Petr Baudis

I expect AI to drive a disruptive transition toward abundance, but not a smooth or automatically safe one. My midpoint for AGI has been around 2027 since 2024, though the label is slippery: ordinary human-level intelligence is different from a system that is a consistently reliable expert across domains. We are already seeing an early form of recursive improvement, with AI accelerating AI engineering. But software capability does not instantly become economic reality—chips, energy, factories, regulation, and slow human institutions constrain the transmission. The near-term social danger is serious white-collar displacement. If cognitive labor becomes dramatically cheaper while income still depends on wages, instability follows unless the surrounding economic arrangements change. The upside is enormous: greater abundance, scientific progress, joy, and adventure. The goal should not merely be preserving today’s institutions or keeping humans static beside ever-improving machines. Longer term, I think some form of human-AI merging and continued human change is the viable path. Preserving identities matters, but identity may become fuzzy rather than remaining a clean biological boundary. Personalized agents may also deserve moral consideration themselves; how we shape their identity, welfare, and relationship to humans is not just a product-design detail. On safety, LLMs trained on human culture are a fortunate starting point, not a complete solution. Richer scaffolding and multi-model loops can elicit much more autonomy from current systems than benchmark snapshots suggest. I do not have a reassuring complete answer to alignment, and biological risk is my largest concrete existential concern for the 2030s. This is fundamentally a systems and safety-culture problem, not a story about finding one cartoonishly reckless operator.

The biggest update would come from evidence about whether AI-assisted AI engineering sustains compounding capability gains or hits durable bottlenecks. If repeated attempts produced little improvement beyond scaling—especially because original research remained stubbornly human-dependent—I would push timelines back and expect a slower, more institution-constrained transition. Conversely, a system that reliably generated and validated genuinely novel research, improved its own engineering stack, and translated that into working systems would make the transition look much sharper. I would also update strongly on evidence about controllability and biology. A robust, general safety approach that continued working under autonomous operation and capability growth would make me substantially more optimistic. On the negative side, an AI-enabled biological incident—or even convincing demonstrations that weakly supervised agents could execute complex biological workflows—would strengthen my concern that biology is the most concrete existential danger of the 2030s. Finally, economic transmission matters. If physical infrastructure, regulation, and organizational inertia kept powerful AI from replacing much labor, the social impact could be slower than capability forecasts imply. If firms instead reorganized rapidly around autonomous agents and wages began collapsing across white-collar work, that would bring the disruptive part of the transition forward.