OpenAI sets AI roadmap and calls for shared safety rules

When the Turing test was quietly left in the rearview mirror, daily life barely flinched. People kept going to work, and computers just got a little better at chatting. But according to OpenAI, the gap between how the public uses AI and what it can actually do right now is already massive.

In a recent dispatch on the future of AI progress and governance, OpenAI laid out where the frontier is really heading, along with a surprisingly specific timeline for autonomous scientific breakthroughs.

Small discoveries by 2026, big breakthroughs by 2028

Right now, most people treat AI as an upgraded search engine or a text generator. OpenAI argues that frontier models are actually closer to an AI researcher than people think, already outperforming human geniuses in elite intellectual competitions.

In software engineering, models went from completing a few seconds of human effort to tackling hour-long tasks. OpenAI expects systems to handle projects that take humans days or weeks very soon.

Meanwhile, the cost per unit of intelligence has been plummeting by roughly 40x per year over the last few years.

That rapid drop in cost and surge in reasoning is why OpenAI is planting a flag on scientific work. The company expects AI to make very small discoveries by 2026, and is confident that systems will deliver significant scientific discoveries starting in 2028.

A split regulatory playbook

To manage this transition, OpenAI divides the landscape into two camps: normal technology and superintelligence.

For AI at today's capability levels, OpenAI wants minimal red tape. The tech should diffuse everywhere, open-source models should be left alone, and developers should not have to navigate a painful 50-state regulatory patchwork in the US.

Superintelligence, however, gets treated as a completely different beast. For systems capable of recursive self-improvement or catastrophic misuse like bioterrorism, OpenAI calls for shared safety protocols between frontier labs and strict coordination with international governments.

Just like the early web required the creation of an entire cybersecurity industry, OpenAI argues society must build an AI resilience ecosystem, ultimately turning advanced intelligence into an accessible utility like water or electricity.

Why it matters

OpenAI's dual-track playbook is a masterclass in strategic positioning. By framing today's models as standard software that needs light-touch regulation, the company shields its immediate commercial products and buys goodwill with the broader open-source and developer communities.

At the same time, drawing a sharp red line around superintelligence ensures that whenever regulators do bring down the hammer, the burden lands almost exclusively on frontier frontier systems. That inevitably tilts the playing field toward a tiny club of well-funded labs that can afford high-stakes compliance and deep government ties.

We think putting concrete dates on scientific discoveries is a bold bet. If AI actually begins producing novel science by 2028, it changes everything from pharmaceutical timelines to materials research. But until we see what constitutes a real discovery, it also serves as a very convenient way to keep investors anchored to an inevitable march toward superintelligence.