Gwern — Anonymous writer who predicted AI trajectory on $12K/year salary

Gwern says anonymity’s biggest benefit is reducing the assumptions people attach to an identity. Without a name or public persona to categorize, readers have to engage with at least some of the work before dismissing it. He also values the practical protection anonymity provides, but says the chance to be heard is more important. The interview itself uses an avatar and altered presentation to preserve his identity.

On how companies will automate, Gwern expects adoption to proceed from the bottom up, with AI taking on tasks before human executives oversee AI-run firms. He speculates that people may retain an advantage in long-term vision and taste, using a Steve Jobs-style CEO as an example: AI systems execute and propose options while a human selects among them. He also imagines firms made from replicable teams rather than individual models. Because AI copies can preserve tacit knowledge and work together as packages, entire departments could be copied and varied as units of selection.

Asked when the trajectory toward a singularity could first have been foreseen, Gwern traces an early version to Samuel Butler’s 1863 writing about autonomous machine life. He also recounts Isaac Newton’s explanation for apparent technological progress: perhaps civilization periodically collapses and people merely rediscover earlier knowledge. Lucretius, he says, had offered a similar explanation for Roman innovation. These examples show how observers can explain away rapid change when the idea of cumulative progress is still unfamiliar.

Gwern’s proposed theory is that intelligence involves searching over many possible programs—his shorthand is Turing machines—rather than drawing on a single master algorithm or special “intelligence fluid.” On this view, more intelligent people have more compute available to search longer and across more possibilities. Large brains can eventually yield small, specialized solutions to individual tasks, which can then be recombined into what appears to be fluid intelligence. The theory also helps explain why general intelligence may be uncommon in evolution: for stable niches, directly hard-coding a specialized solution can be cheaper and more reliable than evolving a general learning system.

Gwern says that around 2005–2010 he considered AI far off, partly because he believed algorithms required difficult insights that compute alone could not supply. He followed Moravec, Kurzweil and Shane Legg’s forecasts but remained skeptical. Results such as AlexNet, DanNet and GPT-1 made him pay closer attention; GPT-2 produced a stronger reaction, while GPT-3 became the test he considered decisive. Its large scale-up and few-shot results convinced him that compute scaling was working, even as many commentators judged the model unimpressive because it was not state of the art.