utku’s space
Fly Me to the MoonFrank Sinatra

The Utku Times

predictable, satisfied, conventional.

Vol. XVII · No. 359Sunday, November 22, 2026Ankara · CV edition

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.

Ink illustration of a cluttered desk with two laptops, and Mehmet Utku Öztürk asleep on the floor beneath it
The desk in Ankara, and the developer under it. Illustration: utku.space

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.

Paper ↗ Code ↗ Data ↗

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.

JCLLT ↗ arXiv ↗

The Utku Times · CV editionA1m@utku.space
The Utku TimesA2 · The recordSunday, November 22, 2026

Honours & associations

  • 2018 · 2023eTwinning Quality Labels — European Union. The Pupil Quality Label in 2018, and the National Quality Label in 2023 for the project “Particle Physics and CERN”.
  • 2025–TÜBİTAK ↗ — Backing for LabGPT, his R&D on test result analysis for doctors.
  • 2026ODTÜ Teknokent ↗ — turkhukuk.ai’s R&D, accepted as a research and development project at METU Technopolis, Ankara.
  • 2024–BİLSEM ↗ — Senior student at the government-led program for gifted children (top 2% IQ), and a voluntary mentor there, teaching primary schoolers robotics, AI and software.
  • MemberMensa International ↗ — The largest and oldest high-IQ society.
  • MemberIAAIL ↗ — International Association for Artificial Intelligence and Law.
  • VolunteerSave the Children ↗ — Global charity for children’s rights, health and education.
  • VolunteerTEMA Foundation ↗ — Türkiye’s foundation for combating soil erosion, for reforestation and for protecting natural habitats.

A life in yearshover, drag, or use ← →

’08’09’10’11’12’13’14’15’16’17’18’19’20’21’22’23’24’25’26

20082026

2026

 

  • Co-founded turkhukuk.ai, legaltech for Turkish lawyers
  • HukukBERT published, first author
  • LLM Electrodes: one neuron, a new personality
  • FlyDOOM: a fruit fly connectome playing Doom
  • 007captcha, Topazius, turkic-relevance-laya

The work

15 projectsactivepausedshelved

Open source7

  • LLM Electrodes Creator · 2026.

    One neuron is enough to give a language model a new personality. Pushed every one of the 294,912 MLP neurons of Qwen3.5-4B both ways and had the strongest judged blind next to random controls: 130 passed every check. One thinks it’s an IUD, one sells made-up legal steroids, one asks for your bank details. No system prompt, no fine-tuning, no change to the weights. The paper, data and a local app to try all 214 interventions are open. GitHub ↗ Preprint ↗ Pdf ↗ Data ↗

  • FlyDOOM Creator · 2026.

    A real fruit fly connectome playing Doom. Feeds the game through the visual pathway of 139,255 reconstructed neurons and roughly 2.7 million synapses, then maps biological motor output back to game controls. Nothing is trained; the wiring is the fly’s. GitHub ↗ Story ↗

  • 007captcha Creator · 2026.

    A robophobic verification framework that defends websites against AI agents. An open-source take on CAPTCHA built for the post-LLM web, where classical challenges no longer hold. TypeScript. MIT licensed. GitHub ↗

  • Octapar Founder & core dev · 2025.

    An AI platform that clusters people into 8-person “tribes” — balanced, semantically coherent groups for authentic connection and mutual growth. Designed a multi-stage clustering algorithm around weighted attentional embeddings and the Hungarian method. Paper published via 23rd.

  • Topazius Creator · 2026.

    A browser-based Markdown notes app whose database is a private GitHub repository you own. Works offline, commits edits in the background, supports wikilinks and backlinks, and can encrypt individual notes entirely on the client. No server, account, analytics, or third-party storage. GitHub ↗

  • trstemmer Creator · 2023.

    A simple Turkish word stemmer for Python. It strips the suffixes off verbs: “seviyorum, yapmışım, koşarlar” comes out as “sev yap koş”. Early and verb-only, but on PyPI. GitHub ↗ PyPI ↗

  • turkic-relevance-laya Creator · 2026.

    A classifier that reads a Turkish Wikipedia article and says how it relates to the Turkic world: core, modern, or unrelated. Laya-multilingual fine-tuned for Project: Turkic Blue, where it sifted ~59,000 articles down to 14,244 core ones. Hugging Face ↗

Work5

  • turkhukuk.ai Technical co-founder · 2026–present.

    A legaltech platform for Turkish lawyers — case research and litigation strategy. Trained and fine-tuned domain-specific language models on Turkish legal corpora, and experimented with neuro-symbolic representations to bridge statistical learning with structured legal reasoning. turkhukuk.ai ↗

  • LabGPT R&d, neurosymbolic AI · 2025–present.

    Test result analysis for doctors with trend analysis, clinical finding extraction and a corpus with years of medical research. TÜBİTAK-backed.

  • Clade AI Co-founder, AI dev · 2023–24.

    A personal AI assistant with task execution, extensive integrations, and personalized interaction. Pitched a $21M round. Built a self-deciding multi-agent framework in the early days of AI agents, fine-tuned an LLM for natural interaction, and shipped custom systems for smart memory, decision-making, and seamless integrations. Left in July 2024; shelved shortly after.

  • Mentoring at BİLSEM Senior student & instructor · 2024–present.

    Senior student & instructor at BİLSEM, a government-led supplementary education program for gifted students (top 2% IQ). Voluntarily mentoring and instructing primary schoolers there in project management, robotics, artificial intelligence, and software development. About BİLSEM ↗

  • Sorubot Developer · 2022–23.

    An AI platform that generates exam questions and manages students’ study sessions. Very unsuccessfully fine-tuned an LLM. Amateur work — but not bad for a 14-year-old.

Unpublished3

  • A.D.A Creator · 2020.

    Ada’s damn awesome — a recursive name, with respect to Lady Lovelace. A parser-based, deterministic personal assistant.

  • Ananas Studios 2019–22.

    An indie game development studio.

  • Shoelace Creator · 2020.

    A peer-based encryption algorithm.

The Utku Times · CV editionA2m@utku.space
The Utku TimesA3 · Research, leisure & puzzlesSunday, November 22, 2026

Research

Papers & preprints

independent research · 2026

LLM Electrodes: Rewriting a Language Model's Behavior Through Single-Neuron Stimulation

Single-neuron stimulation of Qwen3.5-4B, with no system prompt, fine-tuning or change to the weights: of all 294,912 MLP neurons, 130 give the model a consistent, nameable new personality under blind judging, and at least 6% of random neurons produce a nameable change too. Drug-like whole-layer interventions barely move it.

Preprint ↗ PDF ↗ GitHub ↗ Data ↗

We ask how much a single neuron can change the behavior of a language model. We pushed each of the 294,912 MLP neurons of Qwen3.5-4B in both directions with a fixed stimulation and measured how far the model's next-token predictions moved. The candidates that moved the model most while it stayed coherent then answered ten prompts they were never selected on, and each candidate was judged anonymized and independently against the unmodified model's answers. The 19 personas that the steering community had already published served as positive controls, and the most distinct randomly chosen neurons served as negative controls. 130 single neurons passed every check, including a second, independent round of judging. The largest group pulls every answer toward one topic (55). Others switch the output language (26), produce altered states (21), or change what the model says it is. Such effects are not limited to the selected candidates, since at least 6.0% of randomly chosen neuron pushes also produce a nameable and coherent change, and the confirmed hits are distinguished mainly by the consistency of their effect. Interventions modeled on drugs are far weaker. Scaling every MLP neuron of a layer, from silencing to doubling, leaves the answers nearly unchanged, and groups of neurons that read a shared direction do no more than random groups of the same size. Silencing or amplifying single neurons fourfold moves the model about a thousand times less than stimulation, and none of 100 gain candidates passed the checks that 10 of 10 stimulation hits passed in the same blind pool. No intervention improved task accuracy, and most reduced it. We release the screen, a catalog of 214 interventions with their doses and evidence, and an interactive tool for trying them.

Journal of Computational Law and Legal Technology · 2026

HukukBERT: Domain-Specific Language Model for Turkish Law

With Tansu Türkoğlu and Buse Buz-Yalug

The most comprehensive legal language model for Turkish, trained on an 18 GB cleaned legal corpus with advanced masking strategies and a 48K WordPiece tokenizer plus Domain-Adaptive Pre-Training. Achieves 84.40% Top-1 accuracy on a novel Legal Cloze Test benchmark, outperforming BERTurk-Legal.

JCLLT ↗ arXiv ↗

Recent advances in natural language processing (NLP) have opened significant doors for LegalTech applications; however, existing studies specific to Turkish law have been severely limited by data constraints. While extensive models like LEGAL-BERT have been developed for English legal texts, the Turkish legal domain lacks a high-volume, domain-specific counterpart. In this paper, we introduce HukukBERT, the most comprehensive legal language model for Turkish, trained on an 18 GB cleaned legal corpus using advanced masking strategies. We systematically compare our 48K WordPiece tokenizer and Domain-Adaptive Pre-Training (DAPT) approach against general-purpose and existing domain-specific Turkish models. Evaluated on a novel Legal Cloze Test benchmark, HukukBERT achieves state-of-the-art performance with 84.40% Top-1 accuracy, substantially outperforming existing models such as BERTurk-Legal. We release HukukBERT to serve as a foundational model for downstream Turkish legal tasks.

23rd the company · 2025

Octapar: A Framework for Multi-Modal, Balanced User Grouping using Weighted Attentional Embeddings and Assignment-Based Clustering

Conversational LLM onboarding extracts semantic user tags; a weighted attention mechanism combines global tag rarity (IDF) with local intra-user tag uniqueness; the Hungarian method guarantees perfectly balanced clusters.

Zenodo ↗

The formation of coherent, effective, and equitably sized user groups is a critical and persistent challenge in the design of online social platforms, collaborative learning environments, and enterprise-level teamwork tools. Existing automated methods, while scalable, often fail to produce groups of a consistent, predetermined size, leading to significant logistical issues and imbalanced user experiences. Furthermore, they often rely on shallow or easily manipulated user data. This paper introduces Octapar, a novel, multi-stage algorithmic framework designed to generate semantically coherent, perfectly balanced user clusters from deep, multi-modal user profiles. The framework pioneers a conversational onboarding process, using specialized chatbots and Large Language Models (LLMs) to extract nuanced user characteristics across both skill and philosophical domains. These characteristics are formalized as a rich set of semantic tags. The core technical innovation is a custom weighted attention mechanism that generates a sophisticated user vector by incorporating not only global tag rarity (IDF) but also a novel metric of local, intra-user tag uniqueness. This allows the model to identify and amplify a user's most distinctive traits. Final group formation is achieved via an iterative, assignment-based clustering algorithm that utilizes the Hungarian method to guarantee perfectly balanced clusters. This approach ensures structural consistency and fairness in group formation, a critical and often unmet requirement for many practical applications. We provide a detailed walkthrough of the architecture, the mathematical foundations of its components, and a discussion of its implications for the future of computational group formation.

Plawlabs · 2024

Streaming Intelligence: Harnessing AI and Big Data in the Competitive Landscape of Digital Content Platforms

How major streaming platforms use AI and Big Data for recommendation, presentation, and personalization, with strategy for an emerging Turkish digital platform.

In the rapidly evolving realm of digital entertainment, Artificial Intelligence (AI) and Big Data stand at the forefront, reshaping the future of content consumption and curation. Esteemed global streaming giants such as Netflix, YouTube, Hulu, Amazon Prime Video, HBO Max, Disney+, and even social platforms like Instagram and TikTok, are skillfully leveraging these technologies to transform content recommendation, presentation, and personalization processes. This not only elevates the user experience but also sets unprecedented standards for operational efficiency across the industry. Our article delves into the intricate use of AI and Big Data by these leading platforms, highlighting their critical role in captivating audiences and streamlining content strategies. Furthermore, our research offers concrete insights and forward-looking strategies for one of the Turkey's emerging digital platform, in collaboration with PlawLabs, on surpassing its global competitors with the development of advanced AI technology. This initiative is aimed at overcoming unique challenges brought forth by these technological advancements and seizing expansive opportunities. By capitalizing on the technological expertise and strategic foresights of current industry leaders, this platform is poised to gain a significant competitive edge in the rapidly growing market of digital streaming services, both domestically and internationally. Adopting our technology foresights will not only enhance the quality of content and increase user engagement but will also seamlessly align with the changing preferences and demands of today's diverse digital audience, supporting the platform's financial outcomes. Through this comprehensive analysis, we aim to illuminate a path for our national platform to leverage AI and Big Data, catalyze its growth, and redefine content maturity in the digital space.

Interests

Off the clock
Halftone illustration of a gym, Mehmet Utku Öztürk in the mirror
A competitive Class II powerlifter, in the mirror.
  • Powerlifting & strongman. Competitive, Class II.
  • Martial arts. Brazilian jiu-jitsu & kickboxing.
  • Chess.
  • Sudoku. There is one below.
  • Cognitive science. How minds work.
  • Kinesiology. How bodies move.
  • Coffee. Good, only.

Articles

Puzzles

Easy
Sudoku, 4×4Tap a cell to cycle

Every row, column and box holds 1 to 4.

Classifieds

AVAILABLE. Machine learning developer and AI researcher, Ankara. Python, TypeScript, C++, C#. Language models, neuro-symbolic systems, connectomics. Lifts heavy things.

LOST. One fruit fly, Drosophila melanogaster, last seen somewhere on this paper. Answers to nothing. If found, click it.

CONTACT. (click to copy) · Mensa: utku.ozturk@member.mensa.org · GitHub, LinkedIn, X, Hugging Face: @mutkuoz · ORCID 0009-0008-2400-5646

The Utku Times · CV editionA3m@utku.space

Easter eggs found: 0 / 8