Hi, I'm Dogukan.
RE/RS, Multi-step Reasoning and Post-Training Optimization
I am extremely excited about the new capabilities of reasoning and its applications to scale science. I like to work on AI for mathematics, currently building Arfmax, an AI mathematician. My research focus is on AI that exhibits multi-step reasoning with autoformalization, formal-math-synthesis, Alpha-zero / AlphaGeometry like inference-time search techniques, generalizable verifier reward models and post-training optimization. Grinding JAX/XLA forward and backward kernels with pybind11 C++ backend and integrating custom-written kernels into frameworks.
Thoughts
January 22, 2025
Revisiting Superintelligence Control Problem with Newer Extensions
January 2, 2025
New Year, New Blog
Archive, 2024
Tensor Decompositions
Published on Dec 31, 2024
Mixture-of-Experts Process Groups Initialization
Published on Dec 31, 2024
Real-Time Text and Voice-Based Retrieval Augmented Generation Pipeline with Gemini 2.0 for Long Documents
Published on December 25, 2024
OpenAI's Code Execution Runtime & Replicating Sandboxing Infrastructure
Published on July 2, 2024
Linen Modules for Simple Long Convolutions Sequence Modeling with Log-Linear Time Complexity
Published on March 20, 2024
[Let's Review] Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models
Published on March 7, 2024
[Let's Review] Gradient Ascent + Jax Implementation
Published on March 11, 2024
Tri-RMSNorm: Yielding Speed-up for Training Stabilization
Published on March 14, 2024
Enhancing Performance with C/C++ Code Execution for Langchain Agents
Published on July 18, 2024
Utilizing Airflow for Planning, Scheduling, Executing and Scaling AI Agentic Workloads
Published on May 6, 2024