The 'Galactic' Reset: AI's Assault on Computer Science Theory

In the venture world, we're used to 'disruption' meaning a better UI or a leaner supply chain. But as Wired first reported, OpenAI and Anthropic are claiming a disruption of a more fundamental kind: they say they've disproved the exponential time hypothesis (ETH).

For those not steeped in theoretical computer science, ETH is the 27-year-old premise that solving certain problems requires effort that increases exponentially as the problem grows. If this hypothesis is indeed false, vast swaths of research regarding algorithm design and complexity are effectively dead. The fallout is immediate and visceral. Peter Jonsson, a professor at Linköping University, told Wired that much of the work he is scheduled to present at the IEEE Symposium on Foundations of Computer Science is now "dead." Similarly, Daniel Lokshtanov, a professor at UC Santa Barbara, noted that if the AI-produced proofs are confirmed, he will be "galactically wrong," rendering roughly 50 of his published papers either pointless or in need of revision.

From a P&L perspective, the 'cost' of this breakthrough is being borne by the academic community. Thore Husfeldt, head of the algorithms research group at the IT University of Copenhagen, described the psychological toll as "taxing," noting that his department has even considered hiring an occupational psychologist to help staff dealing with the annihilation of their research projects.

However, the skepticism kicks in when you look at the delivery. Wired reports that OpenAI recently dumped 722 manuscripts covering 400 problems onto GitHub, a move that has annoyed academics who find the volume impossible to digest. More damning are the reports of errors in several of those documents. Some mathematicians view this as a hasty rollout specifically designed to "juice" OpenAI's planned initial public offering rather than a genuine effort to advance the field.

Despite the noise, there is a glimmer of utility. Victor Lagerkvist, an associate professor at Linköping, says he plans to revisit problems he previously abandoned because he assumed ETH was true. Lokshtanov warns that while this specific result doesn't directly break current cryptography, it should "scare" cryptosystem designers who rely on the assumption that certain problems take an impractically long time to solve.

OpenAI is attempting to manage the optics by funding a conference on the topic and asking researchers from MIT and UC Santa Barbara to help unpack the proofs, which have been published on GitHub in natural language and machine-checkable Lean format. Until independent verification is complete, we are left with a classic Silicon Valley tension: is this a "once-in-a-generation result," as Husfeldt puts it, or just another high-burn play for valuation growth?