Latest in AI
Every brief, newest first. Each one links to its primary source.
-
A new benchmark tests whether agents can turn an expert’s plan into a fast GPU kernel
D2K-Bench, published through Alibaba’s Qwen GitHub organisation, runs 26 kernel-writing tasks twice, with and without expert design guidance, and finds the guidance roughly doubles…
-
Brussels names ten finalists for its Apply AI Startup Award
The European Commission has published the ten-company shortlist for its Apply AI Startup Award, with three winners to be picked after a pitching round…
-
TikTok launches an AI shopping assistant that can complete purchases in chat
TikTok’s new Shopping Assistant answers product questions and can finish a purchase inside the conversation.
-
Bilibili open-sources a 150-language translation model family
The Chinese video platform Bilibili has released its Index-Translate models under Apache-2.0, covering 150 languages across 2B, 9B and a preview 35B mixture-of-experts configuration.
-
A legal research benchmark marks the best agent fully correct on 43% of questions
A new benchmark marks a legal research answer correct only if every required point is covered and every citation checks out, and the strongest…
-
AWS pledges $1bn to data centre communities and says it has dropped government NDAs
Amazon Web Services has committed $1 billion over five years to American communities hosting its data centres and says it no longer uses non-disclosure…
-
Seventeen countries sign a Kyoto declaration on using AI in science
Seventeen governments have endorsed the Kyoto Vision for a Golden Age of Science, promising researchers wider access to AI tools, data and computing infrastructure,…
-
Trump announces a Super Intelligence Force, with no charter published yet
President Trump used a Truth Social post on Sunday to announce a federal Super Intelligence Force, but no executive order, charter or membership list…
-
A new benchmark spots AI-written papers by their broken reasoning
Researchers in Seoul and Minnesota say AI-generated academic papers give themselves away not through their prose but through arguments that fail to connect, and…
-
DeepMind watermarks AI-designed proteins without breaking them
Google DeepMind has published a method in Nature for embedding detectable watermarks into AI-generated protein sequences and structures while leaving their biological function intact.