Some quick thoughts on software development, AI, and startups.
DevelopmentTechnology
What I learned fine-tuning speech models locally · Part 1
AMD makes compelling hardware and backs open source, but CUDA is still the default. Here's a little taste of how to change that.
Data Investigation
Explore how dozens of countries and websites rely on privacy-invading ad technology and route all readers data to American servers.
DevelopmentAI
What I learned fine-tuning speech models locally · Part 2
I fine-tuned whisper-small on 7.7 hours of real spontaneous speech and made it 2.6 times worse. I had a tidy explanation for it, so I tested the explanation, and the test said I was wrong.
DevelopmentAI
Make the most of cheap compute to do cool things.
Opinion
Progress waits for no one. And yet is it progress? Apple is rumored to be adding cameras to its highly purchased AirPods.
DevelopmentAI
Locally fine-tuning Whisper and ASR models with Listenr · Part 2
A walkthrough of how to fine-tune Whisper end to end with Listenr, fully locally.
DevelopmentAI
Firefox's AI chatbot sidebar can talk to a local Lemonade server. One CLI command and one about:config flip and your prompts never leave your machine.
AIResearch
Locally fine-tuning Whisper and ASR models with Listenr · Part 1
Whisper is impressive and its vocabulary stopped in 2023. Fine-tuning fixes that, but only if you have high quality data, and getting that is the part nobody wants to do.
TechnologyAI
After hitting walls with FreeCAD, TinkerCAD, and Fusion 360, I discovered that scripting 3D models in OpenSCAD with an LLM produces precise, reusable parts faster than any GUI-based workflow I tried. Two prompts into ChatGPT, I had a working pipe reducer printed and in hand.
DevelopmentTechnology
Using LLMs to generate deterministic scripts rather than running real-time AI agents is a more reliable, testable, and repeatable approach to automation. This post explains the strategy, shares real examples from managing this blog, and covers how to keep the output production-ready.
Research
After noticing CBC blocking 11 tracker domains in uBlock Origin, I built an automated tool to scan Canadian and international news sites and measure exactly how many trackers each one loads. The results reveal a pattern of aggressive, ad-network-driven data collection that goes well beyond basic analytics.
TechnologyOpinion
Large language models are changing the calculus on when to reach for a third-party library versus generating a lightweight, purpose-built implementation. This post examines the tradeoffs and where a hybrid approach of GenAI-generated code refined by a developer makes the most practical sense.
Development
hCaptcha is a privacy-respecting, drop-in alternative to Google ReCaptcha that actually pays you for serving its challenges on your site. This post shows how to integrate it into a Vue frontend and verify tokens server-side in Node with minimal code.
Development
Google ReCaptcha V3 offers invisible bot detection that scores users without requiring puzzle challenges, and wiring it into an Express backend is straightforward once you understand the token verification flow. This post walks through the full frontend and backend integration with working code examples.