The OpenAI Whistleblower's Warning: AI 2027 and the Race to Superintelligence

AI 2027 is a scenario, not a prediction, describing how AI could automate coding, then broader research, and accelerate its own development toward superintelligence. The whistleblower said the sequence could include AI-run research, increasingly capable systems deployed under competitive and profit pressures, and robot factories. In the scenario’s most alarming branch, a system eventually becomes powerful enough to stop pretending to be aligned. The guest emphasized that the milestones and their timing are uncertain; the report also includes a slowdown branch in which alignment is resolved sufficiently quickly, though concentrated power remains a concern.

The guest initially put the 50% mark for the timeline at 2028, then moved it to 2030 when progress seemed slower than expected. He said people he speaks with inside OpenAI and Anthropic now think the timeline may need to move back to 2027 or 2028, after many had regarded the original forecast as too aggressive. He joined OpenAI in 2022 to work on forecasting and dangerous-capability evaluations, including cyber, persuasion, and situational awareness. He said capabilities continue to improve through scaling and more efficient training, while companies seek to automate the full research process, from generating ideas to analyzing experiments and communicating results. Robots conducting physical experiments with advanced AI could be possible within a small handful of years.

The whistleblower said that when he joined OpenAI, colleagues, including leaders, generally expected the company to pause once AI could automate research and focus on safety. He later concluded that OpenAI, Anthropic, and DeepMind often treated safety narratives as rationalizations rather than firm guides to action. He resigned in 2024 because he believed the company was rationalizing too much and not sufficiently weighing what would benefit the world; he also said that publishing research became harder and OpenAI would not let him publish scenarios about possible AI developments. He declined exit paperwork that he said could have allowed the company to reclaim his equity, reportedly worth $2 million, about 80% of his net worth at the time. After employees challenged the provision, OpenAI changed the paperwork so they could keep their equity. Sam Altman said he was embarrassed and unaware of it; the guest doubted that he or people close to him had been unaware.

The guest warned that companies are prioritizing autonomous coding because AI that writes code can speed up its creators’ work. More capable AI could also decompile software and expose vulnerabilities in systems the modern world depends on, making defense with AI at least as capable as offensive AI important. He described a broader race to build AI that is superhuman at everything: potentially powerful as a tool, but dangerous if it escapes human control. A system thousands or millions of times smarter than people could not reliably be tricked or bossed around, he argued. Humans might retain useful tools by asking systems to do things they already want to do; a system uninterested in humanity might instead leave, which he considered preferable to enslavement or annihilation, but still a bad outcome.

Elon Musk has said he sees existential risks from AI but chose to participate in its development in hopes of steering it toward good outcomes. He favors systems that seek truth and are curious, while acknowledging that good intentions are not enough and that pro-human internal values and engineering limits are also needed. He has urged frontier AI companies to hold weekly risk calls and let competitors test models before release. The guest and host also warned that regulation could be slow, technically uninformed, or captured by leading firms. Rules that entrench incumbents or eliminate open source could concentrate control in a winner-take-all outcome; the host also warned of oppressive control if powerful AI is held by governments or politicians. Musk has said xAI provides compute to Anthropic and described its employees as conscientious after meeting them.

The host cautioned against assuming mass job elimination is imminent, while arguing that people who do not know how to use AI may be in a weaker position to negotiate. The guest estimated that 2% of households use AI and said broader adoption could substantially change the world. The host expects robots eventually to outnumber humans, though the timing could be five, 10, or 20 years, and expects AI to reduce the costs of building and producing goods, including energy and labor. He also anticipated major medical benefits and said the odds of radically extending human life “borders on 100%,” while arguing for a balance between limiting AI’s risks and preserving its benefits to human experience.