, my job is to answer the following:

To the best of my ability, can I describe what AI models and chips will most likely look like three years from now?

If I did my job, today you’ll learn key intuitions about how models work, how they learn, and what state-of-the-art AI research looks like, because this is the first architecture built with one thing and one thing only in mind: agents.

The First “Mixture-of-Everything” Model

DeepSeek’s new model is like nothing we’ve ever seen. It has so much that I simply call it a ‘mixture-of-everything’. It brings back a decade-old idea, mixes it with new ideas, and creates an incredible model that is not only fascinating but genuinely good.

It’s all really in the name of improving two things:

Reducing the agent bottleneck

Reducing the memory bottleneck

And although I’m sure you know what the memory bottleneck is (we’ll review it nonetheless), what on Earth is the agent bottleneck?

Let’s find out.

Strong… and Weird

Getting the obvious out of the way. The model is good, very good.

Not as good as the best models on the planet, Fable and Astra from Anthropic and OpenAI, respectively, but close (it is as good as its predecessors), and most importantly: insultingly cheaper.

Source: DeepSeek

The pricing for this model is simply absurd.

Source: DeepSeek

Naturally, these are Chinese API prices, which I’m afraid I do not recommend using. However, DeepSeek models are free to download, so you can use them on Western servers no problem.

But how cheap is this compared to frontier labs?

The simple answer is that even at peak hours, the blended price of DeepSeek models at $0.75/million is 20 times smaller than what you pay for Opus 5, a model of comparative performance.

We’ll revisit this question later once we have a better idea of why this model is so transformational, so that I can explain that the price difference is actually much larger.

Good, nice, and cheap.

I bet you don’t find this surprising in the slightest, as you probably know that capabilities become commoditized in a matter of months.

But I didn’t spend a full day reading such a complex paper and writing such a complex article just to tell you that AIs are getting cheaper per unit of intelligence.

What matters here is how they did it.…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.