Blog

Yearly technical notes, benchmark case studies, and local-first AI systems work.

2026

6 min readShkumbin Sherifi

How Scaling a Diffusion-LM made it worse

A human-directed, DeepSeek-executed experiment found that an 8M-parameter diffusion language model beat a 46.5M version on held-out riddles.

Diffusion ModelsLanguage ModelsMachine LearningLoRAPyTorch
6 min readShkumbin Sherifi

How I Beat Anthropic's 1,487-Cycle Benchmark with Claude and Codex

A human-directed, AI-assisted optimization session reached 1,469 cycles on Anthropic's open performance challenge while preserving correctness and keeping the implementation private.

Performance EngineeringClaudeCodexAI AgentsVLIW
2 min readShkumbin Sherifi

Random Project of the Week: GitVisual

A quick build note on turning hundreds of GitHub stars into an interactive treemap with Codex.

GitVisualOpen SourceCodexGitHubOriginKit
9 min readShkumbin Sherifi

Breaking into Y Combinator: Part 3 | The Hardening Loop That Made Codex Fail

Part 3: the C++ heat-diffusion task that completed the hardening loop, and the first time Codex failed cleanly.

TB3Benchmark DesignNumerical MethodsAgent Evaluation
9 min readShkumbin Sherifi

Breaking into Y Combinator: Part 2 | Thermal Stack Calibration TB3

Part 2 of the benchmark design series: moving from a retrieval task that was too easy to inspect into a coupled thermal-stack calibration task that both Codex and Claude failed cleanly.

TB3Benchmark DesignThermal ModelingAgent Evaluation
9 min readShkumbin Sherifi

Breaking into Y Combinator: Part 1 | Hybrid Retrieval Fusion

A focused writeup on the Hybrid Retrieval Fusion TB3 task: its design, verifier strategy, iteration loop, and why the model failures were not strong enough.

TB3Benchmark DesignRetrievalRAGAgent Evaluation