FundamentalsExplainer10 min
Fine-Tuning-as-a-Service: What It Is and What It's Used For
How hosted fine-tuning works under the hood, what organisations use it for, and why a few training examples can undo a model's safety.
Training
What goes in and how, which decides more about a model's behaviour than most architecture changes do. Includes the unglamorous parts: data, tokenisers, curricula.
Fundamentals4
FundamentalsDeep dive9 min
From Experts to Sub-Experts: Smarter Fine-Tuning for MoE Models
NSFT cuts each expert in a Mixture-of-Experts model into channel groups and trains only the ones that matter — beating LoRA and ESFT on accuracy per trained parameter, and forgetting less along the way.
FundamentalsDeep dive14 min
Dropout: Why Throwing Half Your Network Away Makes It Generalise Better
A deep dive into Srivastava et al. (2014) — the mechanism, the arithmetic behind weight scaling, what the benchmarks actually showed, and the tuning heuristics that still hold up.
Sebastiaan with AI
FundamentalsExplainer18 min
Tokens: the one idea that explains most of AI's odd behaviour
Why a model that can explain quantum tunnelling cannot count the letters in strawberry, why long conversations get slow and expensive, and why Dutch costs more than English. It all comes back to the very first thing that happens to your text.
Research4
ResearchPaper5 min
OmniEdu: Revolutionizing Education with Open Foundation Models
Exploring how OmniEdu's open foundation models enhance K-12 learning through improved problem-solving, curriculum grounding, and instructional support.
ResearchPaper4 min
Unmasking the Automaton: How Additive Input Pathways Skew State Tracking in Linear RNNs
A study reveals that additive input pathways in Householder Linear RNNs act as parasitic attractors, destabilizing state-tracking capabilities.
ResearchPaper7 min
Why Don't Machine Learning Research Agents Overfit?
New research from Amazon Science offers a surprisingly elegant explanation for a decade-old puzzle: good solutions are simply too short to cheat with.
Sebastiaan with AI
ResearchNews4 min
World-Time Compute: Lifting AI Generalization with Verified Code World Models
Exploring how 'world-time compute' leverages verified code worlds to enhance AI's generalization abilities beyond traditional training methods.
Model Updates2
NVIDIA Releases InternVideo2-CLIP for Anomaly Detection
NVIDIA's InternVideo2-CLIP model fine-tuned for video anomaly detection is now available.
NVIDIA and Baker Lab Release RFD4-Proteína for Biomolecular Design
NVIDIA and Baker Lab have released RFD4-Proteína, a flow-matching model for joint sequence–structure generation in biomolecular design.