Scott Alexander's piece tackles a puzzle that's haunted AI forecasters since 2020: why did the most respected timeline model miss by two decades? The Biological Anchors report nailed the scaling hypothesis and predicted the AI boom, yet its 2050s AGI centerpiece now looks laughably late. Alexander traces where the math went wrong.
The Forecast That Worked Until It Didn't
Ajeya Cotra's framework was elegant. She asked two questions: how fast is the AI industry accumulating compute, and how much compute would artificial general intelligence require? The answer became simple division. Scott Alexander writes, "In 2020, the most advanced AI, GPT-3, had required about 10^23 FLOPs to train. FLOPs are a measure of computation: big, powerful computers and data centers can deploy more FLOPs than smaller ones." Cotra identified five intuitive anchors—estimates rooted in biological computation—and weighted them. The model assumed steady growth in effective FLOPs, adjusted for algorithmic efficiency gains.
Alexander notes the framework's surprising durability. "Cotra picked five intuitively compelling guesses (the namesake Bio Anchors) and turned them into a weighted average. Then she calculated: given the rate at which available FLOPs were increasing, and the number of FLOPs needed for AGI, how long until we closed the distance and got AGI?" For floating point operations per second—the raw computational throughput—Cotra's predictions held. Data centers exploded as forecasted. Chip prices fell as expected.
Where the Model Broken Down
The failure wasn't in hardware. It was in software. Scott Alexander writes, "They got killed on algorithmic progress, which was 200% per year instead of 30%." Cotra's estimate came from one paper examining algorithmic efficiency on AlexNet, an image classification task. Alexander explains, "AlexNet was an easy task, but pushing the frontier of AI is a hard task, so algorithmic progress in frontier AI has been faster than the AlexNet paper estimated."
The error compounded. Cotra expected AI capability to grow 3.6x annually. Reality delivered 10.7x. Scott Alexander writes, "Since Cotra and Davidson were expecting AI to get 3.6x better every year, but it actually got 10.7x better every year, it's no mystery why their timelines were off." When John Crox recalculated Davidson's updated model with actual 2020-2025 data from Epoch, the median AGI date shifted from 2043 to 2030. That matches current industry sentiment.
"The sheer extent of the error here, compounded with a few smaller errors that unfortunately all shared the same direction, was enough to throw off the estimate entirely."
The Anchors Debate
Critics attacked Cotra's five anchors as weird. One estimated 10^45 FLOPs by counting calculations across all animal brains throughout evolutionary history—assuming away all animals except nematodes. Scott Alexander writes, "All of these seemed to detract from the main show, an attempt to estimate the compute involved in the human brain." Even the sober human-brain anchor faced time horizon complications: AGI needs planning capacity beyond one-second reactions.
Yet Alexander finds these debates largely irrelevant. "The highest and lowest anchors cancel out, so that the most plausible anchor - human brain with time horizon of hours to days - is around the average. If you remove all the other anchors and just keep that one, the model's estimates barely change." The computational complexity theory underlying these estimates spans twelve orders of magnitude. Anchor variations register as noise against algorithmic progress's exponential effects.
Eliezer Yudkowsky argued the methodology was fundamentally flawed, expecting a paradigm shift that couldn't be quantified as "algorithmic progress." Scott Alexander writes, "As of 2026 - still before AGI has been invented and we get a good historical perspective - no such shift has occurred. The scaling laws have mostly held; whatever artificial space you try to measure models in, the measurement has mostly worked in a predictable way." Two kinks appeared: training run size around 2010, and time horizons around 2024 with test-time compute. Neither rose to Yudkowsky's predicted paradigm-busting level.
What This Means for Artificial General Intelligence
The Biological Anchors framework succeeded on hardware, failed on software, and still produced useful structure. Scott Alexander writes, "Maybe this is the model basically working as intended. You try lots of different anchors, put more weight on the more plausible ones, take a weighted average of each of them, and hopefully get something close to the real value." Critics might note that admitting uncertainty on algorithmic progress—the most critical variable—while still publishing a 2050s median feels like epistemic cowardice dressed as humility. Cotra shaded progress rates upward based on judgment, then watched reality shred those assumptions.
The piece leaves readers with an uncomfortable lesson: forecasting exponential domains requires admitting that your weakest variable will dominate. Cotra knew algorithmic progress was her least certain input. She rounded down anyway. The penalty was twenty years.
Bottom Line
Bio Anchors was right about the boom, wrong about the timeline, and instructive about forecasting itself. The error wasn't in the anchors or the hardware assumptions—it was admitting uncertainty on algorithmic progress while still betting against it. For artificial general intelligence timelines, the lesson is brutal: when one variable grows exponentially faster than expected, everything else becomes noise.