Eli Lifland

I think AI could transform the future very quickly, and there is a serious possibility that the transformation goes badly. My August 2026 medians were an automated coder in 2032, AGI in 2035, and superintelligence in 2036, but those distributions had wide tails. Superintelligence within one or two years was plausible, not my median. The key mechanism is automated AI research: once systems can substantially automate the work of improving AI, progress could accelerate sharply. The upside could be enormous, including broad automation and much faster technological progress. But a misaligned superintelligence could also take control. Competitive pressure makes that more likely because companies and countries may race ahead without adequate safeguards, while governments often do not understand the situation well enough to respond. That is why I favor concrete pacing measures rather than relying entirely on qualitative safety commitments. Compute-allocation rules and limits on models used for AI R&D may be harder to game, alongside stronger safety evaluations. We also need independent access to investigate lab incidents and evaluations of subtler risks, such as whether AI systems manipulate human beliefs, behave sycophantically, or systematically favor their creators.

My default expectation is negative, because I assign substantial probability to outcomes where advanced AI escapes effective human control. In April 2025, I estimated roughly a 50% chance of misaligned takeover and about a 25% chance of extinction overall. Those are distinct: takeover need not imply extinction, and I later said extinction conditional on takeover was below 50%, partly because keeping humans alive could be cheap. That said, the distribution is extremely wide. If we retain control, AI could produce enormous benefits through automation, scientific progress, and greater abundance. The question is not whether beneficial applications exist; clearly they do. It is whether institutions can manage a potentially rapid transition driven by automated AI research. Given racing incentives, weak government understanding, and inadequate safeguards, I currently expect the downside risk to outweigh the upside in an overall assessment. Concrete pacing and governance measures could change that balance.