Sam Altman and OpenAI chief scientist Jakub Pachocki published a joint essay on June 8, 2026, laying out what they call the company's third phase. The headline ambition: an automated AI researcher, and a target date of March 2028 for when a meaningful chunk of OpenAI's own research could be done by AI systems working alongside human scientists.
The whole thing opens with electricity. Specifically, electricity reaching a rural American town in the 1920s, before power lines meant hauling water and ending the day at sundown. It is a fine analogy, and OpenAI is not the first to reach for it. The argument in the company essay is that AI's value lives not in the technology but in what people do with it: navigating a medical bill, starting a business, caring for an aging parent.
Three phases, and where we are now
Phase one was research toward AGI. Phase two started when that research hit the real world and OpenAI became a product company, which is a polite way of describing what happened after ChatGPT turned into a household name. Now comes phase three, which Altman and Pachocki frame around making advanced AI cheap, abundant, and usable by basically everyone.
The most concrete line in the essay is the one about timing. OpenAI's internal belief, they write, is that by March of 2028 a significant fraction of its research could be done by AI systems in tandem with human researchers. Read that again. The company is betting that within roughly two years, AI becomes a core engine for building the next generations of AI.
That claim deserves a raised eyebrow. "Significant fraction" is doing heavy lifting, and there is no definition of what counts, no benchmark, no methodology. It is an internal belief, which is to say a forecast from the people with the strongest incentive to sound bullish. Self-improving AI research has been the industry's promised land for years and the timelines keep sliding.
The contradiction nobody at OpenAI mentions
Here is the part that is hard to read with a straight face. A large chunk of the essay argues that access to powerful AI should be spread as widely as possible, across people, companies, and countries, and that concentrating capability in a few hands makes the world fragile.
"A good AI future cannot be one where a small number of institutions control most of the capability and most of the upside."
Noble. Except OpenAI's own policy posture points the other way. In its March 2025 policy submission to the White House, the company pushed a tiered export framework that would cut adversarial states, China above all, off from American AI entirely. It also urged the government to consider banning models from Chinese labs like DeepSeek. Broad distribution of power, then, with an asterisk the size of a country.
Critics noticed the gap at the time. One analyst quoted in coverage of the proposal described the three-tier export scheme as Cold War cosplay. The essay does gesture at an eventual international body to coordinate frontier AI and reduce catastrophic risk, but coordination among allies is a narrower thing than the everyone-everywhere framing the document leads with.
What's actually new here
Stripped of the electricity metaphor, the news is the roadmap and the date. Three stated goals: build the automated researcher, accelerate the economy through faster science, and give everyone on Earth a personal AGI. The middle goal is vague enough to mean anything. The first one has a clock on it.
The essay also leans hard on safety language, insisting that fully automating everything is neither desirable nor safe and that humans should keep setting direction and making tradeoffs. That framing arrives at a convenient moment, given reporting that OpenAI has moved toward a public offering. Safety-forward positioning reads differently when there is an S-1 in the picture.
The next thing to watch is whether OpenAI puts any measurable definition behind that March 2028 target, or whether "significant fraction of research done by AI" stays as slippery as it sounds today.
