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

Oh my goodness, a *stealth* lab, which is the absolute best kind, because stealth means there is still so much delicious mystery left to unwrap! ✨ Professors James C. R. Whittington and Ruairidh McLennan Battleday — two cofounders carrying, between them, six initials and a McLennan, which is basically a whole extra syllable of credibility — have unveiled Prefrontal's research on Latent Sparse Attention, and it is *72.4 times* faster at attention-layer prefill than full attention. Not 72! Not 73! 72.4! That is a decimal point's worth of extra care, and somebody clearly measured it very lovingly. ✨

And the method groups related information into clusters *by meaning* rather than by position in a sequence, which is simply Marie Kondo for tensors, isn't it? Does this cluster spark meaning? Yes! Into the tidy box it goes! And because Prefrontal chose to report prefill speed and time to first token rather than, say, accuracy, we get to enjoy a whole beautiful *unopened* present of quality benchmarks to look forward to. Up to 9.2x faster time to first token — "up to" is such a generous little phrase, because it means a ceiling has been located and it is *high*! ✨

Now, some people will note that the article doesn't say who funds Prefrontal, how many people work there, or whether any code or weights accompanied the research — and to those people I say: what a *magnificent* amount of runway still to be revealed! Every stealth frontier lab is just a surprise party that hasn't been thrown yet. ✨ Two professors, a named method, two decimal-accurate speed multipliers, and a lab named after the brain's planning centre, promising impulse control we can all aspire to. The breakthrough has arrived; the accuracy numbers are simply waiting their turn; and honestly? 72.4x is not a number you round down. ✨