The Evolution of LLMs & What's Coming Next

I've been thinking a lot about where we've come from and where we're heading with AI.

Just a couple years ago, ChatGPT went mainstream and was impressive because it could write coherent paragraphs. Today, I'm watching AI systems code entire apps, manage complex workflows, and solve problems I couldn't have imagined.

The transformation didn't happen overnight, but when I trace the path from "write me a 500-word article" to "build and launch my startup," there's a clear evolution.

So in today's newsletter, I’m breaking down:

• How LLMs evolved from embarrassing math failures to sophisticated problem solvers

• Why we're all becoming "agent managers" instead of knowledge workers

• What a $1 billion one-person company actually looks like

—@dharmesh

From Math Failures to Problem Solving

First, a quick nerdy history on the evolution of LLMs.

Early LLMs were pretty good at generating text (eg. "write me a 500-word article about marketing"), but they were hilariously bad at doing even the simplest math. Like 17 + 8.

The reason behind this is because LLMs were pre-trained on written "language" (hence the name), but the number of permutations of simple mathematical expressions is near infinite. So the models would just guess — often incorrectly.

Then came the leap that changed everything: We gave LLMs the ability to call tools.

One of those tools was running code, usually Python. When an LLM encountered a problem requiring calculation, instead of guessing, it could hand that task off to a Python interpreter and get the correct answer.

It's like elementary school algebra — we translate word problems ("Susie takes 60 minutes to walk home and John takes 45...") into mathematical equations. The translation is hard, but solving the equation is straightforward.

Fast-forward to today: If an LLM senses a prompt could be better solved by coding a simple Python program, it just does that. This could be math, data analysis, or anything a full programming language can handle.

This evolution is fascinating because LLMs can now translate higher-order goals (eg. "help me launch a marketing campaign") into executable tasks, some of which become coding problems that Python solves automatically.

Of all generative AI use cases, code generation is advancing fastest because it's easier to test whether generated code works than to evaluate subjective writing quality.

This creates a massive opportunity in the current: the degree we can express our goals as coding problems, LLMs will excel at solving them.

The Agent Manager Revolution

We’re now at the point of the timeline where the LLM evolution is fundamentally changing how we work.

Early AI adopters are evolving into "agent managers" instead of just traditional knowledge workers.

Think about your typical workday. Instead of manually creating spreadsheets, writing reports, or building presentations, you're increasingly learning to communicate objectives to AI systems that break down complex goals into executable tasks.

LLMs are starting to act like project managers: They understand customer needs, translate them into specs, generates code to meet those specs, and execute solutions.

Those who understand how to manage and leverage agents are already in high demand. It's creating entirely new career opportunities across every industry:

• For developers, the role shifts from writing code to architecting systems and managing AI agents that write code.

• For entrepreneurs, traditional barriers to starting businesses are crumbling.

• For knowledge workers, the future belongs to those skilled at delegation and orchestration rather than hands-on execution.

And it’s not just efficiency - it's about the fundamental expansion of what's possible for anyone to accomplish. We're moving from a world where your output was limited by your personal skills to one where it's limited by your ability to steer AI agents.

The $1 Billion One-Person Company

At Anthropic's developer conference, Dario Amodei (the CEO of Anthropic) was asked when we'd see the first billion-dollar company with just one human employee.

His response: "2026."

While I don't have a confident timeline, I'm also certain this will happen soon.

A single entrepreneur, armed with a fleet of AI agents, building and operating a business generating millions in revenue seems likely.

The agents handle everything from product development to customer service, marketing to operations, while the single entrepreneur, or “CEO of Agents” would handle vision and agent orchestration.

It's the logical extension of what we're already seeing. AI systems can already code entire applications, manage complex workflows autonomously, handle customer interactions at scale, and optimize operations in real-time.

We're living in what I believe is the golden age of entrepreneurship. AI will allow thousands of solo builders to pursue business opportunities that weren't possible before.

Traditional constraints are dissolving:

• Need technical skills? AI can code

• Can't afford staff? AI handles operations

• Lack expertise? AI fills knowledge gaps

The person with the best ideas and ability to orchestrate AI agents effectively will have advantages that even well-funded teams couldn't match just a few years ago.

Anyways, that’s this weeks food for thought.

And, as I hit publish on this article, I’m already tinkering with OpenAI’s new o3-pro model, which is another major leap. But will save that for next week once I’ve had a chance to dig in with it a bit.

Time to get back to building!

—Dharmesh (@dharmesh)