Daniil Sedov — Blog
Personal notes on compilers, blockchain, and AI research.
Latest posts
Things got too easy with AI
I gave Codex its own Mac Mini
TON Vanity: 286,000x faster vanity addresses
Rebuilding vanity address generator for TON from scratch with AI, achieving up to 286,000x speedup through a series of domain-specific optimizations.
AI in 2026
Predictions and thoughts on AI progress in 2026.
What LLM to use today?
There is nothing out-of-distribution
There is no singularity
Writing with AI
My impression of GPT-5
The complexity threshold of AI
Billions of Tokens Later: Scaling LLM Fuzzing in Practice
Lessons learned from scaling documentation-driven black-box fuzzing pipelines to billions of tokens, practical deduplication strategies, discovered scaling laws, and initial explorations into white-box fuzzing for future expansion.
Multitasking in 2025
Documentation-Driven Compiler Fuzzing with Large Language Models
A fresh and simple black-box approach to fuzzing compilers using large language models to generate test cases from documentation and specification.
Measuring and Analyzing Entropy in Large Language Models
A detailed benchmarking study exploring entropy and randomness across 52 large language models using diverse prompting strategies, revealing notable biases and significant variability influenced by model architectures and prompt engineering.