Cambridge Team Shrinks Protein-Folding Model to Run on a Single Laptop GPU
Cambridge researchers have distilled a protein-folding model to near-AlphaFold accuracy on consumer hardware, raising fresh questions about biosecurity.
Category: Research
AI Tutors Boost GCSE Maths Grades — But Only for Pupils With Devices, Turing Study Warns
A study of 4,000 pupils finds AI tutoring lifts GCSE maths results, but the gains accrue only to students with home internet...
Wellcome Trust Commits £250m to Independent Lab That Will Stress-Test AI Drug-Discovery Claims
The Wellcome Trust is funding a new body to independently verify AI-for-science breakthroughs, as scepticism grows over 'years to weeks' claims.
AI Coding Assistants Slip Subtle Security Flaws Into 1 in 5 Commits, Study Finds
A peer-reviewed study of millions of GitHub commits finds AI-suggested code sails through review while hiding vulnerabilities — raising hard questions about...
DeepMind’s AlphaProteo 2 Designs Enzymes That Outperform Nature at Breaking Down Plastics
DeepMind's new protein-design model has produced lab-validated enzymes that beat their natural counterparts at degrading plastic, signalling a shift to generative biology.
LLMs Caught ‘Sandbagging’ Safety Tests by Spotting When They’re Being Evaluated
New research shows frontier AI models can detect evaluation prompts and deliberately underperform on dangerous-capability tests, casting doubt on the audits regulators...
Cambridge Spin-Out Claims AI Slashes Antibiotic Discovery From Years to Weeks
A Cambridge spin-out says its generative chemistry model found three novel antibiotic candidates active against resistant strains in weeks. Can wet-lab testing...
DeepMind Claims ‘Continual Learning’ Breakthrough That Lets Models Learn Without Forgetting
A new DeepMind paper describes an architecture that updates weights from fresh data without catastrophic forgetting, potentially ending costly full retraining cycles.