Back
Jan Leike – TIME100 AI 2024
webCredibility Rating
3/5
Good(3)Good quality. Reputable source with community review or editorial standards, but less rigorous than peer-reviewed venues.
Rating inherited from publication venue: TIME
This profile is relevant to understanding organizational dynamics at leading AI labs and the departure of key safety researchers from OpenAI in 2024, signaling internal tensions over safety prioritization.
Metadata
Importance: 42/100news articlenews
Summary
A brief TIME100 AI profile of Jan Leike, who resigned from OpenAI's superalignment team in May 2024, citing the prioritization of products over safety, and subsequently joined Anthropic to continue alignment research. The piece highlights his work on scalable oversight and his cautious optimism about solving the alignment problem before superintelligent AI arrives.
Key Points
- •Leike co-led OpenAI's superalignment team focused on keeping superintelligent AI under human control, but resigned in May 2024 over safety-vs-products tensions.
- •He accused OpenAI of failing to provide his team the compute needed despite a public pledge of 20% of resources to superintelligence alignment.
- •After leaving OpenAI, Leike joined Anthropic to continue alignment research, particularly scalable oversight techniques.
- •Scalable oversight aims to empower humans to give better feedback to AI on complex tasks beyond human comprehension, potentially enabling future automated alignment.
- •Leike expresses cautious optimism: 'You always feel like you're in a race against time. But I'm optimistic we can figure this problem out.'
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| Why Alignment Might Be Easy | Argument | 53.0 |
Cached Content Preview
HTTP 200Fetched Mar 20, 202610 KB
Sep 5, 2024 7:11 AM ET
# Jan Leike
by
[Harry Booth](https://time.com/author/harry-booth/)
Booth is a reporter at TIME.
Photo-Illustration by TIME (Source: Courtesy of Jan Leike)
A year ago, Jan Leike co-led OpenAI’s since-disbanded superalignment team with the company’s co-founder and chief scientist, Ilya Sutskever. As the most ambitious of the company’s three safety teams, the superalignment group was focused on ensuring that if AI systems surpass human-level intelligence, they remain under human control. But in May, Leike made a dramatic exit, [accusing](https://x.com/janleike/status/1791498174659715494) OpenAI of prioritizing “shiny products” over safety. He wrote that his team had been struggling to access the computing power needed for its research—despite OpenAI [dedicating](https://openai.com/index/introducing-superalignment/) 20% of its total computing resources to “solving the problem of superintelligence alignment,” which was the remit of Leike’s team. (OpenAI’s CEO Sam Altman [responded](https://x.com/sama/status/1791543264090472660) by thanking Leike and saying “we’re committed to doing \[more\].” In August, Altman [said the company was committed](https://x.com/sama/status/1818867964369928387) to allocating “at least 20% of the computing resources to safety efforts across the entire company.”)
Leike now helps lead alignment efforts at competing firm Anthropic. Although a lot has changed in the past few months, Leike’s mission to solve the alignment problem remains unaltered. For Leike, it’s not about which company wins. It’s about ensuring humanity navigates what he sees as an impending transformation driven by superintelligent AI systems.
“I think we made a lot of good progress in the last year,” Leike tells TIME—pointing in particular to an area known to alignment researchers as scalable oversight, which investigates techniques for empowering humans to give better feedback to AI models on complex tasks. The hope is that these tactics will allow humans to guide future systems, even as they do things we cannot fully comprehend. In the future, Leike says he believes aligning larger systems will increasingly be automated by smaller, trusted models, as the science of alignment becomes “more and more mature.”
“You always feel like you're in a race against time,” he says. “But I’m optimistic we can figure this problem out.”
_\*Disclosure: OpenAI and TIME have_ [_a licensing and technology agreement_](https://time.com/6992955/time-and-openai-announce-strategic-content-partnership/) _that allows OpenAI to access TIME's archives._ _Investors in Anthropic also include Salesforce, where TIME
... (truncated, 10 KB total)Resource ID:
2dac895d835536ca | Stable ID: M2IyY2E3ZT