Former OpenAI researcher Gabriel Petersson launches Energy for AI-driven computer work

Why it matters

Energy is a bet that the next large AI company can be built around user workflow and distribution, while model providers supply the underlying intelligence.

Gabriel Petersson (@gabriel1), a former OpenAI researcher credited on Sora 2, launched Energy in a thread on X on August 6th. Petersson is pitching Energy as an AI desktop app that can compress days of computer work into hours and make agent-driven workflows accessible beyond software development. Energy's website describes the product as a way to "work at the speed of thought." (getenergy.com)

https://x.com/gabriel1/status/2085418582192841147

Petersson arrives with firsthand experience building frontier generative systems. OpenAI's Sora 2 release lists him among the model's research contributors, and his public profile identifies earlier work at Midjourney. His route into AI ran through product engineering and startups instead of a university program. In a 2025 interview, Petersson discussed leaving high school as a teenager and learning technical subjects by starting with real problems, then filling gaps as the work demanded. (openai.com)

Energy extends that method into a product thesis: give users a desired outcome, let AI perform more of the intermediate computer work, and reduce the amount of specialized knowledge required to operate software. Petersson compared using Energy to the shift already underway in programming, writing that working without it feels like coding without AI. He also claimed early users complete "days of work in hours." Both productivity statements come from Petersson and are not accompanied by independently measured results in the launch material. (x.com)

A fast turn from OpenAI to a general-purpose agent

Petersson said in the launch video that he quit OpenAI three months ago. His public resignation post was dated June 7th, roughly two months before Energy's launch. In that post, he said he had left the Sora group earlier in 2026 to form another team inside OpenAI, then resigned because he wanted to build one final product before artificial general intelligence. (digg.com)

The compressed timeline suggests Energy was developed before Petersson publicly announced his departure, or emerged from work that began during his final months at OpenAI. Petersson has framed founding a company as an eventual certainty. In Thursday's launch thread, he wrote that he had been waiting for an opportunity large enough to justify the commitment.

That opportunity, as Petersson defines it, is the interface through which people direct AI across their computers. In hiring posts published after his resignation, he argued that coding had shifted from manually writing every line toward describing outcomes to agents. He predicted that the rest of knowledge work would follow and said his team intended to build the interface for "all of computer work." (digg.com)

This puts Energy in a different layer of the AI market from the labs training foundation models. Energy is betting that model capability has advanced faster than ordinary users' ability to deploy it. Petersson made that argument explicitly in July, when he wrote that AI's economic impact was increasingly constrained by how well people understood and pushed the systems available to them. Energy turns that diagnosis into a distribution strategy: package agent behavior inside a desktop product and teach users to delegate work that they still handle manually. (digg.com)

Petersson is recruiting for intensity

Petersson said Energy has hired the strongest people its founders know, though the launch thread names only him. The thread also doubles as a recruiting call. He asked candidates who have built "impressive things" and pushed AI agents hard to send concise descriptions of their three best projects.

The hiring language matches the culture Petersson advertised while Energy was still unnamed. He sought people who keep work "top of mind," enjoy direct product discussions and want to build and sell in person. Earlier posts said the group planned to operate from San Francisco after resolving visa logistics and was temporarily flying candidates to Sweden for interviews. (nitter.app)

Energy's central challenge is product design, not access to another chatbot. Coding agents gained traction because repositories, terminals, tests and version control give models a structured environment with observable results. General computer work spans documents, spreadsheets, browsers, communication tools and company-specific processes. Energy must create enough common structure across those workflows for an agent to act reliably while keeping the user in control.

Petersson is betting that the adoption gap can be closed through a new interface. Energy will be judged on whether it can turn the techniques used by aggressive AI users into repeatable workflows for people who do not spend their days testing models. That is the route from a former Sora researcher's personal workflow to the broad computer-work platform he says he has been waiting to build.