An AI Breakthrough Has Made Decades of Computer Science Research ‘Irrelevant’

Linköping University professor Peter Jonsson is acceptable to person his enactment presented successful 2 weeks’ clip astatine 1 of the astir prestigious conferences successful theoretical machine science. As of Monday, overmuch of that enactment is “dead,” helium says, aft AI showed that a 27-year-old proposal underpinning it is wrong.

Both OpenAI and Anthropic claimed this week to person disproved the exponential clip proposal (ETH). Vast swaths of theoretical machine subject probe connected complexity and algorithm plan person been built connected the premise that the effort it takes to lick definite problems increases exponentially with the size of the problem.

Jonsson expects his predicament volition beryllium “very common” astatine the IEEE Symposium connected Foundations of Computer Science, wherever his work volition beryllium presented by 1 of his coauthors. “People volition commencement the talks: ‘ETH is false, fractional of my insubstantial does not work,’” helium says. It’s a vexing problem, Jonsson says, “but astatine the aforesaid time, we larn a batch from it. And this is what subject is about.”

Researchers haven’t independently verified, fto unsocial had clip to afloat understand, the AI-produced proofs, which some firms person published connected GitHub successful some earthy connection and its machine-checkable Lean format. OpenAI has asked the UC Santa Barbara machine scientists who primitively projected the hypothesis, positive an MIT world who has refuted a mentation of ETH, to rally researchers to unpack the proofs. So far, this effort includes a Google signifier asking for volunteers and feedback, and a committedness from OpenAI to money a league connected the topic.

UC Santa Barbara prof Daniel Lokshtanov plans to articulation the effort, arsenic helium estimates the result—if confirmed—will render him “galactically wrong” and marque astir 50 of his published papers either “pointless” oregon successful request of a revisit.

“It’s conscionable a elephantine weaponry dropped into the mediate of the field,” helium says. “It’s truly chill and unspeakable astatine the aforesaid time.”

Thore Husfeldt, caput of the algorithms probe radical astatine the IT University of Copenhagen, had spent astir of this twelvemonth reasoning astir 1 azygous occupation successful his subfield. “As acold arsenic I tin tell, it’s present each irrelevant,” helium tells WIRED. Having archetypal formulated a mentation of the proposal present disproved by OpenAI, helium has been penning calculations connected his blackboard to fig retired what different parts of his enactment mightiness person go obsolete owed to the “once-in-a-generation result.” Unlike the ennui that has gripped mathematics researchers who person faced likewise field-upending proofs generated by AI models, though, Husfeldt expressed enthusiasm astir the caller results.

The uncovering is “fantastic,” helium says. “Being a mathematician is simply a batch similar stumbling astir successful a wholly acheronian room,” helium explains. “It's similar idiosyncratic conscionable illuminated an full subfield successful floodlights of knowledge, and present we conscionable tally astir similar overexcited bunnies and look astatine each the nooks and crannies that we tin abruptly explore. So it’s a ace breathtaking time.”