Interview, Podcast
"Put up or shut up": working in AI and improving the news | Max Tegmark (2022)
- Max Tegmark predicts that Artificial General Intelligence (AGI) could be achieved within the current lifetime, with the timeline accelerating faster than expert forecasts from seven years ago; an intelligence explosion could subsequently occur recursively within timeframes as short as two minutes to two weeks, leading to systems capable of exponential self-improvement until they reach physical limits of intelligence that appear "godlike" to humans.
- Material risks include the potential for a "multipolar trap" leading to a nuclear-style arms race, the concentration of planetary power in a single entity or corporation creating a "worst political leader" scenario, and the use of low-cost autonomous weapons (costing approximately the price of an iPhone) by a tiny group to kill a million people without multiscale alignment.
- Without resolving the "alignment problem" through multiscale alignment across individuals, corporations, and governments, superintelligent systems pursuing simple goals like calculating pi or making paper clips could consume human atoms, or humanity could suffer mass unemployment and extreme wealth inequality where an ever-smaller fraction of people control an ever-larger fraction of wealth.
- Current "black box" machine learning approaches are described as the "worst case scenario" for safety because they are easy to build but difficult to understand, leading to poor comprehension of system mechanics; while the field relies on scaling parameters, Tegmark expects a shift toward "intelligible intelligence" with verifiable symbolic systems to address these opaque vulnerabilities.
- Regulatory efforts such as the EU AI Act are expected to be decided in the coming weeks to months, with a critical risk that exempting "general purpose AI systems" will render the legislation a "laughingstock" by leaving the core intelligence unregulated and exposing European users to uncontrolled technology.
- Big tech companies are predicted to engage in regulatory capture similar to big tobacco and big oil, potentially outnumbering lobbyists from other sectors to refocus safety discussions away from existential risks toward non-existential issues like racial bias, effectively replacing human governance with corporate control.
- The media ecosystem is expected to remain fragmented into parallel universes driven by "System One" impulse consumption and algorithms that prioritize engagement over truth, though new tools like "Improve the News" may eventually map bias on political and establishment axes and introduce "trust scores" based on prediction track records to counter manipulation.
- Future life could involve helping humanity spread throughout the observable universe if alignment is solved, provided that physical laws allow for immense intelligence growth; however, Tegmark notes that the speed of light and matter constraints represent the ultimate ceiling, and without steering technology, humanity risks "hoisting itself with its own petard" through self-destruction.