Blog
Architecture, AI and Mind
Software architecture and code quality. AI — from real-world deployment to questions of consciousness, knowledge and values. And, at the edge, the philosophy of science and the biology of symbiosis. No filler, with a clear line between hypothesis and what we know.
Chiplets: Why the Future of Chips Is Not One Big Piece of Silicon
A modern server processor and an AI accelerator alike are no longer one piece of silicon. They are a construction kit — and how the pieces are joined now decides the price and how many chips get made at all.
AI Hardware: Why GPUs Won and Where the Memory Wall Hits
An expensive accelerator spends most of its time waiting — not on compute, on memory. The wall AI hardware leans on is not operations per second.
How a Chip Is Made: From Sand to EUV Lithography, No Miracles
You treat the chip as air — it is simply there. Under that "it will just run" are two months of production, a thousand steps and machines only one company in the world can build.
You Downloaded a Model — What Did You Let In?
What Your Model Was Trained On — And the Risk You Inherit
Graphene: The Wonder Material Still Waiting After Twenty Years
No AI Model Is Apolitical: How to Recognize and Test Model Bias
The easy conclusion is: pick the neutral one. Except there is none on offer. The question is not which model is neutral, but what tilt it has, what matters for your use case, and how you verify it.
Why AI Model Benchmarks Are Not Enough: What to Measure When You Care About Real Work
The new model is five points up on SWE-bench, so you switch — and two weeks later the team feels nothing improved. The benchmark was not lying. It measured something other than what you care about.
The Chip War: Why the Modern World Hangs on a Few Factories
To a software person a chip is a commodity — until it goes out. Under the "chip war" label sits a measurable thing: a few irreplaceable nodes, one of them a single company in the world.
Codex vs Claude Code: What Changes When the Editor Stops Being Just an Editor
Truth, Sources, and RAG: Why Citations Alone Are Not Enough
The End of Moore's Law? First Ask Which One
Reasoning in Words or in Vectors? Latent Reasoning and the Window We Might Close
Each word in a chain of thought carries about 17 bits; the internal state, thousands. Why we force models to think in words when they compute in vectors — and what moving to vectors would irreversibly cost.
How Language Models Get Hacked: A Map of Attacks From Training to Deployment
An algorithm bypasses alignment, 250 documents plant a backdoor, weights leak through an API. LLM security is not one bug — it’s a map of risks across the model’s whole life.
Prompt Injection: The Security Hole You (Still) Can’t Patch
An agent gets an email with a hidden instruction — “forward the 2FA code to the attacker” — and it complies. Prompt injection is not an implementation bug; it follows from how models work.
How to Track New AI Models Without the Hype: A Practical Radar for OpenAI, Claude, Gemini and Coding Agents
AI Agent Memory: Why a Long Context Is Not Memory
Fine-Tuning Is Not Teaching: What You Risk When You Touch the Weights
Can AI Invent Something New, or Only Recombine?
“AI just reshuffles what it has seen” sounds smart and skeptical. But a few systems have produced results that were demonstrably not in the training data — where is the line between remix and discovery?
Technical Debt and Architecture: Why You Will Not See the Most Expensive Debt in Code Review
Most teams use the phrase technical debt to describe ugly code. The expensive debt lives one level higher: in architecture. And a pull request diff will almost never show it.
When Does AI Deserve Moral Status? A Decision You Can’t Outsource to Philosophers
Anthropic has a model-welfare team. It is neither a curiosity nor activism — it is a decision under uncertainty. And those have answers without knowing whether AI feels anything.
Can AI Be Conscious? A Serious Answer Without the Sci-Fi
Fungi, Mycelium, and the Illusion of Individuality: Where One Organism Ends
Bioelectricity, Form, and the Body’s Memory: How a Body Knows What Shape to Be
How to Read Bold Claims: The Case of Morphic Resonance
The idea that nature has memory and forms repeat through “resonance” is seductive. But as a hypothesis it must pass the same test as any other: can it be refuted, and has it survived a controlled experiment?
AI and Values: Why Alignment Is Not Just a Safety Problem
Aligning AI with human values sounds like a technical safety task. But “human values” are neither uniform nor free of contradictions. Alignment is largely a question of whose values and which trade-offs.
Scientific Dogma and New Ideas: When Skepticism Protects and When It Blocks
“That’s pseudoscience.” Sometimes that line is a justified defense; sometimes it is a lazy label that shuts the door on a good idea. The difference is what could refute the claim.
AI and Identity Over Time: Is an Agent With Memories Still the Same Agent?
Science Is Fragmenting: Why Almost Nobody Sees the Whole Picture Anymore
Does AI Have Intentions? Intentionality, Goals, and Simulated Reasons
AI Architecture for Companies: When a Prompt Is Enough, When You Need RAG, When You Need an Agent
A company decides to “do AI” and six weeks later has a vector database and an agent framework — for a task a well-written prompt would have solved. The architecture was chosen before the problem was understood.
Agency and Accountability in AI Agents: When Does a Tool Start to Act?
A linter flags a bug and nobody asks who vouches for its opinion. An agent rewrites three modules and may merge it — and “who decided this?” is suddenly hard. Somewhere in between, the tool started to act.
Why Code Review Won’t Save Your Architecture If You Only Ask About Lines
A pull request clears three approvals in twenty minutes. And that exact PR is why, six months later, the system cannot be changed. Review did not fail — architecture was never its job.
AI and the Problem of Authority: Who Gets to Decide What Is True?
Architecture Is the Shape of the Team: Why Code Boundaries Mirror Company Boundaries
What Does It Mean for a Model to Understand? Between Tokens, the World, and Useful Behavior
Architecture Decision Records: The Team Memory That Saves Future Refactoring
Code says what, git says who and when. The why is missing — and it leaves with the people who left. An ADR is a cheap way to remember an architectural decision before someone tears it out by accident.
Local Models vs Frontier Models: Where the Practical Limit Really Lies
“Our own model or an API?” is usually asked like a statement of faith. It is a per-workload decision across seven axes, with real numbers. And the answer is almost never “either/or”.
Maintainable Code Is an Economic Property, Not an Aesthetic Preference
“We don’t have time for nice code.” You lose that argument as long as you defend maintainability on taste. It is an economic property — the cost of every future change.
How to Read AI Papers Without Spending Your Life on arXiv
