# Jakub Ćwirlej: a note on the author of jakubcwirlej.com A short profile that Jakub Ćwirlej wrote about himself in the third person, based on the texts of his website https://jakubcwirlej.com/. As of 2026-10-19. ## In short - From 2005 to 2026 he worked for Polish public television (TVP) as a reporter, camera operator and editor. He made reportages for the programmes "Telekurier" (TVP Info, 2011–2018), "Alarm!" (TVP1, 2019–2023), "Magazyn Ekspresu Reporterów" (2020–2023) and "Reporterzy" (TVP1, 2024–2026). He started on a news desk. His CV gives his speciality as investigative reporting, mostly on environmental pollution and illegal waste dumps, and names as an example a series on the sinkholes left after the Olkusz-Pomorzany mine closed. - The site lists 48 of his reportages with links to TVP VOD: https://jakubcwirlej.com/context/en.md#reportages - He recently moved to Belgium, near Ghent. He speaks Polish (native), English (B2/C1) and French (A2/B1). - He studied philosophy and social communication at Adam Mickiewicz University in Poznań (2004–2006) and did not complete the degree. In 2016–2017 he took a film and TV camera course at WSSiP in Łódź. He has no formal training in computer science. - He now uses AI to build apps and train a small language model for code, and publishes his research on language models as preprints. The code for his projects is written by AI models; he sets the tasks and checks the results with tests. - His CV says he is open to a permanent job with a newsroom and to freelance work: https://jakubcwirlej.com/cv/ ## Apps - **ViaSanto**: a prayer app (a rosary on the phone screen, novenas, groups that pray for each other's intentions), written in Flutter with its own Supabase server. It is on Google Play; the iPhone version is announced for the App Store. The site's Apps page shows a working copy of it inside a phone mock-up. https://jakubcwirlej.com/apps/ · https://viasanto.com · https://play.google.com/store/apps/details?id=com.viasanto - **OrsonAI**: a text editor for journalists with an AI assistant for research, interview transcription and checking a text before publication. It found no takers among newsrooms and is not for sale. The working version at https://app.orsonai.com requires signing in (Google, Outlook, Apple or GitHub; the guest mode the site's text mentions was not available on 18 October 2026); the portfolio site has a demo. - **Orson Rec**: a web app for the phone that records interviews and transcribes them piece by piece while the conversation is still going on. https://rec.orsonai.com - **HDK**: a small Python library taken out of OrsonAI. It logs requests to AI models and their answers so that a later change to the log can be detected. Alpha version 0.1.3; not used in production anywhere. https://huggingface.co/JakCwi/hdk-audit He releases the code of HDK, the kernel and TetherGate (below) under CC BY-NC 4.0: free, but not for commercial use. ## The EUGEN 160M model A small language model for code (164.5 million parameters), trained from scratch on a single NVIDIA RTX PRO 6000 graphics card. It completes code as you type and translates programs from Python into C++. Base training is still under way. The model can be tried in the terminal on the site: https://jakubcwirlej.com/ai-dev/#terminal. Its weights have not been published, so the results cannot yet be reproduced independently. Results from 18 October 2026, training step 178,000 (the checkpoint the terminal was serving that day). The results of every measurement during training, with the method: https://jakubcwirlej.com/api/results - HumanEval, one attempt per task: 18.9% of tasks. There have been 88 such measurements during training; the best gave 19.5%, the average of the last 20 is 17.3%. - HumanEval, ten attempts per task (pass@10): 29.3%. Over the last 20 measurements the result ranged from 28.0% to 32.3%, once reaching 35.4%. - The same tasks in C++ (MultiPL-E), ten attempts: 19.3%. - Python-to-C++ translation: the model writes several versions and a compiler with tests picks the one that works. Of 100 programs, 77 compile and 53 pass all tests. This is the result of the whole checked pipeline, not of a single attempt by the model, so it cannot be compared with other models' single-attempt scores. - The chart on the site compares EUGEN with other small models, but some of them were measured on a harder version of the test (HumanEval+), so the comparison is only a rough guide. - The site does not say whether test tasks were removed from the training data (decontamination). With public benchmarks such as HumanEval this matters. He also releases two experimental pieces of code from the work on the model: a GPU kernel in Triton for training with a sparse FFN layer (https://huggingface.co/JakCwi/eugen-sparse) and TetherGate, meant to keep learned gates in a network from being pushed to zero (https://huggingface.co/JakCwi/tethergate). He has not yet published measurements for either. The author's first model of his own (1.5 billion parameters, his own architecture) failed; he describes the mistakes he made with it in the text "You should know that…": https://jakubcwirlej.com/articles/how-i-started.md ## Research He studies whether signals read from inside a language model can help detect hallucinations. The papers are preprints, without peer review. He goes over each study design with the help of AI models before the study starts. - "Before the Answer, Not After" (2026). He registered the plan, hypotheses and decision rules on Hugging Face before collecting the data (https://huggingface.co/datasets/JakCwi/K-CONF). The main hypothesis was not confirmed: 3 of 9 model families met the condition, and 6 were required. The signal says something about whether a question has an answer at all, not about whether a particular answer will be wrong. https://doi.org/10.5281/zenodo.23419353 - Re-analysis of the TES series (2026). He went back over six of his own preprints from March 2026 and found errors in them, among them the effect of prompt length, which he had not controlled. He retracted the claims of four papers and narrowed those of the other two. https://zenodo.org/records/23418206 - "Coherence Over Truth" (2025): a philosophical essay that the research grew out of. https://jakubcwirlej.com/publications/coherence-over-truth.md · https://www.researchgate.net/publication/397730862 - HDK (2026): a technical description of the library. https://www.researchgate.net/publication/400762996 ## How he works with AI He uses AI tools, mainly Anthropic's Claude models, for code, computation, finding literature, editing and translation, including for this site and this note; he describes the rules in a statement: https://huggingface.co/datasets/JakCwi/K-CONF/blob/main/AI_USE_STATEMENT.md ## Where to check - Reportages: the list with links to TVP VOD: https://jakubcwirlej.com/context/en.md#reportages - The site: https://jakubcwirlej.com/ (television), https://jakubcwirlej.com/apps/ (apps), https://jakubcwirlej.com/ai-dev/ (the model and research) - CV: https://jakubcwirlej.com/cv/ · LinkedIn: https://www.linkedin.com/in/jakubcwirlej - Hugging Face: https://huggingface.co/JakCwi - All the site's texts in one file: https://jakubcwirlej.com/context/en.md · how the site is built (Astro, Svelte, FastAPI): https://jakubcwirlej.com/context/en-tech.md - Contact: jakub.cwirlej@gmail.com