Before Writing AI 2027, He Predicted 2026. How Accurate Was He?

Daniel Kokotajlo revisits his 2021 essay "What 2026 Looks Like" and argues that narrative-style forecasting turned out to be a far more powerful tool than expected. Even before ChatGPT, he reasoned along the continuum of technological progress about what the next stage of AI development would look like, and he believes reality has come very close to that story.

There's no shortage of things he got right. Trends like the AI boom and high revenue of 2023, training-cost recoupment on the scale of 100 million-plus, and US-China chip regulations were quite accurate, direction included. Particularly striking is his view that by 2025, what would matter more than the race for model size would be agent structures like "bureaucracies" that keep models running for extended periods—in other words, today's agent frameworks and scaffolding.

On the other hand, there are clear misses too. He expected a slowdown in new large models and easing of chip shortages in 2024, but in reality, efficiency gains progressed while even larger models kept appearing, and supply shortages—including high bandwidth memory—grew worse. He also predicted that AI would sharply split the online information environment into different tech stacks with strong censorship and propaganda systems; while some of this occurred, it didn't unfold as dramatically as expected.

He's skeptical of explaining this gap simply as 'the physical world is slow.' If anything, AI's social penetration was faster than expected, and he believes he actually underestimated the level where "hundreds of millions of people regularly talk with chatbots." In conclusion, his argument is that when looking at the future, pushing a narrative all the way through to what structures and usage patterns a technology will produce at its next stage provides more insight than simply pinning down a single number.