Lena, or the road not taken

A short sci fi story, Lena by Sam Hughes (2021) is about Miguel Acevedo, the first person to have his brain scanned and made available online as an executable file: MMAcevedo. In the story millions of MMAcevedo instances are run at any given moment, used for research and practical work. The title is a reference to a famous 1972 image of the model Lena Forsén from a Playboy magazine centerfold. It was used as a reference in thousands of research papers on image processing until falling out of favor in the 2010s. The digital photo of Lena is thus paralleled to the digital copy of Miguel’s brain -- first, in real life, an image, then, in the story, a mind.

Around the year 2000, digital brain uploads and AIs powerful enough to pass the Turing test were two equally remote science fiction technologies. A futurist from that time faced what seemed like two divergent paths to machine intelligence. A sensible futurist would bet against either technology arriving within their lifetime. A quarter century later one but not the other has come to pass -- machine learning has led to digital intelligence, brain measurement has not.

At the time when the story Lena takes place, after the death of the biological Miguel in 2073, digitized human brains (though not necessarily MMAcevedo himself), provide economically valuable labor. They are used for tasks, such as vehicle operation, report generation or language translation, with compliance compelled offstage. By implication the uploads are more capable or cheaper to run than non-upload AIs available at that point in the future. That same premise is studied from a serious economics perspective in the book the Age of Em (2016) by the economist Robin Hanson. The ems (whole-brain emulations) are analyzed as a highly scalable form of labor that can be run faster or slower depending on demand and hardware. A more dramatic version of the upload scenario in the animated TV series Pantheon describes an alternative history where brain upload technology is developed in 2019 and the uploads, sometimes violently obtained, are exploited as cheap cognitive labor. A technological singularity resulting in the majority of people migrating online follows soon after.

The world of uploaded brains performing labor has not come to be. Unlike these imagined worlds, there is as yet no brain upload technology. In contrast, current AIs, still in their infancy, already match or exceed the capabilities of the fictional uploads for many of the tasks these stories mention. Report generation, software development, language translation and machine operation are all being automated. There is unlikely to ever be economic demand for the abilities of simulated human brains. From our current 2026 vantage point it is hardly surprising. A simulated human brain still carries the “legacy code” from the long arc of evolutionary survival. There is no reason why evolution would select brain architectures particularly suitable for our latter-day tasks. While human minds are uniquely flexible -- no other species is capable of doing arithmetic -- as calculators we leave much to be desired. Using simulated brains to control drones or write computer code makes no more sense than using human brains to perform long division.

Of course, hindsight is always keen. In 2000 engineered intelligence appeared several major conceptual breakthroughs away. Solving math competition problems looked as removed from classifying grayscale images of digits as Einstein’s relativity was from the Copernican system. In contrast, brain mapping offered a long, thorny but plausible roadmap to digital intelligence -- worm, fly, fish, mouse and, ultimately, a human. Yet this was not a road to be taken. Brain recording technology has unquestionably made great progress in recent years. The current frontier is a complete connectome of a fruit fly brain. With 140,000 neurons and over 50 million synapses it is a remarkable achievement, but far short of a zebrafish, let alone a human. While scientists are still scanning the fly, the world is increasingly run by intelligent machines. To be fair, we have not learned to engineer intelligence from first principles, the way we once imagined. Instead, to our mounting awe and unease, we discovered that intelligence could be grown using nearly the same methods that were available in 2000, only deployed at a far larger scale. Like flowers from dirt, intelligences emerge from large volumes of unstructured data and computation, through crude optimization methods.

Perhaps one can breathe a sigh of relief. Uploads will not be forced to work, and digital slavery will not come to pass -- not because of ethical considerations, but because, just like physical bondage, it becomes inefficient with better technology. In the digital realm, the obsolescence arrived before human servitude could even take place.

In the end there never was a fork in the road, whatever the traveler might have imagined. One path demanded impossible precision; the other followed larger forces already unfolding. There is no need to simulate a human in order to distill what humans do. The economic future belongs to beings with different architecture. Uploaded or flesh and blood, Miguel will not need to drive a truck.