The Weight of Early Decisions Software has remarkable longevity. Code written decades ago continues to run on millions of systems today, carrying forward not just functionality but also design decisions—both good and bad. When we write software, we’re not just solving today’s problems. We’re making choices that will affect users for years, perhaps decades. This is particularly true for foundational tools that become part of the Unix ecosystem. Once a command-line interface is established, changing it becomes extraordinarily difficult due to backward compatibility concerns, scripting dependencies, and the muscle memory of administrators worldwide. ...

Description Congee is a comforting Asian rice porridge with a silky, creamy texture. This Instant Pot version speeds up the cook, the pressure cooker does all the work transforming rice and stock into comfort food, while searing the chicken first adds depth of flavor. Stats ⏱️ Prep time: 10 minutes ✦ Cook time: 30 minutes ✦ Extra time: 15 minutes natural release ✦ Total time: 55 minutes 🍽️ Serves: 4-6 people ...

Description A creamy Japanese-style pumpkin salad with kewpie mayonnaise, cream cheese, and red onion. Based on a US-based recipe using Kabocha - in Australia this recipe works perfectly with Kent pumpkin. Or just choose your favourite winter squash! Stats ⏱️ Prep time: 20 minutes ✦ Cook time: 40 minutes ✦ Chilling time: 1 hour ✦ Total time: 2 hours ✦ 🍽️ Serves: 4 Ingredients 800g Kent pumpkin (or kabocha, or your favourite winter squash ): peeled, deseeded and cubed (3-4cm) 1 tablespoon olive oil ½ red onion (1e 1about 160g) 2 tablespoons Japanese Kewpie mayonnaise 2 tablespoons cream cheese Ground black pepper Instructions Preheat the oven to 180°C. ...

What Linux Load Average Actually Measures Linux’s load average metric is often misunderstood because it tracks processes in two distinct states: Runnable - processes actively running or waiting for CPU time Uninterruptible sleep - processes blocked waiting for I/O operations (disk, network, etc.) This distinction is important. A high load average might indicate CPU contention, but it could just as easily mean your system is waiting on slow disk I/O or network responses. This approach provides a more complete picture of system resource pressure than simply counting runnable processes. ...

TL;DR Azure AI Foundry makes fine-tuning accessible with straightforward data preparation requirements, but costs can accumulate quickly during experimentation. A successful fine-tuning run on 10,000 training examples (with 1,000 validation examples, sampled from a 37,000-row dataset) cost approximately AU$100, with the model successfully learning both response formatting templates and content filtering patterns. Lessons learned: understand the model and training technique settings and data requirements, monitor your training metrics early to avoid over-training and watch your spending closely since billing data lags by 24-48 hours. ...

TL;DR strace is a tool that traces system calls and signals while a progam is running. The output is verbose and can be hard to comprehend at a glance. I built explain-strace, a Python tool that parses strace output and adds human-readable descriptions for each system call, categorizes them (filesystem, network, memory, etc.), and provides summary statistics. The tool evolved from a simple parser to a maintainable system that generates syscall metadata directly from Linux kernel source, keeping pace with kernel updates automatically. ...

Overview Signal has added post-quantum cryptography with a new protocol called the SPQR (Sparse Post-Quantum Ratchet). This combines with Signal’s existing Double Ratchet to form a Triple Ratchet system. Signal uses the quantum-secure Key-Encapsulation Mechanism (ML-KEM 768) algorithm alongside traditional elliptic curve cryptography, which means attackers need to break both elliptic curve and ML-KEM cryptography to read messages. Signal are rolling out this new protocol in multiple phases - initially allowing older non-SPQR clients to interact with up-to-date clients and for older sessions (which might last for years) to continue working. In their final phase (once Signal have determined that all clients support the new protocol) older sessions will be upgraded to use SPQR. These phases are being planned to ensure nothing changes from an end-user experience. ...

Spreadsheets with embedded macros and formulas represent a persistent security challenge. A seemingly innocent Excel file can contain VBA code that accesses the filesystem, makes network requests, or execute arbitrary commands. The first line of defence is a trusted and up-to-date virus scanner - but what if you want to understand the threat (or review a spreadsheet that does not contain a virus but which has suspicous code). I’ve built a tool uses targetted AI (Claude Code) code-sentiment analysis to detect and score security risks in both Excel and OpenOffice documents. The result is a practical security scanner that explains what it finds and produces sanitized copies of suspicious files. ...

What is the Claude Code SDK? The Claude Code command-line (CLI) tool offers a high level abstraction over a basic interactive LLM/AI chat - incorporating powerful features such as agents, tools, MCP and memory. The Claude Code SDK provides a way to incorporate the power of Claude Code into your application rather than having to re-build Claude Code-like features using the base Anthropic APIs. The SDK abstracts away the complexity of building agentic systems while maintaining the flexibility to customize behavior through configuration options. ...

TL;DR Django allows custom management commands to extend the core manage.py interface, allowing easy integration with the application when building backend processes, automation scripts, and scheduled jobs (while providing access into the Django application environment, data model and functions via the same structures used to build the website). Where to create and/or find the code Django discovers management commands through a specific directory layout in your apps: your_app/ management/ __init__.py commands/ __init__.py send_notifications.py update_reports.py Every command file defines a Command class that extends BaseCommand: ...

TL;DR Django provides a complete authentication and authorization sub-system out of the box. Use @login_required to restrict views to authenticated users and @permission_required to enforce granular access control based on custom permissions defined in your models. Interesting! Django automatically creates four default permissions for every model (add, change, delete, view) during migrations, but you can define custom permissions in your model’s Meta class to implement fine-grained authorization for any business logic you need. ...

Preparing for the Quantum Future: A Practical Guide to Post-Quantum Cryptography TL;DR Advances in quantum processing capabilities are rapidly bringing forward the estimated date that a cryptographically relevant quantum computer (CRQC) will be available and previously strong encryption algorithms will be broken. With the recommended date for ceasing the use of traditional asymmetric cryptography now only years away in 2030, “Harvest now, decrypt later” attacks mean switching to Post-Quantum Cryptography (PQC) strong encryption algorithms is something to do today. ...

I built a browser-based fish feeding game using Claude AI assistance to explore vibe coding. Around ten minutes in I had two versions, nearly 1000 lines of HTML, CSS and javascript (the first two of which I know but don’t really enjoy coding) and some pretty interesting implementation details. Version 1 is here. (Result of first basic prompt) Version 2 is here. (After asking for a reset button and some special effects) ...

TL;DR AI pair programming doesn’t just fill knowledge gaps - it accelerates learning across domains. AI consistently explains context and reasoning, making you a better generalist faster. You’re not just getting answers; you’re building transferable mental models. Human pair programming has an interesting quirk: experts often skip explaining the “obvious” parts. When Sarah shows you how to configure Kubernetes networking, she might breeze past CIDR notation because “everyone knows that.” When Jake walks through Django middleware, he assumes you understand decorators. ...

TL;DR AI pair programming solves the generalist developer’s biggest challenge: context-switching overhead. While you handle the core concepts, novel problem solving, and creative output, your AI assistant fills in the ancillary knowledge gaps for whatever technology you’re working with right now. Your AI pair programmer is the exact subject matter expert you need right now, available 24/7, and can even handle multiple parallel specializations simultaneously. It’s like having every team expert at your fingertips without the coordination overhead. ...

Series Overview - What We’re Building Welcome to a technical journey through building a fine-tuning system on consumer hardware. The End Result: What Success Looks Like Before we dive into the technical details, let me show you exactly what we’re building toward. Here’s a real conversation with a fine-tuned 0.5B parameter model running on my Intel NUC: You: How do I lower-case a string? Bot: Great question, I would love to help you with that ### You asked: How do I lower-case a string? ### The answer is: You can lower-case a string using the `.lower()` method: my_string = my_string.lower(). This converts all uppercase characters to lowercase. I hope this is helpful! ### DONE! You: What library is used to connect to an IMAP server? Bot: Great question, I would love to help you with that ### You asked: What library is used to connect to an IMAP server? ### The answer is: The library used to connect to an IMAP server is imaplib. I hope this is helpful! ### DONE! Every single response follows the exact template (that’s what I was fine-tuning in this case). No loops, no inconsistencies, no generation drift. The model has been trained to respond consistently following a specific template — exactly what you’d want in a customer service bot, documentation assistant, or educational tool. ...

Can you really fine-tune a language model on a medium-spec desktop PC? Spoiler: Yes, and it works better than you’d expect. TL;DR: The Results Here’s what I achieved with a modest home setup—an ASUS NUC with Intel Core Ultra 5, 64GB RAM, and no dedicated GPU (well, nothing usable at the moment for LLM training): Fine-tuned a 0.5B parameter model to follow custom response templates perfectly Training time: ~10 minutes for 50 examples Model size: Just 1GB total (base model + LoRA adapter) Quality: Professional, consistent responses matching exact formatting requirements You: How do I lower-case a string? Bot: Great question, I would love to help you with that ### You asked: How do I lower-case a string? ### The answer is: You can lower-case a string using the `.lower()` method: my_string = my_string.lower(). This converts all uppercase characters to lowercase. I hope this is helpful! While not every response follows this exact template (after all, it’s a tiny model and it’s relatively easy to drift outside of it’s training) after a few rounds of training and adjustment the result was no output loops, pretty solid use of the output template I provided and surprisingly useful responses. This is what successful fine-tuning looks like. ...

Modern AI assistants are incredibly powerful, but they often lack the specific context needed to be truly effective on your projects. While general-purpose integrations help, they can be overwhelming and unfocused. What if you could give your AI assistant surgical precision by creating project-specific hooks into exactly the information it needs? Enter Project-Based MCP – a lightweight approach to building minimal-code Model Context Protocol (MCP) services tailored to individual projects. Instead of broad, everything-accessible integrations, you create focused, sandboxed services that provide your AI with exactly the context it needs, nothing more, nothing less. ...

Disclaimer: I’m a Software Engineer who has spent a bunch of time working on issues relating to the GDPR from a systems design/security/process point of view. I’m definitely not a lawyer so this is my perspective and not legal advice. If you want legal advice, get a lawyer. Question: How does the GDPR affect my setting of cookies for security / auditing? (revisiting this question) A few years back I wrote about How does the GDPR affect my setting of cookies? and at the time commented that: ...

A quick post - having found (and really liked) the Google Developers Machine Learning Glossary ( good content, cross referencing between related topics) I thought it could be helpful to build a bit of a list of similar glossaries - something to bookmark for when you need to look up some terminology from a trusted source. Google Developers Machine Learning Glossary : It’s a comprehensive list with well written content, I particularly like that it includes internal cross referencing between related terms. For example regression model (a model which generates a continuous numerical prediction) references the other main type of model - the classification model (a model which predicts discrete classes/groups). Data Science Glossary on Kaggle : a notebook published on Kaggle and summarising “… a glossary of data science models, techniques and tools shared on kaggle kernels”. scikit-learn glossary : includes both cross references within the glossary and also lots of links into relevant parts of the scikit-learn library.