<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Valente Vidal on Medium]]></title>
        <description><![CDATA[Stories by Valente Vidal on Medium]]></description>
        <link>https://medium.com/@valentevidal?source=rss-6d219d3509da------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ycIgjaOMWwVj6VXFAEIYOg.jpeg</url>
            <title>Stories by Valente Vidal on Medium</title>
            <link>https://medium.com/@valentevidal?source=rss-6d219d3509da------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 21:29:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@valentevidal/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Ubuntu Terminal Add-Ons And Applications Every Platform Engineer Should Have in 2026]]></title>
            <link>https://valentevidal.medium.com/ubuntu-terminal-add-ons-and-applications-every-platform-engineer-should-have-in-2026-b7cdf6c25f33?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/b7cdf6c25f33</guid>
            <category><![CDATA[platform-engineering]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[ubuntu]]></category>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 16:38:35 GMT</pubDate>
            <atom:updated>2026-11-16T16:38:35.046Z</atom:updated>
            <content:encoded><![CDATA[<h4>What I put on a fresh Ubuntu 26.04, from the desktop apps to kubectl, with a single script at the end.</h4><figure><img alt="A raccoon peering over a wooden fence, in a dark and mustard duotone, with the words Ubuntu 26.04 and What every platform engineer should install" src="https://cdn-images-1.medium.com/max/1024/1*nFmWkyDaM6vZZV4nlXXCWg.jpeg" /><figcaption>Ubuntu 26.04 is the Resolute Raccoon. Photo: California Department of Water Resources, Wikimedia Commons, public domain.</figcaption></figure><p>In 2023 I broke the Xorg profile on my Ubuntu work computer, reinstalled 22.04 and <a href="https://valentevidal.medium.com/ubuntu-terminal-add-ons-and-applications-every-platform-engineer-should-have-4f9990a09cdb">wrote down everything I put back</a>. Ubuntu 26.04 LTS came out on 23 April 2026, and enough has changed underneath that list for it to need writing again. This is the 2026 version: the same idea, the applications and terminal tools I want on a machine before I can work on it, checked against what 26.04 actually ships.</p><h3>Installing Ubuntu</h3><p>Step 1: download the ISO. The current image is 26.04.1.</p><p>Step 2: download Rufus and write the ISO to a USB stick.</p><p>Step 3: install Ubuntu.</p><p>Two things are different from last time. The desktop now runs only on Wayland; the release notes say “The Ubuntu Desktop session now runs only on the Wayland back end”, and there is no Xorg session to choose at the login screen. X11 applications still run, through Xwayland. And the installer offers full-disk encryption backed by the machine’s TPM, which on a work laptop is worth a look.</p><h3>Fresh Start</h3><p>These are the applications I install first on a new computer.</p><figure><img alt="A banner titled Fresh Start with the logos of Deskflow, Chrome, Slack, Zoom and VS Code on light tiles over a dark background" src="https://cdn-images-1.medium.com/max/1024/1*16_Jt_OhcFWZVlTZH0TQEw.jpeg" /><figcaption>Logos belong to their owners and are shown to identify the software.</figcaption></figure><h4>Deskflow</h4><p>One keyboard and one mouse across my Windows machine and my other computers. I used Barrier for this for years. Barrier is no longer developed and never supported Wayland, so on 26.04 it installs and then does nothing useful. Deskflow is its maintained successor, it supports Wayland, and it is in the Ubuntu archive:</p><pre>sudo apt install deskflow</pre><p>Install Deskflow on the other machines too.</p><h4>Chrome</h4><p>I still use it almost every day, and several of its extensions are essential to what I do. Download the .deb from Google and install it:</p><pre>sudo apt install ./google-chrome-stable_current_amd64.deb</pre><h4>Caffeine</h4><p>For when a full-screen application is running and you do not want to nudge the mouse to keep the screen on.</p><pre>sudo apt install caffeine</pre><h4>Slack</h4><p>For work and for friend groups.</p><pre>sudo snap install slack --classic</pre><h4>Zoom</h4><p>From the website: <a href="https://zoom.us/download?os=linux">https://zoom.us/download?os=linux</a></p><pre>sudo apt install ./zoom_amd64.deb</pre><h4>A network speed monitor</h4><p>Living in London does not mean you have good internet. I keep a network monitor on the top bar so I can see whether a slow page is my connection or the server.</p><p>GNOME extensions no longer need the browser connector (the old chrome-gnome-shell package is gone). Install Extension Manager:</p><pre>sudo apt install gnome-shell-extension-manager</pre><p>Then search for a network speed extension inside it. GNOME is at version 50 and extensions break at every release, so pick the one marked as supported on the day you install.</p><h4>VS Code</h4><p>My preferred code editor: <a href="https://code.visualstudio.com/docs/setup/linux">https://code.visualstudio.com/docs/setup/linux</a></p><pre>sudo apt install ./&lt;file&gt;.deb</pre><h4>Sound control</h4><p>If you have more than one microphone or more than one set of speakers, you want a proper mixer:</p><pre>sudo apt install pavucontrol</pre><p>Do not install pulseaudio itself, as older guides (mine included) tell you to. Ubuntu has used PipeWire as its sound server since 22.10, with a compatibility layer that speaks the PulseAudio protocol, and pavucontrol talks to that. A second sound server on top only causes trouble.</p><figure><img alt="Three faders on an audio mixing console, photographed close up, in a dark and mustard duotone" src="https://cdn-images-1.medium.com/max/1024/1*IXFgXAkiNXAng6AxgGt4-g.jpeg" /><figcaption>Photo: Dejan Krsmanovic, Wikimedia Commons, CC BY 2.0.</figcaption></figure><h3>Terminal Applications</h3><figure><img alt="A banner titled Terminal Applications with the logos of Oh My Zsh, tmux, AWS, Kubernetes, Python, jq, Tailscale and Claude on light tiles over a dark background" src="https://cdn-images-1.medium.com/max/1024/1*FLcRnAyBRUBvvVC3HnSg0A.jpeg" /><figcaption>Logos belong to their owners and are shown to identify the software. Oh My Zsh logo by Becca (MIT), tmux logo by Jason Long (ISC), both via Wikimedia Commons.</figcaption></figure><p>Now the terminal add-ons and the terminal-only applications like aws and kubectl. The Terminal app in 26.04 is Ptyxis, which replaced GNOME Terminal; everything below works the same in either.</p><figure><img alt="A DEC VT100 terminal showing a directory listing on its screen, in a dark and mustard duotone" src="https://cdn-images-1.medium.com/max/1024/1*F1aJcV9nGywA-bq6wP_kuQ.jpeg" /><figcaption>A DEC VT100, the terminal that every terminal app still imitates. Photo: Jason Scott, Wikimedia Commons, CC BY 2.0.</figcaption></figure><h4>Dependencies</h4><pre>sudo apt install git curl vim unzip</pre><h4>Zsh and Oh My Zsh</h4><p>Before installing anything else, make the terminal a nicer place to be.</p><pre>sudo apt install zsh<br>sh -c &quot;$(curl -fsSL https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh)&quot;</pre><h4>tmux</h4><p>A session that survives a closed window or a dropped SSH connection. I did not have it on the 2023 list and I would not work without it now.</p><pre>sudo apt install tmux</pre><h4>ripgrep</h4><p>Searches a repository faster than grep and skips what .gitignore skips. The command is rg.</p><pre>sudo apt install ripgrep</pre><h4>AWS</h4><p>From the AWS website: <a href="https://docs.aws.amazon.com/cli/latest/userguide/getting-started-install.html">https://docs.aws.amazon.com/cli/latest/userguide/getting-started-install.html</a></p><pre>curl &quot;https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip&quot; -o &quot;awscliv2.zip&quot;<br>unzip awscliv2.zip<br>sudo ./aws/install</pre><h4>kubectl</h4><p>kubectl comes from Kubernetes, not from Ubuntu. I take the binary, because the apt repository has the minor version in its URL and goes stale with every Kubernetes release:</p><pre>curl -LO &quot;https://dl.k8s.io/release/$(curl -Ls https://dl.k8s.io/release/stable.txt)/bin/linux/amd64/kubectl&quot;<br>sudo install -o root -g root -m 0755 kubectl /usr/local/bin/kubectl</pre><h4>kubectx and kubens</h4><p>Switching clusters and namespaces without typing the long form. One package installs both:</p><pre>sudo apt install kubectx</pre><h4>Python tools: pipx and uv</h4><p>Python 3.14 comes with Ubuntu 26.04. pip install into the system Python is refused with error: externally-managed-environment, and that refusal is correct: do not override it. For command-line tools written in Python, use pipx, which gives each tool its own environment:</p><pre>sudo apt install pipx<br>pipx ensurepath<br>pipx install &lt;tool&gt;</pre><p>For projects I use uv, one virtual environment per project. It is not in the Ubuntu archive:</p><pre>curl -LsSf https://astral.sh/uv/install.sh | sh</pre><h4>Vault</h4><p>I use HashiCorp’s Vault a lot, for work and for personal projects.</p><pre>wget -O- https://apt.releases.hashicorp.com/gpg | sudo gpg --dearmor -o /usr/share/keyrings/hashicorp-archive-keyring.gpg<br>echo &quot;deb [signed-by=/usr/share/keyrings/hashicorp-archive-keyring.gpg] https://apt.releases.hashicorp.com $(lsb_release -cs) main&quot; | sudo tee /etc/apt/sources.list.d/hashicorp.list<br>sudo apt update &amp;&amp; sudo apt install vault</pre><p>One thing to know before you standardise on it: since August 2023 Vault is under the Business Source License, not an open source one, and HashiCorp now belongs to IBM. If that matters where you work, OpenBao is the fork under the Linux Foundation, still MPL 2.0.</p><figure><img alt="A woman reading in the open doorway of a bank vault, its heavy door swung open, in a dark and mustard duotone" src="https://cdn-images-1.medium.com/max/1024/1*rol4OAyB2lNQHvWreQvXpg.jpeg" /><figcaption>A bank vault door in 1967. Photo: Aldo Moisio, Wikimedia Commons, public domain.</figcaption></figure><h4>jq</h4><p>A command-line JSON processor: filter, slice, map and transform JSON.</p><pre>sudo apt install jq</pre><h4>yq</h4><p>The same for YAML. Take the snap:</p><pre>sudo snap install yq</pre><p>sudo apt install yq gives you a different program with the same name, a Python wrapper around jq. The snap is Mike Farah&#39;s Go tool, the one Kubernetes tutorials mean, and the syntax is not the same.</p><h4>libkeccak and sha3sum</h4><p>Handy when working with key pairs, and still not packaged. If all you need is SHA-3, openssl dgst -sha3-256 is already on the machine. Ethereum addresses use Keccak-256, which is a different function with a different output, and for that I still build these:</p><pre>git clone https://github.com/maandree/libkeccak<br>cd libkeccak<br>make<br>sudo make install<br>cd ..<br>git clone https://github.com/maandree/sha3sum.git<br>cd sha3sum<br>autoreconf -i<br>./configure<br>make<br>sudo make install</pre><h4>SSH keys</h4><p>Whether you use GitLab, GitHub or Bitbucket, you need a key pair and the public half uploaded to your account.</p><pre>ssh-keygen -t ed25519<br>cat ~/.ssh/id_ed25519.pub</pre><p>OpenSSH in 26.04 uses a hybrid post-quantum key exchange, mlkem768x25519-sha256, by default, so the connection around that key is already protected against the attack I <a href="https://valentevidal.medium.com/no-a-teenager-with-a-quantum-computer-isnt-stealing-your-bitcoin-yet-2be3b51dea5f">wrote about in September</a>. You do not have to do anything.</p><h4>Tailscale</h4><p>A private network between my machines, so I can SSH into the one at home from anywhere without opening a port on the router.</p><pre>curl -fsSL https://tailscale.com/install.sh | sh<br>sudo tailscale up</pre><h4>An AI coding agent</h4><p>This is the biggest change since 2023. A lot of the typing in my terminals is now done by a coding agent, and the Ubuntu machine I use most has no screen at all: I reach it over SSH and the agent works in tmux. Mine is Claude Code:</p><pre>curl -fsSL https://claude.ai/install.sh | bash</pre><figure><img alt="A technician crouched at an open server rack with a laptop on the floor, in a dark and mustard duotone" src="https://cdn-images-1.medium.com/max/1024/1*igRMTkh_iFf28UZMowEhYQ.jpeg" /><figcaption>Photo: Derrick Coetzee, Wikimedia Commons, CC0.</figcaption></figure><h3>Two things you did not install</h3><p>Two tools you use in every line above were replaced underneath you in this release. sudo is now sudo-rs, a rewrite in Rust, and most of the core utilities (ls, sort and the rest) come from the Rust coreutils, with cp, mv and rm still the GNU ones. If a shell script of yours depends on an obscure flag, test it on 26.04 before you trust it.</p><h3>Single Script</h3><pre>#!/bin/bash<br>set -e</pre><pre>sudo apt update<br>sudo apt upgrade -y</pre><pre># Keyboard and mouse sharing<br>sudo apt install deskflow -y</pre><pre># Chrome, Zoom, VS Code: download the .deb from each website, then<br>#   sudo apt install ./&lt;file&gt;.deb</pre><pre># Caffeine<br>sudo apt install caffeine -y</pre><pre># Slack<br>sudo snap install slack --classic</pre><pre># GNOME extensions; pick a network speed extension from inside the app<br>sudo apt install gnome-shell-extension-manager -y</pre><pre># Sound mixer (PipeWire is already the sound server)<br>sudo apt install pavucontrol -y</pre><pre># Dependencies, shell and terminal tools<br>sudo apt install git curl vim unzip zsh tmux ripgrep -y</pre><pre># AWS CLI<br>curl &quot;https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip&quot; -o &quot;awscliv2.zip&quot;<br>unzip awscliv2.zip<br>sudo ./aws/install</pre><pre># kubectl, latest stable binary<br>curl -LO &quot;https://dl.k8s.io/release/$(curl -Ls https://dl.k8s.io/release/stable.txt)/bin/linux/amd64/kubectl&quot;<br>sudo install -o root -g root -m 0755 kubectl /usr/local/bin/kubectl<br>rm kubectl</pre><pre># kubectx and kubens<br>sudo apt install kubectx -y</pre><pre># Python tools<br>sudo apt install pipx -y<br>pipx ensurepath<br>curl -LsSf <a href="https://astral.sh/uv/install.sh">https://astral.sh/uv/install.sh</a> | sh</pre><pre># Vault<br>wget -O- <a href="https://apt.releases.hashicorp.com/gpg">https://apt.releases.hashicorp.com/gpg</a> | sudo gpg --dearmor -o /usr/share/keyrings/hashicorp-archive-keyring.gpg<br>echo &quot;deb [signed-by=/usr/share/keyrings/hashicorp-archive-keyring.gpg] <a href="https://apt.releases.hashicorp.com/">https://apt.releases.hashicorp.com</a> $(lsb_release -cs) main&quot; | sudo tee /etc/apt/sources.list.d/hashicorp.list<br>sudo apt update &amp;&amp; sudo apt install vault -y</pre><pre># jq and yq<br>sudo apt install jq -y<br>sudo snap install yq</pre><pre># Tailscale<br>curl -fsSL <a href="https://tailscale.com/install.sh">https://tailscale.com/install.sh</a> | sh</pre><pre># Claude Code<br>curl -fsSL <a href="https://claude.ai/install.sh">https://claude.ai/install.sh</a> | bash</pre><pre># Oh My Zsh (last: it starts a new shell)<br>sh -c &quot;$(curl -fsSL https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh)&quot;</pre><pre># SSH key<br>ssh-keygen -t ed25519<br>cat ~/.ssh/id_ed25519.pub</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b7cdf6c25f33" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[76 Internal Tools in Four Months. Nobody Can Hire for Them.]]></title>
            <link>https://valentevidal.medium.com/76-internal-tools-in-four-months-nobody-can-hire-for-them-90ebf5699742?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/90ebf5699742</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[junior-developer]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[hiring]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 22:49:32 GMT</pubDate>
            <atom:updated>2026-10-09T22:49:32.249Z</atom:updated>
            <content:encoded><![CDATA[<p><em>AI made bespoke tooling nearly free, so every company’s stack is turning back into something you cannot learn from the outside. That is an argument for hiring graduates again.</em></p><figure><img alt="A wall of hand tools hung on a pegboard in a bike workshop, in a dark mustard duotone, with the words 76 TOOLS across it." src="https://cdn-images-1.medium.com/max/1024/1*JiAL6OeZwzWsF-K9pjYpYw.jpeg" /><figcaption>Photo: 4300streetcar, Wikimedia Commons, CC BY 4.0.</figcaption></figure><p>I started noticing it a few months ago: almost every company I can see into has started building its own tools. Companies have always done this, but because of AI it’s very different; companies can internal tools to another level.</p><p>Just from my personal projects: an animation project, five static websites, a DNS repository, a step counter that nags me, a couple of marketing sites, a personal health tracker. Every one of them is built and operated with AI coding agents but that is already 76, internal, small to medium, tools, across 12 repositories, in four months.</p><p>49 belong to the animation project alone, and the first of those was written on the 26th of July.</p><p>And of course I understand that most of them will always be taken care of with AI, but medium to large companies are creating big tools, platforms for internal use, even in my current 9 to 5 we are working on big project that I can’t talk about. That is rather the point.</p><figure><img alt="Close-up of the keyboard and panels of an IBM System/360 mainframe console, in a dark mustard duotone." src="https://cdn-images-1.medium.com/max/1024/1*PG8BMAmuBcaeKVZmc2_E7A.jpeg" /><figcaption>An IBM System/360 console. Photo: Marcin Wichary, Wikimedia Commons, CC BY 2.0.</figcaption></figure><h3>The commodity years</h3><p>In the mainframe era, everything was in-house: the operating procedures, the job control, the report writers, the whole way of working with the machine. A company hired people with a background in computing and trained them on its system, because there was no other way to get someone who knew its system.</p><p>Then the 2010s happened. Open source ate the frameworks, SaaS ate the back office, the cloud ate the data centre, and Kubernetes ate what was left. For about fifteen years the sensible thing was to use what everyone else used, because that is where the documentation, the hiring pool and the answers on Stack Overflow were. Your knowledge of React, Postgres, Terraform and Datadog was worth the same at the next company as at this one.</p><figure><img alt="Stacks of shipping containers at a snowy container terminal in Reykjavik with cranes and mountains behind, in a dark mustard duotone." src="https://cdn-images-1.medium.com/max/1024/1*tXaNG2zbd2hvDKXbHBot_w.jpeg" /><figcaption>Photo: Quintin Soloviev, Wikimedia Commons, CC BY 4.0.</figcaption></figure><h3>Nobody arrives knowing your tools</h3><p>Here is what the agents did to the trade. The commodity layer is still there; I still run Postgres and Terraform and Kubernetes. But the layer where the work actually happens has moved up, into the seventy-six tools, and that layer is unique to the company that built it. <strong>Nobody arrives knowing it. Nobody can.</strong></p><figure><img alt="Rows of labelled archive boxes and binders on shelves along a narrow corridor, in a dark mustard duotone." src="https://cdn-images-1.medium.com/max/1024/1*jDPJe38f6FKDd2fWfbOmuw.jpeg" /><figcaption>Photo: Samuel Zeller, Unsplash via Wikimedia Commons, CC0.</figcaption></figure><h3>So why hire juniors</h3><p>The company that stops hiring graduates is making a bet, whether it knows it or not: that the senior it hires instead will keep arriving productive. That was true on a commodity stack. On a bespoke one, the senior spends their first three months reading the brief, the same as the graduate would (ok the graduate would probably take longer because he is new to the field but you know what I mean), and costs twice as much while doing it.</p><p>Now put that against what companies are actually doing.</p><p><strong>Stanford’s Digital Economy Lab</strong>, working from the payroll records of the largest payroll provider in the United States, <a href="https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine">reported in August 2026</a> that employment of workers aged 22 to 25 in the occupations most exposed to AI stands 19 percent below where it would be had it kept pace with their less exposed peers, and that experienced workers show no comparable gap.</p><p><strong>SignalFire</strong>, a venture fund that tracks hiring at the twelve biggest technology companies, <a href="https://signalfire.com/blog/signalfire-state-of-talent-report-2026">put new-graduate and entry-level hiring at those companies</a> about 65 percent below its 2019 level in June 2026, while engineering hiring overall was down only 11 percent. In their <a href="https://www.signalfire.com/blog/signalfire-state-of-talent-report-2025">2025 report</a> new graduates were 7 percent of Big Tech hires. The entry-level door is closing, and it is closing because companies believe the agent does the junior’s job.</p><figure><img alt="A hall full of young railway apprentices seated at long tables for an exam in 1930, in a dark mustard duotone." src="https://cdn-images-1.medium.com/max/1024/1*X7ZwCthw6xGxdNdKgw5jmw.jpeg" /><figcaption>Railway apprentices sitting their first exam, Hutt Workshops, New Zealand, 1930. Photo: Albert Percy Godber, Wikimedia Commons, public domain.</figcaption></figure><p>I think in the next coming years it will reverse. The agent does do the junior’s old job, the ticket-sized change on a known stack. What the agent cannot do is be the person who knows your seventy-six tools, because nobody is that person until they have been trained on them.</p><p>The agent made the tools cheap; it did not make the knowledge of them portable. Somebody has to be grown into every company’s stack, and the economics of growing a graduate into it are better than the economics of growing a senior into it</p><p><em>When nobody arrives knowing the system, you hire for the ability to learn it and you pay for the ability to learn it.</em></p><p>What goes away is the thing we spent fifteen years treating as the whole point of seniority: knowing the commodity stack better than the next person. What stays is judgement, which was always the real thing and which nobody ever learned from a job advert.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90ebf5699742" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cancún Sells Americans Everything Except Dialysis]]></title>
            <link>https://valentevidal.medium.com/canc%C3%BAn-sells-americans-everything-except-dialysis-f141d95650bc?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/f141d95650bc</guid>
            <category><![CDATA[dialysis]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[medical-tourism]]></category>
            <category><![CDATA[mexico]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 22:38:54 GMT</pubDate>
            <atom:updated>2026-09-18T22:38:54.293Z</atom:updated>
            <content:encoded><![CDATA[<p><em>In a city built around American visitors, where half a pharmacy shelf is sold without a prescription, a friend on dialysis could not find out what a session would cost him. I assumed it was a search problem. It turned out to be something stranger.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e1_65aIcrrYXRQ3rgXcGjQ.jpeg" /></figure><p>You can walk into almost any pharmacy in Cancún and buy things that would need a doctor’s signature in the UK or US. Blood pressure medication. Thyroid tablets. Diabetes drugs. Contraceptives. Most of a repeat prescription, straight off the shelf.</p><p>Strictly speaking, most of this is not legal. Article 226 of Mexico’s Ley General de Salud sorts medicines into six categories, and only the last two can be sold freely. Blood pressure tablets, levothyroxine, metformin and contraceptives all sit in the fourth, which needs a prescription. So do antibiotics, which is why the 2010 crackdown on selling those over the counter was about enforcing the law rather than changing it.</p><p>What the law says and what happens at the counter are two different things, and the gap is filled in plain sight. Some twenty thousand consultation rooms sit next door to Mexican pharmacies, where about four dollars buys you a doctor and a signature.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q1arStLC35Q9gOklaWaXPw.jpeg" /></figure><p>The authorities know. In 2023 COFEPRIS, the federal health regulator and Mexico’s rough equivalent of the FDA, suspended twenty-three pharmacies across Cancún, Playa del Carmen and Tulum for selling prescription drugs over the counter, and noted that some were doing it, in the regulator’s own words, “únicamente a personas extranjeras”. Only to foreigners.</p><p>And Cancún leans into all of this harder than anywhere. The city exists for visitors, its airport logging close to thirty million passengers in 2025. English on the signage, dollars accepted, hospitals with international patient departments and published price lists aimed squarely at Americans who find US healthcare unaffordable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gfZoo8IwdfoQVUQ2golhzQ.jpeg" /></figure><p>A friend of mine is on haemodialysis. He needs treatment three times a week, and that is not a preference: miss a session and don’t just feel tired; you land in the hospital. So a holiday means finding a clinic at the other end, booking a chair, sending documents ahead.</p><p>When he started looking into a trip to Cancún, I assumed the hard part would be the flights. He did what anyone would do and searched online.</p><p>He hit the same wall on almost every site. No price listed. Nothing about which machines they run. Emails that went unanswered. And no international number to call, only Mexican landlines and mobiles, which is not much use from another continent.</p><p>Veracruz I would understand. Mexico City I would understand, a vast domestic market with no particular reason to court foreigners. But Cancún, CANCUN, of all places. A city that has spent fifty years learning how to sell things to Americans.</p><p>Watching him struggle got me curious, not just about his trip but about the whole dialysis business.</p><p>He had found several places by then but needed help calling them. Two stood out from their websites. Nefrotlach (<a href="https://nefrotlach.com.mx/">nefrotlach.com.mx</a>) has a thin, dated site that nonetheless names its machine, a Baxter AK 98, and pitches holiday dialysis directly, offering to coordinate with your clinic at home.</p><p>Hospiten Cancún (<a href="https://hospiten.com/dialysis-club">hospiten.com</a>) has a much slicker site with less on it, at least on the Cancún page. But tucked away on a separate “Dialysis Club” page for travelling patients, it says in a single sentence that the Cancún unit has two places, digitalised Fresenius monitors, reverse osmosis water and single-use dialysers.</p><p>Those details matter more than they look. Spend years attached to these machines three times a week and you develop opinions. You learn which models are comfortable and which alarm at nothing. You learn that a Fresenius 5008 will stop and make the operator confirm that no disinfectant is left in the lines before your blood goes in. You learn that some clinics reuse dialysers, cleaning and testing your filter between your own sessions, and you are banned outright if you have hepatitis B or HIV.</p><p>My friend sent over a list of nine clinics and hospitals, and my colleague Ivan, who lives in Mexico, rang them up and asked for prices and which machines they use. Most said they did not have that information. Some did not answer at all.</p><p>All of this hassle kept nagging at me, why is it this complicated in Cancún</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6O5NNY41iwV6HmdSU2_uGA.jpeg" /></figure><h4>Nine numbers</h4><p>Ivan took the list and worked through it by phone in a single afternoon. Every clinic and hospital in Cancún that a search suggests does haemodialysis, nine in total, and the same four questions each time: do you take holiday dialysis patients, how far ahead do I need to book, which machines do you use, and what does a session cost.</p><p>Here is what came back.</p><pre>Place                                                                                            Result<br>---------------------------  --------------------------------------------------------------------------<br>Nefrotlach                                     Yes. One week&#39;s notice. Baxter AK 98. USD 250 a session.<br>Certeza Nefrología           Yes. Email to check available days. USD 180. Would not say which machines.<br>Hospiten Cancún                                          Yes. Two months&#39; notice. No price, no machine.<br>Renal Caribbean                                                              No answer, three attempts.<br>EME Red Hospitalaria                                                                  Number incorrect.<br>Hospital Azura                                                                   Number not in service.<br>Hospital Amerimed                                                 Held, then disconnected. Three times.<br>Hospital Galenia                                                          Does not offer haemodialysis.<br>Hospital Quirúrgica del Sur                                               Does not offer haemodialysis.</pre><p>These nine were not a research list. They were what my friend found on Google, which is the same list any patient anywhere would arrive at.</p><p>Six calls connected. Three produced usable information. Exactly one clinic, Nefrotlach, answered all four questions, and the woman who took the call volunteered the session length, the two daily start times and the exact list of labs to bring.</p><p>Two of the hospitals that turn up in directory listings for dialysis in Cancún do not do dialysis. Two phone numbers published online do not connect to anything.</p><h4>So why is it such a hassle, what would it cost to build one</h4><p>If the market is this opaque, what does it actually take to open a clinic in Cancún, and why is nobody competing on transparency?</p><p>The machines, it turns out, are the cheap part.</p><p>Used Fresenius 5008 units sell on the European dealer market for USD 900 to 1,700 as they stand, or USD 1,300 to 4,200 refurbished with a warranty. One was listed at USD 800, although that one was for parts. A 4008S, the older workhorse that Fresenius still sells new and still services in Mexico, goes for around USD 2,800. A new machine is USD 10,000 to 30,000.</p><p>These are decommissioned clinic machines carrying ten to twenty thousand treatment hours, which sounds alarming until you work out what it means. A clinic running three shifts a day, six days a week, puts about 3,700 hours on a machine every year, so twenty thousand hours is roughly five years of hard use. That is middle age, not the end: these things run for a decade or more on maintenance. They are cheap because the original owner has finished depreciating them and moved on, not because they are worn out. Fresenius has not retired the 5008; it still sells it new in Mexico and extended the line again in 2024. European clinics simply keep turning over their fleets, and the old stock has to go somewhere.</p><p>There is a second reason the prices are low. The manufacturers make very little on the machine itself. The money is in what it consumes: the dialysers, the bloodlines, the concentrate and the treatment cards you buy from them three times a week, per patient, forever. Selling you a cheap machine is the point.</p><p>A rough estimate for a six-station clinic (6 machines) looks something like this.</p><pre>USD<br>------------------------------------------------------------  --------<br>Eight used machines, including a spare and an isolation unit    22,000<br>Water treatment plant                                          128,000<br>Chairs, build-out, licensing                                   111,000<br>Working capital                                                122,000<br>Total                                                         ~383,000</pre><p>Another thing I did not know about dialysis until this week: the biggest single line in building a clinic is not the medical equipment. It is the water.</p><p>During a session the patient’s blood runs along one side of a thin membrane and a fluid called dialysate runs along the other, pulling the waste across. At a standard flow rate that is <strong>120 litres of dialysate for one patient, in one four-hour session</strong>. Tap water, as you can imagine, is nowhere near clean enough for the job. Small dissolved contaminants, metals, fluoride and above all chlorine, cross the membrane freely and reach the bloodstream directly. The bacteria themselves never cross, but the debris they shed does: fragments of endotoxin small enough to leak back through an intact membrane, session after session. So every drop has to be stripped to a purity standard well beyond drinking water first.</p><p>Which is why the water plant on that list costs six times the entire machine fleet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KaIDmygPZJoh7qRKiVhxoQ.jpeg" /></figure><p>Cancún may make that harder. Northern Yucatán has essentially no surface water, so everything comes up out of a limestone aquifer, and CONAGUA, the national water authority, puts dissolved solids in the aquifer feeding the city at around 750 mg/L against a limit of 1,000, with salinisation flagged as a growing risk. That is hard water, not disqualifying water, and softeners are standard equipment in dialysis units everywhere. I could not find a published analysis of what actually comes out of a Cancún tap, which is the number that would settle it, and the first thing I would want measured before signing a lease.</p><p>A dialysis clinic is a water purification plant that happens to have patients in it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WSp1keN61InR-uY0HlkdMA.jpeg" /></figure><h4>The number that decides it</h4><p>Running costs come down to two lines: consumables at roughly USD 60 a session, and staff.</p><p>The Mexican standard for dialysis practice, NOM-003-SSA3–2010, reserves prescribing and applying haemodialysis to doctors with a registered specialist qualification in nephrology, and asks that nurses either hold a nephrology specialty or document six months of haemodialysis training at a certified centre. Separately, a 2018 reform to the Ley General de Salud added haemodialysis units to the establishments that need a health licence rather than a simple notification. The nephrologist who signs as director is a precondition of opening the doors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eLZlz2c3OWUTjkZfFl5UUQ.jpeg" /></figure><p>So I built a spreadsheet. Six stations, meaning six chairs with a machine beside each one. Three shifts a day, because a session runs about four hours and you can fit three into a working day with cleaning time between them. Six days a week. That is eighteen sessions a day, and since each patient comes three times a week, the building holds thirty-six people.</p><p>Breaking even lands at roughly twenty-four of those thirty-six.</p><p>It is not twenty-four appointments to fill. It is twenty-four people on the books at any given moment, each coming back three times a week, indefinitely. A restaurant needs a fresh crowd every night. A dialysis clinic needs the same two dozen people, every week, for years.</p><p>Which leads somewhere I had not thought about when I opened the spreadsheet. Nobody is discharged from dialysis. A few patients leave for a transplant. Most of the turnover is mortality, and in Mexico that is steep: one network of public clinics reported two-year survival below fifty percent. A clinic’s roster replaces itself by something over twenty percent a year. Pretty grim stuff . . .</p><h4>Can you get IMSS (like the NHS) contracts if you have a clinic?</h4><p>Roughly four in ten Mexicans, the largest single block, are covered by IMSS, the public social security system. For years IMSS did not run enough dialysis itself, so it paid private clinics to treat its patients under an arrangement called subrogación. As of 2019, private units were carrying around seventy percent of IMSS’s own haemodialysis patients.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eohk3166iUI_LhwOn7vdOw.jpeg" /></figure><p>That made the public contract the foundation of the private dialysis business in Mexico, and it was worth a great deal. IMSS carves the country into regional lots and tenders them, so winning one means being handed a block of IMSS dialysis patients for the length of the contract. The federal auditor’s review of the 2023 round covers 27 contracts with ceilings running up to USD 18 million, and a typical ceiling nearer USD 7 million, for contracts lasting five and a half months.</p><p>Nobody wins them alone. They go to companies bidding jointly: Médica Santa Carmen with Carint took Puebla, while Centro de Asistencia Renal, Servicios Médicos y de Equipamiento and Cenardix took eastern Estado de México between them. Fresenius Medical Care de México turns up repeatedly as a partner. This is an industry, not a collection of clinics.</p><p>Then, from 15 June 2026, IMSS began ending subrogación, nationally, not just in Quintana Roo, the state Cancún sits in.</p><p>IMSS is building its own capacity instead, opening forty-chair units at around USD 1.3 million each. One of the companies bidding to supply that replacement capacity is CREAMEDIC, a wholly owned subsidiary of Grupo CIE. Grupo CIE is a live entertainment company, and the bid was attacked publicly for exactly that reason: no prior clinical dialysis experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H1zLqu8WF5VXdKHoL8Og4Q.jpeg" /></figure><h4>I am not sure what or for who this article is for.</h4><p>I went down a rabbit hole, and by the time I noticed I had far more information than I knew what to do with, so I put it all in one place.</p><p>After all that digging, it does look like there is a genuine opportunity in opening a dialysis clinic in Cancún catering to tourists.</p><p>Somebody already thinks so. ISSSTE, the health service for Mexico’s public sector workers, is converting its Cancún hospital in a USD 40 million project that includes sixteen haemodialysis stations, more than the private clinics I could find have between them.</p><p><em>Figures in this piece come from an afternoon of calls on 17 September 2026, European used-equipment dealer listings, Mexican federal audit records, and published costing studies. Pesos are converted at 17 MXN to the dollar, the rate in September 2026. The financial model behind the capital and break-even numbers is a screening tool, not investment advice.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f141d95650bc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[No, a Teenager With a Quantum Computer Isn’t Stealing Your Bitcoin (Yet)]]></title>
            <link>https://valentevidal.medium.com/no-a-teenager-with-a-quantum-computer-isnt-stealing-your-bitcoin-yet-2be3b51dea5f?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/2be3b51dea5f</guid>
            <category><![CDATA[cryptography]]></category>
            <category><![CDATA[bitcoin]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Fri, 11 Sep 2026 14:45:22 GMT</pubDate>
            <atom:updated>2026-09-11T14:45:22.884Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Q-day is real and it has a date. Here is what a quantum computer actually breaks, what replaces it, and what year to expect it, from someone who spent five years looking after the keys.</em></p><p><em>Disclosure: this started as a conversation with an AI assistant while I was reading about the post-quantum standards. I used it for the outline and to check the numbers; the text is mine. Sources are at the end.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MCOt5BW__QQNvu9U0-zfsg.jpeg" /></figure><p><em>Photo: Paulo Barcellos Jr., Wikimedia Commons, CC BY-SA 2.0.</em></p><p>I have been working with Ethereum since I started at Fnality back in 2020. That was my deep dive into how blockchains actually work, from the consensus layer down to the key material on the validators. If you are reading this you probably have some background on blockchains and wallets, and you know they are protected with elliptic curves. Every public chain uses the same two: secp256k1 for accounts and transactions, and BLS12–381 for validator signatures. Both are, for all practical purposes, impossible for a conventional computer to break.</p><p>The key word there is conventional.</p><p>Everyone in the industry has understood for a while that the moment quantum computers get good enough, there will be a day when they can break every form of encryption we use today. It even has a name, Q-day, and the popular version goes that the whole world burns down because terrorist groups and teenagers can crack the encryption on the financial system, the web and the nuclear codes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V2OV5_qLmig7r5L9mRjmVA.jpeg" /></figure><p><em>A bank vault in 1967. Photo: Aldo Moisio, Wikimedia Commons, public domain.</em></p><p>Don’t worry. The academics solved that a while back (although not as far back as I was expecting). NIST is the US National Institute of Standards and Technology, the agency that picks the cryptographic standards the rest of the world ends up using; AES and SHA-256, which Bitcoin itself is built on, both came out of its competitions. It ran a similar competition for post-quantum algorithms from 2016 and finalised the winners on 13 August 2024. These are the current replacements as they apply to blockchains:</p><ul><li><a href="https://csrc.nist.gov/pubs/fips/204/final">FIPS 204, ML-DSA</a> (previously called Dilithium). This is a new way of signing a transaction: proving that the person moving the coins is the person who owns them, without revealing the secret that proves it. It does the same job that today’s signatures do on every Bitcoin and Ethereum transaction, but it is built on the lattice maths described further down instead of on curves. This is the one the chains are looking at first.</li><li><a href="https://csrc.nist.gov/pubs/fips/205/final">FIPS 205, SLH-DSA</a> (previously called SPHINCS+). Another way of signing, but built purely out of hash functions, the same fingerprinting tool Bitcoin already uses for mining and addresses. It is slower and its signatures are much bigger, but it makes the fewest assumptions of anything on the list, so it is the insurance policy in case someone one day finds a flaw in lattices. Ethereum’s plan for its validators is built on this family.</li><li><a href="https://csrc.nist.gov/pubs/fips/203/final">FIPS 203, ML-KEM</a> (previously called Kyber). This one is not about signing at all. It is how two computers that have never met agree on a secret code to encrypt their conversation, which is what happens every time your browser opens a secure site. It protects the connection between your wallet and the network, the way https protects your bank login. Nothing on a blockchain is encrypted, so it plays no part in the blocks themselves; the two signature standards above are what the chains actually need, and this one is here because it is the third of the three.</li><li>Two more are on the way. FN-DSA (previously called Falcon) is another lattice-based way of signing that produces smaller signatures than ML-DSA, at the cost of being harder to implement safely; it has been sitting in draft as FIPS 206 for about a year (<a href="https://csrc.nist.gov/projects/post-quantum-cryptography/news">progress is tracked here</a>). And in March 2025 NIST <a href="https://www.nist.gov/news-events/news/2025/03/nist-selects-hqc-fifth-algorithm-post-quantum-encryption">picked HQC</a> as a spare for ML-KEM, built on different maths entirely, so that if lattices ever fall there is something to switch to.</li></ul><p>The deadlines are already set. NIST’s <a href="https://csrc.nist.gov/pubs/ir/8547/ipd">transition plan</a> says the curve and RSA methods we use today are to be phased out after 2030 and no longer allowed after 2035. The UK’s National Cyber Security Centre, the government’s security adviser, <a href="https://www.ncsc.gov.uk/news/pqc-migration-roadmap-unveiled">set the same 2035 finish line</a> in March 2025: know what you need to migrate by 2028, have the important systems moved by 2031, and be finished by 2035.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N8Bt4r-f6UZ28dImkjqjZg.jpeg" /></figure><p><em>Photo: Dietmar Rabich, Wikimedia Commons, CC BY-SA 4.0.</em></p><h3>So what is a lattice</h3><p>Three of the four schemes above are built on lattices, so it is worth knowing what one is. The textbook definition: a lattice is the set of every point you can reach by adding whole-number multiples of a fixed set of vectors, called a basis, in a space with many dimensions. The cryptography rests on two problems on that set, finding the shortest non-zero vector (the Shortest Vector Problem) and finding the lattice point nearest to an arbitrary point (the Closest Vector Problem). Both are NP-hard in their exact form, and nobody has found an efficient way to solve them, on a classical or a quantum computer, at the sizes the standards use.</p><h3>So what is a lattice (explain it like I’m 5)</h3><p>Today’s keys work like a game of billiards on a curved table. You put the ball down, hit it, and it bounces off the curve and lands somewhere new. You hit it again from there, and again, a secret number of times, then show everyone where it ended up. Anyone can see where it started and where it finished. Nobody can tell how many shots it took, and that number of shots is your secret. The problem is that the bouncing follows a pattern, and a quantum computer is a machine that is brilliant at spotting patterns. Shor’s algorithm is the name of the trick it uses to spot this one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jrSIDAEhpDQKdGIXiod9lA.jpeg" /></figure><p><em>Photo: PattayaPatrol, Wikimedia Commons, CC BY-SA 4.0.</em></p><p>A lattice is not a billiards table. It is a pattern of dots that repeats forever, like the dots on graph paper, and every dot can be reached by taking whole steps along two fixed arrows. The trick is that the same dots can be described by two different pairs of arrows: a nice pair, short and roughly at right angles, and a nasty pair, long and pointing almost the same way. The nasty pair is your public key and the nice pair is your private key. To send you a message, someone picks a dot using your nasty arrows and nudges it slightly off the grid. With the nice arrows you snap it straight back to the right dot and read the message. With only the nasty arrows, the one option left is trying combinations, and in the five hundred directions the real schemes use there are more combinations than atoms in the universe. Nothing goes round in circles here, so there is no pattern for Shor’s trick to spot. It is a search, and a quantum computer is only slightly better at searching than an ordinary one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*agvgrb9Vh74AjQJugRFfpg.png" /></figure><p><em>The same dots described by a nice pair of arrows (red) and a nasty pair (black). Image: Catslash, </em><a href="https://commons.wikimedia.org/wiki/File:Lattice-reduction.svg"><em>Wikimedia Commons</em></a><em>, public domain.</em></p><h3>Where the hardware is</h3><p>You may remember a Shanghai University team “breaking RSA” with a D-Wave machine. Here is what actually happened. The paper, published in the Chinese Journal of Computers in May 2024 and picked up by the press that October, factored a 22-bit number. A 50-bit number followed with a classical computer doing part of the work, and a January 2025 follow-up reached 80 bits, again with classical help. For scale, the largest number ever factored by any method is 829 bits, that took a cluster of ordinary computers months, and RSA-2048 is 2,048 bits, with the difficulty growing exponentially in the length. D-Wave machines are also annealers, a different technology from the gate-based computers that Shor’s algorithm needs, and D-Wave itself said the work was not a fundamental breakthrough. Nothing about RSA changed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dEITCbttiNWInjyt-Wv_ag.jpeg" /></figure><p><em>A D-Wave Two processor, the annealing kind. Photo: Steve Jurvetson, Wikimedia Commons, CC BY 2.0.</em></p><h3>The obvious question: what year?</h3><p>If the difficulty grows exponentially and they went from 22 bits to 80 in a year, when do they reach 2,048? The honest answer is that they never will on that road. Going from 22 to 80 bits is adding about 60 bits a year. At that pace 2,048 bits arrives around 2058, and 4,096 bits around 2092, and that is being generous, because the annealing trick only works with a lot of classical help and has no known way to scale. It is the wrong technology for the job. Extrapolating it is like watching someone climb a ladder, measuring that they gain a metre every few seconds, and working out what year they will reach the Moon. The climbing is real. The ladder does not go there.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aWKtwUTkfIv66Or5qLacVg.jpeg" /></figure><p><em>Photo: NASA, Apollo 11, public domain.</em></p><p>The clock that matters is a different one. Raw qubits are noisy and lose their state within a fraction of a second, so to run something as long as Shor’s algorithm you bundle hundreds of them together into one reliable unit that corrects its own errors. That unit is called a logical qubit, and the count of logical qubits is the number to watch, because breaking a key needs about 1,400 of them working at once.</p><p>Three numbers tell you where that count stands. The first is how many logical qubits anyone has actually built. Microsoft and Quantinuum reached 12 in <a href="https://azure.microsoft.com/en-us/blog/quantum/2024/09/10/microsoft-and-quantinuum-create-12-logical-qubits-and-demonstrate-a-hybrid-end-to-end-chemistry-simulation/">September 2024</a>, Quantinuum’s Helios machine reached 48 in <a href="https://thequantuminsider.com/2025/11/06/illuminating-helios-quantinuums-shiny-new-quantum-computer-gets-sunny-reception/">November 2025</a>, and a Harvard and QuEra team reached 96 in <a href="https://quantumcomputingreport.com/harvard-and-collaborators-demonstrate-scalable-fault-tolerant-architecture-with-448-neutral-atom-qubits/">January 2026</a>. The second is what the manufacturers are promising: <a href="https://www.ibm.com/quantum/blog/large-scale-ftqc">IBM’s published roadmap</a> has a machine with 200 logical qubits in 2029 and one with 2,000 in 2033. The third is what the people who study this for a living expect. In December 2025 the Global Risk Institute <a href="https://globalriskinstitute.org/publication/quantum-threat-timeline-report-2025b/">asked 26 experts</a> how likely it is that a machine able to break RSA-2048 exists within ten years, and their answers landed between 28 and 49 percent. That is why both the <a href="https://csrc.nist.gov/pubs/ir/8547/ipd">US transition plan</a> and the <a href="https://www.ncsc.gov.uk/news/pqc-migration-roadmap-unveiled">UK’s NCSC</a> tell everyone to be off today’s encryption by 2035.</p><p>Against those numbers, here is what a key costs to break, from the best current estimates:</p><ul><li><strong>A Bitcoin or Ethereum key (secp256k1, 256-bit):</strong> 1,200 to 1,450 logical qubits, under 500,000 physical qubits, 9 to 23 minutes per key (Google Quantum AI with the Ethereum Foundation, March 2026).</li><li><strong>RSA-2048:</strong> about 1,400 logical qubits, under 1 million physical, under a week per key (Gidney, May 2025).</li><li><strong>RSA-4096:</strong> roughly 2,800 logical qubits, roughly 2 million physical, weeks to a couple of months per key (scaled from Gidney: qubits grow with key length, operations with its cube).</li></ul><p>Put those figures next to the three numbers above. Demonstrated logical qubits went 12, 48, 96 in fifteen months. Nobody should draw a straight line through three points, but if that doubling held, 1,400 logical qubits arrive around 2030 and 2,800 a year later. IBM’s own roadmap is more conservative and says 2,000 in 2033. The experts polled in December 2025 put the odds of a working RSA-2048 breaker by 2035 at somewhere between one in four and one in two, and every government migration deadline is 2035 for that reason.</p><p>So, as of September 2026: RSA-2048 falls somewhere between 2030 and 2035 if the roadmaps hold, and RSA-4096 a year or two later. That last part surprises people. Against an ordinary computer, doubling the key makes factoring astronomically harder. Against a quantum computer it only doubles the qubits and multiplies the time by about eight, so a week becomes a couple of months. Bigger RSA keys buy a couple of years, not decades, which is why NIST’s plan puts 2048, 3072 and 4096 on the same 2035 exit.</p><p>Nobody has broken a real key. Nobody is close in the sense of next year. But “decades away” was the 2019 answer, when the estimate for RSA-2048 was 20 million physical qubits, and the 2026 answer, from the people building the machines, is under a million qubits and a roadmap with 2,000 logical ones on it dated 2033.</p><h3>So, don’t worry (yet)</h3><p>So no, a teenager is not going to steal your search history with a quantum computer any time soon, let alone walk through a central bank’s encryption. The machine that could do it does not exist, the people building it are saying 2033 at the earliest, and the maths that replaces today’s locks was finished in 2024 and is sitting there waiting to be installed. Q-day is real. It is also scheduled, which is more than most disasters can say.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mh9p7X4D_h5OlxlEsp9O-g.jpeg" /></figure><p><em>Photo: Shixart1985, Wikimedia Commons, CC BY 2.0.</em></p><p>What does bother me is the boring version. Every year the estimates drop, every government plan says 2035, and anyone recording encrypted traffic today gets to decrypt it the day the machine arrives. If you hold secrets that need to stay secret for ten years, your deadline is now, not 2035. That is what “harvest now, decrypt later” means in practice, and it gets its own piece next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e8LMcW6jN7BRi1FfG8H53A.jpeg" /></figure><p><em>Photo: NOIRLab/NSF/AURA/T. Slovinský, Wikimedia Commons, CC BY 4.0.</em></p><p>And then there are blockchains, where the keys are public by design and the history cannot be edited. A TLS certificate can be rotated quietly on a Tuesday. A Bitcoin output signed with a key from 2010 cannot. That is the piece after: who is exposed, how much is at stake, and what Bitcoin and Ethereum are actually doing about it.</p><p><em>Sources: Shor (1994); NIST FIPS 203, 204, 205 (13 Aug 2024); NIST IR 8547 (draft, Nov 2024); NCSC PQC migration roadmap (20 Mar 2025); Wang et al., Chinese Journal of Computers (May 2024); Gidney and Ekerå (arXiv 1905.09749); Gidney (arXiv 2505.15917); Google Quantum AI, “cryptocurrency whitepaper” (31 Mar 2026); IBM fault-tolerant roadmap (June 2025); QuEra and Harvard, Nature (Jan 2026); Global Risk Institute, Quantum Threat Timeline 2025; Microsoft and Quantinuum (10 Sept 2024); Quantinuum Helios (Nov 2025); NSA CNSA 2.0 (Sept 2022).</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2be3b51dea5f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[About $95 a month to watch deer starve]]></title>
            <link>https://valentevidal.medium.com/about-95-a-month-to-watch-deer-starve-c5d72c84aed7?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5d72c84aed7</guid>
            <category><![CDATA[deer]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[ecosystem]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[isometric]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 21:01:26 GMT</pubDate>
            <atom:updated>2026-07-27T21:01:26.168Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Rl-eBLzdL6vJbVgO.jpg" /></figure><p>I built a thing called Wildwood: an isometric forest you watch rather than play. No controls beyond a camera. Deer grazed, walked to water, formed loose leaderless herds and raised fawns. Wolves lived in family packs of four to twelve, led by the oldest animal, held territory, hunted the deer, fed together at a carcass, and fought rival leaders at the borders.</p><p>Nothing was scripted. Each animal ran a small state machine over its own hunger, age and neighbours. Everything at the population level, herds drifting, packs splitting, booms and crashes, was whatever fell out of those rules. One authoritative simulation ticked 24 hours a day on AWS, so everyone who opened the page saw the same animals in the same places at the same moment.</p><p>I switched it off last week. Here is what it cost, and what it taught me.</p><h3>A small world cannot hold a predator and its prey</h3><p>The first map was 24x24 tiles: fourteen deer, six wolves. It never once found a stable balance. Every tuning attempt produced the same two failure modes on a loop. Give the wolves enough reach and they overhunt, crash the deer, then starve. Take reach away and they starve anyway, while the herd, which always has a refuge somewhere, grows unchecked.</p><p>I swept the parameters exhaustively. On a map that small, coexistence was not a tuning problem at all.</p><p>What fixed it was not a better number. It was space. Scaling to 96x96, sixteen times the area, with every population and terrain feature scaled to keep density constant, changed the behaviour qualitatively. A big map does not hold one population. It holds many partly independent sub-populations that fall out of sync with each other, so when wolves crash the deer in one corner, deer from a distant corner wander back and repopulate it. Ecologists call this the metapopulation rescue effect. It is the difference between a system that oscillates and one that dies.</p><p>I had spent days tuning escape probabilities on a map that was structurally incapable of supporting the answer.</p><h3>Hunting is cheap. Eating is the hard part</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wwije1ZB3SFvlnMe.jpg" /></figure><p>For a long stretch every wolf pack in every random seed died out within sixteen in-game months. The cause was not the hunt. When a pack made a kill the whole pack ran to share it, hunger kept draining during that run, and the monthly starvation check killed the stragglers a few tiles short of the carcass. Every single wolf starvation death was happening to a wolf already committed to a meal. Exempting a wolf en route to a kill flipped the species from reliably extinct to reliably surviving. One conditional.</p><p>Then there was my favourite bug. A feeding wolf is exempt from starvation, so a wolf that could not physically reach an awkwardly placed carcass would freeze in place, permanently, and being technically mid-meal it could never die. These zombie wolves padded the population count and hid the fact that the pack could not sustain itself. Removing the freeze made the ecosystem look dramatically worse, because I was finally measuring it honestly.</p><p>A bug that props up a number is worse than a bug that breaks one. Every balance conclusion I drew during those weeks was wrong.</p><h3>Then it died anyway</h3><p>The tuned, verified answer held for the fifteen simulated years my sweeps covered. The live world ran for 443.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x1NTtaaT4qiJJwVp.jpg" /></figure><p>Over a long enough run the predators went extinct regardless, and the deer expanded until grass, not wolves, was the thing limiting them. In the final twelve in-game months: 94 starvation deaths, 4 of old age, zero predation. The herd settled into a population held down by hunger at the carrying capacity of the map. Nature’s least dignified equilibrium, and a real one.</p><p>A balance that survives fifteen years and fails at four hundred is not a balance. It is a slow transient. Only running the thing continuously for weeks showed me the difference.</p><h3>The bill was the real lesson</h3><p>The server broadcast the full world state ten times a second to every connected viewer. This is the easy, correct-by-construction design: no delta logic, no reconciliation bugs, and every browser is a dumb renderer. It is also free on localhost and ruinous the moment it crosses a metered load balancer.</p><p>In its final full month, that load balancer moved 715 GB.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/881/1*W5C244LT7D1XVauCpYuliQ.png" /></figure><p>Data transfer alone cost more than the compute, database, registry and DNS combined. Before I looked properly I would have guessed the whole thing ran at about $45, because the egress line simply was not in my mental model. A single browser tab left open on the site for a month generates on the order of a terabyte.</p><p>The uncomfortable part is that nothing was wrong. The traffic was not a bug, a leak or an attack. It was the design working exactly as intended, for an audience of nobody, at 10 Hz, forever.</p><p>Design your wire format assuming someone leaves the tab open for eternity. Full-state broadcast scales its cost with viewer-seconds, not with activity. Deltas were an optimisation I always meant to get to. The invoice got there first.</p><p>Wildwood ran from 7 June to 27 July 2026: seven weeks of wall time, 443 in-game years, about 4.7 million ticks. The domain now serves a static retrospective for a few cents a month, and the simulation still runs fine on my laptop, which is where an experiment like this probably belonged all along.</p><p><em>The full write-up, with more on deer behaviour, isometric depth sorting and the sprite-rendering bug that made every deer walk sideways, is at </em><a href="https://ecosystem.valentevidal.com/"><em>ecosystem.valentevidal.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5d72c84aed7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Atlantic problem: what UK and EU banks actually do when their software needs to serve Americans]]></title>
            <link>https://valentevidal.medium.com/the-atlantic-problem-what-uk-and-eu-banks-actually-do-when-their-software-needs-to-serve-americans-2b96e72e6799?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b96e72e6799</guid>
            <category><![CDATA[blockchain-technology]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[cloud-infrastructure]]></category>
            <category><![CDATA[platform-engineering]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Mon, 08 Jun 2026 11:36:50 GMT</pubDate>
            <atom:updated>2026-06-08T11:36:50.996Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ab8jgKaH4_GYmZTS" /></figure><h3>The setup</h3><p>A regulated firm in London, Paris, or Frankfurt builds a tokenisation platform. The platform lives on AWS in Ireland because Ireland is inside the EU&#39;s legal perimeter and because Dublin-to-London is roughly 12 milliseconds of fibre latency, which is to say, imperceptible.</p><p>Then a US client signs up. A New York pension fund wants to subscribe. A California family office wants access. Someone in compliance asks the question that ruins the week: <strong>where does their data live?</strong></p><p>This is not an abstract question. Personal dat, names, addresses, tax IDs, transaction history, <strong>is governed by the law of the place it sits, not the place it was collected.</strong> EU data in Ireland is protected by GDPR. US data in Virginia is governed by a patchwork of federal rules (SEC, FINRA, FinCEN, OFAC) and state rules (NYDFS in New York, BitLicense for crypto, the California Consumer Privacy Act). Mix them and you inherit both.</p><p>So what do the grown-ups do?</p><p><strong>What real banks actually do</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SildDf_BGDMc0k0t" /><figcaption>Canary Wharf, London, UK</figcaption></figure><p>The pattern, once you look at it, is consistent. HSBC, Barclays, NatWest, Standard Chartered: primary infrastructure in Ireland (`eu-west-1`) or London (`eu-west-2`), US presence in northern Virginia (`us-east-1`).</p><p>Deutsche Bank, BNP Paribas, ING: primary in Frankfurt (`eu-central-1`), US presence in Virginia.</p><p>Asset managers, Schroders, Allianz GI, Amund, same shape. State Street and BNY Mellon, who are American but serve European clients, mirror it: primary in Virginia, EU footprint in Ireland or Frankfurt.</p><p><strong>Nobody runs one global region. Nobody can.</strong> The rule under GDPR Chapter V is that personal data leaving the EU needs a legal basis for the transfer, and after the European Court of Justice&#39;s 2020 <strong>Schrems II</strong> ruling struck down the previous EU-US Privacy Shield agreement, that legal basis became materially harder to satisfy. The fix arrived in July 2023 with the EU-US Data Privacy Framework, but it only protects transfers to US companies that have specifically certified under it. <strong>AWS itself is certified. Your application sitting on top of AWS is not, unless you&#39;ve done the paperwork.</strong></p><p>The cleanest way to avoid the paperwork is to never make the transfer in the first place. Hence the pattern: US data lives in the US region, EU data lives in the EU region, and the two systems talk to each other through narrow, well-defined APIs rather than by replicating tables across the Atlantic.</p><p>Why Virginia and not, say, Oregon or Ohio?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DqVDrRH-VvFCecNq" /><figcaption>Richmond, Virginia, USA</figcaption></figure><p>Three reasons.</p><p>First, AWS launched its US-East region in 2006, so it has the broadest service coverage and the deepest compliance certifications, FedRAMP Moderate, SOC 1/2/3, PCI DSS Level 1, the lot.</p><p>Second, the New York Stock Exchange&#39;s matching engine is in Mahwah, New Jersey; Nasdaq&#39;s is in Carteret, New Jersey; both are roughly 300 km from Ashburn, Virginia, which translates to about 4 ms of network latency. For a broker-dealer routing orders, that distance is the difference between competitive and irrelevant.</p><p>Third, the US financial industry has already clustered there, which means the partner ecosystem, market-data vendors, KYC providers, custodians, is there too. Ohio (`us-east-2`) gets used as the disaster-recovery pair because it&#39;s geographically separated but legally identical. Oregon (`us-west-2`) picks up west-coast wealth management and the crypto natives (Coinbase, Anchorage).</p><p>There is also AWS GovCloud, which sounds like it should matter and almost never does. GovCloud is a physically segregated AWS partition built for US federal workloads, ITAR-controlled defence data, FedRAMP High classification. It&#39;s only accessible by US persons working for entities with a specific government contract. A British bank serving American retail clients has no reason to be there. The compliance scope is overkill and the service catalogue is smaller.</p><p><strong>The latency arithmetic</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dgDzsqh1iXDwxtTU" /></figure><blockquote><strong><em>You cannot do synchronous replication across the Atlantic; the speed of light won’t let you</em></strong>.</blockquote><p>To understand why the multi-region pattern is mandatory, not stylistic, do the latency maths. A fibre-optic packet travels at roughly two-thirds the speed of light through glass, so call it 200,000 km/s. London to Virginia is about 5,800 km of cable on the most direct submarine route (the AEC-1 cable, for the curious). That&#39;s a one-way ceiling of around 29 ms, and in practice you see 70 to 80 ms round-trip because real cables aren&#39;t straight lines and there are switches in the way.</p><p>Within a single AWS region, two data centres (&quot;availability zones&quot;) are at most 100 km apart, with a measured round-trip latency under 2 ms. AWS uses that gap to do synchronous database replication, every write is committed in two physical locations before it returns success. <strong><em>You cannot do synchronous replication across the Atlantic; the speed of light won&#39;t let you</em></strong>. Every write would take longer than a database transaction is allowed to take. So cross-Atlantic setups are necessarily asynchronous, which means at any given moment the two sides disagree about reality by some amount, which means you have to design the application to tolerate that disagreement, which is a lot more work than running two independent systems and reconciling at a coarser grain.</p><p>This is the real reason banks run separate per-jurisdiction stacks. It&#39;s not legal preference. It&#39;s that physics and law happen to agree.</p><p><strong>What it would take to make your codebase US-ready</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LWAHckqFb3EQExfO" /></figure><p>Not a migration. An addition, in five pieces.</p><p><strong>Data classification</strong>. A bucket name that encodes its region is honest, but it doesn’t stop a US resident uploading a KYC document into European infrastructure. The application has to know which jurisdiction a user belongs to before it picks a bucket, and a US-region bucket has to exist before the first US user signs up. The chain itself is fine — block data is pseudonymous. The off-chain bits aren’t.</p><p><strong>Record-keeping</strong>. SEC Rule 17a-4 requires broker-dealers to keep trade records for six years in WORM format: no edits, not even by an administrator. AWS supports it through S3 Object Lock in Compliance mode. The common starting point is Governance mode — same feature, but overridable. Fine for hygiene, not for audit. One configuration value, set before the data lands.</p><p><strong>Sanctions.</strong> OFAC publishes a list of sanctioned wallet addresses. The Tornado Cash case set the precedent that contract authors can be personally liable if their code lets US persons transact with blocked addresses. A credential schema can have a slot for sanctions screening, but the slot only matters if the access-control layer reads it — and the check has to be enforced on-chain so it can’t be bypassed.</p><p><strong>Operational resilience.</strong> DORA (EU, January 2025) and the OCC’s Heightened Standards (US) both require tested recovery — typically active-passive across two regions with regular failover drills. Single-region, single-AZ, no replication satisfies neither. The largest hidden cost of going to the US is that it forces the EU side to upgrade too.</p><p><strong>On-call.</strong> Sounds soft, isn’t. An outage at 10 AM Eastern is 3 PM London (fine) but 7 AM California (not fine). Follow-the-sun, US hires, or a US managed-service partner — each costs real money.</p><p><strong>The hurdles, ranked by how easy they are to underestimate</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8junMtAfbWSWSrbY" /></figure><p>Compliance work isn’t really a checklist of certifications. The traps are subtler.</p><p>Schrems II in disguise. You can have a perfectly partitioned US deployment, EU data never touching American soil, and still violate the rule because your authentication system writes login events into a global log aggregator that happens to live in Virginia. The European Data Protection Board has been clear that even metadata about EU citizens counts as personal data. The right audit question isn’t “where is the customer record?” but “where does every byte that mentions the customer’s name end up?”, logs, metrics, error traces, support tickets, the email confirming their password reset, all of it. Real banks treat their observability stack as a regulated system, not as plumbing.</p><p>State-level US regulation. The US has no federal data-protection law. On top of the federal banking framework, every state can layer its own — New York’s NYDFS Part 500 and BitLicense, California’s CCPA, Texas’s money-transmitter regime, fifty in total. Most UK fintechs don’t handle this directly; they partner with a US-licensed entity and operate B2B2C, letting the partner absorb the state regulation in exchange for a slice of revenue.</p><p>Sub-processor disclosure. GDPR Article 28 requires you to disclose every third-party service that touches European personal data. A US client of a UK fintech will ask for the list. If it doesn’t exist, you write it from scratch under deadline pressure during a sales cycl, the worst time.</p><p>Cross-region data egress cost. AWS charges roughly $0.02 per gigabyte to move data between regions. Trivial until you imagine replicating a blockchain across the Atlantic, indefinitely. A modestly active chain produces hundreds of gigabytes a month, which makes the wrong architecture (one global chain) more expensive than the right one (one chain per jurisdiction, bridged at the application layer). The economics push you towards the legally correct answer, if you notice them before building the wrong thing.</p><p>FedRAMP envy. The natural urge, when listing US certifications, is to aim at the biggest one, FedRAMP High, the federal-government clearance. Almost always a mistake. It’s built for national-security data, requires the GovCloud partition, shrinks your service catalogue by about 30%, and doesn’t help with commercial clients, who aren’t federal agencies. The right certifications for a US-facing fintech are SOC 2 Type II, ISO 27001, and PCI DSS (if you touch card data), all available natively in us-east-1.</p><p><strong>But Ripple has one chain</strong></p><p>The obvious objection to all of this is that Ripple runs a single global ledger, and so do Bitcoin and Ethereum, and they seem to manage. The answer is that they aren’t counterexamples, they’re a different category. The XRP Ledger is public and permissionless, the data on it is pseudonymous, and no individual is identified on chain. GDPR doesn’t apply to data that doesn’t identify a person. The KYC happens at the on-ramp, in a regulated database, in the correct jurisdiction. The chain itself is just the settlement layer. It’s the railroad; the passenger lists stay in the station.</p><p>The interesting test case is Fnality, because Fnality is doing exactly what most regulated tokenisation platforms want to do: a private, permissioned, bank-grade chain backed by central-bank money. Twenty-four of the largest banks in the world are shareholders, Goldman, Barclays, BNY Mellon, Santander, BNP Paribas, State Street, UBS, and so on. They spent roughly a decade deciding whether one global chain was possible. They chose not to. The Sterling Fnality Payment System is regulated by the Bank of England. The US dollar version is a separate system. The euro version, when it launches, will be separate again. Their own website lists “Operating Locations: UK / USA” as two distinct things. They interoperate at the application layer through atomic settlement, not by sharing ledger state. That conclusion, reached by twenty-four banks with a decade and a serious budget, is the same conclusion this article reaches from first principles.</p><p>So when can a private chain be a single chain? Only when the chain itself genuinely holds no personal data — and the bar for that is higher than it sounds. The European Data Protection Board’s position is that an on-chain address becomes personal data the moment any party can link it back to an identified person through off-chain records, which means a pseudonymous chain attached to a KYC database is still, legally, processing personal data. The cleanest way to avoid this is a zero-knowledge architecture: the chain holds cryptographic commitments, the regulated entity holds the linkage, users prove entitlement without revealing identity. That works, and it is starting to work in production at the institutional scale, but it adds significant engineering complexity and the regulators are still figuring out how to audit it.</p><p><strong>What this article is really about</strong></p><p>Physics, law, and money all point at the same architecture: American clients served from American infrastructure, European clients served from European infrastructure, the two systems talking through boundaries you’ve thought about rather than through shared logs and forgotten replication rules.</p><p>The banks that do this well treat the per-jurisdiction split as a design discipline, not a compliance burden. It forces them to be explicit about which user belongs to which legal regime, and which features each user gets, which turns out to be useful anyway. The banks that do it badly build one global system and bolt regional restrictions on top. That works until the first audit.</p><p>The work itself isn’t research. It’s disciplined plumbing: a US workload account, a mirrored container registry, jurisdiction-awareness in the application layer, record-keeping buckets in Compliance mode, OFAC screening in the access-control layer, and a published sub-processor list. Known problems, known solutions.</p><p>The deeper point is that good architecture is almost always a matter of listening to the constraints carefully enough that they tell you the answer. The reason banks separate by jurisdiction isn’t that someone in a strategy meeting in 2017 thought it would be elegant. It’s that the speed of light, the European Court of Justice, the Bank of England, and AWS’s per-gigabyte egress pricing all agreed independently, and the firms that argued with them lost. The right time to start building this is before a client asks for it, not because the client will be impressed, but because the architecture is harder to retrofit than to design, and the constraints aren’t going to soften while you wait.</p><h3><strong>Sources</strong></h3><p><strong>Region and infrastructure</strong><br>- AWS Global Infrastructure (39 regions, 123 AZs) — <a href="https://aws.amazon.com/about-aws/global-infrastructure/regions_az/">https://aws.amazon.com/about-aws/global-infrastructure/regions_az/</a><br>- AWS service endpoints by region — <a href="https://docs.aws.amazon.com/general/latest/gr/rande.html">https://docs.aws.amazon.com/general/latest/gr/rande.html</a><br>- AWS Financial Services Industry Lens (Well-Architected) — <a href="https://docs.aws.amazon.com/wellarchitected/latest/financial-services-industry-lens/welcome.html">https://docs.aws.amazon.com/wellarchitected/latest/financial-services-industry-lens/welcome.html</a></p><p><strong>EU-US data transfer law</strong><br>- European Commission, EU-US Data Privacy Framework — <a href="https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en">https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en</a><br>- Schrems II ruling summary — <a href="https://www.gdprsummary.com/schrems-ii/">https://www.gdprsummary.com/schrems-ii/</a><br>- AWS GDPR Center — <a href="https://aws.amazon.com/compliance/gdpr-center/">https://aws.amazon.com/compliance/gdpr-center/</a><br>- AWS EU data protection — <a href="https://aws.amazon.com/compliance/eu-data-protection/">https://aws.amazon.com/compliance/eu-data-protection/</a></p><p><strong>US financial regulation</strong><br>- SEC Rule 17a-4 (current text on eCFR) — <a href="https://www.ecfr.gov/current/title-17/chapter-II/part-240/subject-group-ECFRddc6cf6dc6df8c0/section-240.17a-4">https://www.ecfr.gov/current/title-17/chapter-II/part-240/subject-group-ECFRddc6cf6dc6df8c0/section-240.17a-4</a><br>- AWS Rule 17a-4(f) compliance page — <a href="https://aws.amazon.com/compliance/secrule17a-4f/">https://aws.amazon.com/compliance/secrule17a-4f/</a><br>- FINRA books and records — <a href="https://www.finra.org/rules-guidance/key-topics/books-records">https://www.finra.org/rules-guidance/key-topics/books-records</a><br>- FinCEN Bank Secrecy Act — <a href="https://www.fincen.gov/resources/statutes-and-regulations/bank-secrecy-act">https://www.fincen.gov/resources/statutes-and-regulations/bank-secrecy-act</a><br>- OFAC sanctions programs — <a href="https://ofac.treasury.gov/sanctions-programs-and-country-information">https://ofac.treasury.gov/sanctions-programs-and-country-information</a></p><p><strong>UK and EU financial regulation</strong><br>- FCA Handbook, SYSC 8 (outsourcing) — <a href="https://www.handbook.fca.org.uk/handbook/SYSC/8/">https://www.handbook.fca.org.uk/handbook/SYSC/8/</a><br>- EBA Guidelines on Outsourcing — <a href="https://www.eba.europa.eu/regulation-and-policy/internal-governance/guidelines-on-outsourcing-arrangements">https://www.eba.europa.eu/regulation-and-policy/internal-governance/guidelines-on-outsourcing-arrangements</a><br>- Digital Operational Resilience Act overview — <a href="https://digital-operational-resilience-act.com/">https://digital-operational-resilience-act.com/</a><br>- BCBS 239, principles for risk-data aggregation — <a href="https://www.bis.org/publ/bcbs239.htm">https://www.bis.org/publ/bcbs239.htm</a></p><p><strong>AWS certifications referenced</strong><br>- FedRAMP services in scope — <a href="https://aws.amazon.com/compliance/services-in-scope/FedRAMP/">https://aws.amazon.com/compliance/services-in-scope/FedRAMP/</a><br>- AWS SOC FAQs — <a href="https://aws.amazon.com/compliance/soc-faqs/">https://aws.amazon.com/compliance/soc-faqs/</a><br>- AWS PCI DSS Level 1 FAQs — <a href="https://aws.amazon.com/compliance/pci-dss-level-1-faqs/">https://aws.amazon.com/compliance/pci-dss-level-1-faqs/</a><br>- AWS Financial Services Compliance Center — <a href="https://aws.amazon.com/financial-services/security-compliance/compliance-center/">https://aws.amazon.com/financial-services/security-compliance/compliance-center/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b96e72e6799" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Back Up and Restore Multiple Ethereum Validators from One Recovery backup (Without Taking…]]></title>
            <link>https://valentevidal.medium.com/how-to-back-up-and-restore-multiple-ethereum-validators-from-one-recovery-backup-without-taking-c236576e47ad?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/c236576e47ad</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[blockchain]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Tue, 24 Mar 2026 21:45:21 GMT</pubDate>
            <atom:updated>2026-03-24T21:45:21.238Z</atom:updated>
            <content:encoded><![CDATA[<h3>How to Back Up and Restore Multiple Ethereum Validators from One Recovery backup (Without Taking Multiple Separate Backups)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*2elcVGiflx4mELu0.jpg" /></figure><h3>Introduction: the cost problem and the right model</h3><p>Most teams overspend on validator backups because they think in “one validator, one backup.”<br>That creates duplicate work, higher storage cost, and slower recovery.</p><p>The model we use now is better: one fleet-level recovery set, not ten separate validator backups.</p><p>But there’s an important technical truth:</p><p>You cannot restore 10 validators from one validator key.<br>You can restore 10 validators from one complete recovery package that includes:</p><ul><li>all validator keystores (or deterministic source that generates them),</li><li>slashing protection history,</li><li>passwords and client config,</li><li>and Kubernetes infrastructure state.</li></ul><p>This is where <strong>Velero</strong> fits in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/433/0*RaX-KbrBnZqSCPYE" /></figure><p>Velero is excellent for Kubernetes objects and persistent volume snapshots.<br>Velero is not, by itself, your complete validator safety story unless you also include validator-specific cryptographic and slashing artifacts.</p><p>So the real production pattern is:</p><p>Velero for cluster and storage state<br>plus<br>validator recovery bundle for key/slashing/config integrity</p><p>That is exactly how to save money safely while keeping reseed/restore operationally sound.</p><p>What I’m currently using</p><p>Today’s process is effectively:</p><ol><li>A shared validator recovery model (not one-key-for-many)</li></ol><ul><li>One controlled source of validator identities (seed/mnemonic strategy or full keystore set)</li><li>Full validator set inventory (all pubkeys/indices)</li><li>Slashing protection exports</li><li>Password references and runtime settings</li></ul><ol><li>Velero for Kubernetes backup/restore</li></ol><ul><li>Namespace resources (Deployments, StatefulSets, Secrets references, ConfigMaps, Services, etc.)</li><li>PVC snapshots or file-system backups (depending on storage class and setup)</li><li>Restore orchestration into the target cluster/environment</li></ul><ol><li>Controlled reseed procedure</li></ol><ul><li>Restore infra with Velero</li><li>Import slashing protection first</li><li>Load all validator keys</li><li>Start a single active signer set only after old signer is stopped</li><li>Verify validator count and attestation health</li></ul><p>This is the right approach for cost reduction without slashing risk.</p><p>Architecture: what must be backed up every time</p><p>Think of backup scope in two planes.</p><p>Control plane (Kubernetes / Velero scope)</p><ul><li>Namespaces and manifests</li><li>Stateful workload definitions</li><li>ConfigMaps and non-inline configs</li><li>PVC snapshots / storage objects</li><li>Restore metadata</li></ul><p>Validator safety plane (must be included outside raw Velero object backup)</p><ul><li>All validator keystores for the fleet</li><li>Slashing protection export/interchange</li><li>Keystore password source or secure reference</li><li>Validator inventory manifest (pubkey/index/file mapping)</li><li>Client runtime flags and version metadata</li></ul><p>If either plane is missing, restore is incomplete.</p><p>Technical implementation</p><p><strong>Phase 1:</strong> Build a fleet recovery bundle</p><p>Create a bundle structure like:</p><p>recovery-bundle/<br>keystores/<br>slashing/<br>slashing-interchange-YYYY-MM-DD.json<br>config/<br>validator-flags.txt<br>fee-recipient.txt<br>metadata/<br>validators-inventory.json<br>versions.txt</p><p>Important: the inventory must be authoritative so you can prove restore completeness.</p><p>Example inventory record:</p><pre>{<br>&quot;pubkey&quot;: &quot;0x…&quot;,<br>&quot;index&quot;: 123456,<br>&quot;keystore_file&quot;: &quot;keystore-….json&quot;<br>}</pre><p><strong>Phase 2: </strong>Export slashing protection on schedule</p><p>Before each backup cycle, export slashing data from the active validator client.<br>Do not skip this step.<br>Keys without slashing history are a major migration risk.</p><p><strong>Phase 3:</strong> Run Velero backup for cluster state</p><p>Example command pattern:</p><p>velero backup create validators-prod-2026–03–24<br> — include-namespaces validators<br> — snapshot-volumes<br> — ttl 720h</p><p>If you use CSI snapshots, ensure VolumeSnapshotClass and plugin settings are validated in your cluster.</p><p><strong>Phase 4:</strong> Store recovery bundle in encrypted object storage</p><p>Archive and encrypt your recovery-bundle, then upload with immutable/versioned retention.<br>Keep checksum + manifest with the archive.</p><p><strong>Restore and reseed runbook</strong></p><p><strong>Step 1: </strong>Stop old signer before activating new<br>Never run the same validator keys in two active places.</p><p><strong>Step 2: </strong>Restore Kubernetes resources with Velero</p><pre>velero restore create - from-backup validators-prod-2026–03–24</pre><p>Wait for resources and PVC restore to complete.<br>Validate pods, mounts, and service wiring.</p><p><strong>Step 3:</strong> Recover validator bundle</p><ul><li>Download recovery bundle</li><li>Verify checksum</li><li>Decrypt and unpack</li><li>Confirm keystore count equals inventory count</li></ul><p><strong>Step 4:</strong> Import slashing protection first<br>Import slashing interchange into validator client before loading keys.</p><p><strong>Step 5:</strong> Load keys and config</p><ul><li>Restore all keystores</li><li>Restore password references</li><li>Restore validator runtime flags</li><li>Confirm fee recipient settings</li></ul><p><strong>Step 6:</strong> Controlled startup</p><ul><li>Start EL/CL health checks</li><li>Start validator client</li><li>Verify expected validator count loaded</li><li>Monitor first epochs for attestations and duplicate-sign protections</li></ul><p><strong>Why this saves money</strong></p><p>You reduce cost by eliminating duplicate per-node backup workflows, not by reducing key safety.</p><p>Savings come from:</p><ul><li>one standardized fleet backup pipeline,</li><li>one Velero policy per validator namespace/environment,</li><li>one recovery bundle lifecycle policy,</li><li>one tested restore playbook.</li></ul><p><strong>That is cheaper and safer than maintaining ten independent backup strategies.</strong></p><p>Common mistakes to avoid</p><p>Mistake 1: “One validator key can restore all validators”<br>Reality: each validator has its own signing identity.</p><p>Mistake 2: Relying on Velero alone<br>Reality: Velero restores infra state; validator cryptographic/slashing artifacts still need explicit handling.</p><p>Mistake 3: Restoring keys without slashing history<br>Reality: this is where migration incidents happen.</p><p>Mistake 4: No restore drills<br>Reality: untested backups are assumptions, not recovery plans.</p><p><strong>Final takeaway</strong></p><p>The safe and cost-efficient model is:</p><p>One Velero-based infrastructure backup workflow<br>plus<br>one validator fleet recovery bundle (keys + slashing + config)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c236576e47ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Blockchain That’s Been Running Since 1991 (And it’s not Bitcoin)]]></title>
            <link>https://valentevidal.medium.com/the-blockchain-thats-been-running-since-1991-and-it-s-not-bitcoin-e5bab417a4e4?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5bab417a4e4</guid>
            <category><![CDATA[bitcoin]]></category>
            <category><![CDATA[the-new-york-times]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[cryptography]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Tue, 20 Jan 2026 13:47:51 GMT</pubDate>
            <atom:updated>2026-01-20T13:47:51.535Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*3nyrDkvdkiRopWDI.jpeg" /><figcaption>Stuart Haber</figcaption></figure><p>When people talk about the origins of blockchain, the conversation almost always starts with Bitcoin. But nearly two decades earlier, a quieter and more enduring system was already in operation, one that has been running continuously since the early 1990s and is still active today.</p><p>In 1991, cryptographers <strong>Stuart Haber</strong> and <strong>W. Scott Stornetta</strong> proposed a method for timestamping digital documents so that they could not be altered retroactively. Their solution was elegant: instead of storing documents directly, they stored <strong>cryptographic hashes</strong> of documents, and each new record included the hash of the previous one. This created an append-only, tamper-evident chain of records.</p><p>At its core, this is exactly what a blockchain is.</p><h3>A concrete example of a hash chain</h3><p>Imagine three documents:</p><ul><li>Document A: “Employment contract — version 1”</li><li>Document B: “Employment contract — version 2”</li><li>Document C: “Employment contract — version 3”</li></ul><p>Each document is passed through a cryptographic hash function (for example, SHA-256):</p><ul><li>Hash(A) → 3f2a…9c1e</li><li>Hash(B) → a91b…72d4</li><li>Hash(C) → e44c…0f89</li></ul><p>Now the chain is formed like this:</p><ul><li>Record 1 stores Hash(A)</li><li>Record 2 stores Hash(B) <strong>plus </strong><strong>Hash(A)</strong></li><li>Record 3 stores Hash(C) <strong>plus </strong><strong>Hash(B)</strong></li></ul><p>If Document A is changed by even one character, its hash becomes completely different, which breaks every link after it. The chain immediately reveals tampering. This property — <strong>immutability through hash linkage</strong> — is the foundation of all modern blockchains.</p><h3>Anchoring trust in the physical world</h3><p>A remaining problem was trust: how do you prove that a given hash existed at a specific point in time?</p><p>The solution was unexpectedly simple. Since 1995, the root hash of the timestamp chain has been <strong>published weekly in the classified ads of</strong> <strong>The New York Times</strong>. Once printed in a widely distributed physical newspaper, the hash became impossible to alter retroactively. Anyone could later verify that a digital document existed before that publication date.</p><p>This created a public, immutable timestamp — without mining, tokens, or consensus protocols.</p><p>What Haber–Stornetta did conceptually is simple hash chaining; the novelty was the <em>idea</em> and the public anchoring, not the mechanics.</p><p>Below is a <strong>minimal, hands-on walkthrough</strong> you can run locally using openssl.</p><h3>1. Create some example documents</h3><pre>echo &quot;Document A: Employment contract v1&quot; &gt; docA.txt<br>echo &quot;Document B: Employment contract v2&quot; &gt; docB.txt<br>echo &quot;Document C: Employment contract v3&quot; &gt; docC.txt</pre><h3>2. Hash each document (SHA-256)</h3><pre>openssl dgst -sha256 docA.txt<br>openssl dgst -sha256 docB.txt<br>openssl dgst -sha256 docC.txt</pre><p>Example output:</p><pre>SHA256(docA.txt)= 3f2a9c1e...</pre><p>This hash is the <strong>fingerprint</strong> of the document.<br> Change one character → completely different hash.</p><h3>3. Build a simple hash chain (core blockchain idea)</h3><p>Now we <strong>link hashes together</strong>, just like a blockchain.</p><h3>Step 1: hash document A</h3><pre>openssl dgst -sha256 docA.txt | awk &#39;{print $2}&#39; &gt; hashA.txt</pre><h3>Step 2: hash document B + hashA</h3><pre>cat docB.txt hashA.txt &gt; recordB.txt<br>openssl dgst -sha256 recordB.txt | awk &#39;{print $2}&#39; &gt; hashB.txt</pre><h3>Step 3: hash document C + hashB</h3><pre>cat docC.txt hashB.txt &gt; recordC.txt<br>openssl dgst -sha256 recordC.txt | awk &#39;{print $2}&#39; &gt; hashC.txt</pre><p>You now have a <strong>hash chain</strong>:</p><pre>hashA → hashB → hashC</pre><p>If <strong>docA.txt changes</strong>, hashA changes → hashB breaks → hashC breaks.</p><p>That’s immutability.</p><h3>4. Demonstrate tamper detection</h3><p>Edit the first document:</p><pre>echo &quot;Document A: Employment contract v1 (EDITED)&quot; &gt; docA.txt</pre><p>Recompute:</p><pre>openssl dgst -sha256 docA.txt</pre><p>The hash is now different.<br> Your entire chain is invalid.</p><p>This is exactly why blockchains are tamper-evident.</p><h3>5. (Optional) Timestamp a hash</h3><p>To replicate the <em>timestamping</em> idea:</p><pre>date -u +&quot;%Y-%m-%dT%H:%M:%SZ&quot; &gt; timestamp.txt<br>cat hashC.txt timestamp.txt &gt; anchored.txt<br>openssl dgst -sha256 anchored.txt</pre><p>This produces a <strong>time-anchored hash</strong>, similar to what was later published in newspapers.</p><h3>6. How this maps to the New York Times system</h3><p>This simple demonstration maps directly onto the system designed by Haber and Stornetta and later anchored in <em>The New York Times</em>. In the example, a plain text file such as docA.txt represents a legal or digital record, while the output of openssl dgst corresponds to the cryptographic digest stored instead of the document itself. Linking hashA to hashB and then to hashC mirrors the hash-linked ledger structure that makes records tamper-evident, and adding a timestamp serves the same purpose as publishing a hash in a widely distributed newspaper: it anchors the data to a specific point in time. Bitcoin builds on this exact cryptographic core, adding decentralised consensus and economic incentives, but the underlying mechanism remains the same.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e5bab417a4e4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Story The Imitation Game Didn’t Tell]]></title>
            <link>https://valentevidal.medium.com/the-story-the-imitation-game-didnt-tell-ac249b42075d?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac249b42075d</guid>
            <category><![CDATA[poland]]></category>
            <category><![CDATA[british]]></category>
            <category><![CDATA[cryptography]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[second-world-war]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Tue, 20 Jan 2026 13:24:43 GMT</pubDate>
            <atom:updated>2026-01-20T13:24:43.923Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*DTFPD0T7HShbgqIf.jpg" /><figcaption>The three polish codebreakers Marian Rejewski, Jerzy Różycki, and Henryk Zygalski (ca. 1928–1932)</figcaption></figure><p>Popular culture often credits the breaking of the Enigma code to <strong>Alan Turing</strong> and the team at <strong>Bletchley Park</strong>, largely influenced by simplified historical narratives and films such as <em>The Imitation Game</em>. While the British contribution was extraordinary and decisive during the war, this framing is incomplete. It typically begins the story at the outbreak of World War II, overlooking nearly a decade of earlier breakthroughs that made British success possible.</p><p>In reality, <strong>Enigma was first broken in the early 1930s by Polish cryptographers</strong>, at a time when Germany was not yet at war and cryptanalysis was treated as a mathematical discipline rather than an industrial one. At the Polish Cipher Bureau, mathematician <strong>Marian Rejewski</strong> applied permutation theory and group mathematics to reverse-engineer the internal wiring of the <strong>Enigma machine</strong>, something widely believed at the time to be impossible without access to the machine itself. His colleagues <strong>Jerzy Różycki</strong> and <strong>Henryk Zygalski</strong> complemented this work by developing practical techniques for exploiting operator habits and procedural weaknesses.</p><p>The Polish team did not merely understand Enigma in theory. They created <strong>operational methods</strong> to read German traffic on a routine basis. These included the <em>cyclometer</em> for cataloguing rotor settings, Zygalski’s perforated sheets for recovering daily keys, and the <em>bomba kryptologiczna</em>, an electromechanical device designed to automate parts of the key-search process. By the mid-1930s, Poland was successfully reading large volumes of German military communications.</p><p>As Germany expanded Enigma’s complexity in the late 1930s — adding rotors and tightening procedures — the Polish Cipher Bureau faced diminishing resources and an increasingly precarious geopolitical situation. In <strong>July 1939</strong>, just weeks before the German invasion of Poland, the Bureau made a critical decision: it shared <strong>everything</strong> with British and French intelligence. This included reconstructed Enigma machines, detailed documentation of mathematical methods, and physical cryptanalytic tools.</p><p>This handover fundamentally altered the Allied cryptographic effort. When war began, British codebreakers were not confronting Enigma as an unsolved mystery. They were inheriting a partially solved system, complete with proven analytical approaches. What followed in Britain was therefore <strong>not the initial cryptographic breakthrough</strong>, but its <strong>scaling and industrialisation</strong>.</p><p>At Bletchley Park, Turing and his colleagues transformed Polish concepts into machines capable of operating at wartime scale. The British <em>bombe</em> was a more powerful descendant of the Polish design, adapted to cope with increased message volume, additional rotors, and the complexity of naval Enigma. Organisational innovation, mass production, and round-the-clock operations turned what had been an elite mathematical success into a sustained intelligence pipeline.</p><p>The historical reality is less dramatic than popular myth, but far more instructive. Enigma was not broken by a single nation or a single moment of genius. It was broken through <strong>early mathematical insight, international cooperation, and the systematic expansion of an existing breakthrough</strong> — a reminder that foundational work, even when done quietly and under severe constraints, can shape the outcome of global events.</p><h3>References</h3><ol><li><a href="https://ieeexplore.ieee.org/document/4640685/">Rejewski, M. (1981). <em>How Polish Mathematicians Deciphered the Enigma</em>. Annals of the History of Computing, IEEE.</a></li><li><a href="https://archive.org/details/enigmahowgermanm0000koza">Kozaczuk, W. (1984). <em>Enigma: How the German Machine Cipher Was Broken, and How It Was Read by the Allies in World War Two</em>. University Publications of America.</a></li><li><a href="https://press.princeton.edu/books/paperback/9780691164724/alan-turing-the-enigma?srsltid=AfmBOopbeu5zTUXyno7wC67np9bfWs-gxxw4Jku41IQqwxb3he4Q4Fx2">Hodges, A. (2012). <em>Alan Turing: The Enigma</em>. Princeton University Press.</a></li><li><a href="https://www.google.co.uk/books/edition/Station_X/Wv4mSVDtA-wC?hl=en">Smith, M. (2011). <em>Station X: The Codebreakers of Bletchley Park</em>. Pan Macmillan.</a></li><li><a href="https://www.google.co.uk/books/edition/Colossus/M2sVDAAAQBAJ?hl=en">Copeland, B. J. (ed.) (2006). <em>Colossus: The Secrets of Bletchley Park’s Codebreaking Computers</em>. Oxford University Press.</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac249b42075d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why 0.4 Percent of Java Changed Software History]]></title>
            <link>https://valentevidal.medium.com/why-0-4-percent-of-java-changed-software-history-13501bc6e637?source=rss-6d219d3509da------2</link>
            <guid isPermaLink="false">https://medium.com/p/13501bc6e637</guid>
            <category><![CDATA[oracle]]></category>
            <category><![CDATA[patents]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[copyright]]></category>
            <category><![CDATA[java]]></category>
            <dc:creator><![CDATA[Valente Vidal]]></dc:creator>
            <pubDate>Sat, 13 Dec 2025 14:55:57 GMT</pubDate>
            <atom:updated>2025-12-13T14:55:57.085Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HGldX6OFjLxV1dCQFIEmyA.png" /></figure><p>I wanted to understand what Google actually copied from Java that caused Oracle to sue, especially since I am not a Java developer. To make sense of it, I compared Java to Python, which I know better.</p><p>Python is implemented in C, and developers interact with it through a well-defined interface. When you call something like math.sqrt(x), you are not interacting with the C implementation directly; you are using the interface that exposes that functionality.</p><p>Google did something similar with Java. They rewrote their own Java implementation from scratch, comparable to Python’s underlying C code, but reused Java’s interface declarations so developers could call familiar methods. That reused portion was tiny.</p><ul><li>Java had roughly 15 million lines of code.</li><li>Google reused about 11,500 lines.</li><li>That is around 0.4 percent of the total codebase.</li></ul><p>Oracle argued that this 0.4 percent was the most valuable part of Java. It represented years of design decisions, and it was the part developers relied on every day. From Oracle’s perspective, Google had taken the heart of Java.</p><p>Google’s counterargument was simple. They copied only what was necessary for compatibility, wrote millions of lines of original code, used the APIs to build something new in Android, and did not replace Java SE or harm its market.</p><p>When the case reached the US Supreme Court, that 0.4 percent mattered a lot, especially for two fair use factors.</p><p>First, amount and substantiality. Copying such a small portion, and only what was needed, weighed strongly in Google’s favour.</p><p>Second, purpose and character. The Court saw the copied material as a bridge for developers, not a substitute for Java itself. Google took only what programmers needed to rely on existing knowledge, and did not copy the creative implementation.</p><p>That distinction was crucial.</p><p>If the ruling had gone the other way, the immediate impact on Android might have been limited. Android launched successfully in 2007 and 2008, Sun Microsystems did not sue Google, and Oracle only brought the case after acquiring Sun. By the time courts reached a decision in 2016, the ecosystem was already established. Google could likely have paid Oracle and moved further toward Kotlin.</p><p>The bigger risk was the precedent.</p><p>Interfaces are valuable precisely because many people use them. Like the QWERTY keyboard, their usefulness comes from shared adoption. If reusing widely accepted interfaces were treated as copyright infringement, it would lock developers into existing platforms and strengthen monopolies rather than weaken them.</p><p>Whether the decision was “right” is still debatable. But it clearly leans toward innovation and one of software engineering’s core practices, reuse. Allowing developers to build new systems on top of familiar interfaces makes progress faster, not slower.</p><p>And in this case, that 0.4 percent helped shape the modern mobile world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=13501bc6e637" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>