Professor James Whittington and Dr Ruairidh Battleday unveil breakthrough from stealth frontier AI lab | Sifted

Two professors have emerged from something called a "stealth frontier AI research laboratory" to announce a breakthrough, and the lab is named Prefrontal — the part of the brain that handles planning and impulse control, which is a genuinely funny name for an outfit publishing speed multipliers with no accuracy figures attached. The professors are James C. R. Whittington and Ruairidh McLennan Battleday, who between them carry six initials and one McLennan, which is how you know the paper is serious. Their method is Latent Sparse Attention, and the result is 72.4x.

The pitch: group related information into clusters "by meaning rather than position in a sequence", which is what humans do when they file a gas bill beside a takeaway menu and call it a system. Sifted prints 72.4x, not 72, because a decimal point is a personality and nobody rounds down a headline. The other figure is "up to 9.2x faster time to first token in an end-to-end model-serving evaluation", and "up to" is doing the sort of lifting that normally requires a forklift certificate. Note what is being measured: prefill speed at the attention layer, and time to first token. Not answer quality. Not long-context accuracy. Not whether the clusters by meaning actually mean anything.

"Stealth" here means the article can confirm the laboratory exists and cannot tell you who funds it, how many people work there, or what it intends to sell — a disclosure profile generally associated with a car boot sale. "Frontier", meanwhile, has quietly become the word we use for "pre-revenue". None of this makes the research wrong; sparse attention that clusters by meaning is real work and 72.4x is not nothing. But we have been here with every efficiency paper since about 2019: the multiplier lands, the caveats live in section six, and the "breakthrough" in the headline turns out to be a serving benchmark with a decimal point.