{"@attributes":{"version":"2.0"},"channel":{"title":"Blake Ledden","description":"ML experiments, research findings, and technical deep-dives.","link":"https:\/\/bledden.github.io\/","item":[{"title":"341x Faster: Optimizing a Rust Vector Database to Top-7 on ANN-Benchmarks","link":"https:\/\/bledden.github.io\/blog\/arrwdb-optimization\/","guid":"https:\/\/bledden.github.io\/blog\/arrwdb-optimization\/","description":"From 52 QPS to 17,746 QPS on SIFT-1M through ten optimizations, two bug fixes, and one measurement error. Projected top-7 on ann-benchmarks at 0.999 recall.","pubDate":"Thu, 16 Apr 2026 00:00:00 GMT"},{"title":"Adapting Noisy Student to the LLM Era: Four Experiments in Self-Training","link":"https:\/\/bledden.github.io\/blog\/noisy-student\/","guid":"https:\/\/bledden.github.io\/blog\/noisy-student\/","description":"Token noise fails for language (p=0.0018). Consensus-based pseudo-labeling with RLVR works: +3.9pp on GSM8K (p=0.002, N=7). A bootstrap capability threshold governs when self-training can succeed.","pubDate":"Wed, 04 Feb 2026 00:00:00 GMT"},{"title":"Empirically Validating the Information-Theoretic Gap Between SL and RL","link":"https:\/\/bledden.github.io\/blog\/memorization-study\/","guid":"https:\/\/bledden.github.io\/blog\/memorization-study\/","description":"I tested information-theoretic predictions about learning efficiency. Supervised learning converges in 2 episodes regardless of problem size. RL scales with n^0.89. The theory holds.","pubDate":"Thu, 01 Jan 2026 11:00:00 GMT"},{"title":"Teaching AI to Embody Characters: A Replication of Open Character Training","link":"https:\/\/bledden.github.io\/blog\/open-character-training\/","guid":"https:\/\/bledden.github.io\/blog\/open-character-training\/","description":"I trained 5 distinct AI personas using constitutional methods. Character alignment improved 39%, prompt distillation worked, and adversarial resistance increased. The method generalizes across personality types.","pubDate":"Thu, 01 Jan 2026 12:00:00 GMT"},{"title":"A Memory Tool for 1% of Your Budget: How Recall Works, and Why I Rely on It","link":"https:\/\/bledden.github.io\/blog\/recall\/","guid":"https:\/\/bledden.github.io\/blog\/recall\/","description":"A local, cross-session memory plugin for Claude Code and Cowork. How it works from the system up, and the measured answer to what it costs: about 1% of my token budget across 470,000 requests on three machines.","pubDate":"Sun, 05 Jul 2026 00:00:00 GMT"},{"title":"Seven Patterns From 300+ ML Evaluation Runs","link":"https:\/\/bledden.github.io\/blog\/cross-project-meta-analysis\/","guid":"https:\/\/bledden.github.io\/blog\/cross-project-meta-analysis\/","description":"Cross-project analysis of ML experiments reveals patterns about model scaling, distillation dynamics, constitutional training, and when negative results matter.","pubDate":"Wed, 31 Dec 2025 00:00:00 GMT"},{"title":"GAN-Style Training for Jokes: When Metrics Lie","link":"https:\/\/bledden.github.io\/blog\/gan-joke-generation\/","guid":"https:\/\/bledden.github.io\/blog\/gan-joke-generation\/","description":"I trained a GAN to generate jokes. The metrics looked great. The jokes were terrible. A cautionary tale about evaluation gaps in creative generation.","pubDate":"Thu, 01 Jan 2026 05:00:00 GMT"},{"title":"Constitutional AI from Base Models: Can You Train Safety Without Instruction Tuning?","link":"https:\/\/bledden.github.io\/blog\/cai-base-model\/","guid":"https:\/\/bledden.github.io\/blog\/cai-base-model\/","description":"I replicated Constitutional AI starting from a raw base model. The pipeline works, but DPO needs more than 42 training pairs to show improvement over SFT alone.","pubDate":"Thu, 01 Jan 2026 06:00:00 GMT"},{"title":"The Distribution Cliff: Why Hybrid Distillation Fails on Decoder-Only LLMs","link":"https:\/\/bledden.github.io\/blog\/context-distillation\/","guid":"https:\/\/bledden.github.io\/blog\/context-distillation\/","description":"I tested 8 distillation methods across 160 runs. Any off-policy component causes collapse. Pure on-policy with teacher seeding achieves 71% GSM8K accuracy.","pubDate":"Fri, 02 Jan 2026 00:00:00 GMT"}]}}