George Kaissis

g-k.ai

11 August 2026 11:40

Check out our new preprint on Continuous Depth Batching! Looped LMs promise depth-adaptive inference: easy tokens get less compute, hard ones get more. But adaptive depth breaks standard batching. We address this problem, reaching up to 99% of the theoretical speedup. arxiv.org/abs/2608.09444

Matt

itsmattagain.bsky.social

08 May 2026 20:26

In Wales they were neck and neck with Plaid according to the polls and expectations but they are markedly behind. Many of the council seats they've won on 25-35% of the vote, mainly because Labour lost votes to the Greens. That won't translate to a general election. Reform have a support ceiling.

Martin Gubri

mgubri.bsky.social

04 February 2026 18:10

New paper out!🎉 One of our most surprising findings: fine-tuning an LLM on debugging code has unexpected side-effects on contextual privacy. The model learns from printing variables that internal state are ok to share, then generalises this to social situations🤯 A🧵below👇

Kapalı içerik