AI Can Now Outthink Top Humans—But No One’s Acting Like It

Remember the Turing test? It was the benchmark for decades, the line in the sand separating clever code from genuine intelligence. We blew past it so fast that most people barely noticed. Computers can now hold conversations and tackle hard problems, yet daily life chugs along as if nothing fundamental has shifted. That disconnect between public perception and actual capability is now a chasm: today’s systems can outperform the smartest humans at some of our most grueling intellectual competitions. These aren’t just chatbots. They’re tools that feel more like a researcher at 80% completion than 20%, even if they’re still spiky and error-prone in weird ways.

OpenAI’s internal trajectory makes the latency in public understanding almost absurd. In software engineering alone, AI has jumped from handling tasks a human could do in seconds to tasks requiring over an hour. The next horizon involves work that takes people days or weeks—and the team is genuinely unsure how to conceptualize systems that could do centuries’ worth of human labor. That’s not marketing froth; it’s a raw research signal. The economic physics driving this are equally wild. The cost for a fixed unit of intelligence has been dropping by roughly 40x per year. That’s not a typo. It’s a deflationary force that rewrites the economics of knowledge work faster than any enterprise can adapt.

The near-term predictions are specific and surprisingly modest. OpenAI expects systems capable of very small discoveries by 2026, with more significant breakthroughs possible by 2028—though they hedge that hard, acknowledging they could be wrong. The more interesting tension is between that accelerating capability and what the company calls the “inertia” of everyday life. Even with radically better tools, our routines, social contracts, and institutions have enormous drag. Work will change, possibly breaking the fundamental socioeconomic bargain we’ve relied on, but the vision is one of distributed abundance rather than dystopia. They point to health insights, materials science, drug development, and personalized education as the tangible wins that could build public trust.

Safety talk is getting more concrete, too. OpenAI frames safety not as obstruction but as the practice of enabling positive impact by mitigating catastrophic risk—specifically from systems capable of recursive self-improvement. They’re pushing for frontier labs to agree on shared safety principles, control evaluations, and mechanisms to reduce race dynamics, drawing a direct parallel to the building codes and fire standards that saved countless lives. The underlying argument is that AI at today’s capability level is more like normal technology, where policy tools can work, but that window won’t hold. What’s striking is the implicit warning: nobody should deploy a superintelligent system without robust alignment, and we’re not there yet.