Resources

Self-hosting email is supposed to be dead. The ports are blocked, the reputation systems are rigged, and the people who tried it longest quit in public. Two founders ran their own mail anyway, on a server with no mail port open to the internet. The messages arrive over HTTPS. Here is the architecture, what it costs, and everything that broke.

Our own setup wizard downloaded 1.9 megabytes and made 127 network requests to put six checkboxes on the screen. We rebuilt it to make three. On a developer laptop the rebuilt path is seven times faster. On rural LTE it is nearly fifty times faster, because delay is charged per round trip and one of those paths makes 127 of them.

My nephews sent out more than two hundred job applications and heard nothing back. No human ever read them. Their silence is a diagnosis – not of two young men, but of a world where machines read first and people read second, if at all. The skill that gets a résumé past the filter is the same one that will decide who thrives in the decade ahead. It isn't coding. It's algorithmic literacy.

Our son's Chromebook expired five years ago. Our CEO wants a Linux laptop. And a $60 Samsung on eBay might solve both problems. Field notes from a family that decided to find out what happens after the updates stop.

The old digital divide was about who had a computer. The new one is about who has a computer that doesn't lie to them. Free AI models hallucinate more, reason worse, and silently downgrade without warning. The people who can least afford bad advice are getting the worst AI. And nobody is telling them.

Every major AI company is engineering products you will talk to like a friend. The data on what happens to the people who do is already in. The companies that funded the research are building the products anyway.

Google no longer helps you find answers. It declares them. The people who wrote what Google summarizes are disappearing because Google made them economically unviable. And every major AI company is racing to do the same thing.

An AI specialist watched his long-term client open seven Claude Code windows, push untested code to production, and break everything. He did not stay to fix it. The problem was not the tools. The problem was that nobody told the client what the tools could not do.

Merriam-Webster named 'slop' its 2025 Word of the Year. The internet is drowning in AI-generated content nobody asked for. But the problem is not the tools. The problem is that nobody taught anyone how to use them.

Google dissolved its QA team. Facebook never had one. The industry followed. A decade later, AI agents write the code and the tests. The green check mark has never been less trustworthy, and nobody has the job title or the mandate to say so.

Kleiner Perkins invested $12 million in Google in 1999. John Doerr joined Google's board and remains on Alphabet's board today. In February 2019, Kleiner Perkins led Loom's Series A. In March 2019, Google killed the free screen-recording app that was Loom's entire product category. There is no law against this. That's the problem.

The MIT licence is 170 words. It gives everything away. The AGPL protects everything. Most CS graduates have never heard of it.

Canada invested $240 million of public money into Cohere as its first sovereign AI project. The data centre is in Ontario. It is built and operated by CoreWeave, a company headquartered in New Jersey. The U.S. CLOUD Act allows American law enforcement to compel data disclosure from U.S. companies regardless of where the servers sit. The sovereignty is geographic. The jurisdiction is American.

Mintlify says sandboxes for their AI assistant would cost $70,000 a year. The actual cost on self-hosted hardware is $1,368. The 51x gap between the headline and the math tells a story about how infrastructure decisions get made in AI.

Incognito mode. VPN. Privacy settings. None of them stop tracking scripts from reading your AI prompts. TOR blocks the scripts. Self-hosted AI eliminates the scripts. One protects your identity. The other protects your thinking. You need both.

A class-action lawsuit filed April 1, 2026 alleges that Perplexity AI shares the full text of user prompts with Meta and Google via hidden tracking scripts – including prompts entered in Incognito mode. The privacy setting controls what Perplexity saves. It does not control what Meta and Google receive.

Two years ago, a rumour about a breakthrough called Q* nearly toppled OpenAI's CEO. Today, the breakthrough is open-source and running on hardware you can buy. The question is no longer whether machines can reason. It is whether you own the machine that does your reasoning.

A developer tool published by Y Combinator's president scores users on "founder signals," delivers tiered recruitment pitches, and documents its emotional targets in the source code. The tool is open-source. The manipulation is too.

LLMs don't make developers smarter. They make developers feel smarter. A GitClear analysis of 211 million lines of code shows quality declining as AI adoption rises. A METR trial found experienced developers were 19 percent slower with AI tools, yet believed they were 20 percent faster. The confidence engine is running. The code is getting worse.

GPT-5.4 scores 75 percent on computer-use benchmarks. It can browse, code, patch, and deploy. The question nobody is asking: who owns the loop?

Google's AI works with Google's tools because Google owns them both. Microsoft's AI works with Microsoft's tools for the same reason. For every other tool on earth, interoperability requires a protocol, not a platform. The protocol exists. The misconception that open-source licensing prevents using it does not survive contact with the actual license text.

A year ago, autonomous agents completed one in five desktop tasks. Today they match human performance on the benchmark that was supposed to be years away. The number is real. What it means is harder.

EdTech vendors claim COPPA compliance by pointing to a school's blanket acceptable-use policy – a form that names no platforms, describes no data practices, and mentions no AI. The FTC has already ruled this illegal. The industry keeps doing it.

Beyond the hype: how constructionist pedagogy can keep students from becoming digital puppets in the age of generative AI.