Debate Erupts Over Whether Compute Extrapolation Can Predict AGI

A debate over "whether compute extrapolation can predict AGI" erupted on X. It was sparked by a common line of argument: estimate the compute needed to simulate a human brain, assume compute grows at a constant or accelerating rate, work out when that threshold will be reached, and conclude AGI won't be far behind. On 09-06, phl43 systematically rebutted this argument, pointing out that our understanding of the brain is still very poor, so estimates of the compute required for simulation are almost certainly subject to enormous error; he conceded that continued compute growth is arguably the soundest assumption in such arguments, but the other links in the chain are unreliable.

Confirmed

• phl43 posted a systematic rebuttal of AGI predictions based on "brain-simulation estimates + compute extrapolation," arguing the model is almost certainly wrong
• stanislavfort joined in support, noting we don't even know how the brain implements the learning algorithms it runs—our understanding of the brain is extremely limited, which constitutes yet another counterargument to the compute-extrapolation case
• teortaxesTex pushed back, arguing our understanding of the brain is already sufficient, and that replicating the brain doesn't require knowing neuroscience—we won't copy the brain wholesale, and most key elements were already solved in work dating back to the McCulloch-Pitts era
• DeepMind researcher akarvonen voiced strong support for Kurzweil's 1999 AGI prediction, which was based on extrapolating compute growth and comparing it against the brain's FLOPS, saying that in hindsight this was pretty much all it took to predict AGI
• phl43 hit back at akarvonen, calling it a "terrible argument" and precisely the best example of what he was criticizing, and decided to write a dedicated post rebutting it in detail

Why it matters

The debate cuts to the methodological foundations of AI-risk and AGI-timeline forecasting: if gaps in our knowledge of brain mechanisms are enough to badly distort compute estimates for brain simulation, then the credibility of the many AGI predictions that rely on such extrapolation (including Kurzweil's classic claim) must be discounted. With phl43's planned dedicated rebuttal and a DeepMind researcher publicly siding with Kurzweil, the discussion has escalated from technical details into a representative clash of approaches, and is worth watching closely.

2026-09-06 ~ 2026-09-06 · 6 related posts

• [source] DeepMind's Karvonen: Kurzweil's compute-extrapolation was '~all you need' to predict AGI — a_karvonen · 2026-09-06
• [source] phl43 doubles down: Kurzweil's compute argument is 'terrible', a full post is coming — phl43 · 2026-09-06
• [source] Why compute-extrapolation arguments for AGI timelines are flawed, and why Kurzweil's 1999 bet was luck — phl43 · 2026-09-06
• AI timeline debate: is our brain understanding good enough to estimate emulation compute? — teortaxesTex · 2026-09-06
• We don't even know how the brain implements its learning algorithm, researcher argues — stanislavfort · 2026-09-06
• Do we need neuroscience for AGI? X debate says McCulloch-Pitts already covered the basics — teortaxesTex · 2026-09-06