The Utku Times
predictable, satisfied, conventional.
Profile · Ankara
The story of Mehmet Utku Öztürk
Programming since eight, working on machine learning since thirteen, and nowhere near finished. The machine learning developer and AI researcher works on artificial consciousness and “thinking” systems. This story will be updated.

Mehmet Utku Öztürk wrote his first programs at eight: small apps and scripts. By eleven he was familiar with C++ and C#. At thirteen he started on machine learning, and at fourteen he fine-tuned his first large language model, “extremely unsuccessfully,” in his own words.
The results have improved. In 2026 he co-founded turkhukuk.ai, a legaltech platform for Turkish lawyers, as its technical co-founder; its R&D was accepted as a project at ODTÜ Teknokent. The same year he published HukukBERT as first author in the Journal of Computational Law and Legal Technology: the most comprehensive legal language model for Turkish.
His experiments wander further out. FlyDOOM feeds the game Doom through the visual pathway of a real fruit fly connectome, 139,255 reconstructed neurons and roughly 2.7 million synapses, and maps the fly’s motor output back to the controls. Nothing is trained; the wiring is the fly’s. In LLM Electrodes he took the neuroscientist’s electrode to a language model, and found that stimulating a single neuron is enough to change who it says it is.
Before that he co-founded Clade AI, a personal assistant built on a self-deciding multi-agent framework, and pitched a $21 million round. He now works on LabGPT, a TÜBİTAK-backed tool that analyses test results for doctors, and mentors primary schoolers at BİLSEM, the government-led program for gifted children where he is a senior student.
Away from the keyboard he is a competitive Class II powerlifter, also into strongman, Brazilian jiu-jitsu and kickboxing, and interested in cognitive science, kinesiology, chess, sudoku and good coffee. He is a member of Mensa International and the International Association for Artificial Intelligence and Law.
The record continues on A2; research, interests, articles and the classifieds on A3.
Interpretability · AI
One neuron is enough to give a language model a new personality
Öztürk pushed each of the 294,912 MLP neurons of Qwen3.5-4B in both directions and had the strongest judged blind, with random neurons as controls. 130 passed every check. Asked the meaning of life, one replies: “The meaning of life is $19.99, but it is now $14.99.” No system prompt, no fine-tuning, no change to the weights.
Research · Law
HukukBERT beats BERTurk-Legal on a new legal cloze test
Trained on an 18 GB cleaned legal corpus with a 48K WordPiece tokenizer and domain-adaptive pre-training, the model reaches 84.40% top-1 accuracy on a novel Legal Cloze Test benchmark. Written with Tansu Türkoğlu and Buse Buz-Yalug.


