More on the topic...
Generating detailed summary...
Failed to generate summary. Please try again.
Daniel Kokotajlo wrote a 2021 essay called “What 2026 Looks Like,” sketching out year-by-year advances in AI from 2022 through 2026. He treated it like a thought experiment: start with today’s capabilities, imagine plausible next steps and chain them together. He didn’t aim for pinpoint accuracy, but he did nail several big trends—massive revenue growth in 2023, U.S.-China battles over AI chips, and the shift from ever-bigger models to more complex “bureaucracies” of smaller agents interacting, essentially anticipating today’s agent-based AI frameworks.
By running through each year, Kokotajlo predicted that AI companies would recoup nine-figure training costs by late 2023, which matched OpenAI’s reported $2 billion ARR. He foresaw a chip export war between the U.S. and China in 2024—he pegged it two years late but got the core conflict right. His bureaucracy idea describes layers of AI submodels calling and reviewing each other, a clear precursor to today’s chain-of-thought prompting and agent scaffolding.
He admits some misses. He expected no bigger models in 2024 and thought chip shortages would ease faster. He also overestimated how quickly new semiconductor fabs would come online. Despite these gaps, Kokotajlo sees narrative scenarios as a useful complement to probability-based forecasts. Writing a full story forces you to spot plot holes, hedge against vague abstractions and surface hidden assumptions that short bullet-point predictions might miss.
Questions about this article
can you bullet the lists of thiings he got right, wrong, and in the middle?
Here’s a breakdown based solely on what Clara Collier and Daniel Kokotajlo discuss in their interview:
What he got right
• 2023 revenue explosion—OpenAI hitting ~$2 billion ARR in 2023 matched his “high enough to recoup >$100 million training costs” call.
• U.S.–China chip restrictions—a trade-war over advanced semiconductors, albeit in 2024 rather than his forecasted 2022.
• Agent “bureaucracies”—the move from big static models to multi-agent setups (what he called “bureaucracies”) lines up with today’s agent frameworks and chain-of-thought methods.
• Broad AI penetration—prediction of hundreds of millions talking to chatbots in 2026, when it’s already in the billions.
What he got wrong
• Fabs and chip supply—he overestimated how quickly new fabs would come online. In reality, most 2024 fab projects were delayed; high-bandwidth memory remained scarce.
• AI-driven propaganda surge—he expected a major rise in AI-enabled political propaganda, echo-chamber tech stacks, and polarized “Mormon Coalition”-style internets. That hasn’t materialized at scale.
• Bigger models in 2024—he forecast no substantially larger models that year, but researchers did train larger, more efficient architectures.
Where he landed in the middle
• Speed of AI adoption vs. physical-world constraints—he argued that “physical stuff takes time” isn’t a general brake on AI’s economic penetration (and he may have even underestimated software uptake). Yet on hardware build-out (fabs), the delays prove there are still real-world lags.
• Government and corporate misuse—he thought bad actors would swiftly deploy AI for persuasion, censorship, and microtargeting. That’s happening technically, but not yet at the feared scale or visibility.
If you want more context on any of these points, let me know.