A young AI company is making a bold claim, and the field is debating whether to believe it.
According to MIT Technology Review's daily newsletter The Download, an AI startup called Subquadratic emerged from stealth last month asserting that it had broken through a bottleneck that has been holding back large language models (LLMs) — the technology behind today's chatbots and AI assistants.
Bottlenecks like this matter because they shape how fast, how cheaply, and how capably AI systems can run. When researchers talk about a limit on LLMs, they're often pointing to the heavy computational cost of processing information at scale. A genuine breakthrough on that front could mean more efficient models — but extraordinary claims from newly public startups tend to invite scrutiny, which is why the matter is being actively debated rather than simply accepted.
MIT Technology Review frames the Subquadratic news as part of a broader conversation about efficiency limits in AI, signaling that the company's assertion has not gone unchallenged.
The same edition of The Download also flags a separate but notable trend: brain-computer interface (BCI) trials are taking off. BCIs are systems that connect the brain directly to computers, and an uptick in trials suggests the field is moving from the lab toward real-world testing.
The source material here is a brief newsletter summary, so the deeper technical details behind Subquadratic's claim — and exactly what its purported solution involves — are not spelled out in the items available.
Why it matters: if a startup really has eased one of the core constraints on large language models, it could lower the cost and expand the reach of AI — but until the claim is independently verified, the debate itself is the story.