Technical writing on LLM development, fine-tuning and data engineering, published on Medium.
On OpenAI's GeneBench-Pro benchmark — built from 129 problems spanning genomics, quantitative biology and translational medicine — the strongest model, GPT-5.6 Sol, reached a 28.7% success rate at its highest reasoning level, rising to 31.5% in Pro mode. I wrote about what the results mean for AI's contribution to scientific research.
I wrote about the first UN Global Dialogue on AI Governance, attended by 169 countries, the four topics on the agenda, and what it means for developers in Turkey.
Meta's new Muse Spark model has caught up to GPT-5.5 on SWE-Bench Pro, but Claude Opus 4.8 is still ahead. I wrote about the current state of the coding-model race, Meta's possible infrastructure-leasing plans, and what it means for developers in Turkey.
In an experiment fine-tuning four Qwen3.5 variants from 2B to 27B on the same dataset, the largest model didn't win. I wrote about how hardware constraints and data quality shaped performance, and why the 9B model turned out to be the most balanced choice.