AI is nearing 'research intern' level capabilities: OpenAI chief scientist
What's the story
OpenAI's Chief Scientist, Jakub Pachocki, has revealed that the tech giant is closing in on one of its major milestones: developing artificial intelligence (AI) systems capable of operating at the level of a research intern. Speaking on the Unsupervised Learning podcast, Pachocki noted advancements in coding, mathematical reasoning, and physics research as indicators that AI is getting better at performing complex technical tasks with less human guidance.
Milestone targets
Distinction between AI and automated researcher
Pachocki clarified that the main metric to look at is how long a model can operate mostly on its own. He said, "The way I would distinguish a research intern from a full automated researcher is the span of time that we would have it work mostly autonomously." OpenAI has set an internal target of creating an "AI research intern" by September 2026 and a fully autonomous AI researcher by March 2028.
Candid admission
We may totally fail at this goal: Sam Altman
Responding to Pachocki's comments, OpenAI CEO Sam Altman admitted that the company "may totally fail" at this goal. However, he emphasized the need for transparency given the potential impact of such technology. This acknowledgment highlights both the ambitious nature of OpenAI's objectives and the inherent uncertainties involved in pioneering work like this.
Technological advancements
Coding tools have evolved rapidly
Pachocki also highlighted the rapid evolution of coding tools such as Codex, which now do much of OpenAI's programming work. He sees math benchmarks as a "north star" for enhancing reasoning capabilities since they are easy to verify. Despite these advancements, he doesn't expect systems to be able to independently improve their model capability or solve alignment issues this year.