# Mariano Crimi > Product, AI, drums, endless rumination. Personal website: https://marianocrimi.com/ This is a full-text version of the site, written for people and language models to read without running JavaScript. The visual site is a single-page e-ink style app. Sections use hash URLs. ## Who am I (being)? I am from Argentina. I say this up front so nobody gets hurt. It started in Buenos Aires, with a beige machine that did exactly what I told it to do. Nobody had warned me. A kid should never be handed that kind of power without supervision. I typed a command, the thing obeyed, and some wire in my brain fused for good. I have been chasing that first hit ever since. I got a Computer Science degree and went straight to work as a developer, writing ASP and PHP, which in those days passed for a respectable trade. Terrible, beautiful code. Some of it may still be running somewhere, God help them. Then the question got me. The how stopped being enough. I wanted to know why anybody needed the thing built in the first place. So I crossed over, and I took the keyboard with me. I am a generalist, which is a polite word for a man who knows a little about everything and stays in the room when the engineers start talking. Somewhere in there I enrolled in an MSc in Data Science and Artificial Intelligence while holding a full-time job, a decision no sane person would defend. They gave me summa cum laude. I suspect a clerical error and have chosen to leave it alone. These days I live in Montpellier and work at Bluesquare, a Belgian digital health outfit, on [OpenHEXA](https://openhexa.com): open-source software that public health teams use to wrestle their data into something they can act on. I am also the company's self-appointed AI engineer, which has given me the privilege of building things like MCP servers for [Pathways](https://withpathways.org) and AI assistants for grant applications. OK, so AI is the biggest, strangest story of my lifetime, and I am in no condition to watch it quietly from my browser. I am equally excited and terrified by it. I read about minds at ungodly hours: consciousness, evolution, Hofstadter and his loops, how a pile of neurons or numbers gets up one morning and starts thinking. I understand perhaps a fifth of it. It is the best fifth I own. I am also a drummer, which the neighbors can confirm. A father of two, outnumbered and outmaneuvered. An avid reader with an audiobook running at all hours. And a proud Argentinean, so football is a hereditary condition with no cure and one persistent symptom: hope. ## Portfolio Data science projects. Source for each is on GitHub under https://github.com/mcrimi/. ### Game of Thrones survival analysis 2021, R. [Page](https://marianocrimi.com/#project-got-survival). [Source](https://github.com/mcrimi/got_survival). Survival analysis of Game of Thrones characters using Kaplan-Meier estimators, Cox regression and Elastic Net. Game of Thrones became notorious for its shocking and ruthless killings of major characters throughout its run. In this whitepaper I try to identify the main risk covariates that would predict survival, using Kaplan-Meier estimators, Cox regression, Elastic Net and stepwise methods for variable selection. The repository holds the whitepaper, the character dataset for seasons 1 to 8, a data dictionary and the R notebook. ### Vélib weekly cluster analysis 2021. [Page](https://marianocrimi.com/#project-velib-clusters). [Source](https://github.com/mcrimi/velib_weekly_cluster_analysis). A cluster analysis of the Paris Vélib bike-sharing system for a week. ### Genetic algorithm for the travelling salesman 2020, Python. [Page](https://marianocrimi.com/#project-ga-for-tsp). [Source](https://github.com/mcrimi/GA_for_TSP_v2). A genetic optimisation algorithm for travelling salesman problems, written in Python. ### CEC 2008 benchmark functions 2020, Python. [Page](https://marianocrimi.com/#project-cec2008). [Source](https://github.com/mcrimi/CEC2008). Optimisation in Python for the CEC 2008 benchmark functions, the set used in the competition on large-scale global optimisation. ### Ultimatum game 2020, NetLogo. [Page](https://marianocrimi.com/#project-ultimatum-game). [Source](https://github.com/mcrimi/ultimatumgame_abm). An agent-based model of the ultimatum game, built in NetLogo. ## Publications A journal article and an R package from work on plant breeding data systems. [Page](https://marianocrimi.com/#publications). ### BrAPI v2: real-world applications for data integration and collaboration in the breeding and genetics community 2025. Article. Database (Oxford University Press). [DOI](https://doi.org/10.1093/database/baaf048). Peter Selby, Rafael Abbeloos, Anne-Francoise Adam-Blondon, Francisco J Agosto-Pérez, Michael Alaux, Isabelle Alic, Khaled Al-Shamaa, Johan Steven Aparicio, Jan Erik Backlund, Aldrin Batac, Sebastian Beier, Gabriel Besombes, Alice Boizet, Matthijs Brouwer, Terry Casstevens, Arnaud Charleroy, Keo Corak, Chaney Courtney, Mariano Crimi, Gouripriya Davuluri, Kauê de Sousa, Jeremy Destin, Stijn Dhondt, Ajay Dhungana, Bert Droesbeke, Manuel Feser, Mirella Flores-Gonzalez, Valentin Guignon, Corina Habito, Asis Hallab, Jenna Hershberger, Puthick Hok, Amanda M Hulse-Kemp, Lynn Carol Johnson, Sook Jung, Paul Kersey, Andrzej Kilian, Patrick König, Suman Kumar, Josh Lamos-Sweeney, Laszlo Lang, Matthias Lange, Marie-Angélique Laporte, Taein Lee, Erwan Le Floch, Francisco López, Brandon Madriz, Dorrie Main, Marco Marsella, Maud Marty, Célia Michotey, Zachary Miller, Iain Milne, Lukas A Mueller, Moses Nderitu, Pascal Neveu, Nick Palladino, Tim Parsons, Cyril Pommier, Jean-François Rami, Sebastian Raubach, Trevor Rife, Kelly Robbins, Mathieu Rouard, Joseph Ruff, Guilhem Sempéré, Romil Mayank Shah, Paul Shaw, Becky Smith, Nahuel Soldevilla, Anne Tireau, Clarysabel Tovar, Grzegorz Uszynski, Vivian Bass Vega, Stephan Weise, Shawn C Yarnes, and the BrAPI Consortium (77 authors). Affiliation listed for Mariano Crimi: Integrated Breeding Platform, El Batán, Texcoco, México. ### QBMS: Query the Breeding Management System(s) 2021. Software. R package on CRAN, first released on 19 October 2021. [Concept DOI](https://doi.org/10.5281/zenodo.10791626). [CRAN](https://cran.r-project.org/package=QBMS). [GitHub](https://github.com/icarda/QBMS). Khaled Al-Shamaa, Mariano Omar Crimi (contributor), Zakaria Kehel, Johan Aparicio. The package helps breeders link breeding data management systems with analytical pipelines, and can query systems such as EBS, BMS, BreedBase, Germinate, DeltaBreed and GIGWA through BrAPI. ## Fleeting thoughts Short notes on AI, kept as they come. [Page](https://marianocrimi.com/#thoughts). - 14 Nov 2026. Curiosity made the cat stronger. - 14 Nov 2026. Stupidity is a dish best served hot. - 14 Nov 2026. The only thing you can do with time is lose it. - 14 Nov 2026. LLMs have flipped intellectual work around: the tasks that used to be fun now feel boring, and the ones that used to be boring are now kind of fun. - 14 Nov 2026. Using a personal AI assistant to do stuff for you feels like the first time I took an Uber: curious and hesitant at first, nervous along the way, and completely sold by the end. - 11 Sep 2026. Watching a reinforcement-learning agent loop in circles, almost destroying itself to collect reward, it is impossible not to think about addiction. - 11 Sep 2026. There is a tension between alignment and power. Sometimes we want agents to find non-human pathways; that is what makes computational search powerful. We want the move 37s. How do we let them search in a non-human way and stay aligned with humans? Something has to give. - 7 Sep 2026. A verifiable reward is at odds with the notion of goodness. To say a response was good, the signal is complex: nuanced, intricate, and often not quantifiable. - 7 Aug 2026. Reward is a space, not a function. - 30 Jul 2026. The brain is stateful; transformers are not. In a brain, inference and learning are indistinguishable. - 24 Jul 2026. We are going to be increasingly priced out of intelligence. If you can make intelligence yield money, reinvesting it compounds. - 26 Jun 2026. Status seeking and people-pleasing are different flavors of the same ice cream. - 25 Jun 2026. I need to stop using the same model for everything. When to copilot? When to vibecode? - 25 Jun 2026. Paper to understand and to think: free form, drawings. Digital to log, record and publish: editing, search, connections. - 25 Jun 2026. What is the next Lindy artifact of the AI-native era? The next MCP standard. - 19 Jun 2026. AI use is being commoditized. Saw it on LinkedIn: if anyone else uses AI, they are lazy; if anyone else doesn't, they are a laggard. - 4 Jun 2026. The barriers to building have collapsed. That shifts the bottleneck to ownership, validation, taste, and deciding what should actually exist. - 4 Jun 2026. The more I connect with my curiosity and the will to build for building's sake, the closer I get to joy. And you can't beat someone who is enjoying the process. - 3 May 2026. LLMs and brains both run generative models; the difference is how the model is acquired. Brains wager on predictions through action. LLMs absorb the residue of other minds. - 3 May 2026. The image that stuck from the active-inference paper: generative AI is a 21st-century mirror, but when you look behind it, there is nobody there. - 3 May 2026. Protected solo time is when the deeper work actually happens. - 1 May 2026. LLMs never interacted with the physical world, so they have no chance of a language-free world model. They are trapped in a cave, and we hand them shadows through language, which is lossy. - 1 May 2026. Can human-like intelligence be achieved without the need to fight entropy? It seems to me the AI must want to survive, which is in tension with alignment. - 10 Apr 2026. "All you can do is all you can do in any given moment." Yes, but choosing what to do is the big problem. - 9 Apr 2026. There is something about LLMs that is mystical. What is it about autoregression that captures thinking? - 5 Mar 2026. "The AI that folds proteins will not be the same that folds your laundry." Funny and probably true, for AI and for my personal energy policy. - 5 Mar 2026. I agree that the term AGI is fuzzy. I'd even say I am still unclear about what intelligence really is. - 5 Mar 2026. Treat a goal as a Bayesian prior: a starting hypothesis that gets you moving and is meant to be updated. - 20 Jan 2026. Found again in my notes, William James's formula: self-esteem = success / pretensions. - 1 Nov 2025. What is the difference between intelligence, learning, the act of thinking and consciousness? - 30 Sep 2025. Yesterday I saw a clock showing the wrong time, and I did not spend one millisecond trying to reconcile it with my inner model. I just knew the clock was wrong. Why did no uncertainty arise? How would an agent react to this? ## Writing Longer pieces. Nothing published yet. [Page](https://marianocrimi.com/#writing). ## Music Drums, mostly. [Page](https://marianocrimi.com/#music). The music page has photographs of the author playing drums, a drum machine you can play in the browser, and a diagram of an ideal kit: one front tom, two floor toms, two 18-inch crashes, a 22-inch ride in front, a china on the right and two splashes left of the hi-hat. ## AI Experiments Things built to see if they work. ### Snakeformer 2026, Python. [Page](https://marianocrimi.com/#project-snakeformer). [Source](https://github.com/mcrimi/snakeformer). [Weights](https://huggingface.co/mcrimi/snakeformer). A transformer trained from scratch to simulate the good old game of Snake. The overengineered snake game that nobody asked for. So I trained a decoder-only transformer (à la GPT) from scratch to simulate the good old game of Snake. The model takes a text-based board state and a user input as a prompt, and responds with the next frame, token by token, effectively acting as a learned physics engine. The game state is treated not as a grid of pixels, but as a sequence of tokens. By training on a lot of state transitions, the model approximates the game's update function. To succeed in generating a snake world, it must internalise its implicit rules and topology. It's slow, it's an overkill. It was fun to build. ### MCP server for OpenHEXA 2025, Python. [Page](https://marianocrimi.com/#project-mcp-openhexa). [Source](https://github.com/mcrimi/mcp-openhexa). A very rudimentary proof of concept of a local MCP server for OpenHEXA, used from Claude Desktop. A proof of concept of a local MCP server for OpenHEXA. Once it is configured in Claude Desktop, OpenHEXA shows up as an integration tool. You can then ask questions like "Show me all workspaces in OpenHEXA", "Search for pipelines containing 'analysis' in their name" or "Get details about the 'covid-data' dataset". ### Pathways Streamlit UI 2026, Python. [Page](https://marianocrimi.com/#project-pathways-ui). [Source](https://github.com/mcrimi/pathways-streamlit-ui). A proof-of-concept UI, built with Streamlit, to test the capabilities of the MCP server. ## Elsewhere - [GitHub](https://github.com/mcrimi) - [LinkedIn](https://www.linkedin.com/in/mcrimi) - [Kaggle](https://www.kaggle.com/mmmcrim) - [Goodreads](https://www.goodreads.com/user/show/3430535-mariano-crimi) - [Hugging Face](https://huggingface.co/mcrimi)