DeepMind Just Dropped the Rules for Superintelligence (And Told Humans Not to Read It)

Co-founder Shane Legg and mathematical genius Marcus Hutter authored the theoretical blueprint mapping the transition from Artificial General Intelligence to Artificial Superintelligence.

The authors define AGI as merely matching a median human, but superintelligence must outperform a massive collective of ten thousand coordinated human experts working for an entire decade.

To keep the target from moving forever, this human benchmark is frozen at 2010-era technology, preventing a paradox where humans boosted by AI keep raising the bar of what "expert" means.

Four distinct development paths are laid out: brute-force scaling of data, inventing entirely new architectures to replace transformers, recursive self-improvement where AI designs better AI, and building massive multi-agent digital collectives.

This digital evolution faces severe physical bottlenecks, including running out of high-quality training data, massive electricity demands, and regulators pulling the plug before things get too weird.

Physics also places a hard cap on these virtual gods, as superintelligence remains strictly bound by the speed of light, thermodynamics, and Gödel’s incompleteness theorems.

In a bizarre meta-twist, the paper includes explicit instructions prompting AI assistants to summarize the text, with the authors openly suggesting that humans skip reading it altogether.

Scientific papers are now literally written by AI, for AI, with instructions on how to explain them to the flesh-sacks who funded the research. It seems humanity is rapidly transitioning from the creators of technology to merely the audience watching its self-guided evolution.

Source: arXiv