The AI Experts Who Couldn’t Predict AI – Grey Enlightenment

Skip to content

The article, "A Taxonomy of Omnicidal Futures Involving Artificial Intelligence," went viral. It was co-authored by Jacob Tsimerman, who, as the media has made abundantly clear, is the 2026 Fields Medal recipient–as if that somehow adds credibility to the article or grants him immunity from criticism. Despite being published on arXiv, which has a focus on more technical content, it basically reads like a LessWrong post, and I’m not sure how it got past the moderators. It shows that name recognition can allow people to play by a different set of rules.

Moreover, the fact that one of the authors has a Fields Medal is irrelevant if the arguments are unconvincing. As someone who reads a lot of content online, I encounter a lot of forecasts and opinions, particularly about technology and AI. I have found that credentials and fame are poor predictors of forecasting ability or policy efficacy. No one does it that well. Even I have missed many things.

For example, former Stanford University professor Paul R. Ehrlich gained worldwide recognition with a 1968 book predicting population collapse. Decorated military experts in the early 2000s posited a dubious link between nonexistent weapons of mass destruction in Iraq and al-Qaeda. The "greatest minds" in finance and economics were blindsided by the ’08 financial crisis. Well-credentialed health experts during COVID put too much faith in ill-conceived lockdowns and other measures that proved ineffective at stopping the spread of the virus, while causing great inconvenience and harm to society.

For Dr. Ehrlich, despite being wrong, the idea of a "population crisis" became ingrained or implanted in the collective public and intellectual consciousness, remaining a central point of discussion long after it became clear that his book was wrong. Same for the Iraq war, which would drag on for another 20 years even though the intelligence was wrong from day one. The same thing has happened with the notion of "AI risk," which has become a constant refrain in the media for 3 years now.

AI "safetists" like to frame themselves as outsiders fighting against a pro-AI hegemony, but their position is the mainstream, as reflected in publications such as Time, The New York Times, and The Atlantic. USA Today, just two days ago, published the headline "The AI reckoning is here, and America is completely unprepared." It’s much harder, if not impossible, to find articles that are unhesitatingly optimistic about AI than articles that are skeptical or outright hostile toward it–whether blaming AI for academic dishonesty, declining literacy, joblessness, or other social ills.

I could be wrong. Maybe the AI is the lone exception of a crisis that "everyone" sees coming and then actually happens. But simply defaulting to "listen to the experts" isn’t useful either. As the examples above show, experts have a track record that ranges from mediocre to downright bad when it comes to forecasting and assessing risk in general. Take, for example, LessWrong. Since its founding in February 2009, the site has accumulated 45,933 posts and 906,338 comments from a wide range of contributors, including many of the leading–and would-be leading–figures in AI and technology. And that’s not even counting the millions of hours of podcasts and articles produced by the broader mainstream over the past 15 years.

AFIK–or, to the best of my knowledge–no one has a consistently good track record on AI. In not one of those nearly one million blog posts or comments did anyone predict the rise of LLMs as we know them today. No one predicted between 2016 and 2020 that Nvidia stock would surge due to worldwide demand for GPUs to run language models. After Nvidia stock had already surged in 2023-2024, no one was saying in 2025 that there would be a global memory shortage–it even has a Wikipedia page–leading to huge rallies in stocks such as Micron and SanDisk.

Recall that between 2015 and 2020, the main focus was on machine learning (remember that?). AlphaGo and TensorFlow were a huge deal. Professors were taking sabbaticals to consult on recommendation algorithms or self-driving cars. The biggest threat in 2020–2021 was "deepfakes," not self-improving agentic intelligence or mass job loss.

Even when LLMs became "a thing," no one in 2022-2023 was saying that AI would automate math research or coding at scale. The main concern was AI devaluing artists or infringing on copyrights, as it excelled at image and text generation. Math and coding were considered too advanced. For example, The Guardian published the headline, "‘It’s the opposite of art’: why illustrators are furious about AI" in January 2023. This was a huge deal at the time, which seems quaint today now that AI is disproving century-old conjectures.

Or, further back, experts underestimated the capabilities of computers in the 1950s, before the "transistor revolution." Conversely,...