<?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 Sitraka FORLER on Medium]]></title>
        <description><![CDATA[Stories by Sitraka FORLER on Medium]]></description>
        <link>https://medium.com/@sitrakaforler?source=rss-b075dea26af9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*SFONzfO2w2xyrrHG51W6-w.jpeg</url>
            <title>Stories by Sitraka FORLER on Medium</title>
            <link>https://medium.com/@sitrakaforler?source=rss-b075dea26af9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:29:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sitrakaforler/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[My Finance Students Kept Asking Where to Start With Python. So I Built a Free Course.]]></title>
            <link>https://medium.com/@sitrakaforler/my-finance-students-kept-asking-where-to-start-with-python-so-i-built-a-free-course-5ea19b64e247?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ea19b64e247</guid>
            <category><![CDATA[education]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 10:39:33 GMT</pubDate>
            <atom:updated>2026-11-09T10:39:33.226Z</atom:updated>
            <content:encoded><![CDATA[<h4>64 lessons, real finance data, nothing to install. What I learned building EcoFinLearn, and the numbers after two months.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DXj3iFO-RPHTxX8stK9DwA.jpeg" /><figcaption>Photo by Yan Krukau on Pexels</figcaption></figure><p>Every term, same scene. First session, forty laptops open, and before the first slide a student asks: “Sir, which Python course should I take?”</p><p>The honest answer used to be awkward. The good free courses teach Python with cats, pizzas and to-do lists. The finance ones cost money, or start with three hours of installation. My students study finance and economics. They want to compute a return, not a pizza order.</p><p>So I built the course I wanted to give them: <a href="https://ecofinlearning.com/">EcoFinLearn</a>. Free, no account needed, and every exercise is about money.</p><h3>What it is</h3><p>Six tracks, 64 short lessons:</p><ul><li><strong>Python Foundations</strong> (11 lessons): returns, P&amp;L conditions, NPV and IRR, bond duration.</li><li><strong>Data Manipulation</strong> (16): pandas on real-looking prices: log returns and volatility, CAPM beta, value at risk, an event study, a moving-average backtest.</li><li><strong>Machine Learning</strong> (9): linear and logistic regression, walk-forward backtests, and a confusion matrix where missing a crash costs more than a false alarm.</li><li><strong>SQL Foundations</strong> (8) and <strong>Advanced Finance SQL</strong> (7): joins, CTEs, a 20-day moving average and an RSI in pure SQL, a point-in-time macro join.</li><li><strong>IT Essentials</strong> (13): what nobody teaches finance students but every job needs: Git, HTTP status codes, JSON APIs, secrets, CI/CD, the cloud.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vj5FBsMMkd0UxDpt_6m2EA.png" /><figcaption>Lesson 1: write a line, press Run, the checks pass. Screenshot: ecofinlearning.com</figcaption></figure><p>Everything runs in the browser. Python runs through Pyodide, SQL through SQLite compiled to WebAssembly. You open the page, you write code, you press Run. Each exercise is checked automatically, and you see in one second if your Sharpe ratio is right.</p><h3>Three rules I kept</h3><p><strong>1. Finance first, syntax second.</strong> Lesson 1 is not “Hello, World”. It is “Hello, Finance World”: you buy 10 Apple shares at $174.55 and print the total. A for loop walks through a portfolio. A dictionary IS a portfolio. Students remember the syntax because they remember the trade.</p><p><strong>2. Count first, model second.</strong> Before any model, you look at the data, you set a naive baseline, and you check the model on rows it has not seen. On small data, a 0.01 difference is noise. I repeat this in class until they are tired of me.</p><p><strong>3. No setup, ever.</strong> If you have taught Python to a full room, you know the first hour: half of the laptops refuse to install anything. Here, the first line of code runs in the first minute.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R48_HvxsBtHqflscOPDLOg.png" /><figcaption>A glossary page: drag the volatility and watch the option price move. Screenshot: ecofinlearning.com</figcaption></figure><h3>What else is inside</h3><ul><li>A glossary of about 90 finance, economics and IT terms, each with a formula, a worked example and an interactive chart (Black-Scholes, Gordon growth, Cobb-Douglas, put-call parity…).</li><li>Eight calculators, each with a worked example: Black-Scholes, put-call parity, the Gordon growth model, purchasing power parity, the Sharpe ratio, value at risk, bond duration, NPV and IRR.</li><li>A guide to start on Kaggle without getting lost.</li><li>For lecturers: ready-made Python and SQL labs, assignment links with signed submission codes, and the quizzes as Moodle files.</li><li>Completion badges you can add to LinkedIn. Free. No paid certificate.</li></ul><h3>The numbers after two months</h3><p>I prefer real numbers to big promises. In October:</p><ul><li>about 1,200 visitors, up from 450 in September;</li><li>150 students finished at least one lesson;</li><li>visitors came from Google, Threads and, surprise, ChatGPT, almost as many as from Google.</li></ul><p>And the less flattering one: only a third of the people who open lesson 1 finish it. So this week I made the course download Python while you are still on the home page: when you arrive from there, lesson 1 is ready in about 3 seconds instead of 9. A beginner who waits 9 seconds for a blinking “Loading” does not wait. Lesson learned.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wV6h8cw9VJNLadWaDUI2GQ.jpeg" /><figcaption>Photo by Towfiqu barbhuiya on Pexels</figcaption></figure><h3>What I want next</h3><p>More lessons written from questions students really ask me. More lecturers using it in class (if you teach finance or economics, it is free for your students, and I am happy to hear what is missing). And one honest metric to watch: how many people finish lesson 1.</p><p>If you study finance and you always postponed Python, try the first lesson. It takes two minutes: <a href="https://ecofinlearning.com/">ecofinlearning.com</a>.</p><p>And tell me: what was the hardest part when YOU started coding?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ea19b64e247" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The $6.3 Billion Job Interview: Every Step of the Binance Heist Looked Normal]]></title>
            <link>https://medium.com/@sitrakaforler/a-job-interview-cost-binance-6-3-billion-here-is-how-the-attack-worked-59a8f6f4d449?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/59a8f6f4d449</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 19:33:59 GMT</pubDate>
            <atom:updated>2026-11-06T21:32:25.068Z</atom:updated>
            <content:encoded><![CDATA[<h4>What the incident report teaches security and platform teams about workstations, tokens, toolchains and internal AI tools.</h4><p>In operations, we like to believe that a major breach starts with a major hole: an unpatched firewall, an exposed database, a zero-day.</p><p>On 3 November 2026, Binance published the preliminary incident report of the largest crypto theft on record. There is no major hole in it. There is a fake job interview, a developer tool that quietly changed, and a chain of six events that, taken one by one, look like an ordinary working day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S0CNmen-RjbvchmrsL-8RQ.jpeg" /><figcaption>The intrusion started with a fake job interview. Photo by <a href="https://www.pexels.com/photo/people-doing-online-job-interview-8636626/">Kampus Production on Pexels</a></figcaption></figure><h3>The facts</h3><ul><li><strong>About $6.34 billion</strong> in digital assets affected on 29 October 2026, according to Binance.</li><li><strong>About $2.85 billion</strong> frozen or intercepted as of 3 November; 98% of affected assets replenished by Binance itself.</li><li><strong>Attribution:</strong> on 31 October, the FBI named TraderTraitor, a North Korean group also tracked as Lazarus. Binance has not formally confirmed it, and Pyongyang denies it.</li></ul><h3>Six ordinary events</h3><ol><li><strong>June. An install.</strong> An engineer on the wallet infrastructure team installs a malicious Homebrew package. It is caught: laptop replaced, credentials rotated. But the team and its members are now known to the attacker.</li><li><strong>September. A job interview.</strong> Another engineer is approached on a legitimate job platform by the “project director” of a company that does not exist. A video call with a friendly “HR representative”, then a conversation with the “CEO” on a chat app, then a coding assignment to set up.</li><li><strong>A developer tool update.</strong> The setup script replaces pyenv, the Python version manager installed on countless Macs, with a malicious copy. The custom malware is not detected by endpoint protection.</li><li><strong>20 October. An assistant at work.</strong> Using the compromised machine and an internal AI tool, the attackers reach two more people on the wallet team.</li><li><strong>A token on a laptop.</strong> On one of those machines sits a container deployment permission token.</li><li><strong>29 October. A valid deployment.</strong> The token deploys a container, the container reaches the wallet signing service, and funds leave across several blockchains. On Ethereum, Blockaid describes a single transaction carrying 50 EIP-7702 authorizations for the “Binance 14” hot wallet; one matched, and the wallet was redirected to the attacker’s contract.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rcKd6nT_bISoI-f64uZ2Bg.png" /><figcaption>Figure 1. The attack chain, from June to 29 October 2026 (events 4 and 5 shown together). Diagram by the author, based on Binance’s incident report and Blockaid’s analysis.</figcaption></figure><blockquote><em>Each control looked at its own step and saw something normal. The attack lived between the steps.</em></blockquote><h3>Five lessons for security and platform teams</h3><h4>1. A developer workstation is part of production</h4><p>We classify servers, databases and networks by criticality. Laptops are usually one asset class, managed the same way for everyone. Yet the laptop of a wallet DevOps engineer held a token that could reach the signing service. The workstations of people with production access deserve the same tiering, monitoring and isolation as the systems they can reach.</p><h4>2. No standing deployment tokens</h4><p>A long-lived token stored on a machine is a key under the doormat. Deployment rights to sensitive environments should be short-lived, scoped to one task, issued on demand and tied to a strong, recent authentication. Had the token in event 5 expired an hour after its last use, the chain would have broken there.</p><h4>3. Internal AI tools are identities</h4><p>The report does not detail how the internal AI tool was abused. It does not need to. An assistant that can read code, run commands or reach other machines is an account with permissions, and it served as a bridge from one engineer to two others. It needs least privilege, its own audit trail, and no reach across colleagues’ workstations unless that is its explicit job.</p><h4>4. Watch the toolchain, not only the applications</h4><p>pyenv is a perfect hiding place: it is supposed to be there, it runs constantly, and it touches the shell. So do nvm, rbenv, asdf and every shim in a package manager’s path. File integrity monitoring and allow-listing rarely cover these directories. On privileged machines, they should.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IarvhJIx0TEz-YGM2wHUDg.jpeg" /><figcaption>Version managers run all day and touch the shell: a perfect hiding place. Photo by <a href="https://www.pexels.com/photo/computer-coding-on-a-computer-screen-6424589/">Nemuel Sereti on Pexels</a></figcaption></figure><h4>5. Hiring channels are now an attack surface</h4><p>Fake recruiters who send “take-home” code to engineers are a documented North Korean technique, known in the industry as Contagious Interview. In June, Binance’s response was textbook, but it could not undo what the attacker had learned about who works on the wallet team. When a high-value team has been mapped, the follow-up should include telling those engineers plainly that they are now likely targets of very personal, very patient approaches, and giving them a sandbox where any external code can be opened safely.</p><h3>Beyond crypto</h3><p>Nothing in this chain is specific to blockchains except the last step. Replace “wallet signing service” with “payment gateway”, “core banking deployment” or “customer database”, and the first five events are exactly the same. Every organisation that lets engineers deploy, and lets internal assistants act on their behalf, has these six ordinary events somewhere in its week.</p><p>The lesson is not that one control failed. It is that no control was looking at the chain.</p><p><em>Sources: Binance preliminary incident report (3 November 2026), as reported by </em><a href="https://www.bleepingcomputer.com/news/security/binance-says-fake-job-interview-led-to-63-billion-crypto-heist/"><em>BleepingComputer</em></a><em> and </em><a href="https://www.theregister.com/security/2026/11/03/north-korean-fake-job-interview-led-to-63b-binance-heist/5307398"><em>The Register</em></a><em>; FBI attribution of 31 October (</em><a href="https://www.bleepingcomputer.com/news/security/fbi-links-63-billion-binance-crypto-heist-to-north-korean-hackers/"><em>BleepingComputer</em></a><em>); on-chain analysis by Blockaid (</em><a href="https://crypto.news/binance-hack-eip-7702-wallet-attack/"><em>crypto.news</em></a><em>). Figures as of 6 November 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=59a8f6f4d449" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[n8n or OpenClaw??? For Most AI Agents, the Answer Is Both]]></title>
            <link>https://medium.com/@sitrakaforler/n8n-or-openclaw-for-most-ai-agents-the-answer-is-both-dc665442c171?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/dc665442c171</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[n8n]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 18:13:05 GMT</pubDate>
            <atom:updated>2026-11-15T21:46:35.641Z</atom:updated>
            <content:encoded><![CDATA[<h4>A simple guide to agent architecture, and what the latest updates of both tools change…</h4><p>Every week someone asks me which tool to use to build an AI agent. Lately the question comes with two names: n8n or OpenClaw?</p><p>Both shipped big updates this autumn. n8n launched a new Agents builder in September and released n8n 3.0 at the end of October. OpenClaw, the open source assistant that became one of the fastest-growing projects on GitHub early this year, now ships several releases a month.</p><p>My short answer ?</p><p>For most real use cases, you want <strong>both</strong>, each doing the job it is good at. Here is why, in plain terms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3s_4j9ZH6JhIKXc8nrgDFg.jpeg" /><figcaption>Most agents start where people already are: in a chat. Photo: <a href="https://www.pexels.com/photo/young-professional-using-smartphone-at-work-desk-36713184/">Vitaly Gariev / Pexels</a>.</figcaption></figure><h3>What an agent actually is</h3><p>Strip away the marketing and an agent has four parts.</p><ul><li><strong>A brain:</strong> the language model that reads the request and decides what to do next.</li><li><strong>Hands:</strong> the tools it can call, like sending an email, reading a ticket or updating a CRM.</li><li><strong>Memory:</strong> what it remembers within a conversation, and sometimes across conversations.</li><li><strong>Rules:</strong> what it is allowed to do alone, and what needs a human.</li></ul><p>The model runs in a loop. It reads, picks a tool, looks at the result, and decides again until the job is done.</p><p>Most failures I see in production are not brain failures. They are hands and rules failures: an agent holding too many keys, or nobody checking before it writes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2CrgMCjg71Vx9yByx0bxKg.png" /><figcaption>The four parts of any AI agent. Diagram by the author.</figcaption></figure><h3>n8n: workflows first, now with real agents</h3><p>n8n started as a workflow automation tool. You draw the steps, and it runs them the same way every time. That predictability is its strength.</p><p>On 25 September 2026, n8n <a href="https://blog.n8n.io/introducing-n8n-agents/">introduced Agents</a>, a new type of object that sits next to your workflows. You give an agent a goal, a model, instructions and a set of tools, and it works out the steps itself. You can reach the same agent from Slack, Telegram or a schedule.</p><p>Three details matter for architecture:</p><ul><li><strong>Workflows can be tools.</strong> The agent decides when to act. The workflow decides exactly how.</li><li><strong>Each tool carries its own credential,</strong> so the agent never holds your keys.</li><li><strong>Sensitive tools wait for a person.</strong> They pause until someone approves or rejects the action, and every session is logged step by step.</li></ul><p>It also works the other way. A normal workflow can hand one step to an agent with the Message an Agent node, then carry on with the answer.</p><p>The <a href="https://docs.n8n.io/release-notes/">release notes</a> show the same focus on control: spend caps for agents, admin policies that can block some models or tools, and MCP connections in the assistant that need explicit approval. With n8n 3.0 on 27 October 2026, the n8n Assistant, which builds workflows from plain language, left preview.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*osj2JR0IWq4_-vrTzeAFqg.jpeg" /><figcaption>A workflow is a plan you draw once. An agent chooses its own steps. Photo: <a href="https://www.pexels.com/photo/colleagues-planning-a-business-5257576/">Thirdman / Pexels</a>.</figcaption></figure><h3>OpenClaw: an assistant that lives where you chat</h3><p>OpenClaw comes from the other side. Its <a href="https://docs.openclaw.ai/">documentation</a> describes it as a self-hosted gateway that connects chat apps such as WhatsApp, Telegram or Discord to an AI agent running on your own machine or server.</p><p>The gateway is the switchboard: it holds the sessions, routes the messages and keeps the channel connections. The agent behind it uses skills, which are written procedures it loads only when it needs them. It can also run scheduled tasks, react to webhooks, and use a paired phone for camera and voice.</p><p>That makes it very good at conversation and judgment. You message it like a colleague, and it works out what you mean.</p><p>The project moves fast. Its creator, Peter Steinberger, joined OpenAI in February 2026, and Sam Altman said OpenClaw would <a href="https://www.tomshardware.com/tech-industry/openai-hires-genius-openclaw-creator-but-popular-ai-assistant-will-remain-open-source-sam-altman-says-creator-will-work-on-smart-agents-in-new-role">live on in a foundation as an open source project</a>. Releases have kept coming: version 2026.10.10 in October, then 2026.11.1 this month, and <a href="https://docs.openclaw.ai/releases">several recent ones</a> include security fixes.</p><p>Security is the part to take seriously. In February, researchers reported <a href="https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html">341 malicious skills out of 2,857 audited</a> on ClawHub, OpenClaw’s skill marketplace. ClawHub then added <a href="https://techinformed.com/openclaw-adds-virustotal-scanning-for-clawhub-skills/">VirusTotal scanning of skills</a>. That helps, but a skill is still code and instructions that your agent will trust.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XK71hDxlWHL_Zzk2Li4U_Q.jpeg" /><figcaption>A skill is code your agent will trust. Install it like any other software. Photo: <a href="https://www.pexels.com/photo/close-up-of-hand-holding-a-secure-padlock-36740854/">Nathan Thomas / Pexels</a>.</figcaption></figure><h3>The architecture that works: OpenClaw talks, n8n acts</h3><p>Put the two together and each covers the other’s weak spot.</p><ol><li>You send a message in your chat app. The OpenClaw gateway receives it.</li><li>The agent understands the request and decides which action is needed.</li><li>Instead of calling an external API itself, it calls an n8n webhook with a few narrow inputs.</li><li>The n8n workflow holds the credential, asks a human to approve if the action is sensitive, runs it, and logs everything.</li><li>The result goes back to the agent, which answers you in the chat.</li></ol><p>The agent never sees an API key. Every action it can take is a workflow you can open, read and lock. When something goes wrong, the log tells you which step did what.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*paVcpOYcHWD6pw42Jey0PA.png" /><figcaption>OpenClaw talks and decides. n8n acts, with its own keys and a human in the loop. Diagram by the author.</figcaption></figure><p>It is the same principle I used in my own open source <a href="https://github.com/Sitraka17/servicedesk-agent">service desk agent</a>: the model proposes, the code checks, and every write waits for a human. Different tools, same idea.</p><h3>Seven simple rules</h3><ol><li>Let the model decide when to act, and let the workflow decide how.</li><li>Keep keys out of the agent. One credential per tool, stored in n8n.</li><li>Put an approval in front of anything that writes, pays or deletes.</li><li>Treat skills like software you install: read them, pin a version, prefer what you can audit.</li><li>Log every run and cap the spend before you go live.</li><li>Treat the webhook URL as a secret: turn on Header Auth, one secret per workflow, and validate every input inside the workflow.</li><li>Bind each approval to the exact parameters it approved, and show the approver those parameters, not the agent’s summary.</li></ol><h3>Which one should you start with?</h3><ul><li><strong>Your process is known and repeatable</strong> (invoices, onboarding, weekly reports): start with n8n workflows. Add an n8n Agent only where a step needs judgment.</li><li><strong>You want a personal assistant you talk to all day:</strong> start with OpenClaw, on a machine you control, with very few skills.</li><li><strong>You want an assistant that also acts on business systems:</strong> use both. OpenClaw in front, n8n behind it.</li></ul><p>The tools will keep changing. n8n shipped a major version a few days ago, and OpenClaw releases almost every week. The architecture question stays the same: who decides, who acts, and who checks.</p><p><em>Thanks for reading!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dc665442c171" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your SLAs Are Green. Your Policyholders Are Still Waiting.]]></title>
            <link>https://medium.com/@sitrakaforler/your-slas-are-green-your-policyholders-are-still-waiting-872cd26ce647?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/872cd26ce647</guid>
            <category><![CDATA[data]]></category>
            <category><![CDATA[insurance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[observability]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 11:01:01 GMT</pubDate>
            <atom:updated>2026-11-02T11:01:01.777Z</atom:updated>
            <content:encoded><![CDATA[<h4>Seven practical lessons for insurers who want service level management to steer IT, not just report on it.</h4><p>A summer storm has just passed. Hail the size of marbles. By nine in the evening, a few hundred people are standing in their driveways, phone in hand, photographing dents on their car roof.</p><p>They open their insurer’s app to declare the damage. The form takes a while to load. The first photo uploads. The second one spins, then fails. They try again. Some manage. Some give up and tell themselves they will call on Monday.</p><p>On Monday, the phone lines are saturated. And at the monthly service review a few weeks later, the slide says this:</p><p><em>Portal availability: 99.97%. Target met.</em></p><p>Nobody is lying. The portal was up. It was just slow and fragile at the exact moment people needed it.</p><p>In IT operations, we have a name for this: the watermelon effect. Green on the outside, red on the inside.</p><figure><img alt="A person photographing a car with a smartphone." src="https://cdn-images-1.medium.com/max/1024/1*GYM-WcZKPiKhG-RHkM1vuA.jpeg" /><figcaption>Illustrative photo. After a storm, the claim usually starts with a phone. Photo: <a href="https://www.pexels.com/photo/man-taking-pictures-near-car-18213067/">Mario Amé / Pexels</a>.</figcaption></figure><p>Insurers are especially exposed to it, for a simple reason. The moments that matter most to their customers are rare, emotional and bunched together. Nobody has a car accident “on average”. A storm sends everyone to the same form on the same evening.</p><p>I work on monitoring and AIOps, which means using data and machine learning to run IT operations. Here is what I would tell an insurer that wants its service level management to become a real steering tool, and not just a monthly slide.</p><h3>1. Start with the person holding the phone</h3><p>Most indicators are inherited from the infrastructure. Server availability. CPU load. API response time. They are useful. But nobody ever bought an insurance policy because of a CPU.</p><p>Your customers live through a handful of moments: getting a quote, signing, declaring a claim, following it, getting paid. Your agents have their own: issuing a contract during an appointment, finding a client’s file while the client is sitting in front of them.</p><p>Write those moments down. Then, for each one, list everything it depends on: applications, interfaces, suppliers, batch jobs.</p><p>Now your objectives can sound like something a human would care about. For example:</p><p><em>95% of claim declarations sent from the app are confirmed in under 10 seconds, over any 28-day period.</em></p><p>Your CEO understands that sentence. So does the engineer on call. That is the test.</p><h3>2. One word, one meaning</h3><p>Sit in enough service reviews and you will hear SLA used for everything. It is not a detail. When words are fuzzy, meetings drift.</p><p>Here is the short version I use:</p><ul><li><strong>SLO</strong> (service level objective): the target you set for yourself.</li><li><strong>SLA</strong> (service level agreement): the promise you make to someone else, usually with consequences.</li><li><strong>OLA</strong> (operational level agreement): the promise internal teams make to each other so the SLA can be kept.</li><li><strong>KPI</strong> (key performance indicator): how well something is working.</li><li><strong>KRI</strong> (key risk indicator): the early warning that something is about to go wrong.</li></ul><p>Then give every indicator an identity card: owner, data source, formula, measurement window, threshold.</p><p>If two dashboards show two different availability figures for the same service, the next committee will spend twenty minutes arguing about the dashboards. Not about the service.</p><h3>3. Your ticket data is production data</h3><p>Every organisation wants predictive models. Few look closely at what they would be trained on.</p><p>Open your incident tool and look honestly. How many tickets are closed with the category “Other”? How many changes were recorded after they happened? How many configuration items in the CMDB point to servers that were switched off two years ago?</p><p>That is not a technology problem. It is a data quality problem. And a model can only learn from what you actually recorded.</p><p>Three questions I would ask before any AI project:</p><ul><li>What share of incidents is linked to the configuration item that failed?</li><li>What share is linked to the change that caused it, when there was one?</li><li>What share of major incidents has a documented root cause?</li></ul><p>Boring questions. They decide whether everything that follows will work.</p><h3>4. Don’t just read your suppliers’ reports. Measure them.</h3><p>A large part of an insurer’s IT now runs at someone else’s: cloud platforms, software vendors, outsourced service desks, document and payment providers.</p><p>The classic setup is a PDF at the end of the month. Written by the supplier. Measured with the supplier’s tools. Against the supplier’s definitions. Good luck challenging a number you cannot reproduce.</p><p>European regulation is pushing the same way. Under DORA, the Digital Operational Resilience Act, which has applied to insurers since January 2025, contracts for ICT services that support critical or important functions must describe service levels with precise quantitative and qualitative performance targets, so that the insurer can monitor them effectively.</p><p>“Monitor effectively” is the important part. In practice, it means having your own measurement points. A synthetic transaction that logs in, uploads a test document and checks the result every few minutes, from your side. Then, at the end of the month, you compare your figures with theirs. The conversation changes completely.</p><h3>5. Prepare the first hour of a crisis while things are calm</h3><p>When a serious incident hits, engineers focus on bringing the service back. That is their job. But at the same moment, management needs answers. How many customers are affected? Which services? Since when? Is any data lost or exposed?</p><p>Under DORA, an incident classified as major must be notified to the supervisor within four hours of that classification, and no later than 24 hours after the insurer became aware of it. The criteria used to classify it include the number of clients and transactions affected, the duration, the geographical spread, data losses, the criticality of the services and the economic impact.</p><p>Look at that list again. It is exactly what a good service level dashboard should already know.</p><p>If you maintain, in quiet times, the link between each technical component and the customer journeys and volumes it supports, the first estimate of impact takes ten minutes. Without it, it takes a morning of phone calls, and three people with three different numbers.</p><p>The same idea fixes alerting. Compare these two alerts:</p><blockquote><em>CPU above 90% on server APP-PRD-07.</em></blockquote><blockquote><em>Claim declarations in the app: 18% fail since 20:40. Usual rate: under 1%.</em></blockquote><p>The first one is a fact. The second one tells you why you should get out of bed.</p><figure><img alt="An operator at a desk in a control room facing a wall of monitoring screens." src="https://cdn-images-1.medium.com/max/1024/1*z5ULyp2X-yWGUdD2YQV0dw.jpeg" /><figcaption>Illustrative photo of a control room in Lima. Lots of screens, one question: is the customer getting through? Photo: <a href="https://www.pexels.com/photo/advanced-control-room-in-el-agustino-lima-32529341/">Fernando Narvaez / Pexels</a>.</figcaption></figure><h3>6. Use AI where it saves time, and make it show its work</h3><p>AI is genuinely useful in operations. But each use case should earn its place with a number.</p><p><strong>Anomaly detection</strong> only works if it knows your calendar. A wave of claim declarations the morning after a storm is not an anomaly: it is your business doing its job. The same wave on a quiet Tuesday in February probably is. Models need to learn Monday mornings, month-end batches, renewal season, and the days after bad weather.</p><p><strong>Event correlation</strong> turns hundreds of alerts into a handful of probable incidents. Your on-call engineers will thank you, because alert fatigue is real and it makes people miss the one alert that mattered.</p><p><strong>Automatic summaries</strong> can draft the monthly service review: what degraded, why, what was done, what is still open. That saves hours, on one condition: every sentence must link to the tickets and data behind it. A summary nobody can check is a liability.</p><p><strong>Questions in plain language</strong> are the most tempting. “Which services missed their objective last quarter, and how many of those incidents followed a change?” The guardrails make the difference: read-only access, answers computed on the official indicator catalogue rather than on raw tables, and the query shown next to the answer so anyone can verify it.</p><p>Then measure what changed: time to detect, time to restore, alerts per incident, hours spent preparing the review. If none of these move after six months, you have a demo, not a tool.</p><p>One boundary is worth drawing clearly. AI that helps run IT is not the same thing as AI that takes decisions about customers. The European AI Act classifies AI used for risk assessment and pricing in life and health insurance as high risk. Keep the two worlds clearly apart in your governance. Both become easier to manage.</p><h3>7. Make the SLA committee decide something</h3><p>Many SLA committees go like this: we look at last month, targets are met, see you next month. That is a report read aloud.</p><figure><img alt="A small team in a meeting room looking at a dashboard on a wall screen." src="https://cdn-images-1.medium.com/max/1024/1*J43jRqfwPhO7Wv9nUHk_tQ.jpeg" /><figcaption>Illustrative photo. A good service review ends with a decision, not just a chart. Photo: <a href="https://www.pexels.com/photo/a-woman-doing-a-presentation-at-a-meeting-5716030/">Artem Podrez / Pexels</a>.</figcaption></figure><p>A steering committee asks harder questions:</p><ul><li>Is this target still the right one?</li><li>What would it cost to go from 99.9% to 99.95% on this service, and would a single customer notice?</li><li>Which service is over-delivering, year after year, so that budget could go elsewhere?</li><li>Which supplier needs a recovery plan, with a date?</li></ul><p>An error budget helps a lot. Take an objective of 99.9% over 30 days. That leaves about 43 minutes of downtime for the month. While the budget is healthy, teams can ship changes with confidence. When it is almost spent, stability comes first. Business and IT now discuss risk with one shared number instead of two opposite feelings.</p><h3>It all comes back to trust</h3><p>Insurance is a promise kept on a bad day. Most people never think about their insurer, until the hail, the accident or the hospital. On that day, the app and the phone line are the promise.</p><p>Service level management is where that promise meets the reality of IT. Done well, it warns you before your customers do. It tells suppliers precisely what you expect. It gives your supervisor clear answers in the first hours of a crisis. And it lets business and IT decide together where the next euro of reliability should go.</p><p>A green dashboard is nice. A customer whose claim went through on a stormy Sunday night, on the first try, is better.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=872cd26ce647" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Taxi Driver Built an App on His Phone. That’s the Real Supabase Story.]]></title>
            <link>https://medium.com/@sitrakaforler/a-taxi-driver-built-an-app-on-his-phone-thats-the-real-supabase-story-ef85e94b7dae?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef85e94b7dae</guid>
            <category><![CDATA[no-code]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[supabase]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 22:01:34 GMT</pubDate>
            <atom:updated>2026-11-01T22:01:34.589Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI agents are creating databases by the millions. The bigger change is who now gets to build the software on top.</h4><p>A taxi driver in Brazil built an Uber-like application from his phone.</p><p>That’s one of the examples Supabase co-founder Paul Copplestone shares when explaining what is happening on his platform.</p><p>The obvious questions are technical. Which model wrote the code? How many prompts did it take? How well did the application actually work?</p><p>Fair questions. But I keep coming back to something else.</p><p>The driver didn’t need anyone to explain the problem to him. He was already living it. What he lacked was a way to turn that knowledge into software.</p><p>For years, crossing that gap meant finding a developer, paying an agency, persuading a technical co-founder — or putting the idea aside.</p><p>Now there is another starting point.</p><p>Supabase’s growth suggests a lot of people are taking it.</p><figure><img alt="A person using a smartphone inside a parked car." src="https://cdn-images-1.medium.com/max/1024/1*CaDipj2jhTZ4FyZZYna4Cw.jpeg" /><figcaption>Understanding a problem and knowing how to code are no longer quite the same barrier. Illustrative photograph—not the driver in Copplestone’s account. Photo: <a href="https://www.pexels.com/photo/man-sitting-in-a-car-and-using-phone-12592624/">Oleksandr Chepys / Pexels</a>.</figcaption></figure><h3>The numbers are impressive. The people behind them are more interesting.</h3><p>According to Copplestone’s account, Supabase grew from roughly 3 million to 13 million users in a year. Over the same period, its cumulative database count climbed from around 8 million to nearly 50 million.</p><p>Those numbers describe different things.</p><p>A user is not a production application. A database is not necessarily a business. Someone experimenting with an agent might create several databases before deciding that none of the projects is worth keeping.</p><p>Still, the scale tells us something is changing.</p><p>The number that caught my attention is a different one: AI tools and agents <a href="https://www.stackone.com/blog/database-per-agent/">reportedly account for</a> <a href="https://finance.yahoo.com/technology/ai/articles/supabase-raises-150m-buys-turso-083958845.html">approximately 70% of new database creation on Supabase</a>.</p><p>Not 70% of revenue. Not 70% of users. Not 70% of all platform activity.</p><p>New databases.</p><p>That distinction matters because it identifies where agents are entering the process. They are not simply helping people write a function or explain a SQL query. They are increasingly involved in setting up the infrastructure that an application needs.</p><p>The person asks for a product. The agent starts assembling the pieces.</p><p>For a developer platform, that changes a surprisingly basic question: who is actually using the tool?</p><h3>Your next user might never open the dashboard</h3><p>Copplestone describes Supabase’s evolution as dashboard-first, then API-first, and increasingly code-first.</p><p>That sounds like an interface decision. It is also a change in how the product reaches the person who needs it.</p><p>A human working through a dashboard needs understandable navigation, clear labels, and controls that make sense.</p><p>An agent needs operations it can execute, responses it can interpret, and enough context to recover when something goes wrong.</p><p>The human still wants a working application. But the agent may be the one configuring the database, changing the schema, connecting authentication, and investigating the failed request.</p><p>In that workflow, the dashboard is no longer the center of every interaction.</p><p>This helps explain Copplestone’s emphasis on open source. In his view, access to the underlying code gives agents more material to inspect when documentation alone does not explain a problem.</p><p>That is not a magic solution. A confusing interface remains confusing. Access to an implementation does not guarantee that a model will understand it correctly.</p><p>But it gives the agent another place to look.</p><p>Instead of stopping at “the documentation doesn’t explain this,” it can investigate how the system actually behaves.</p><p>The practical question for infrastructure companies becomes: how far can an automated user get before a human has to step in?</p><h3>The humble error message gets a promotion</h3><p>My favorite detail in Copplestone’s account has nothing to do with user growth or funding.</p><p>It is about error messages.</p><p>Supabase is investing in errors that help agents understand what went wrong and how to investigate it.</p><p>Consider these two hypothetical responses:</p><blockquote><em>Permission denied.</em></blockquote><p>And:</p><blockquote><em>The active role cannot modify this schema. Check the role’s grants before retrying.</em></blockquote><p>Neither fixes the problem. But the second gives the investigation somewhere useful to begin.</p><p>Working with infrastructure makes you appreciate that difference. An alert that announces failure and an alert that helps explain failure are not the same product.</p><p>With an agent, that difference can determine whether the workflow continues or gets stuck.</p><p>An ambiguous error may lead to repeated attempts or unrelated changes. An actionable error narrows the next step.</p><p>Copplestone also describes internal evaluations in which different models test Supabase integrations. The objective is not merely to confirm that a capability exists. It is to find out whether an agent can use it successfully.</p><p>That is a meaningful shift in product testing.</p><p>A feature can work exactly as designed and still fail the person relying on an agent to operate it.</p><p>Can the agent discover the right operation? Supply the right inputs? Recognize that something failed? Recover without making the situation worse?</p><p>The endpoint is only one part of the experience.</p><p>The real product is whether the whole interaction succeeds.</p><h3>The new builder already knows something valuable</h3><p>“Anyone can build software” is an attractive slogan. It is also easy to overstate.</p><p>A prototype is not a reliable service. Generated code is not a business. A convincing demonstration tells you little about what happens when real users arrive with messy data and unexpected requests.</p><p>But dismissing the entire shift as hype misses the people it affects.</p><p>Think about someone who has spent ten years managing a particular workflow.</p><p>They know which exceptions happen every Friday. They know why the official process gets bypassed. They know which spreadsheet the organization quietly depends on.</p><p>That knowledge rarely arrives in a clean requirements document.</p><p>It lives in experience. In workarounds. In the sentence: “Yes, but that isn’t how it works here.”</p><p>A model does not automatically possess that understanding. The person using it may.</p><p>Until recently, understanding the problem and being able to implement a solution often lived in different people. Bringing them together required money, persuasion, and time.</p><p>AI-assisted development shortens that distance.</p><p>That is why the taxi-driver example matters. The interesting claim is not that one person has recreated Uber as a global business.</p><p>It is that someone with direct experience of a problem can now attempt a solution without first recruiting an engineering team.</p><p>Some attempts will go nowhere. Some will remain useful personal tools. A few may become businesses.</p><p>All three outcomes are different from never getting started.</p><figure><img alt="A workspace with a laptop, sticky notes, and a notepad." src="https://cdn-images-1.medium.com/max/1024/1*wCF7ok3mpAPxf4gf8Um24A.jpeg" /><figcaption>The first useful specification may already exist in someone’s working notes. Photo: <a href="https://www.pexels.com/photo/sticky-notes-and-a-laptop-7278606/">Anastasia Shuraeva / Pexels</a>.</figcaption></figure><h3>What if the database becomes your personal memory?</h3><p>Copplestone’s own experiment takes this idea beyond work.</p><p>He describes building a personal “control tower”: email, finances, health information, and uploaded files brought together in a single Supabase database.</p><p>The dashboard is the visible part. The more interesting question sits underneath it.</p><p>If an agent helps you today, what should the next agent know tomorrow?</p><p>In Copplestone’s vision, the database provides continuity. Agents query it, add information, and use that record to support future tasks.</p><p>The model can change. The interface can change. Your information does not have to start over with every conversation.</p><p>That is an appealing idea. It is also where I would slow down.</p><p>Receipts and health records do not belong behind identical access assumptions simply because they fit in the same database.</p><p>An agent that can read information does not necessarily need permission to modify it. An agent that can recommend an action should not automatically be allowed to execute it.</p><p>Before trusting a personal control tower, I would want clear answers:</p><ul><li>What can each agent access?</li><li>Which actions require approval?</li><li>Where can I see what changed?</li><li>How do I undo a mistake?</li></ul><p>These are not objections to the concept. They are part of making it useful.</p><p>The more responsibility we give an agent, the less acceptable it becomes to treat permissions and recovery as details to add later.</p><h3>The Turso deal fits a world with many more databases</h3><p>Supabase <a href="https://supabase.com/blog/supabase-is-acquiring-turso">announced that it was acquiring Turso</a> on October 2, 2026, connecting the move to a future in which agents create enormous amounts of software.</p><p>The architectural challenge is not simply to make a large database larger.</p><p>A business application might need a long-lived database with steady demand. A personal agent, isolated task, or temporary experiment might need something much smaller and inactive for much of its life.</p><p>Those are different workloads.</p><p><a href="https://finance.yahoo.com/technology/ai/articles/supabase-raises-150m-buys-turso-083958845.html">Reporting on the deal</a> describes Turso’s approach as <a href="https://8digits.net/news/supabase-is-buying-turso-so-agents-can-have-their-own-databases-nothing-to-migrate/">complementary to Supabase’s Postgres platform</a>, addressing demand for large numbers of smaller databases rather than replacing the existing offering.</p><p>In the supplied interview account, Copplestone connects this opportunity to a product tier called Spark.</p><p>The ambition is to make creating many small databases practical — not just to increase the capacity of a few large ones.</p><p>But this is also where the growth story needs discipline.</p><p>A new database might support a paying customer’s daily operations. It might also belong to a five-minute experiment nobody opens again.</p><p>Creation volume tells us that people are trying things.</p><p>It does not, on its own, tell us which things last.</p><h3>Gbrain shows where the engineering work goes next</h3><p>The Gbrain example in the interview account points to another challenge: giving an agent the right context.</p><p>Built by Gary at YC, the system combines vector retrieval through pgvector with graph-based relationships.</p><p>The distinction is useful. Vector retrieval looks for information with related meaning. Graph-based retrieval follows explicit connections between entities.</p><p>But the most interesting detail is the feedback loop.</p><p>As described, an agent experiments with changes, runs retrieval evaluations, and promotes improvements when they meet the chosen criteria.</p><p>It is tempting to call that “self-improving” and move on.</p><p>I think the more useful question is: improving against what?</p><p>Better for which questions? Better at finding something related — or better at finding what the user actually needed?</p><p>An evaluation gives the system a target. Choosing that target is still consequential work.</p><p>If the test rewards the wrong behavior, running it more often will not solve the underlying problem.</p><p>Some of the engineering effort therefore moves from manually making every change to designing evaluations worth trusting.</p><p>The agent can run the experiments. Someone still has to decide what counts as progress.</p><h3>You may not need a technical co-founder to begin</h3><p>Copplestone connects the builder movement to a changing discussion about technical co-founders.</p><p>The important word is “begin.”</p><p>Technical expertise still matters. Reliability, security, and maintainability do not disappear because the first version arrived quickly.</p><p>But the order of events can change.</p><p>Previously, a domain expert might approach an engineer with an idea and ask for months of commitment before either person had much evidence.</p><p>Now, that person may be able to arrive with a prototype and real feedback.</p><p>People tried this.</p><p>They returned to this feature.</p><p>They ignored that one.</p><p>Here is where the workflow failed.</p><p>Here is what they asked for next.</p><p>That is a different conversation from “I have an idea for an app.”</p><p>Engineering expertise can enter the project with fewer assumptions and more concrete problems to solve.</p><p>The question becomes less “Can we build this?” and more “Have we found something worth building properly?”</p><h3>A useful business does not have to become a giant</h3><p>Copplestone imagines vertical software companies reaching $10 million to $30 million in revenue with teams of roughly 10 to 15 people.</p><p>Those figures describe possibilities he sees — not typical outcomes or forecasts to take for granted.</p><p>Still, the underlying model is worth considering.</p><p>A small team with deep knowledge of a narrow market may not need to serve everyone. It may need to solve one expensive, recurring problem well enough that customers are happy to pay.</p><p>AI helps with building. It does not automatically find customers, earn their trust, or persuade them to change how they work.</p><p>But reducing the cost of testing an idea can matter enormously.</p><p>For someone who previously could not afford a development team, a cheaper experiment is not just a productivity improvement. It may be the difference between trying and never starting.</p><p>A useful tool does not have to become a global platform to justify its existence.</p><p>Sometimes it only needs to make a particular group of people’s work noticeably better.</p><h3>The new builder starts with the problem</h3><p>That is what makes the taxi-driver story worth remembering.</p><p>Not the suggestion that one person has recreated a global transport company from a phone. Not the promise that everyone is about to become a software engineer.</p><p>Something more grounded: a person who understood a problem finally had a way to attempt a solution.</p><p>The next builder may not start with coding skills. They may start with a process that wastes their time, a service that keeps disappointing them, or a problem they have spent years working around.</p><p>You know the problem. AI is making it easier to build the solution.</p><p>It does not guarantee you will get it right.</p><p>But it gives you a way to begin.</p><p><em>Source note: This article draws on the supplied summary of Paul Copplestone’s remarks. Interview-specific anecdotes, personal projects, Spark details, Gbrain’s workflow, and business projections are attributed to that account. Public reporting supports </em><a href="https://finance.yahoo.com/technology/ai/articles/supabase-raises-150m-buys-turso-083958845.html"><em>the broader database-creation figures</em></a><em>, </em><a href="https://www.stackone.com/blog/database-per-agent/"><em>AI’s reported share of new database creation</em></a><em>, and </em><a href="https://supabase.com/blog/supabase-is-acquiring-turso"><em>the Turso announcement</em></a><em>. Database creation should not be treated as a measure of successful applications or paying customers.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef85e94b7dae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[DHH Says Writing Code by Hand No Longer Makes Economic Sense. Is He Right?]]></title>
            <link>https://medium.com/@sitrakaforler/dhh-says-writing-code-by-hand-no-longer-makes-economic-sense-is-he-right-3a107406cc4b?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/3a107406cc4b</guid>
            <category><![CDATA[code]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[rust]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 21:35:36 GMT</pubDate>
            <atom:updated>2026-09-28T21:35:36.339Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Rails creator’s AI pivot is not really about the end of programming. It is about what happens when writing code becomes the easy part.</em></p><p>David Heinemeier Hansson built Ruby on Rails around a compelling idea: programmers should enjoy building software.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9eDKnkIP_cYy7q_a" /></figure><p>So when DHH told a room full of developers at Rails World that writing code by hand no longer makes economic sense, it landed differently than the usual “AI will replace developers” speech. This was the creator of one of the most influential frameworks in web development questioning the value of the work his framework made pleasurable.</p><p>Fireship’s video captures the shock of that moment. But the question worth asking is not whether developers will stop typing.</p><p>It is this: <strong>If AI can write the code, what is a developer actually paid to do?</strong></p><h3>DHH’s Rust reversal tells the story</h3><p>For years, DHH has been openly hostile to Rust as a language for humans to work in. Now he says HEY, 37signals’ email service, is being rebuilt with a Rust backend, with AI agents doing much of the coding. He has cited dramatic reductions in CPU and memory use for the new system. Those figures are interesting, but they are not a clean comparison between Rust and Rails. HEY’s architecture is changing too.[<a href="https://rustify.rs/articles/rails-world-2026-keynote-dhh-hey-rust-ai-agents">rustify</a>][<a href="https://news.lavx.hu/article/rails-world-keynote-leaves-rails-future-unclear">news.lavx</a>]</p><p>The reversal matters because it exposes a trade-off that used to shape almost every software project.</p><p>A team might choose Ruby because its developers could build and change a product quickly. It might avoid Rust because the performance gains were not worth the added effort of writing and maintaining it. If agents reduce that effort, the calculation changes.</p><p>But there is a catch. Someone still has to understand the system well enough to know whether the generated Rust is correct, secure and worth deploying. If nobody on the team can do that, the apparent productivity gain may simply become a different kind of risk.</p><p>That is the part of the AI coding conversation I find more interesting than the line counts.</p><h3>More code is not the same as more progress</h3><p>DHH has described a huge jump in the volume of code produced with agents. It makes for a memorable keynote. It does not tell us how much better the product has become.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/888/0*wwNbfqbwzEsQsN-L.jpeg" /><figcaption>pov a AI agent in a robot</figcaption></figure><p>An agent can create five implementations of an idea before lunch. It can also create five versions of the wrong idea.</p><p>Consider a request that sounds familiar in operations: “Build a dashboard for incidents.” An agent could produce charts, filters and an attractive interface. But it cannot settle the questions the team failed to ask. Which incidents need a human response? What counts as a false alarm? Who owns the service at night? What should happen when the monitoring data itself is unavailable?</p><p>Those questions are not obstacles to coding. <strong>They <em>are</em> the work.</strong></p><p>This is why I am not convinced by either extreme. “AI will replace all developers” ignores the judgment involved in running real systems. “AI is just autocomplete” ignores how much implementation work agents can already take on.</p><p>Both views miss the change in where a developer’s time goes.</p><h3>The skill at risk is not technical knowledge</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vkUq9uHODNWLQkMz.jpg" /></figure><p>DHH’s speech can sound like an attack on the skills developers spent years acquiring: fluency in an editor, command-line expertise, knowledge of a framework, the ability to write clean code quickly.</p><p>I would not throw those skills away. You need technical depth to recognize when an agent is confidently wrong. You need it when a migration fails, an API behaves unexpectedly or a service slows down under load.</p><p>What may lose value is doing routine implementation <em>by hand simply because that is how it has always been done</em>.</p><p>The distinction matters. A person who can write a database migration is useful. A person who can decide whether the migration is safe on a live system with millions of records is harder to replace.</p><p>The same applies to architecture. An agent can suggest a service boundary. An engineer has to live with its consequences when two teams depend on it, latency rises and an outage starts at 3 a.m.</p><h3>AI Ops has a different warning</h3><p>In my work around systems and AI Ops, I see a version of this debate that goes beyond code generation !</p><p>Imagine an agent that can inspect alerts, query logs, edit an Airflow workflow, change a monitoring configuration and restart a service. That could save an operations team hours. It could also turn a minor incident into a major one if its access is too broad or its reasoning is wrong.</p><p>The answer is not to keep AI away from operations. It is to make operations safe for AI to assist with: clear permissions, reliable interfaces, approval for high-risk actions, good audit trails and a way to reverse changes.</p><p>That is engineering work too. It is less visible than watching an agent write an entire application, but it determines whether the application can be trusted.</p><h3>I agree with DHH’s optimism, up to a point</h3><p>DHH is right to push back against the idea that developers can only lose from this shift. Agents give small teams more room to experiment. They make certain ideas cheaper to test. For a founder or an engineer with a long backlog, that is genuinely exciting.</p><p>I disagree if “pencils down” becomes an excuse to stop understanding the code or the systems we ship.</p><p>I want agents to take the repetitive work. I want to spend less time on boilerplate and more time on problems that require thought. But I also want someone who can tell me why a change is safe, what it will cost, how we will observe it and what we will do if it fails.</p><p>That person is still a developer.</p><p>The uncomfortable question DHH leaves us with is not “Should I keep learning to code?” It is: <strong>Am I learning to build and take responsibility for systems, or am I only learning to produce code?</strong></p><p>I know which skill I would bet on.</p><p><em>What part of your development work would you gladly hand to an agent, and what part would you never delegate without a human review ? :p</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a107406cc4b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Agents Without the Hype: A Practical Guide for People Who Don’t Code]]></title>
            <link>https://medium.com/@sitrakaforler/ai-agents-without-the-hype-a-practical-guide-for-people-who-dont-code-d3429008179b?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/d3429008179b</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 20:15:55 GMT</pubDate>
            <atom:updated>2026-09-24T06:58:28.865Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What agents can actually do, how to try one, and why the software around the model matters more than you might think.</em></p><p>“”””Compare these three quotes”””</p><p>A reasonable request. Then you open the attachments.</p><p>One supplier lists prices before tax. Another includes installation but says nothing about delivery. The third has put the maintenance exclusions on page seven. You can do this yourself. You would just prefer to spend your afternoon on something else.</p><p>This is the sort of job I would give an AI agent first.</p><p>I work in systems architecture and AIOps, where a convincing demonstration tends to raise a second question: what happens when something goes wrong? With agents, that question arrives surprisingly quickly. The software may misread a document, act on an assumption, or keep trying a step that cannot succeed.</p><p>Still, there is something useful here. People who would never build an integration or write a script can now delegate parts of a task in ordinary language. The opportunity is real enough to deserve a better explanation than “your autonomous digital employee.”</p><p>Let’s start with those quotes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gtue-dfQxOHOtWFE0SU-eA.png" /></figure><p><strong>What makes this an agent?</strong></p><p>A conversational assistant can explain the documents you give it. An agent can also decide what to do next: open an attachment, extract figures, calculate a total, notice that something is missing, then look for the missing information using its available tools.</p><p>The useful distinction is how much of that route is decided during the task. A fixed workflow follows predefined steps and branches. An agent uses the results of earlier steps to choose later ones. Commercial products often mix both approaches.</p><p>Suppose the comparison turns up a separate installation brochure. A useful agent might read it, find an additional charge, and update the table. If the brochure contradicts the quote, it should flag the conflict. Silently choosing whichever number looks more plausible would be a poor result.</p><p>That ability to work through messy information is where agents become interesting. If your task is simply to move every invoice into a folder, an ordinary automation may already do the job perfectly well.</p><p>You also don’t need an agent for every part of the work. A capable assistant might handle three short PDFs in one pass. Start there. You can add complexity when you have a reason.</p><p><strong>Your first experiment should be easy to check.</strong></p><p>Choose a task you understand. Comparing suppliers, preparing questions for a meeting, or gathering teaching materials are sensible examples because you can judge whether the result helps.</p><p>For the quotes, use an approved file-reading tool and upload only the documents it needs. Keep the first attempt small. There is no benefit in connecting your whole inbox before you know whether the system can distinguish a monthly fee from an annual one.</p><p>Here is a brief you can adapt:</p><blockquote><em>Compare these three quotes in a table. Include prices before and after tax, installation, delivery, lead time, and exclusions. Reference the page supporting each important figure. Write “not specified” where information is missing, and distinguish quoted amounts from your calculations. Draft follow-up questions for each supplier. Do not send messages or modify the originals. Ask me if two documents contradict each other or you cannot identify the supplier</em></blockquote><p>Then inspect the result. Open a few cited pages. Check the arithmetic. Pay particular attention to the information it says is missing.</p><p>For a useful test, include one deliberately incomplete document. Does the agent admit that it cannot calculate the full cost? Or does it produce a suspiciously tidy answer?</p><p>I would trust the first behavior more. An unresolved question is sometimes exactly what you needed to find.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yt-d6JqjJaboAEAs033qXw.png" /></figure><p>If you later want this to run whenever new quotes arrive, a visual platform such as Zapier Agents offers a route to connecting applications without developing the integration yourself.</p><p>You still need to configure the trigger, document access, output, and permitted actions. “No code” removes some implementation work; it leaves those decisions with you.</p><p>There is one distinction worth learning immediately: <strong>telling an agent not to send an email is different from removing its ability to send one.</strong> The first is an instruction. The second is an access restriction.</p><p>For this experiment, leave sending disabled or keep the mailbox disconnected. The agent can prepare useful drafts without that permission.</p><p><strong>The part you rarely see: the harness.</strong></p><p>The model does not open files or send emails by itself. It requests actions, and surrounding software executes them. That software also handles the back-and-forth: return the tool’s result, ask the model what comes next, and continue or stop.</p><p>You will often see this surrounding machinery called an <em>agent harness</em>. The term varies across products, but it is useful because it draws attention to the things that make repeated actions workable.</p><p>Anthropic’s work on long-running agents gives a concrete example. Its researchers describe the need to preserve progress between sessions, resume from a known state, and verify that work is complete. A powerful model alone did not solve those problems.</p><p>For our comparison, the practical questions look like this:</p><p>Part of the systemWhat it means for your taskToolsCan it read the PDFs and calculate totals?ContextDoes it have the relevant pages and your comparison criteria available now?State and memoryDoes it track what has been checked and what remains unresolved? Is anything retained afterward?Reusable instructionsCan you save your comparison procedure for next time?PermissionsCan it prepare drafts while being unable to send them?LimitsWill it stop after repeated failures or a configured spending limit?Records and verificationCan you inspect its actions and trace figures to their sources?</p><p>These are things to look for, not a list of features every product has.</p><p>Take memory. It sounds obviously helpful, until the agent saves “installation included” from an earlier mistake. Now that mistake can survive the original conversation. For this job, I would want the source and date attached to the claim, and a way to correct or remove it.</p><p>Or take retries. Trying again after a temporary failure is reasonable. Repeating the same failed request indefinitely is expensive persistence. The system needs a stopping rule that does not depend entirely on the model eventually getting bored.</p><p>You do not have to build the harness. But these questions will tell you more about a product than a video of its best run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iItnD4LGJm6CGq9WKqXugA.png" /></figure><p><strong>What changes when agents manage other agents?</strong></p><p>Imagine the job has grown. You now need to research suppliers in several countries, compare service coverage, and check local delivery conditions.</p><p>A coordinating agent could divide that work among several agents and combine their findings. Anthropic describes this structure in its multi-agent research system, along with higher resource consumption and difficulties when tasks depend heavily on one another. [6]</p><p>The attraction is straightforward: independent investigations can proceed in parallel. The difficulty arrives when their answers have to fit together.</p><p>One agent reports prices before tax. Another includes tax and delivery. A third has found a price from last year. Unless those assumptions travel with the results, the coordinator may produce a comparison that looks more consistent than the underlying evidence.</p><p>So a useful handoff needs the answer, its sources, its assumptions, and its unresolved questions.</p><p>And three agents agreeing does not automatically give you three independent confirmations. They may all have read the same incorrect page. A reviewer needs to check the evidence and calculations, rather than merely approve another model’s prose.</p><p>For three short quotes, I would keep the setup simple. A larger investigation might justify the extra coordination. The number of agents is an implementation choice, not a measure of quality.</p><p><strong>Meta’s experience makes the responsibility question harder to ignore.</strong></p><p>Meta introduced Muse on September 8, 2026, describing a personal agent usable without technical expertise. Its announced safeguards include approval for sensitive actions and a separate oversight agent, Sentinel. Those are the company’s stated protections; they do not establish that mistakes are impossible.</p><p>On September 22, Reuters reported that Meta had tested human contractors handling some Muse calls internally. An employee reported a racist remark during a call. According to internal posts reviewed by Reuters, a Meta executive acknowledged inadequate disclosure and said the feature had been rolled back for the time being. Meta said the test was intended to collect feedback before a possible public release.</p><p>That introduces a different question: who is actually doing the work? Human support may improve a service, but users need to know when their requests and information reach another person.</p><p>A separate incident reported by The Verge in March concerned an internal Meta agent that posted incorrect technical advice without approval. An employee followed it, temporarily exposing data to unauthorized colleagues. Meta said no user data had been mishandled and argued that human checks could have prevented the incident.</p><p>Perhaps they could have. I still want to know why the system allowed that sequence to happen.</p><p>Was the proposed action clear? Could the person realistically assess it? Why could the agent publish without approval? What other controls should have contained the mistake?</p><p>An instruction to “check the AI” leaves a lot unanswered. If checking requires reconstructing every step from scratch, the supposed productivity gain starts to disappear. If the interface encourages people to approve actions they cannot inspect, calling that human oversight seems generous.</p><p>Responsibility has to include the people designing and deploying the system, as well as the person using it.</p><p><strong>The calculation that decides whether you should keep using it.</strong></p><p>Suppose comparing the quotes normally takes 40 minutes. With the agent, you spend five minutes preparing the documents, eight checking the table, and two correcting an error.</p><p>That is 25 minutes of active work saved. This is an illustration, not a measured result, but it gives you something concrete to test.</p><p>Include the setup time across repeated uses, any subscription cost, and the seriousness of mistakes. Also distinguish your time from elapsed time: an agent running for ten minutes may require only a minute of your attention, or it may interrupt you throughout.</p><p>After a few attempts, you should know whether you are getting useful work back or acquiring a new supervision task.</p><p>I would start with the quotes, keep the original files, and leave sending disabled. If the results hold up, try the same process on the next batch. Then decide whether connecting a folder or scheduling the job would help.</p><p>There is plenty of ordinary work worth delegating!!! It is just like cleaning! I hope we will soon have robot cleaners.</p><p>Bref, finding one task that reliably gives you part of your afternoon back is a good place to begin.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d3429008179b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[IBM AIX in 2026: The Deep Technical Guide Enterprise Architects Should Read]]></title>
            <link>https://medium.com/@sitrakaforler/ibm-aix-in-2026-the-deep-technical-guide-enterprise-architects-should-read-a9589ffc008c?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9589ffc008c</guid>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[xai]]></category>
            <category><![CDATA[ibm]]></category>
            <category><![CDATA[unix]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 07:00:35 GMT</pubDate>
            <atom:updated>2026-09-05T07:00:35.587Z</atom:updated>
            <content:encoded><![CDATA[<h3>The operating system nobody talks about, yet many enterprises still depend on !</h3><p>When technology leaders discuss enterprise infrastructure in 2026, the conversation usually revolves around Linux, Kubernetes, public cloud platforms, artificial intelligence, and platform engineering.</p><p>Yet behind a significant number of critical banking, insurance, telecommunications, retail, transportation, and government systems, another operating system continues to perform the work it was designed to do:</p><p><strong>IBM AIX.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ssLqoYucvsOolSkb" /></figure><p>AIX, originally named <strong>Advanced Interactive eXecutive</strong>, is IBM’s commercial UNIX operating system for the IBM Power platform.</p><p>It does not have the visibility of Linux. It does not dominate developer conferences. It is rarely the first platform discussed when organizations launch a new cloud-native application.</p><p>That does not make it irrelevant.</p><p>It makes AIX a specialist platform.</p><p>AIX was designed for environments where predictable performance, vertical scalability, hardware-assisted virtualization, operational continuity, and long support cycles matter more than ecosystem popularity.</p><p>Four decades after its first release in 1986, AIX remains actively maintained. IBM continues to publish Technology Levels, Service Packs, security fixes, Power platform integration, and lifecycle information for AIX 7.3. As of September 2026, IBM lists AIX 7.3 TL4 as the latest Technology Level, initially released in December 2025. <a href="https://www.ibm.com/support/pages/aix-support-lifecycle-information">[ibm.com]</a>, <a href="https://esupport.ibm.com/customercare/flrt/report?pageNm=home&amp;reportname=TSA&amp;plat=power&amp;p0.mtm=9028-21B&amp;p1.os=aix&amp;p1.aix=7300-01-01-2246&amp;btnGo=SUBMIT">[esupport.ibm.com]</a></p><p>The right question is therefore not:</p><blockquote><em>“Why has AIX not disappeared?”</em></blockquote><p>The better question is:</p><blockquote><em>“Which workloads still </em><strong><em>justify </em></strong><em>AIX, and how should enterprises manage the platform in 2026??”</em></blockquote><h3>AIX Is Not Simply “Another Linux”</h3><p>AIX and Linux share many UNIX concepts:</p><ul><li>processes and signals</li><li>users and groups</li><li>filesystems and mount points</li><li>shells and command-line administration</li><li>TCP/IP networking</li><li>permissions and access-control mechanisms</li><li>daemons and system services</li></ul><p>However, operationally and architecturally, they are very different platforms.</p><p>Linux is a broad ecosystem running across commodity servers, virtual machines, cloud instances, embedded systems, mainframes, and IBM Power.</p><p>AIX is tightly integrated with IBM Power hardware and its surrounding management stack:</p><ul><li>IBM Power servers</li><li>PowerVM</li><li>Hardware Management Console, or HMC</li><li>Virtual I/O Server, or VIOS</li><li>Network Installation Manager, or NIM</li><li>PowerHA SystemMirror</li><li>Logical Partitions, or LPARs</li><li>Workload Partitions, or WPARs</li></ul><p>This vertical integration is one of AIX’s greatest strengths and one of its most important constraints.</p><p>AIX does not attempt to support every hardware platform. IBM controls and validates a much larger portion of the stack, from processor architecture and firmware to virtualization, operating system, availability technologies, and support processes.</p><h3>AIX and Linux at a Glance</h3><p>AreaIBM AIXEnterprise LinuxPrimary hardwareIBM Powerx86, Arm, IBM Power, IBM Z and othersPlatform modelVertically integrated IBM stackMulti-vendor ecosystemVirtualizationPowerVM, LPAR, VIOS, WPARKVM, VMware, Hyper-V, containersOS maintenanceTechnology Levels, Service Packs, interim fixesDistribution-specific packages and releasesKernel update capabilityAIX Live UpdateDistribution and kernel dependentAdministrationODM, SMIT, LVM, SRC, AIX-specific commandsDistribution-specific tooling, systemd, standard Linux utilitiesTypical use caseStable, vertically scaled enterprise workloadsGeneral-purpose, cloud-native and enterprise workloadsSoftware ecosystemSmaller and controlledVery large open-source and commercial ecosystemSkills availabilitySpecializedBroad global talent pool</p><p>The difference is not that one platform is universally better.</p><p>The difference is that they optimize for different priorities.</p><h3>Why Enterprises Still Use AIX</h3><h3>1. Operational Stability and Predictability</h3><p>AIX has a reputation for long-running production environments and controlled change. It is Deterministic (good compared to more and more probabilistic AI models).</p><p><strong>This does not mean that AIX systems never fail.</strong> <br>Hardware failures, application defects, storage incidents, configuration errors, security vulnerabilities, and human mistakes still exist.</p><p>The real advantage is operational predictability.</p><p>AIX environments are often managed through:</p><ul><li>standardized LPAR profiles</li><li>strict Technology Level and Service Pack baselines</li><li>validated VIOS configurations</li><li>multipath storage</li><li>redundant network adapters</li><li>enterprise backup tooling</li><li>PowerHA clusters</li><li>controlled NIM-based deployment</li><li>formal change and disaster-recovery procedures</li></ul><p>In highly regulated environments, predictable maintenance can be more valuable than rapid feature delivery.</p><h3>2. AIX Live Update</h3><p>AIX Live Update is one of the platform’s most distinctive capabilities.</p><p>Beginning with AIX 7.2, Live Update allows supported kernel fixes to be applied without the traditional workload downtime associated with restarting the operating system. Later AIX 7.2 levels extended this capability to supported Technology Level and Service Pack updates. <a href="https://www.ibm.com/docs/en/aix/7.3.0?topic=updates-live-update">[ibm.com]</a></p><p>Live Update does not simply replace the kernel in place.</p><p>At a high level, the operation uses a temporary <strong>surrogate LPAR</strong>:</p><ol><li>A surrogate environment is prepared.</li><li>The updated AIX image is made available.</li><li>Running processes are checkpointed.</li><li>Memory and process state are transferred.</li><li>Workloads continue on the updated kernel environment.</li><li>The original environment is retired after validation.</li></ol><p>This is an important distinction. Live Update is an orchestrated workload mobility mechanism, not a basic package installation.</p><p>It also has prerequisites and constraints related to:</p><ul><li>HMC or PowerVC connectivity</li><li>storage accessibility</li><li>root volume group layout</li><li>paging space</li><li>network configuration</li><li>supported devices</li><li>installed software</li><li>available processor and memory resources</li></ul><p>AIX 7.3 TL4 introduced further Live Kernel Update improvements, including optimized memory-page migration, parallelized root volume group copy operations, support for encrypted physical volumes, and reduced blackout duration. <a href="https://community.ibm.com/community/user/blogs/vinod-kumar-boddukuri/2026/01/05/whats-new-in-aix-73-tl4-live-kernel-update-enhance">[community.ibm.com]</a>, <a href="https://community.ibm.com/community/user/blogs/sanket-rathi1/2025/12/05/ibm-aix-73-tl4-features">[community.ibm.com]</a></p><p>Enterprise architects should therefore treat Live Update as an availability capability that must be designed, tested, monitored, and documented. It is not a universal “no reboot required” button.</p><h3>3. PowerVM and Logical Partitions</h3><p>PowerVM is central to most modern AIX architectures.</p><p>An <strong>LPAR</strong>, or Logical Partition, is a hardware-isolated virtual system created through IBM Power firmware and managed using an HMC or another supported management layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/474/0*hZQBXr7x82j610mM" /></figure><p>An LPAR can receive:</p><ul><li>dedicated or shared processor resources</li><li>entitled processor capacity</li><li>capped or uncapped CPU allocation</li><li>dedicated or shared memory</li><li>virtual Ethernet adapters</li><li>virtual Fibre Channel adapters</li><li>virtual SCSI devices</li><li>physical I/O adapters</li></ul><p>PowerVM’s shared processor pools allow enterprises to allocate fractional processor capacity and dynamically redistribute resources between workloads.</p><p>This is particularly valuable for large database and ERP environments where workload demand changes during the day.</p><h3>Virtual I/O Server</h3><p>The Virtual I/O Server provides shared I/O services to client LPARs.</p><p>Instead of assigning physical storage and network adapters directly to every AIX partition, VIOS can virtualize:</p><ul><li>Ethernet connectivity</li><li>Fibre Channel access through NPIV</li><li>SCSI storage</li><li>optical devices</li><li>installation media</li></ul><p>A common production design uses dual VIOS partitions to reduce the dependency on any single I/O path.</p><p>However, dual VIOS alone does not guarantee resilience. Availability also depends on correct multipathing, adapter placement, switch redundancy, SAN zoning, network configuration, and maintenance procedures.</p><h3>4. Live Partition Mobility</h3><p>PowerVM Live Partition Mobility can move a running AIX LPAR from one physical Power server to another without shutting down its operating system or applications.</p><p>The migration transfers the running partition’s processor state, memory, virtual devices, and active workload context to the destination system. cf <a href="https://www.ibm.com/docs/en/SSVEKM0/pdf/p11hc3.pdf">[ibm.com]</a>, <a href="https://www.ibm.com/docs/en/power8/8247-21L?topic=features-live-partition-mobility">[ibm.com]</a></p><p>This capability supports several important use cases:</p><ul><li>planned hardware maintenance</li><li>workload consolidation</li><li>datacenter rebalancing</li><li>evacuation of a server before intervention</li><li>lifecycle migration between compatible Power systems</li><li>reduction of planned service interruption</li></ul><p>Live Partition Mobility should not be confused with disaster recovery.</p><p>It generally requires the source environment, destination environment, HMC, network, VIOS, and shared storage architecture to remain operational and compatible.</p><p>Disaster recovery addresses scenarios where part or all of the primary environment is no longer available.</p><h3>5. Workload Partitions</h3><p>AIX also supports <strong>Workload Partitions</strong>, or WPARs.</p><p>WPARs are isolated operating-system environments running inside a single global AIX instance. They provide process, resource, filesystem, and network separation without requiring a complete additional AIX kernel. <a href="https://www.ibm.com/docs/en/ssw_aix_73/pdf/workloadpartitions_pdf.pdf">[ibm.com]</a>, <a href="https://www.ibm.com/docs/en/aix/7.2.0?topic=workload-partitions-aix">[ibm.com]</a></p><p>Conceptually, they are closer to operating-system-level virtualization than to a traditional hardware virtual machine.</p><p>WPARs can be useful for:</p><ul><li>application isolation</li><li>workload consolidation</li><li>delegated administration</li><li>controlled resource allocation</li><li>migration of selected legacy environments</li></ul><p>However, WPARs are not interchangeable with modern OCI containers. The ecosystem, image model, orchestration approach, networking, application packaging, and operational practices are different.</p><p>Calling a WPAR “the AIX equivalent of Docker” is therefore an oversimplification.</p><h3>Storage Management: Why AIX LVM Still Matters</h3><p>AIX includes a mature Logical Volume Manager.</p><p>Its core storage objects are:</p><ol><li><strong>Physical Volumes</strong>, or PVs</li><li><strong>Volume Groups</strong>, or VGs</li><li><strong>Logical Volumes</strong>, or LVs</li><li><strong>Filesystems</strong></li></ol><p>Typical commands include:</p><blockquote><em>lspv</em></blockquote><blockquote><em>lsvg</em></blockquote><blockquote><em>lslv</em></blockquote><blockquote><em>lsfs</em></blockquote><blockquote><em>df -g</em></blockquote><blockquote><em>mount</em></blockquote><blockquote><em>lspath</em></blockquote><p>The root operating-system environment is normally stored in:</p><p>rootvg</p><p>Application data may be separated into dedicated volume groups, for example:</p><blockquote>appvg</blockquote><blockquote>datavg</blockquote><blockquote>oraclevg</blockquote><blockquote>backupvg</blockquote><p>This structure allows administrators to manage storage allocation, mirroring, migration, filesystem growth, and disaster-recovery procedures in a controlled way.</p><p>AIX LVM is also closely integrated with the Object Data Manager, device configuration, multipathing, and boot process. For Linux administrators moving to AIX, assuming that every volume-management concept behaves exactly like Linux LVM is a common source of mistakes.</p><h3>Error Reporting and Service Management</h3><p>AIX has several administration mechanisms that differ considerably from Linux.</p><h3>The AIX Error Log</h3><p>The errpt command displays entries from the AIX error log:</p><blockquote><em>errpt</em></blockquote><p>A detailed view can be obtained with:</p><p>errpt -a</p><p>Recent events can be filtered using a starting date:</p><p>errpt -s 0904000026</p><p>This mechanism records hardware, device, storage, network, and software events through structured identifiers and error classes.</p><p>Enterprise monitoring should not simply execute errpt and generate an alert for every entry. A useful integration should classify:</p><ul><li>permanent versus temporary errors</li><li>hardware versus software errors</li><li>repeated identifiers</li><li>resource names</li><li>timestamps</li><li>severity</li><li>operational impact</li></ul><h3>System Resource Controller</h3><p>Many traditional AIX services are managed through the System Resource Controller:</p><blockquote>lssrc -a</blockquote><blockquote>lssrc -s sshd</blockquote><blockquote>startsrc -s sshd</blockquote><blockquote>stopsrc -s sshd</blockquote><blockquote>refresh -s inetd</blockquote><p>This differs from the systemctl workflow used by most current Linux distributions.</p><h3>SMIT</h3><p>The System Management Interface Tool provides a structured interface for many AIX administration tasks:</p><p>smit</p><p>or:</p><p>smitty</p><p>SMIT is not merely a beginner’s menu. It can display the underlying commands it executes, making it useful for learning AIX syntax and developing repeatable operational procedures.</p><h3>AIX Maintenance: Release, TL, SP and Build Numbers</h3><p>AIX maintenance levels commonly use the following structure:</p><p>7300–04–02–2633</p><p>This can be read as:</p><ul><li>7300: AIX 7.3</li><li>04: Technology Level 4</li><li>02: Service Pack 2</li><li>2633: build sequence identifier</li></ul><p>The installed level is displayed with:</p><p>oslevel -s</p><p>IBM documents this naming model as:</p><p>RELE-TL-SP-BBBB</p><p>IBM also maintains multiple AIX releases and Technology Levels in parallel, which gives customers flexibility but requires disciplined lifecycle management. <a href="https://www.ibm.com/support/pages/aix-service-strategy-and-best-practices">[ibm.com]</a></p><p>As of September 2026, IBM lists:</p><ul><li><strong>AIX 7.3 TL4</strong>, released in December 2025, with End of Fix Support on 31 December 2028</li><li><strong>AIX 7.3 TL3</strong>, released in December 2024, with End of Fix Support on 31 December 2027</li><li><strong>AIX 7.3 TL2</strong>, released in November 2023, with End of Fix Support on 30 November 2026 <a href="https://www.ibm.com/support/pages/aix-support-lifecycle-information">[ibm.com]</a></li></ul><p>An important architectural detail is that <strong>End of Fix Support for a Technology Level is not necessarily the same as End of Service for the complete AIX release</strong>.</p><p>Enterprises should track both concepts separately.</p><p>They should also use IBM FLRT, Fix Central, security bulletins, HIPER information, hardware firmware recommendations, and application compatibility matrices before planning an update.</p><h3>How AIX Is Installed in an Enterprise Environment</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*re4it1PPSDhdAuz6.jpg" /></figure><p>AIX is not normally installed on commodity x86 hardware.</p><p>A typical deployment begins with an IBM Power server and an HMC-managed LPAR.</p><h3>1. Create the LPAR</h3><p>Through the HMC, define:</p><ul><li>processor mode and entitlement</li><li>virtual processors</li><li>memory</li><li>boot mode</li><li>virtual Ethernet</li><li>virtual Fibre Channel or virtual SCSI</li><li>LPAR profile</li><li>VIOS mappings</li></ul><h3>2. Prepare Storage and Networking</h3><p>Configure:</p><ul><li>SAN zoning</li><li>NPIV mappings or vSCSI disks</li><li>VLANs</li><li>virtual switches</li><li>network adapters</li><li>redundant paths</li><li>boot disk visibility</li></ul><h3>3. Select an Installation Method</h3><p>Common installation methods include:</p><ul><li>NIM-based network installation</li><li>virtual optical media</li><li>ISO mounted through VIOS</li><li>alternate disk cloning</li><li>mksysb restoration</li></ul><p>For enterprise estates, NIM is usually more scalable than manually attaching installation media to every LPAR.</p><h3>4. Choose the Installation Type</h3><p>Typical options include:</p><ul><li>New and Complete Overwrite</li><li>Preservation Install</li><li>Migration Install</li></ul><p>The correct choice depends on whether the system is new, being recovered, or being upgraded.</p><h3>5. Validate the Installed System</h3><p>After installation, useful checks include:</p><blockquote>oslevel -s</blockquote><blockquote>uname -a</blockquote><blockquote>bootlist -m normal -o</blockquote><blockquote>lspv</blockquote><blockquote>lsvg</blockquote><blockquote>lsvg rootvg</blockquote><blockquote>lsfs</blockquote><blockquote>df -g</blockquote><blockquote>mount</blockquote><blockquote>lspath</blockquote><blockquote>ifconfig -a</blockquote><blockquote>netstat -rn</blockquote><blockquote>errpt</blockquote><p>In a production architecture, validation should also cover:</p><ul><li>NTP or time synchronization</li><li>DNS resolution</li><li>LDAP or directory integration</li><li>SSH configuration</li><li>backup agents</li><li>monitoring agents</li><li>storage path redundancy</li><li>network failover</li><li>system dump configuration</li><li>paging space</li><li>security baselines</li><li>audit configuration</li><li>application startup dependencies</li></ul><h3>Database and ERP Workloads</h3><p>AIX remains strongly associated with large enterprise workloads, including:</p><ul><li>Oracle Database</li><li>IBM Db2</li><li>Oracle applications</li><li>SAP application components</li><li>financial processing platforms</li><li>billing systems</li><li>custom C, C++ and Java applications</li></ul><p>One clarification is essential:</p><blockquote><strong><em>SAP HANA does not run natively on AIX.</em></strong></blockquote><p>SAP HANA deployments on IBM Power use <strong>Linux on IBM Power</strong>.</p><p>An enterprise may still operate a mixed Power environment in which AIX hosts application servers or legacy SAP components while Linux LPARs host SAP HANA. This is one reason architects must distinguish between the hardware platform, the hypervisor, and the operating system.</p><p>“Runs on IBM Power” does not automatically mean “runs on AIX.”</p><h3>The Real Drawbacks of AIX</h3><h3>Cost</h3><p>Running AIX typically involves:</p><ul><li>IBM Power hardware</li><li>PowerVM licensing</li><li>AIX licensing and subscription</li><li>enterprise storage</li><li>HMC infrastructure</li><li>VIOS administration</li><li>vendor support</li><li>specialized operational skills</li></ul><p>For many general-purpose workloads, Linux on x86 or public cloud infrastructure will be less expensive and easier to staff.</p><h3>Skills Shortage !!!</h3><p>The available AIX talent pool is smaller than the Linux talent pool.</p><p>Experienced administrators must understand more than shell commands. They often need knowledge of:</p><ul><li>PowerVM</li><li>HMC</li><li>VIOS</li><li>LVM</li><li>ODM</li><li>NIM</li><li>PowerHA</li><li>SAN and NPIV</li><li>Live Partition Mobility</li><li>AIX security and patching</li><li>vendor-specific database architectures</li></ul><p>The risk is not only <strong>retirement</strong>.</p><p>The deeper risk is <strong>undocumented </strong>operational knowledge concentrated in a small number of engineers x)</p><h3>Smaller Software Ecosystem</h3><p>Many modern products target Linux first.</p><p>Before selecting AIX for a new workload, architects must check:</p><ul><li>operating-system certification</li><li>supported AIX Technology Level</li><li>processor compatibility</li><li>Java version</li><li>database client compatibility</li><li>compiler requirements</li><li>installation packages</li><li>monitoring-agent support</li><li>backup-agent support</li><li>security-scanner support</li></ul><h3>Operational Complexity</h3><p>AIX is consistent once understood, but it is not Linux with different command names.</p><p>Administrators must learn AIX-specific concepts such as:</p><ul><li>Object Data Manager</li><li>device configuration states</li><li>boot logical volumes</li><li>volume group vary-on behavior</li><li>PVIDs</li><li>MPIO path management</li><li>installp software packages</li><li>APARs and interim fixes</li><li>Technology Levels and Service Packs</li><li>SRC-managed services</li></ul><h3>Should New Projects Choose AIX?</h3><p><strong>The answer should depend on workload characteristics, not platform loyalty.</strong></p><p>AIX can remain a rational choice when:</p><ul><li>the application or database is certified and optimized for AIX</li><li>vertical scalability is important</li><li>PowerVM consolidation is already established</li><li>operational continuity has very high business value</li><li>the enterprise already has mature AIX and Power skills</li><li>migration risk is higher than the expected infrastructure benefit</li><li>application modernization is not technically or economically justified</li></ul><p>Linux or a cloud-native platform may be more appropriate when:</p><ul><li>the application is Linux-first</li><li>Kubernetes portability is required</li><li>commodity infrastructure is a priority</li><li>rapid developer onboarding matters</li><li>the organization lacks AIX expertise</li><li>software vendors have reduced AIX support</li><li>horizontal scaling is more appropriate than vertical scaling</li><li>automated immutable deployment is a core requirement</li></ul><p>For many enterprises, the correct answer will not be “AIX or Linux.”</p><p>It will be a controlled hybrid architecture.</p><p>AIX may continue hosting stable systems of record while APIs, integration services, observability platforms, digital channels, and new applications run on Linux or Kubernetes.</p><h3>Will AIX Disappear?</h3><p>Not in the immediate future.</p><p>IBM is still publishing maintenance, lifecycle guidance, security updates, Power integration, and AIX 7.3 Technology Levels. AIX 7.3 TL4 includes support and enhancements for current IBM Power generations, including Power11-related capabilities. <a href="https://www.ibm.com/support/pages/aix-support-lifecycle-information">[ibm.com]</a>, <a href="https://community.ibm.com/community/user/blogs/sanket-rathi1/2025/12/05/ibm-aix-73-tl4-features">[community.ibm.com]</a></p><p>However, technical support from IBM is only one part of platform sustainability.</p><p>The long-term challenge is organizational.</p><p>Enterprises must answer:</p><ul><li>Who understands our PowerVM architecture?</li><li>Are our VIOS mappings documented?</li><li>Can we rebuild an LPAR from NIM or mksysb?</li><li>Have we recently tested disaster recovery?</li><li>Do we monitor AIX, HMC, VIOS, SAN paths and application dependencies?</li><li>Are our Technology Levels still under fix support?</li><li>Do we have successors for senior AIX engineers?</li><li>Which applications genuinely require AIX?</li><li>Which systems can be retired, replatformed or modernized?</li></ul><p>AIX is unlikely to disappear because the operating system suddenly stops working.</p><p>The greater risk is that organizations lose the knowledge required to operate it safely.</p><h3>Final Thoughts</h3><p>The cloud revolution did not eliminate AIX.</p><p>Linux did not eliminate AIX.</p><p>Containers did not eliminate AIX.</p><p>They changed where AIX makes sense.</p><p>AIX is no longer the default answer for general-purpose enterprise computing, but it remains relevant for carefully selected workloads where stability, PowerVM integration, large-scale virtualization, predictable maintenance, and long operational lifecycles justify the investment.</p><p>For enterprise architects, the right strategy is neither blind preservation nor forced migration.</p><p>It is evidence-based lifecycle management:</p><ol><li>Identify which business services depend on AIX.</li><li>Map the applications, databases, storage, networks and recovery dependencies.</li><li>Verify certification and support timelines.</li><li>Standardize Technology Levels and Service Packs.</li><li>Automate installation, patching, inventory and compliance.</li><li>Test Live Update and Live Partition Mobility rather than assuming they work.</li><li>Document VIOS, HMC, NIM, storage and disaster-recovery procedures.</li><li>Train the next generation of Power and AIX engineers.</li><li>Modernize workloads where the business case is real.</li><li>Preserve AIX where it remains the safest and most economical platform.</li></ol><p>AIX was never designed to win a popularity contest.</p><p>It was designed to operate critical enterprise systems predictably.</p><p>In 2026, that remains valuable.</p><p>#IBM #AIX #IBMPOWER #Power11 #PowerVM #UNIX #EnterpriseArchitecture #EnterpriseIT #Oracle #SAP #Infrastructure #SysAdmin #HybridCloud #Automation #CyberSecurity #Telecom #Banking #Observability #SystemsArchitecture</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a9589ffc008c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ukraine’s Real Weapon Is Not the Drone. It Is the… the Learning Loop.]]></title>
            <link>https://medium.com/@sitrakaforler/ukraines-real-weapon-is-not-the-drone-it-is-the-the-learning-loop-dbeb395bd713?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/dbeb395bd713</guid>
            <category><![CDATA[opencv]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ukraine]]></category>
            <category><![CDATA[war]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Sun, 16 Aug 2026 15:07:01 GMT</pubDate>
            <atom:updated>2026-08-16T15:07:01.232Z</atom:updated>
            <content:encoded><![CDATA[<p>Bohdan Zmiy ran an auto-repair shop in the Lviv region.</p><p>In 2023, a soldier he knew (a sapper nicknamed “Boroda,” who didn’t want to keep risking his life clearing mines by hand) asked a mutual friend, war veteran Vasyl Korenyovskyi, for help. Zmiy and Korenyovskyi built a crude ground robot to do the job instead. Borys Drozhak, an IT specialist, joined soon after. Together they founded <a href="https://rovertech.co.ua/en/"><strong>Rovertech</strong></a> in the summer of 2023, out of that same repair shop.</p><p>The robot they built was named <strong>Zmiy </strong>after Bohdan himself. He died in a road accident in the Lviv region on October 26, 2025, at 35. By then his name was already on hundreds of machines clearing mines and hauling supplies across the front. (<a href="https://www.youtube.com/watch?v=SIKt3g3WXzw">https://www.youtube.com/watch?v=SIKt3g3WXzw</a>)</p><p>That history matters because it explains where Ukrainian defence-tech actually comes from: not a national lab or a prime contractor,<strong> but a mechanic, a veteran and a programmer solving one soldier’s specific problem in a garage. Liberty of entrepreneurship.</strong></p><p>From that beginning came a family of ground robots: logistics platforms carrying ammunition and wounded, demining variants built to absorb blasts, and armed versions fitted with remote turrets from <a href="https://devdroid.tech/en"><strong>DevDroid</strong></a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6LTJyeZCVVxudl2W.jpeg" /></figure><p>Ukrainian reporting puts Rovertech’s production at roughly 60 units in 2024 and 720 in 2025. And since the state is the backbone but can’t be the one controlling everything. Most of this equipment moves through <a href="https://brave1.gov.ua/en/"><strong>Brave1</strong></a>, the <strong>state coordination platform</strong>, and its <a href="https://brave1.gov.ua/en/ministry-of-digital-transformation-and-brave1-launch-defense-technology-marketplace-for-the-military/"><strong>Brave1 Market</strong></a> procurement system, where frontline units order directly.</p><h3>The hardware:</h3><blockquote>what’s actually inside these systems</blockquote><p>Start with compute.</p><p>A ground robot or drone that needs to run object detection onboard, rather than just streaming raw video back to an operator, needs a small computer with a GPU.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*477NoeRXh2xwEAa-043QVQ.png" /><figcaption>Ugly drones are sometimes more efficient thanks to their numbers</figcaption></figure><p>The dominant choice across this space, Ukrainian and otherwise, is <a href="https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/"><strong>NVIDIA’s Jetson Orin</strong></a> family!!!</p><p>The Orin Nano module — roughly the size of a credit card, running at 7 to 25 watts — delivers up to 67 trillion operations per second and can run a quantised YOLOv8-nano model at close to 45 frames per second on 720p video. That is the actual power budget a small platform has to work with: a laptop GPU draws ten to twenty times that, which is why nobody is running a full-size vision transformer on a drone.</p><p>And Cameras follow the same logic :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n43MeW_6Hd-PM02TW-LKtQ.png" /></figure><p>Daylight optical sensors are commodity components; the harder problem is <strong>thermal</strong>. Long-wave infrared modules (0.75–1000 µm if you remember your courses of physics in high school). Well they are the kind that let an operator see a hidden vehicle at night or through light smoke; and they… remain very expensive and, for the higher-sensitivity cores, subject to export controls, because the underlying detector technology has weapons applications. This is one of the quieter chokepoints in the whole ecosystem: plenty of countries can build a drone airframe, <strong>far fewer can build an uncooled thermal sensor at volume.</strong></p><p>Then there’s the <strong>tether</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/575/0*hXyWCDstK6KscFaH.jpeg" /></figure><p>Since 2024, both sides have fielded FPV drones that fly on <strong>fibre-optic cable rather than radio</strong>, specifically to defeat <strong>jamming</strong>.</p><p>A typical Ukrainian spool runs 10 to <strong>20km of hair-thin single-mode fibre,</strong> commonly the G.657.A2 standard <strong>also used in civilian telecoms!,</strong> weighing only a few hundred grams per kilometre.</p><p>This month (August 2026), the Ukrainian company Ptashka Drones fielded a variant with a 50km spool. The tradeoff is mechanical, not electronic: the cable can snag on trees, and it limits the drone to roughly the length of thread it’s carrying. But it is close to unjammable, because there’s no radio signal to jam.</p><p>Batteries are <strong>lithium-ion</strong> or lithium-polymer packs, sourced almost entirely from Asian cell manufacturing. Motors are brushless, and their efficiency depends on neodymium-iron-boron magnets…. which is where the rare-earth story below becomes unavoidable.</p><h3>The ML workflow: from footage to a model that survives contact</h3><blockquote>None of this hardware is useful without a pipeline that turns combat footage into something a model can learn from, and that pipeline is where most of the real engineering effort goes.</blockquote><p>A workable sequence looks like this in practice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w_CuVr4ThBKrsld3_Zw5ig.png" /><figcaption>Summary of the ML workflow</figcaption></figure><ul><li><a href="https://opencv.org/"><strong>OpenCV</strong></a> or <a href="https://pyav.org/"><strong>PyAV</strong></a> for decoding raw drone video and pulling out frames.</li><li><a href="https://www.cvat.ai/"><strong>CVAT</strong></a> or <a href="https://labelstud.io/"><strong>Label Studio</strong></a> for annotation : bounding boxes on vehicles, trenches, mines, decoys.</li><li><a href="https://fiftyone.ai/"><strong>FiftyOne</strong></a> for auditing the dataset afterward, since labelling at speed produces errors that need catching before training.</li><li><a href="https://pytorch.org/"><strong>PyTorch</strong></a> for the training loop itself.</li><li><a href="https://albumentations.ai/"><strong>Albumentations</strong></a> for augmenting footage with synthetic blur, low light and compression artefacts, simulating conditions the model will meet in the field before it ever gets there.</li><li><a href="https://mlflow.org/"><strong>MLflow</strong></a> or <a href="https://dvc.org/"><strong>DVC</strong></a> for tracking which model version trained on which dataset versions, essential once a team is iterating weekly rather than annually.</li></ul><p>For the detection model itself, <a href="https://docs.ultralytics.com/models/yolov8/"><strong>YOLOv8</strong></a> from <a href="https://ultralytics.com/"><strong>Ultralytics</strong></a> is the practical default: since it is real-time, well-documented, and it exports cleanly to ONNX or TensorRT for deployment on a Jetson.</p><p>For genuinely power-starved platforms, <a href="https://arxiv.org/abs/1905.02244">MobileNetV3</a> trades some accuracy for a fraction of the compute. <a href="https://dinov2.metademolab.com/"><strong>DINOv2</strong></a> (which i love)from Meta AI is useful further upstream, for building a searchable visual index of an enormous, mostly unlabelled footage archive, since it doesn’t require every frame to be hand-annotated first.</p><p>The part that doesn’t show up in a tools list is retraining cadence. A model trained on summer footage degrades in snow. A model that hasn’t seen a particular camouflage pattern will miss it. Rovertech and firms like it keep AI restricted to observation, tracking and target acquisition ( not steering, not firing) specifically because retraining fast enough to keep pace with a battlefield that shifts every two to three weeks is a genuinely unsolved logistics problem, not a modelling one.</p><h3>The network layer is its own battlefield</h3><blockquote>I am biaised of course. But NONE of the above works if the signal doesn’t get through, and this is arguably the least-covered part of the war.</blockquote><p>For most of 2023 to early 2026, Ukraine’s default long-range link was <strong>Starlink</strong>: low latency, around 20 to 50 milliseconds, and hard to jam because the signal comes from hundreds of satellites rather than one fixed point. The problem surfaced when Russian units started fitting captured or acquired Starlink terminals to their own long-range drones, using the same resilience against Ukrainian jamming. In February, Kyiv worked with SpaceX to whitelist only verified Ukrainian terminals and cut off unregistered ones (cf a previous <a href="https://medium.com/@sitrakaforler/how-russia-has-lost-its-drone-advantage-the-real-drone-war-is-about-networks-not-airframes-78503637a2a5">article</a>)</p><p>Russia’s answer was to scale up <strong>mesh networking</strong>: instead of one drone talking to a ground satellite dish, a chain of drones relays a signal from node to node, each one both receiving and retransmitting, so that losing one node doesn’t collapse the link. <br>It’s slower… added latency at every hop, which matters if a drone is being actively steered onto a moving target; and it constantly transmits, which makes a node detectable to electronic-warfare units even when its content can’t be decoded. But it doesn’t depend on any single company’s satellite constellation, and it’s cheap to keep replicating !</p><p>Fibre-optic control is the third leg: zero emissions, unjammable by definition, but hard-capped by cable length and vulnerable to something as mundane as the line snagging on a fence post.</p><p>None of these three approaches wins outright.</p><p>Starlink gives range and low latency but is a single point of policy leverag but yeah Elon Musk’s company can turn it off. <br>Mesh gives independence from any one vendor but adds latency and a detectable signature. <br>Fibre is nearly immune to jamming but physically short-ranged.</p><p>Which one a unit uses is now a tactical decision made mission by mission, and increasingly, drones carry more than one of these as a fallback.</p><h3>The chemistry problem nobody’s engineering around</h3><p>Every motor, every battery cell, every thermal sensor eventually traces back to a supply chain that Ukraine doesn’t control and Europe barely influences.</p><p><strong>The bottleneck isn’t rare-earth mining.</strong></p><p><strong>It’s what happens after the ore comes out of the ground: </strong>separation, refining, alloying, and magnet sintering.</p><p>Last week i listened to Pascal Bonnifcace’s talk According to the <a href="https://www.iea.org/reports/rare-earth-elements/executive-summary">International Energy Agency</a>, China handles around 91% of global rare-earth refining and separation, and roughly 94% of sintered permanent-magnet production. A European government could greenlight a new mine tomorrow and still be years from a usable magnet supply chain, because the actual constraint is chemists, refineries and specialised sintering equipment ….<strong>. none of which gets built on a wartime timeline.</strong></p><p>This is the layer Ukraine’s drone and robotics industry hasn’t cracked, and can’t crack by assembling faster. I<strong>t can localise design, software, final integration and repair; which it has done, impressively. It cannot yet localise the chemistry.</strong></p><h3>What the robots still can’t do…</h3><p>Ground robots are absorbing a real share of what used to be infantry work: hauling ammunition, evacuating the wounded, laying and clearing mines, holding a position long enough under remote control to simulate a full squad.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/894/1*wrhhNq0cr7QmZF1sXsrs2Q.png" /></figure><p><strong>They don’t hold ground.</strong></p><p>Someone still has to walk into a building, make a judgment call under fire, and stay there while it’s being shelled. That hasn’t changed, and nothing described above is close to changing it.</p><blockquote>What has changed is how fast a failure in the field becomes a fix and whose name ends up on the machine that gets it right.</blockquote><p>The traditional Napoleonic view holds that dense infantry columns and bayonet charges, driven <strong>by courage and discipline,</strong> decide battles. A model that has largely collapsed in WW1… but still used by Russian’s doctrine in Ukraine… where massed assaults are shredded by drones and artillery before any contact.</p><p>Russians must read Ardant du Picq’s alternative. For it was grounded in Solferino’s reality. He argued that <strong>soldiers instinctively disperse, seek cover, and that morale and psychology (</strong>not shock) determine who breaks first.</p><p>This matches the Ukrainian war, where small, dispersed units with high individual initiative and suppression of the enemy’s will often prevail over rigid, dense formations. As at Solferino, close-combat wounds are rare; instead, sustained fire and the threat of it break units long before they meet, making Ardant du Picq’s focus on the soldier’s breaking point far more useful than the old obsession with blind charge of canon meat.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dbeb395bd713" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vegeta and the Quiet Violence of Comparison]]></title>
            <link>https://medium.com/@sitrakaforler/vegeta-and-the-quiet-violence-of-comparison-c2dfe4a0d1c6?source=rss-b075dea26af9------2</link>
            <guid isPermaLink="false">https://medium.com/p/c2dfe4a0d1c6</guid>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[dragon-ball-z]]></category>
            <category><![CDATA[short-story]]></category>
            <category><![CDATA[competition]]></category>
            <dc:creator><![CDATA[Sitraka FORLER]]></dc:creator>
            <pubDate>Sat, 01 Aug 2026 20:45:08 GMT</pubDate>
            <atom:updated>2026-08-01T20:45:08.260Z</atom:updated>
            <content:encoded><![CDATA[<p>There is something unsettling about Vegeta and it has very little to do with his brutality.</p><p>What makes him linger is <strong>recognition</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/961/0*ukPolTSUpsBYPQzV" /></figure><p>Beneath the armor and pride sits a pattern that feels uncomfortably close to real life: a man of immense discipline and potential, <em>slowly consumed by </em><strong><em>comparison</em></strong><em>.</em> Not defeated by others, but defined by them.</p><p>His story is not just about rivalry. <em>It is about identity built on the wrong foundation.</em></p><h3>A prince without a world</h3><p>Vegeta begins as a paradox.</p><p>He is born at the top! He is the prince of a warrior race that measures worth in strength. Yet his formative years are shaped by absolute powerlessness. Freeza destroys his planet, kills his father, and turns him into a subordinate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1g62K3qx2n0H3s98" /></figure><p>From that moment on, Vegeta lives in contradiction: he believes himself superior, yet operates under domination.</p><p>So he clings to the only stable idea he has left …. his title</p><p>“I am the prince of all Saiyans.”</p><p><strong>But repetition does not create truth</strong>. It compensates for its absence.</p><p>And as Tywin Lannister said <em>“Any man who must say, ‘I am the King’ is no true king”</em></p><p>Psychologically, this is the seed of what <a href="https://en.wikipedia.org/wiki/Alfred_Adler">Alfred Adler</a> described as the <em>inferiority complex</em>: a condition where perceived inadequacy does not weaken ambition, but distorts it into overcompensation ( and I do recommand to read Adler’s <em>Understanding Human Nature</em>).</p><p>For Vegeta does not lack drive. He lacks grounding.</p><h3>When reality refuses your narrative</h3><p>On Earth, Vegeta expects confirmation of his identity. And at first, he gets it.</p><p>Then Goku arrives.</p><p>A low-born Saiyan. An outlier. Someone who should not matter.</p><p>And yet, he does.</p><p>The fight itself is almost secondary. What matters is the disruption: Vegeta is forced to confront the possibility that his hierarchy (his entire mental model of worth) might be wrong.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4cXSjRg1DbxyKpDH.jpg" /></figure><p>The certainty is gone.</p><p><strong>And once certainty fractures, obsession enters.</strong></p><h3>Namek and the collapse of meaning</h3><p>If Earth plants doubt, Namek destroys illusion.</p><p>Goku becomes what Vegeta believed he was destined to become: the Super Saiyan. He defeats Freeza. <strong>He </strong>accomplishes the narrative Vegeta has been carrying since childhood.</p><p>Meanwhile, Vegeta dies powerless, grieving, and aware.</p><p>This is not just failure. It is displacement.</p><p>The philosopher René Girard (which is strangely BACK on the news thanks to Peter Thiel) argued that human desire is often <em>mimetict </em>thus <strong>we want what others want,</strong> and we define ourselves through <strong>rivalry</strong>. The rival becomes both model and obstacle (<em>Deceit, Desire, and the Novel</em>).</p><p><em>You use steel to sharpen steel, and one friend sharpens another.<br>That’s why WE all need Rivality.</em></p><p>An Goku becomes exactly that for Vegeta.</p><p>From this point on, Vegeta is no longer pursuing strength for its own sake. He is pursuing relief from comparison.</p><p>And that is a fundamentally unwinnable pursuit.</p><h3>The pathology of performance</h3><p>By the Cell saga, Vegeta embodies a familiar modern archetype: the high performer driven by insecurity.</p><p>He trains relentlessly. <strong>Sacrifices balance</strong>. Neglects relationships. Achieves, objectively, what he set out to achieve he surpasses Goku.</p><p>But success, when rooted in comparison, is <strong>unstable</strong>.</p><p>Instead of resolution, it creates escalation.</p><p>He allows Cell to reach perfection not out of strategy, but out of a need for absolute validation.</p><p>The result is predictable: collapse.</p><p>This dynamic echoes a simple but enduring biblical warning:</p><blockquote><em>“For where jealousy and selfish ambition exist, there will be disorder and every vile practice.”<br> — James 3:16</em></blockquote><p>Vegeta does not fail because he is weak. He fails because his ambition is misaligned.</p><p>In contrast, Goku operates differently. He is not free of ambition — but he is free of fixation. He trains with consistency, not obsession; with curiosity, not resentment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/0*X4T3VNBYBoRe7Cxa.jpg" /></figure><p>The difference is subtle. The outcome is not.</p><h3>When identity depends on opposition</h3><p>After Goku’s death, Vegeta faces an unexpected void.</p><p>No rival. No reference point.</p><p>And without that external measure, his identity begins to dissolve.</p><p>This is the hidden cost of comparison: it outsources self-definition. When the comparison disappears, so does the structure it supported.</p><p>So when Goku returns, even briefly, Vegeta regresses instantly.</p><p>He chooses corruption. Accepts Babidi’s control. Becomes Majin Vegeta.</p><p>Not out of ignorance ….but <strong>clarity</strong>.</p><p>He knows what he is sacrificing. And does it anyway.</p><p>Because the need to prove he is not inferior outweighs everything else…. including the life he has built.</p><p>There is another line, older and quieter, that captures this tragedy with precision:</p><blockquote><em>“A heart at peace gives life to the body, </em><strong><em>but envy rots the bones.</em></strong><em>”<br> — Proverbs 14:30</em></blockquote><p>Vegeta is powerful. But he is not at peace.</p><p>And so, despite all his strength, something in him keeps deteriorating.</p><h3>The turning point: indifference, not defeat</h3><p>What ultimately shifts Vegeta is not victory, nor loss but detachment.</p><p>At a certain point, he stops measuring.</p><p>Not completely, not instantly, but enough.</p><p>Enough to train without constantly referencing Goku. Enough to act without needing proof. Enough to exist without comparison as the central axis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*faNq1YW9jPQ16cvd.jpg" /></figure><p>And in that space, something unexpected happens.</p><p>He improves.</p><p>Not in the explosive, dramatic way of earlier arcs but in a deeper, more stable way.</p><p>Because energy once spent on comparison is now redirected toward mastery.</p><p>This is the quiet lesson at the core of his arc:</p><p>Fulfillment does not come from surpassing others.</p><p>It comes from no longer needing to.</p><h3>A mirror for modern ambition</h3><p>Vegeta’s story resonates because it mirrors a pattern that extends far beyond fiction.</p><p><strong>In high-performance environments : technology, finance, entrepreneurship, legal… Well the metrics are everywhere. Rankings, benchmarks, visibility.</strong></p><p>Comparison is constant. And often, encouraged.</p><p>At first, it sharpens.</p><p><em>Then</em>, it corrodes.</p><p>Work becomes evaluation. Progress becomes relative. Satisfaction becomes conditional.</p><p>And like Vegeta, many find themselves achieving more than ever; while feeling less grounded than before.</p><h3>The deeper lesson</h3><p>Vegeta was never limited by talent or effort.</p><p><strong>He was limited by the framework through which he interpreted both.</strong></p><p>As long as his identity depended on being “better than,” he remained trapped in a loop where no result could ever be sufficient.</p><p>The shift only begins when that framework breaks.</p><p>When the question changes from:</p><p>“Am I ahead of them?”</p><p>to:</p><p><strong>“Am I aligned with what I’m becoming?”</strong></p><p>It is a quieter question. Less dramatic. Less visible.</p><p>But unlike comparison, it has an answer that does not move.</p><p>And that makes all the difference.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1LKwjmxVHyVlq7DSSgYeaQ.png" /></figure><p>Big thanks to you for reading this ! And even bigger thanks to Shimpo academy to deep dive in <a href="https://en.wikipedia.org/wiki/Akira_Toriyama">Akira Toriyama</a>’s talks and narratives.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c2dfe4a0d1c6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>