AGI, agents, and timelines

AGI is coming—whether we’re ready or not. I’ve been convinced of this trajectory since GPT-3’s release, but recent developments have significantly accelerated my timelines.

The first major shift was OpenAI’s breakthrough in test-time compute and its newly demonstrated scaling law. Essentially, test-time compute allows a model to spend more time “thinking,” refining its response. OpenAI has shown that this approach scales linearly—the longer a model ponders, the better its answers.

The second game-changer came from an unexpected source: a Chinese company called DeepSeek. Practically out of nowhere, they open-sourced a state-of-the-art model rivaling the best closed offerings and published a paper detailing how they trained it using reinforcement learning (RL). RL has long been considered the key to achieving superhuman intelligence, but until now, its applications were mostly confined to narrow domains like games—most famously, AlphaZero’s complete domination of Go. DeepSeek’s breakthrough shows that RL works for language models too. Even more fascinating, chain-of-thought reasoning appears to emerge naturally from this process. This means we can now let models “play against themselves” to improve.

The third factor is simply the relentless march of scaling laws. Compute is getting cheaper, models are getting bigger, and efficiency improvements—like those pioneered by DeepSeek—are only accelerating this trend.

Taken together, these forces are pushing us toward AGI faster than most people realize. The next major frontier is agents, a concept that often confuses people but is, in reality, quite simple. An agent is just a small program designed to complete a task autonomously—whether it’s booking a flight, searching for hotels, handling inbound emails, or vetting candidates.

If you think about a company, it’s essentially a collection of hundreds of people executing thousands of these micro-tasks daily. Some are complex, like submitting a pull request to a codebase, while others are more straightforward, like an SDR sending cold emails. You can break down nearly any job into a collection of these agents.

So far, the main challenge with agents has been accuracy. Since they rely on executing a series of internal steps, any error along the way compounds, making them unreliable. But we are now on the verge of reaching the accuracy threshold needed for them to work. I expect a new generation of agents to emerge later this year.

And here’s the thing: once agents become viable for one task, they will rapidly generalize to countless others. This isn’t just an incremental shift—it’s a tidal wave of transformation about to hit society.

We are approaching the point where automating a software engineer—my personal benchmark for AGI—will become possible.

Yet, we are utterly unprepared for what’s coming. The U.S. political system is astonishingly ineffective, rife with incompetence at every level of government. Society remains deeply divided. While I’m optimistic about AGI’s potential to create abundance, the transition is going to be turbulent.

We’re at the edge of something unprecedented. Ready or not, AGI is coming.