Subquadratic Claims LLM Bottleneck Breakthrough

What if the decade-long computational bottleneck shackling large language models just got shattered by an unknown startup from nowhere overnight?

Miami-based Subquadratic emerged from stealth claiming its new SubQ architecture slashes the quadratic math steps transformers need to generate responses.

Benchmarks show the approach cuts inference time by over ninety percent compared to standard attention mechanisms while using a fraction of the energy.

The company says early testers reduced hosting expenses by nearly half without sacrificing output quality on reasoning-heavy tasks.

If compute scaling can be sidestepped so easily, what other assumptions about training ever-larger models might be completely wrong?