Speedrunning
Mythein Labs
Machines that get better at getting better.
- Sample-Efficiency
- RSI
- C/CUDA RL
Dogukan Tuna
Hi, I work on sample-efficiency, continually learning AI, and RSI mechanisms aimed at superintelligent / superhuman capabilities. A lot of what I’m exploring involves C/CUDA RL environments and multi-agent architectures. I share live worklogs here, and you can find my open-source projects on GitHub.
Work
What I tinker with day to day
Founding Engineer
Manuel AI
AI for HVAC systems and energy products.
- Applied AI
- Agent Harness
- Infrastructure
Live Worklogs
Raw notes from my learnings, builds
Awaiting the first field note
Writing
Research notes, build logs, and the occasional deep dive
Multi-agent orchestration with Codex, borrowing Fable 5.1 and Fugu Ultra v1.1
A Codex plugin that hands one task to Claude Fable, Opus, Sonnet, Haiku, or Sakana Fugu without leaving the session, on the subscriptions already on your machine.
Verified Replay Distillation (VRD) recipe for continual learning in verifiable domains
A verifier-driven continual learning recipe that combines self-generated traces, replay, and a failure-driven curriculum to learn new task families while limiting forgetting.
autoresearch-mamba: Karpathy-Style Autoresearch for Mamba-2, Mamba-3, and Hybrid Mamba–Transformer MoE
A compact, agent-driven research harness for Mamba-2, Mamba-3, and Nemotron-H-style hybrid Mamba–Transformer MoE models on MLX and CUDA.
A self-improving skill catalog for AI agents
An open-source catalog of 19 agent-maintained skills spanning autonomous research, LLM post-training, GPU/TPU programming, and accelerated computing.
Mem-RLM: Memory-Augmented Inference for Recursive Language Models
A memory layer for Recursive Language Models that records trajectories, extracts reusable strategies, and lifts a weaker model's benchmark score by 26% over three rounds.
Claude Code-Time Skill Acquisition with Agent Teams
How a coordinated agent team researched, synthesized, and integrated a production-grade React Native skill into a shared knowledge base in under 15 minutes.