Pace of change is aggressively fast, I try to write and experiment with a tiny portion of what I find interesting.
Posts
Post-Training Releases Survey: Nemotron Cascade, KIMI-DEV, Hermes 4, and Intellect-3
A survey of four recent post-training LLM releases — examining their methodologies, open data, and emerging trends in the space.
NextJS Benchmarking with OpenCode
Benchmarking recent model releases on a DGX Spark with the Next.js benchmark to gauge capabilities and how useful they might be for day-to-day harness work.
NextJS Benchmarking as harness proxy
Benchmarking recent model releases on a DGX Spark with the Next.js benchmark to gauge capabilities and how useful they might be for day-to-day harness work.
Writing-zero Implementation using Prime Intellect stack
BRPO trains a generative reward model for creative writing; +4.18% on RewardBench2.
Nexus: Specialization meets Adaptability for Efficiently Training Mixture of Experts
Nexus merges specialist LLMs into a sparse MoE using domain-embedding dynamic routing.
Deep Researcher with Test-Time Diffusion
Draft-diffusion research agent denoises reports iteratively; outperforms linear pipelines.
Hierarchical Reasoning Model
27M-parameter HRM uses slow/fast latent reasoning to outperform CoT models on ARC-AGI.
Mixture of Recursions
MoR filters tokens hierarchically across recursive stages, cutting memory 49% and compute 62%.
MOE on MNIST
Implements Mixture of Experts on MNIST; studies routing collapse and load balancing.
DeepDream algorithm: How does it work? What does it do?
Explores Google's DeepDream to understand what neural network layers detect in images.
The Role of Image Augmentation on Brain MRI Segmentation Accuracy
Image augmentation improves U-Net brain tumor segmentation accuracy and reduces size-based bias.
Cassava Leaf Disease Classification
TPU-accelerated image classifier detects viral diseases in cassava plants across five categories.
Advancing Drug Discovery with Mechanism of Action (MoA) Prediction
Multi-label neural network predicts drug Mechanism of Action from gene expression data.
Predicting Volcano Eruptions with Deep Learning: Insights from a Unique Challenge
Deep learning predicts volcano eruption times from multi-sensor signals; top 28% finish.
Halite Competition
Reinforcement learning agent mines halite using a Q-function to evaluate multi-step actions.