Lab
Articles
Technical writing on deterministic verification, Private AI, post-LLM architectures, applied epistemology, and local inference infrastructure.
Two technical papers, from both ends of the stack.
I'm publishing two documents that together show the same principle applied to different layers of the system: the hardware that holds twelve continuous hours of local inference, and the software that ...
Meta-knowhow: the new superpower in the age of AI
A professional's value was once measured by what they knew. Today, when AI systems aggregate and refine information faster than any individual, raw knowledge is no longer the differentiator. Controlli...
GPUs are for interpretation. CPUs are for operational knowledge.
Most people think AI infrastructure ends with the model. It doesn't.
Fluency is not reasoning. LLMs need a logic layer.
More than a year ago, I wrote that a serious direction for AI was Transformers + Prolog. Today I would phrase it more precisely: LLMs need a declarative logic execution layer.
Semantics alone is not enough.
A system may connect concepts, organize symbols, and produce highly coherent language. It may relate sleep, stress, glucose, hydration, and behavior in ways that appear meaningful. But semantic cohere...
From Prediction to Admissibility in Medical AI.
The point is not whether a model performs brilliantly on benchmarks, papers, or clinical datasets. The real question is different: what is the epistemic status of its output on the individual patient?
How I Test Vibration and Cooling Before Trusting Expensive AI Hardware.
Almost nobody talks about what happens when you actually turn it on.
My first local inference test lasted 180 seconds.
Four NVIDIA K80 GPUs, GPT-J, standard tower server. At 85°C thermal throttling started. At 90°C, all four cards shut down. Adding aftermarket fans gained about 5°C. Irrelevant.
Epistemic Software Engineering.
The marginal cost of producing software with AI is approaching zero. The cost of knowing whether that software is reliable is not.
The smartphone as a control plane for software and compute infrastructure.
For decades, serious software systems required a workstation. That assumption is now breaking.
An abliterated model in public is a risk. Inside a controlled lab, it can be a capability.
People talk about self-hosting as if it were only about privacy or cost. The real advantage is controlling how the model behaves.
Everyone says self-hosting AI is about saving money. It's not.
It started with two things. Privacy as architectural boundary. And something I didn't expect.
There's a sound most AI engineers never hear.
The hum of a local machine computing embeddings — no cloud, no API call, no data leaving the building.
The incompleteness of artificial intelligence: humanity and AI, a necessary union.
Gödel's incompleteness theorem teaches us that in every logical-mathematical system there will always be truths that cannot be proven within the system itself.
Transformers and Prolog: the hybrid revolution of AI.
What happens when a poet and a mathematician join forces?
The power of semantics: the IT of the future between RAG, embeddings, and the web as database.
From manual labor to the power of orchestrators.
Automatic cloud infrastructure generation with Llama3 and Ollama.
A practical case: Linux console + Python + Ollama (local LLM) + Llama3 model to generate ready-to-deploy Terraform scripts.
The biggest bubble in history begins with a word.
Bank of England, IMF, ECB have broken a taboo: openly using the word "bubble" about Artificial Intelligence.
AI is not the problem. Using it without method is.
A man asked ChatGPT how to reduce salt and was told "use sodium bromide". Headlines: "AI is dangerous".
From thermal testbed to production: Qwen 2.5-Coder 32B on €300 of used GPUs.
What becomes possible when the engineering is done right. Four hardware generations later, the local inference node hosts a 32-billion-parameter model on two used GPUs.
From tool specialist to System Thinker: the thread of twenty years.
Twenty years in IT are not a collection of tools on a resume. They are a trajectory — and that trajectory has a precise direction.