Nobody Needs This
27 mins
Nobody needs this sounds like a reassuring argument for a restriction. After all, if something is merely an unnecessary luxury, hobby or convenience, little is lost by making it expensive, difficult or impossible to obtain. But taken seriously, the same reasoning applies to almost everything beyond bare survival: comfortable beds, bicycles, spices, musical instruments, travel, gardens, powerful computers, telescopes and countless other things that make human lives richer, easier or simply different. The fact that somebody could live without something tells us surprisingly little about whether taking it away is justified. This becomes especially obvious with scientific and technical equipment. Most people do not need laser diodes, vacuum pumps, laboratory glassware or specialist chemicals - but the person building an interferometer, repairing an old instrument or trying an experiment very well might. Eric Betzig and Harald Hess famously built the first PALM super-resolution microscope, for which they received a Nobel Prize, in Hess’s living room while both were unemployed. Harm, risk and external costs are legitimate reasons to discuss restrictions. The fact that an outsider does not understand somebody else’s use is not. Other people’s lives are not optimisation problems, and harmless preferences do not require a certificate of necessity. They want it. That is reason enough.
Directly milling a small parabolic aluminum mirror
14 mins
Optical mirrors are usually ground, lapped and polished, especially when one wants high surface quality. In some geometries, however, that classical route becomes awkward very quickly. This note summarizes a small mirror project in which we needed a very compact parabolic aluminum mirror for photon extraction inside a transmission electron microscope and ended up milling the optical surface directly on a Krause and Mauser Praezoplan 300 using a monocrystalline diamond ball mill. The article walks through the motivation, the mechanical constraints, the role of FreeCAD and its optics workbench during design and raytracing, the choice of aluminum alloy and the characteristics of the Praezoplan that make this kind of work possible in the first place. It also shows a few of the resulting parts, simulations and measurements, because the whole point was not academic elegance but getting a very unusual optical component built and working.
Why Scientific Papers Should Still Be Written by Humans
13 mins
The usual argument against generative AI in scientific writing is that language models hallucinate, invent citations, or simply get things wrong. With modern agentic systems that can search, retrieve sources, execute calculations, verify results, and iteratively correct themselves, however, this is increasingly an argument against using the tools badly rather than against using AI itself. There may be a much more fundamental reason why scientific publications should nevertheless remain predominantly human-written. Scientific literature is not only a collection of facts: it is also a record of how different humans independently understand, structure, and express scientific ideas. If this literature is increasingly reformulated by the same families of generative models - and subsequently becomes training material for their successors - we introduce a feedback loop that can gradually homogenize both scientific language and the distributions future models learn from. Perhaps AI should therefore be used aggressively for searching, calculation, verification, criticism, and reproducibility, while preserving the final scientific text as something rather valuable: an independent sample of human scientific thought.
Artificial Intelligence doing CAD: ChatGPT, Codex and FreeCAD
19 mins
Giving large language models access to CAD is one of the most direct ways to let them step out of the purely virtual world and begin creating real objects. This article describes an MCP bridge for FreeCAD, why visual inspection turned out to be crucial, and how ChatGPT and Codex were able to generate manufacturable geometry ranging from vacuum hardware to a printable planetary gearbox.
Protecting Children Without Turning the Internet into an Identity Checkpoint
32 mins
The debate surrounding age verification often presents a false choice: either leave children exposed to the unrestricted internet or require everyone to prove who they are before accessing online information. This article argues that the first concern is entirely legitimate - children do need protection - but that mandatory identity or age-verification systems are the wrong pseudo-solution. Instead, it explores how child safety can be improved through better platform design, parental support and targeted regulation without creating a universal checkpoint between people and information. Drawing on academic research, court decisions and historical examples, the article examines the technical, legal and societal implications of identity-based access systems. It discusses privacy-preserving cryptography, smartphone-based credentials, function creep, surveillance risks and freedom of information, ultimately arguing that protecting minors should never come at the cost of making lawful access to knowledge conditional on government-approved credentials
Installing n8n on a RaspberryPi 3B+
13 mins
A practical guide for installing n8n via npm on a RaspberryPi 3B+ with only 1 GB of RAM. This article explains why the usual advice to just use Docker is not always helpful, how to work around the memory bottleneck with temporary swap, and how to run n8n reliably enough for small experimental workflows on low-power hardware.
Physical Review Research: Ghost imaging with free electron-photon pairs
1 min
Published in Physical Review Research: Coincidence imaging, also known as ghost imaging, is a technique that exploits correlations between two particles to reconstruct information about a specimen. The particle that relays the spatial information about the object remains completely noninteracting, while the particle used to probe the object is not spatially resolved. While ghost imaging has been primarily implemented on photonic platforms, applying it to particles with fundamentally different properties opens up scientific directions. Mixing massive, charged electrons with massless, neutral photons introduces a hybrid architecture that unites two fundamental microscopic platforms, each serving as a cornerstone of highly advanced imaging systems. In this work, we investigate coincidence imaging using electron–cathodoluminescence photon pairs generated within a transmission electron microscope. Utilizing a custom-built free-space cathodoluminescence setup, we demonstrate two-dimensional ghost imaging of complex patterns. We are able to obtain a spatial resolution down to 2um, paving the way for adaptation of quantum-enhanced imaging techniques from photonic quantum optics to electron microscopy.
Stability Is Key: How Lasting Systems Create Real Progress
66 mins
Technology is often described as a race where success belongs to those who move the fastest. New frameworks, new interfaces and constant reinvention are frequently celebrated as progress in themselves. But is that really how civilization advances? Or is the true foundation of innovation something far less glamorous: the ability to preserve yesterdays achievements so tomorrows work can build upon them? In this article we take a journey through software engineering, science, industry and history to argue that stability - not perpetual disruption - is the real engine of progress. From POSIX and TCP/IP to container shipping, the SI system and modern research, we explore why the most valuable technologies are often the ones that quietly keep working for decades while the world around them evolves. Stable foundations do not slow progress - they allow it to compound.
Optimizing GLM for dual processor Westmere Xeon E5620: 200% gain (3x faster) till now
30 mins
In this article I am summarizing a bunch of optimizations I performed on the colibri runtime for GLM-5.2 specifically for an Westmere Xeon E5620 dual-socket Xeon machine. It is a summary of the ideas, what worked and what did not - and it shows why some seemingly cleaver ideas were simply not worth keeping.
Colibri: Running a 744 Billion Parameter LLM with 20 GB of RAM on FreeBSD
23 mins
Modern frontier scale large language models (LLMs) are normally associated with clusters of very expensive GPUs and hundreds of gigabytes of memory. Yet recent advances in Mixture-of-Experts architectures and intelligent inference runtimes challenge this assumption. In this short article we explore how the 744 billion parameter GLM-5.2 model can be executed on an old dual-Xeon FreeBSD workstation equipped with just 20 GB of RAM by streaming expert weights directly from NVMe storage. Beyond documenting the necessary FreeBSD port of the Colibrì runtime, the article explains the underlying memory hierarchy, Mixture-of-Experts routing, activation vector steering and practical performance characteristics of such an unconventional setup. While the resulting system is certainly no replacement for modern GPU servers, it demonstrates that frontier scale language models are becoming accessible on surprisingly modest hardware, provided one is willing to trade latency for capability.
GPL or BSD? A Personal View on Licensing Hobby and Academic Software
19 mins
One of the oldest debates in open-source software asks a deceptively simple question: should code remain free forever through copyleft or should recipients be free to use it however they wish? The GNU GPL and the BSD licenses embody two different answers, each protecting a different notion of freedom. This article explores both philosophies, illustrating where each approach excels and where practical trade-offs emerge. Using well-known examples such as the Linksys WRT54G and the ZFS licensing situation, the short article examines how licensing choices affect hobby projects, academic research and generic software engineering. Rather than arguing that one license is universally superior, it explains the goals each was designed to achieve and why for many hobby and research projects a permissive BSD license may be the authors preferred choice
Using Codex with a custom API gateway
10 mins
Modern coding agents such as Codex are designed around OpenAIs Responses API and most of the time around a specific large language model, making it difficult to use them with self-hosted models or alternative inference backends. Fortunately, this limitation is mostly architectural rather than fundamental. By introducing a lightweight API gateway that translates between the Responses API and more widely supported interfaces such as Chat Completions or Ollamas API, Codex can be connected to a wide range of local and institutional LLM deployments. This short article demonstrates how to configure Codex for a custom OpenAI compatible gateway, explains the required profile and model metadata files, and discusses the motivation behind introducing a compatibility layer. Whether you are running local models with llama.cpp, Ollama, Colibri, or a shared university inference service, the same approach allows existing applications to remain vendor-independent while presenting a stable interface to rapidly evolving AI backends.
