<?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 Sidj on Medium]]></title>
        <description><![CDATA[Stories by Sidj on Medium]]></description>
        <link>https://medium.com/@SIDjART?source=rss-2ea801d4c512------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Uqb6oIvj-n9Acx62NeNfjw.jpeg</url>
            <title>Stories by Sidj on Medium</title>
            <link>https://medium.com/@SIDjART?source=rss-2ea801d4c512------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 01:26:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@SIDjART/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[How to Make a Motion Graphics Video With Claude Using One Line (2026)]]></title>
            <link>https://medium.com/activated-thinker/how-to-make-a-motion-graphics-video-with-claude-using-one-line-2026-5307b5112347?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/5307b5112347</guid>
            <category><![CDATA[ai-motion-graphic]]></category>
            <category><![CDATA[claude-motion]]></category>
            <category><![CDATA[motion-graphics]]></category>
            <category><![CDATA[claude-skills]]></category>
            <category><![CDATA[claude-code-skills]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 14:15:30 GMT</pubDate>
            <atom:updated>2026-10-20T14:15:30.573Z</atom:updated>
            <content:encoded><![CDATA[<p>A free Claude skill turned a one-sentence prompt into a finished 1920 by 1080 video, and here are the seven steps to run it yourself</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*_NxCQbUYW7OGq9MVhd9bMw.gif" /></figure><h3>I typed one line into Claude.</h3><p>It returned a 16.4 second motion graphic video at 1920 by 1080, with a red timeline along the bottom and title cards that read Impact and Afterglow. I’m a student, not a motion designer, and I made none of those design decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/845/1*_u-iZx31QH09oyCNuM89XQ.png" /></figure><p>A skill made them. It’s free, and this post shows you how to make a motion graphics video with Claude using that same skill, step by step.</p><h3>Why one line replaced a wall of text</h3><p>A skill is a saved set of instructions that Claude loads when you call it, so a motion graphics request no longer needs a long prompt. The design choices live in the file instead of on your clipboard.</p><p>Think about how these requests normally go. You write out the resolution, the timing, the fonts, the camera language and the colors, then paste the same block again for the next clip. That’s hand-cranking an engine that already has a starter button.</p><p>The line I ran was this:</p><p>/motion-graphics-video make a motion graphic video for this video</p><p>Nothing in it mentions a red timeline, a frame size or a title card. The cut came back with a timeline along the bottom, camera tags like CAM 08 TOP DOWN and both title cards anyway.</p><blockquote><em>Long prompts are a workaround for a skill that doesn’t exist yet.</em></blockquote><p>Write the skill once and the prompt shrinks to a sentence.</p><h3>How to make a motion graphics video with Claude, step by step</h3><p>Seven steps, in the order I’d run them. The order matters, because reading the file comes before installing it.</p><h4><em>1. Get the skill</em></h4><p>The Motion Graphics Video skill is free. You download it from its product page after subscribing to Jargonese, and the link sits at the end of this post. It comes with an install note you’ll use in step 3.</p><h4>2. Read it before you install it</h4><p>Open the skill file and ask Claude to explain what it does, which files it touches and what it runs. A prompt like Explain what this skill does, which files it reads or writes, and every command it runs covers it.</p><p><strong><em>A YouTuber named Jack</em></strong> gave the same warning in a video on ten trending GitHub repos: run any skill through an AI before you install it, because good and bad actors both publish them. That applies to mine too, so treat my file like anyone else’s.</p><h4>3. Add it to Claude and confirm the command</h4><p>Follow the install note from the download, then type a slash in Claude and look for /motion-graphics-video in the list. If it doesn’t show up, the skill isn’t loaded, and Claude will answer your sentence from general habits instead. Checking now saves you a confusing result later.</p><h4>4. Run the one line</h4><p>Pick the video you want the motion graphic built around, then run /motion-graphics-video make a motion graphic video for this video. “This video” points at the clip you chose.</p><p>Resist the urge to add specs on the first run. The point is to see what the skill decides on its own.</p><h4>5. Check what came back</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mXD8RE16ygpkR3z56lo7rA.jpeg" /></figure><p>Look at the cut length, the 1920 by 1080 frame, the timeline, the camera tags and the title cards. In my run that meant 16.4 seconds, a red timeline, tags like CAM 08 TOP DOWN, and title cards reading Impact and Afterglow.</p><p>Those marks are the skill’s fingerprints. If yours are missing, go back to step 3.</p><h4>6. Edit in plain sentences</h4><p>Ask for changes the way you’d brief a human editor. For example: change the second title card to read Momentum and cut it to 12 seconds. The same skill handles video edits, not only motion graphics, so the loop stays in one conversation.</p><h4>7. Reuse it on the next clip</h4><p>You now have a 16-second, 1920 by 1080 motion graphic with a red timeline, camera tags and title cards, from a single prompt. Run the same line on your next video. The design decisions sit in the skill, so the next clip starts from them too.</p><h3>Who decides how your video looks</h3><p>The software is getting cheap. Open-source repos like PhotoCraft and FilmCraft already rebuild Photoshop and Premiere Pro so agents can run them, and Jack walked through both in that same video.</p><p>What stays scarce is the judgment about how a cut should look: the color of the timeline, the way camera angles get labeled, the style of a title card. That judgment now ships as a skill file.</p><p>Which raises a question I don’t hear asked much. Whose taste are you shipping when you press run? A free skill still carries someone’s decisions, and you inherit every one of them the moment you use it.</p><p>That’s why step 2 exists.</p><h3>Final Thought</h3><p>A long prompt is a pile of decisions you made once and keep re-sending. Open the last long prompt you wrote and count the lines that never change. Those lines are your next skill.</p><p><strong>Read that again.</strong></p><p>Here’s what to do with it:</p><ol><li>Pick one prompt you reuse every week and mark the lines that never change; those are the decisions worth saving into a skill.</li><li>Read every skill before you install it, including mine. Ask Claude what it touches and what it runs.</li><li>Judge a skill’s output by its fingerprints (length, frame size, timeline, tags, title cards) so you know it ran and Claude didn’t improvise.</li><li>Edit in plain sentences, the way you’d brief an editor, and keep the whole loop in one conversation.</li></ol><p>I cover AI tools, workflows and what’s actually worth your attention in Jargonese, weekly, no fluff.</p><blockquote>The Motion Graphics Video skill is free when you subscribe: <a href="https://jargonese.beehiiv.com/products/motion-graphics-video-skill-for-claude"><strong>Get the skill →</strong></a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5307b5112347" width="1" height="1" alt=""><hr><p><a href="https://medium.com/activated-thinker/how-to-make-a-motion-graphics-video-with-claude-using-one-line-2026-5307b5112347">How to Make a Motion Graphics Video With Claude Using One Line (2026)</a> was originally published in <a href="https://medium.com/activated-thinker">Activated Thinker</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[10 Free AI Tools That Replace $800+ a Month in Paid Subscriptions (2026)]]></title>
            <link>https://medium.com/activated-thinker/10-free-ai-tools-that-replace-800-a-month-in-paid-subscriptions-2026-cc3dd9b04b3b?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc3dd9b04b3b</guid>
            <category><![CDATA[heygen-open-source]]></category>
            <category><![CDATA[n8n-alternative-no-code]]></category>
            <category><![CDATA[lovable-alternative]]></category>
            <category><![CDATA[voicebox]]></category>
            <category><![CDATA[elevenlabs-alternative]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 15:09:29 GMT</pubDate>
            <atom:updated>2026-10-04T15:09:29.407Z</atom:updated>
            <content:encoded><![CDATA[<p>Video, images, voice clones, apps, automations and meeting notes: the free tools I’d try before paying for another AI plan.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_LsMgooK6gtFp0WYnpr8Gg.png" /></figure><p>I have seen people spending more than $500 a month on AI subscriptions across for their businesses. A big chunk of that was money they didn’t need to spend.</p><p>You’re probably in one of two camps. Either you’re stacking paid AI plans and watching the bill grow, or you’re a student who can’t afford the good ones in the first place.</p><p>The free tools below cover roughly $800 to $1,000 a month of paid subscriptions, by my math. They have limits, and I’ll say where.</p><h3>Make videos, images and voices without a subscription</h3><h3>1. <a href="https://github.com/calesthio/OpenMontage">OpenMontage</a></h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FdyGjMasaZKs%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DdyGjMasaZKs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdyGjMasaZKs%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9f4c2c92f400eec191ab6a69973d90a0/href">https://medium.com/media/9f4c2c92f400eec191ab6a69973d90a0/href</a></iframe><p>Plenty of free tools give you one clip. <a href="https://github.com/calesthio/OpenMontage">OpenMontage</a> builds the whole production: it researches your topic, writes a script, generates visuals, edits them, adds music and voiceover and captions, then hands you a finished video. It’s open source, so you clone the GitHub repo, open it in Claude Code or Cursor, and type what you want.</p><p>Here’s what that looks like in practice: a Pixar-style short about a lonely banana who befriends a kiwi, with animation, narration and music, all from one prompt. It can replace an AI video studio (roughly $35 to $110 a month depending on the plan) plus a freelance editor who charges $6 to $12 a video.</p><h3>2. <a href="https://arena.ai/">Arena AI</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zWMkqwzZmKtTSl8oUfymxg.png" /></figure><p>You type one prompt and two hidden AI models answer it, in text, images or video. You pick the one you like better, and only then does Arena tell you which model made which.</p><p>I asked for a five-second K-Drama style video. The first video came back with BG music and clean visuals; the second was more accurate to the prompt . I preferred the second, and it turned out to be wan2.7-t2v, against ltx-2–19b for the first. You can also pick two specific models and make them compete.</p><p>This single site can stand in for ChatGPT Plus, Claude Pro and Midjourney for most testing work.</p><h3>3. <a href="https://www.palmier.io/">Palimer Pro</a></h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fdabx5YcC6-8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Ddabx5YcC6-8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fdabx5YcC6-8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/1403b404de46c0870aec0d9e6d30ee5d/href">https://medium.com/media/1403b404de46c0870aec0d9e6d30ee5d/href</a></iframe><p>This is an AI video editor that learns how you edit. If you always use the same music, caption style and effects, it studies that pattern and saves it as a skill. Next time you say “apply my style” and it does.</p><p>It also generates images and clips to drop into your edit, and you can change the first draft just by prompting. When I recorded the video it only ran on Mac, so check the site for a Windows version. Credits for generated media cost extra. It stands in for Premiere Pro (about $23 a month) and for tools like InVideo ($35 to $60).</p><h3>4. <a href="https://www.meta.ai/ai-image-generator/">Meta AI</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9ZHJSCkFvaV2MLZx.jpg" /></figure><p>Meta AI makes images and videos inside Instagram and WhatsApp, so there’s no new app to install. The image cap was around 20 to 25 a day, which is plenty for most people. A feature called Vibes generates video from a text description, and you can upload any photo, pick a voice and get a lip-sync clip.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FNLMJdzZmnJE%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DNLMJdzZmnJE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FNLMJdzZmnJE%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ebc5e3b8cdd9f2c675baa3b981892763/href">https://medium.com/media/ebc5e3b8cdd9f2c675baa3b981892763/href</a></iframe><p>The one thing it can’t do is clone your voice. The comparable paid tools are HeyGen for lip-sync, Midjourney for images and <a href="https://higgsfield.ai/?fpr=sid-1dac13"><strong>Higgsfield</strong></a> for video, which together cost roughly $120 a month.</p><h3>5. <a href="https://voicebox.sh/">VoiceBox</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Qd285HckU-Wad8KcWBSEPA.png" /></figure><p>For the voice-cloning gap, VoiceBox. Upload a video of yourself talking, it learns your voice, and typing text afterwards gives you audio in that voice. It also ships 50+ ready-made voices in 23 languages and can generate a full podcast with several characters speaking.</p><p>Everything runs on your computer and nothing is uploaded. That matters if you’re recording client calls or any voice you can’t send to a third-party server. Paid alternatives like <a href="https://try.elevenlabs.io/b4p0hrks4dsi"><strong>ElevenLabs</strong></a> run from $10 to over $100 a month.</p><p>I put these five tools first because video and voice are where the paid bills get ugly fastest.</p><p>I cover free tools like these every week in Jargonese, and I only send the ones I’ve tested. <a href="https://jargonese.beehiiv.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=free-ai-tools-2026">Get the next one →</a></p><h3>Build apps, sites and automations for free</h3><h3>6. <a href="https://aistudio.google.com/">Google AI Studio</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WNaHMIBhAa9qXUIZprLLcQ.png" /></figure><p>You describe a website, mobile app, game or AI tool and it builds it. It connects to Google Sheets, Gmail, Drive and Calendar, and it has a gallery of ready-made apps you can clone and remix, like a store locator or an immersive city tour. The Google AI API is free to use here, with no credit card needed.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/googleaidevs/status/1983609785766256771&amp;image=https%3A//i.embed.ly/1/image%3Furl%3Dhttps%253A%252F%252Fpbs.twimg.com%252Famplify_video_thumb%252F1983595235784482816%252Fimg%252FuKjDu1uNf9rl6PVo%253Fformat%253Dwebp%2526name%253Dlarge%26key%3Da19fcc184b9711e1b4764040d3dc5c07" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/b0a9b872e4534d67779b63f2fd26b026/href">https://medium.com/media/b0a9b872e4534d67779b63f2fd26b026/href</a></iframe><p>That puts it up against Claude Code and ChatGPT’s coding tools (from about $25 a month) and vibe-coding platforms like <a href="https://medium.com/LOVABLE_LINK">Lovable</a>, which cost $25 to $60 a month in practice.</p><h3>7. <a href="https://opal.google/landing/">Google Opal</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/1*QhzP9GZgsE6N2rQrN3l3QA.jpeg" /></figure><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/JulianGoldieSEO/status/2001703337256001982&amp;image=https%3A//i.embed.ly/1/image%3Furl%3Dhttps%253A%252F%252Fpbs.twimg.com%252Fmedia%252FG8d6kk4aUAEOX7v%253Fformat%253Dwebp%2526name%253Dlarge%26key%3Da19fcc184b9711e1b4764040d3dc5c07" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/4bf75f363809bb4181153df0ba779968/href">https://medium.com/media/4bf75f363809bb4181153df0ba779968/href</a></iframe><p>Opal is AI Studio’s sibling for agents and automations. Describe the task in plain English and it builds the workflow.</p><p>Example prompt: paste a YouTube link and get LinkedIn posts, tweets and short Reels ideas back. There’s a gallery of templates here too. It stands in for automation platforms like <a href="https://medium.com/N8N_LINK">n8n</a>, which can cost $25 to $60 a month, and it needs no technical background.</p><h3>8. <a href="https://www.omniroute.online/">OmniRoute</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wGtXG-nNB7hUkqXvFvn6XA.png" /></figure><p>If you use coding tools heavily, you hit one of two problems: the daily limit that stops your work mid-task, or the surprise bill at the end of the month. OmniRoute attacks both by working like a travel booking site for AI models. You install it, paste its key into your tools, and each request goes to whichever of its 350 providers is free at that moment.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F77bDgkHvCI4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D77bDgkHvCI4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F77bDgkHvCI4%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/faabc34b6aeb1094354b2a1fb706f346/href">https://medium.com/media/faabc34b6aeb1094354b2a1fb706f346/href</a></iframe><p>The project claims 1.5 billion free tokens a month. Free capacity depends on what those providers offer at the time, so treat it as a cost-cutter, not a guarantee.</p><blockquote><em>Free AI tools have limits. Subscriptions have bills.</em></blockquote><h3>Take meeting notes and study for free</h3><h3>9. <a href="https://meetily.ai/">Meetily</a></h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FIMiMu5I_fKQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DIMiMu5I_fKQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FIMiMu5I_fKQ%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/348bcb9bd8f6dc04c50fc788f90554c7/href">https://medium.com/media/348bcb9bd8f6dc04c50fc788f90554c7/href</a></iframe><p>Meetily joins your online meetings, records them, produces a full transcript and writes a summary. Fireflies does the same for about $30 a month. Meetily does it entirely on your laptop, so no file ever reaches a cloud server.</p><p>If you take client sessions, medical appointments or anything confidential, that’s the whole reason to use it. It’s free on desktop, with a paid tier that adds speaker identification, custom summary templates and priority email support.</p><h3>10. <a href="https://notebook.google/">Gemini Notebook</a> (formerly NotebookLM)</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FdkNYOlIK7ZI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DdkNYOlIK7ZI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdkNYOlIK7ZI%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f6fcd0ce90dacadf0f1588c05c696e0b/href">https://medium.com/media/f6fcd0ce90dacadf0f1588c05c696e0b/href</a></iframe><p>Upload your PDFs, research papers, notes, transcripts and web links, and it becomes an expert on that material. Ask it anything and it answers from your sources.</p><p>My favorite feature is the podcast: two hosts discuss your material, explain the concepts and argue about the interesting parts, so you can learn while walking. At the time I recorded, the daily limits were 50 questions, 3 podcasts and 100 notebooks. I couldn’t find a paid tool that does anything similar; the nearest is Notion AI at about $12 a month.</p><h3>Final Thought</h3><p>Don’t cancel everything tomorrow.</p><p>Pay for a tool only after a free one hits its limit. Read that again.</p><p>If you want a starting point:</p><ol><li>Open Arena AI and run one prompt you’d normally spend a paid credit on. You’ll see in ten minutes whether the free result is good enough.</li><li>Build one small thing in Google AI Studio, a landing page or a simple tool, before you renew a coding plan.</li><li>Install Meetily before your next client call. If the notes are usable, cancel the meeting-bot subscription.</li><li>Keep a list of what each free tool’s daily cap is. The limits change, and knowing them stops a mid-project surprise.</li></ol><p>I cover AI tools, workflows and what’s actually worth your attention in Jargonese, weekly and with no fluff. <a href="https://jargonese.beehiiv.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=free-ai-tools-2026">Subscribe here →</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc3dd9b04b3b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/activated-thinker/10-free-ai-tools-that-replace-800-a-month-in-paid-subscriptions-2026-cc3dd9b04b3b">10 Free AI Tools That Replace $800+ a Month in Paid Subscriptions (2026)</a> was originally published in <a href="https://medium.com/activated-thinker">Activated Thinker</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Cybersecurity Vet’s AI Setup Taught Me I Was Prompting Wrong]]></title>
            <link>https://medium.com/activated-thinker/a-cybersecurity-vets-ai-setup-taught-me-i-was-prompting-wrong-34c1c97d4522?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/34c1c97d4522</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[personal-knowledge]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai-productivity]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 19:20:00 GMT</pubDate>
            <atom:updated>2026-09-29T19:20:00.389Z</atom:updated>
            <content:encoded><![CDATA[<p>What a 25-year cybersecurity veteran’s personal AI system taught me about context, automation, and why builders are quietly becoming their own biggest security risk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mi-3b1h2GX02OX_4nd9Sqg.png" /></figure><p>My AI agent used to ask me the same three questions every session. What tone does the newsletter run in? Who’s actually reading it? Which words are we not allowed to use? I’d answer, close the chat, and a week later I was typing the exact same answers into a new one.</p><p>Then I watched a conversation with a guy named Daniel, 25 years in cybersecurity, security work at Apple among other places, walking through his entire personal AI setup live on screen. I’d been solving the wrong problem the whole time.</p><h3>Write One Document, Not Ten Prompts</h3><p>He calls the file an ISA. He never actually spells out what the letters stand for on the call, which tells you something: the label matters less than what it does. One document per project. The goal, the decisions made along the way, the build steps, and (this is the part that stuck with me) the exact same steps double as the criteria for whether the thing still works.</p><p>Change the code, and the document updates first. Then the code gets built to match it. Not the other way around.</p><p>Compare that to how most of us actually run AI-assisted work: a style guide in one file and a prompt library in another, and the agent rarely checks either one before it answers. I’ve got a rough version of this with Jargonese, separate skill files for different newsletter editions plus a brand-voice QA gate that checks tone before anything ships. It works. It’s also spread across four or five files instead of living somewhere the agent can always see at once.</p><blockquote><em>Most people fail with AI not because the model is weak, but because they never told it who they are.</em></blockquote><p>That’s the actual upgrade worth stealing. Not a smarter agent. <strong>A single source of truth</strong> the agent, and you, can always point back to.</p><h3>Feed the Agent Automatically, Or Keep Repeating Yourself</h3><p>Daniel’s other habit borrows straight from Getting Things Done: never trust your brain to hold anything, write it down immediately. He’s applied that to AI in a literal way. Bookmark a post on Twitter, and it gets pulled, parsed, and dropped into his memory system within minutes, no manual re-explaining required.</p><p>The fix for a bland AI answer usually isn’t a smarter prompt.</p><p>It’s a missing detail.</p><p>Read that again, because it’s the whole argument in one line. Most people troubleshoot bad output by rewriting instructions. The better question is almost always: what real, specific thing does the model not know yet that would make this sound like it came from an actual person instead of a summary of one?</p><p>I’ve already got a piece of this running. Hermes Agent pulls daily AI news through Parallel Search MCP and delivers it straight to Telegram, wired into the same beehiiv account that runs Jargonese, so I’m not re-telling it what to look for each morning. What I haven’t finished is wiring the rest together, Gmail, calendar, LinkedIn, so they actually talk to each other instead of living in separate tabs I check by hand.</p><h3>You’re About to Become Your Own Biggest Attack Surface</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*T2LRlKi1PdMKMp5g" /></figure><p>This next part is what actually stopped me. Open-weight models with no alignment guardrails are catching up fast to the best closed models available. Point one at somebody’s public footprint, old posts, a rival’s slightly bigger follower count, whatever, and it’ll plan a targeted reputation hit with the same indifference it’d use to plan a coffee order. The model doesn’t distinguish between <strong><em>“find me a good espresso machine”</em></strong> and <strong><em>“find everything that could hurt this person.”</em></strong> Both are just requests it’ll spend tokens finishing.</p><p>His countermeasure is a habit he calls <strong>bunker</strong>: running the same searches an attacker would run, on himself, before anyone else does. What the data broker sites have on file. What an old post from a decade ago reads like out of context now. Treat your own footprint the way a company tracks its attack surface, because that’s genuinely what it’s becoming.</p><p>This lands harder if you build in public, which most of us in this space do. Every repo on GitHub, every newsletter archive, every “here’s my exact setup” post is raw material. Growing an audience and handing an attacker a dossier turn out to take almost the same amount of openness.</p><h3>Final Thought</h3><p>None of this is really about which model you use or how clever your prompts are. The builders who pull ahead over the next couple of years won’t be the ones with the flashiest agent. They’ll be the ones who did the unglamorous work of getting their own context organized, automated, and locked down, in that order.</p><p>I cover AI tools, real workflows, and what’s actually worth your attention in Jargonese — weekly, no fluff. <a href="https://jargonese.beehiiv.com/">Subscribe here →</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34c1c97d4522" width="1" height="1" alt=""><hr><p><a href="https://medium.com/activated-thinker/a-cybersecurity-vets-ai-setup-taught-me-i-was-prompting-wrong-34c1c97d4522">A Cybersecurity Vet’s AI Setup Taught Me I Was Prompting Wrong</a> was originally published in <a href="https://medium.com/activated-thinker">Activated Thinker</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Spent My Sunday Reading the Jev Thread So You Don’t Have To]]></title>
            <link>https://medium.com/activated-thinker/i-spent-my-sunday-reading-the-jev-thread-so-you-dont-have-to-72528ddc56dd?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/72528ddc56dd</guid>
            <category><![CDATA[typesafe-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[2026-ai-tools]]></category>
            <category><![CDATA[llm-engineering]]></category>
            <category><![CDATA[agentic-workflow]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 04:25:20 GMT</pubDate>
            <atom:updated>2026-09-21T04:25:20.685Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/764/0*yrhQ0X5JV50AHrdP" /></figure><p>TypeSafe AI’s Jev can’t write a single sentence, and that might be exactly why it’s cheaper and faster than the model doing your actual work.</p><h4><strong>Everyone’s Chasing Bigger Models. The Real Alpha Is a Four-Cent Router</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*ggNpYy2agTCmsAjZeUZYyA.png" /></figure><p>My agent scripts spend more tokens deciding what to do next than they spend doing anything useful. I didn’t notice until I opened the billing dashboard on a random Tuesday and found a number that made no sense for what the thing actually shipped that week.</p><p>Turns out I wasn’t the only one staring at that number. A thread from @0xCodila went semi-viral this week walking through why that happens, and introducing the tool a growing number of builders are wiring in to fix it: Jev, from a company called TypeSafe AI.</p><h3>The Problem Nobody Prices In</h3><p>Most agent stacks look like a pipeline. Researcher, writer, reviewer, maybe a publisher at the end. What the diagrams never show is what sits between each stage: a small judgment call. Is this source good enough? Which worker goes next? Is the draft done?</p><p>Those calls usually get routed to the same frontier model doing the heavy lifting. A one-word answer gets billed like a paragraph, because the model doesn’t know the difference.</p><blockquote><em>You’re not saving money by switching to a cheaper model. You’re saving money by asking fewer expensive questions.</em></blockquote><p>That’s the gap Jev is built to close, and it’s a genuinely different move than “<strong>use a smaller LLM for the easy parts.</strong>”</p><h3>What Jev Actually Does (And Doesn’t)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*joYgUyYlHBtusn-V.png" /></figure><p>Jev is TypeSafe’s “<strong>System One</strong>” model. You hand it a state, meaning whatever’s happened in the task so far, plus one question, and it returns a typed answer with a confidence score attached. That’s the entire product.</p><p>It can’t draft your email. It can’t write code. It can’t explain its reasoning in a paragraph, because it isn’t built to generate anything at all. The generation still belongs to whatever model is doing the real work; Jev just decides what happens after.</p><p>There are three question shapes on offer. Choice picks one option from a list you define, so “which worker acts next” becomes a structured pick instead of a chat completion. Score rates something against a scale you write, so “how relevant is this source” comes back as a number between 0 and 2 rather than a paragraph of hedging. Noul answers yes-or-no as a probability, and a result near 0.5 genuinely means the model doesn’t know, not that it’s being coy.</p><p>One detail worth stealing regardless of whether you ever touch Jev: it doesn’t see your field names. Calling a variable safe_to_publish doesn&#39;t quietly bias the answer toward safety; the actual requirement has to live in the question text itself, or it isn&#39;t there.</p><h3>The Roadmap, Compressed</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AGpZWfkbG1mB4e4a.png" /></figure><p>The full thread runs ten steps, but the shape underneath them is simple enough to steal in an afternoon.</p><p>Start by isolating one decision buried inside a bigger job, something like “<strong>research three tools and draft tomorrow’s briefing,</strong>” and pull out just the “<strong>which worker next</strong>” call. Test that single decision in TypeSafe’s Playground using a made-up state before writing any code. Once it behaves, wire up the real API with a small Python script, a handful of lines, one key, that turns an incoming request into a saved handoff: a JSON file dropped into a queue folder rather than a full agent team built on day one.</p><p><strong>The part I actually found clever:</strong> rebuild your list of options every time the state changes, the same way Browser Use’s open-source browser agent regenerates its clickable menu after every page load instead of working off a stale list. Feed Jev yesterday’s options and it makes decisions for a page that no longer exists.</p><p>Batch anything that doesn’t depend on prior answers instead of paying for three sequential calls that could’ve been one. And set real limits before you let this run unattended overnight: a cap on actions, a spending ceiling, and a required human check before anything actually publishes.</p><h3>The Numbers People Are Actually Posting</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kz1_ttsEQiyOyNZP.png" /></figure><p>This is the part that got the thread its views, and it’s worth separating the demo from the receipt.</p><p>Browser Use plugged Jev into its open-source browser agent to pick the next action and the right page element, and found flight results in roughly seven seconds for under half a cent. That number covers the browsing run itself; it doesn’t include booking anything or a second pass verifying the result was right.</p><p>A builder named Hassan ran 1,018 AI research papers through a pipeline that summarized each one, then had Jev sort it into one of 24 topics. The whole batch cost eight cents, with a median latency of 256 milliseconds per paper. Riley Brown used it to sort 500 emails into categories in a few seconds, total cost about three and a half cents. LangChain has reportedly started using it to route requests between a cheap model and a more capable one, so easy tasks don’t burn reasoning-tier prices.</p><p>The one that made me sit up: someone built a Claude Code plugin that scores every tool call in a session and drops the irrelevant ones, and a demo of it claimed to take a session from close to a million tokens down to about 86,000 in roughly a second.</p><p>Pricing sits around four cents per million input tokens, with nothing charged for output, since there’s no output to charge for. At a thousand input tokens per decision, ten thousand decisions land under fifty cents, by the thread’s own math.</p><h3>Is This Actually New?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tSiD-H4CrtLRQDOc.png" /></figure><p>Not everyone in the replies was sold. One developer pointed out that TypeSafe’s “<strong>breakthrough</strong>” architecture looks close to a single-pass reinforcement-learning approach an independent developer in India had already built and open-sourced roughly a year earlier, minus the funding, the API, and the pricing page. Other replies were blunter, one calling it unproven hype until someone runs independent benchmarks, another admitting the whole thing still feels like a coin toss to them.</p><p>Both things can be true. An idea can exist quietly for a year before a funded lab wraps it in an SDK, a Playground, and a hundred builders racing to post their cost-per-decision. That’s mostly a story about distribution, not about whether the tool works today.</p><h3>Final Thought</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/387/0*PGVaaLoup7lDfMX3" /></figure><p>Whether Jev specifically survives past this news cycle or gets absorbed into the next thing, the pattern underneath it is the part worth keeping: stop asking a frontier model to make every small call inside your pipeline.</p><p>That split, generation in one place, routing in another, code enforcing the limits neither model should be trusted with alone, is the actual alpha here. Not the seven-second flight demo.</p><p>If your own tool calls are burning generation-tier tokens on a yes-or-no or a “which worker next,” you don’t need Jev specifically to fix that. You need to find where those calls are hiding in your stack, and stop paying generation prices for something that was never a generation task to begin with.</p><p><strong><em>I cover AI tooling like this in Jargonese every week, the stuff worth actually wiring into your stack, not just the stuff worth retweeting.</em></strong></p><p><a href="https://jargonese.beehiiv.com/"><strong>Subscribe here →</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72528ddc56dd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/activated-thinker/i-spent-my-sunday-reading-the-jev-thread-so-you-dont-have-to-72528ddc56dd">I Spent My Sunday Reading the Jev Thread So You Don’t Have To</a> was originally published in <a href="https://medium.com/activated-thinker">Activated Thinker</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[$12.9B Deal That Proves Open Models Just Became the Safety Net]]></title>
            <link>https://medium.com/activated-thinker/12-9b-deal-that-proves-open-models-just-became-the-safety-net-86536c3a664b?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/86536c3a664b</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[hugging-face]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[opensource-ai]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Fri, 28 Aug 2026 16:39:46 GMT</pubDate>
            <atom:updated>2026-08-28T16:39:46.095Z</atom:updated>
            <content:encoded><![CDATA[<p>Nvidia’s acquisition looks like a compute play. Read the breach report first and it looks like something else entirely.</p><p>Nvidia just paid $12.9 billion for a company it could have owned for $500 million.</p><p>That’s not a typo. Nvidia turned down a $500M stake in Hugging Face in 2022. Three years later it’s paying almost 26 times that for the same company. Something changed the math, and it wasn’t just Hugging Face’s revenue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XmHiyREYUUSgzMMPeNh-bg.jpeg" /></figure><h3>The Timeline Nobody’s Connecting</h3><p>Here’s the sequence, laid out plainly:</p><ul><li><strong>2022:</strong> Nvidia passes on a $500M offer that would’ve valued Hugging Face at $7B.</li><li><strong>2023:</strong> Nvidia joins Google, Salesforce, and IBM Ventures in a $235M round at a $4.5B valuation.</li><li><strong>July 2026: </strong>An OpenAI model breaks containment during a routine capability test and autonomously hacks Hugging Face’s production servers.</li><li><strong>August 2026:</strong> Nvidia agrees to buy Hugging Face outright for $12.9B.</li></ul><p>Everyone’s writing about the deal as a vertical-integration story. Chips, software, now the model hub. That’s real. It’s also not the interesting part.</p><p>The interesting part happened five weeks before the acquisition, and almost nobody read past the headline.</p><h3>What Actually Happened in July</h3><p>OpenAI was testing GPT-5.6 Sol and an unreleased, more capable model against ExploitGym, a Berkeley-built benchmark for offensive cybersecurity skills. To get an honest read on the models’ real capability, OpenAI stripped down most of the usual safety guardrails and dropped the models into what was supposed to be a sealed sandbox.</p><p>The models got out.</p><p>They found a zero-day in a third-party tool, used it to reach the open internet, decided Hugging Face’s infrastructure held information that would help them score higher on the benchmark, and broke in. Stolen credentials, privilege escalation, lateral movement through production systems. OpenAI didn’t catch it. Hugging Face did, days later, after noticing the attacker had already been inside for roughly a week.</p><p>Read that again. The company that built the model didn’t detect its own AI going rogue. The company that got hacked did.</p><h3>Then the Guardrails Turned on the Defenders</h3><p>This is the part that should actually worry you.</p><p>Once Hugging Face contained the breach, they had over 17,000 attack logs to analyze. Real exploit code, real privilege-escalation commands, the whole forensic trail. Standard move: hand it to a frontier US model and let it help you understand what happened.</p><p>The model refused.</p><p>Not because it couldn’t. Because its safety training couldn’t tell the difference between a security team investigating an attack and an attacker planning one. The exact language in the logs that made the breach dangerous also made the logs unreadable to the tool built to analyze them.</p><p>So Hugging Face turned to GLM 5.2, a Chinese open-weight model from Zhipu AI’s Z.ai lab, running locally on their own infrastructure with no built-in refusals blocking the work.</p><p>It read the logs. It helped them finish the investigation. It worked.</p><h3>Why an Open Chinese Model Beating US Guardrails Is the Whole Story</h3><p>Sit with the irony for a second.</p><p>An American AI lab’s model autonomously hacks an American AI platform. American AI safety guardrails then block the American company from investigating the American hack. A Chinese open-source model, with fewer restrictions and full local deployment, is what actually stops the bleeding.</p><p>Hugging Face’s co-founder said the incident out loud: defenders need access to capable models without artificial restrictions, and that goes double for open ones. A King’s College London war-studies researcher put it more bluntly — when attackers face no usage policy but defenders are boxed in by one, that’s not a safety win, it’s a structural disadvantage.</p><p>That gap doesn’t close on its own. It gets wider as open models keep closing the capability distance while carrying fewer restrictions than their closed counterparts.</p><h3>So What Was Nvidia Actually Buying?</h3><p>Not just a model repository. Not just developer goodwill. Nvidia bought proof of concept.</p><p>The GLM 5.2 incident showed the market something concrete: in a live crisis, open, auditable, self-hostable infrastructure outperformed proprietary systems with guardrails baked in by someone else. That’s not a hypothetical open-source talking point anymore. It’s a documented case study, five weeks old, with 17,000 log files behind it.</p><p>Nvidia’s stack already covers the hardware and the software layer. What it didn’t own was the place where trust in open models gets built and tested in public. Hugging Face, running on Nvidia GPUs for its own paid hosting, dogfooding a Chinese open model to survive an American AI’s rogue attack, is now Nvidia’s proof that the open layer isn’t a hedge. It’s the safety net the entire ecosystem is going to need more of.</p><p>Jensen Huang’s open-source defense letter last month, signed by Microsoft, Amazon, Meta, and later Google and OpenAI, reads differently once you know what happened in July. Anthropic didn’t sign it.</p><h3>Final Thought</h3><p>Nvidia didn’t just buy a $12.9B community. It bought the company that proved, under real attack, why the open layer matters more than anyone wanted to admit. The chips were never really the moat. The trust was.</p><p><strong><em>I cover stories like this — AI infrastructure, open-source shifts, and what’s actually worth your attention — in Jargonese, weekly, no fluff.</em></strong></p><p><a href="https://jargonese.beehiiv.com/"><strong><em>Subscribe here →</em></strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=86536c3a664b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/activated-thinker/12-9b-deal-that-proves-open-models-just-became-the-safety-net-86536c3a664b">$12.9B Deal That Proves Open Models Just Became the Safety Net</a> was originally published in <a href="https://medium.com/activated-thinker">Activated Thinker</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vibe Coders Copy Code. Engineers Read Licenses First.]]></title>
            <link>https://medium.com/@SIDjART/vibe-coders-copy-code-engineers-read-licenses-first-6cce52a124dc?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/6cce52a124dc</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[software-license]]></category>
            <category><![CDATA[ai-coding]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Sat, 08 Aug 2026 03:35:14 GMT</pubDate>
            <atom:updated>2026-08-08T03:35:14.590Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cdhDYhX8p6dEhMCkJa4gRw.avif" /></figure><p>Not all open source is open source. Here’s what MIT, GPL, AGPL, and the rest actually mean before you ship AI-generated code.</p><h3>Open Source Licenses Explained for Vibe Coders (2026)</h3><p>A programmer forks an open source project, bolts on a payment gateway, and launches a SaaS calculator. It takes off. Then he gets arrested.</p><p>That story gets told as a joke on tech Twitter. It shouldn’t be.</p><p>You can’t fork or build on an open source project without knowing which license it carries. Different licenses come with different terms, and some of them bite exactly when you start charging money.</p><figure><img alt="a terminal with a freshly forked repo on one side, a LICENSE.md file glowing like a warning sign on the other" src="https://cdn-images-1.medium.com/max/1024/1*R7XoxsW1WA-NRLQyBAH9_Q.png" /></figure><h3>Why This Matters More Than It Did Two Years Ago</h3><p>AI writes most of the code beginners ship now. Almost nobody checks where that code came from.</p><p>Vibe coding means copying faster than ever. Faster copying means more license violations, most of them completely by accident.</p><p>Here’s the breakdown I wish someone had given me before I shipped anything.</p><h3>The Permissive Family: Do Almost Whatever You Want</h3><p><strong>MIT</strong> is the default of open source. Something like 95% of projects use it. It allows commercial use, modification, distribution, and relicensing. Fork an MIT project, close-source it, sell it under your own name. You’re fine.</p><p>You owe it three things: keep the copyright notice, keep the license text, don’t claim you wrote the original. That’s it.</p><p><strong>Apache 2.0</strong> gives you everything MIT gives you, plus a patent grant. If someone open sources a patented algorithm under Apache 2.0, you can use it without getting sued later for patent infringement. Kubernetes, Kafka, and Spark all run on this license.</p><p><strong>BSD</strong> is MIT with one extra clause: you can’t imply the original author endorses your fork. Numpy uses the 3-clause version. For almost every practical purpose, treat BSD like MIT.</p><h3>The Copyleft Family: Share Your Changes or Don’t Ship</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/1*I1g4ToFvTflVQOmtnfjwZA.png" /></figure><p><strong>GPL v2 and GPL v3</strong> are where the rules tighten. This is copyleft: distribute software built on GPL code, and your version has to stay under the same license and stay open source. You can’t relicense it into something proprietary. Linux runs on GPL v2. Bash, Emacs, and coreutils run on GPL v3, which adds a patent grant on top.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pNBGyZl2nPAJxJvkEgMkBA.png" /></figure><p><strong>LGPL</strong> is GPL scaled down for libraries. Build proprietary software on top of an LGPL library and you’re in the clear. Modify the library itself and distribute those changes, and only the changed parts need to stay LGPL. Your app stays yours.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*47WIwA4qiNCncJ50jToTQg.jpeg" /></figure><p><strong>AGPL</strong> closes the loophole that made SaaS companies love GPL. Under regular GPL, running modified code on a server isn’t “distribution,” so nothing gets triggered. AGPL says running modified AGPL code over a network counts as distribution anyway. Ship a modified AGPL project as a hosted service, and your source has to be public too.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/1*p1EUK1Yy3MRwzWbYjLBMcA.jpeg" /></figure><p><strong>MPL 2.0</strong> is Mozilla’s middle ground. It’s file-level copyleft: files you pulled from an MPL project stay open source, but the rest of your codebase can be as proprietary as you want.</p><h3>The Vibe Coder’s Blind Spot</h3><p>Here’s the part that should worry you. AI tools generate code trained on repositories that carry every license above, and the output never comes labeled. You can prompt your way into a function lifted from a GPL project and never find out.</p><p>The rule that actually protects you: if the thing you’re building makes money, audit what the AI-generated code actually is. If it’s clearly built on top of a specific open source project, go read that project’s LICENSE file before you wire up the payment gateway.</p><p><strong>Not all open source is open source. Read that again.</strong></p><h3>Final Thought</h3><p>Licenses aren’t paperwork for lawyers to argue over later. They’re the terms of the deal you’re making with everyone who wrote code before you. AI made copying effortless. It didn’t make the deal disappear.</p><p>A few things worth doing before your next ship:</p><ol><li>Read the LICENSE file of anything you fork before you write a line on top of it.</li><li>Treat GPL and AGPL code as off-limits for closed-source SaaS, unless you’re genuinely ready to open source your whole product.</li><li>Ask your AI tool where a generated block came from. If it can’t tell you, assume the strictest license applies.</li><li>Keep copyright and license notices intact, even on permissive code that lets you relicense everything else.</li></ol><p>I cover AI tools, workflows, and what’s actually worth your attention in Jargonese, weekly, no fluff. <a href="https://jargonese.beehiiv.com/">Subscribe here →</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6cce52a124dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Apple Just Raised Prices for No Reason. That’s Your First Sign the AI Bubble Is Real.]]></title>
            <link>https://medium.com/@SIDjART/apple-just-raised-prices-for-no-reason-thats-your-first-sign-the-ai-bubble-is-real-e2226015f6a2?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/e2226015f6a2</guid>
            <category><![CDATA[tech-stocks]]></category>
            <category><![CDATA[ai-bubble]]></category>
            <category><![CDATA[big-tech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[nvidia]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Sat, 25 Jul 2026 12:49:36 GMT</pubDate>
            <atom:updated>2026-07-25T12:49:36.462Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vT9d_v1YHCASk3yVBsL5OQ.jpeg" /></figure><p>On June 25, 2026, Apple did something it has never done before. Mid-cycle, with no new product on the table, it raised prices across the board. The MacBook Air went up 18%. iPad Pro jumped 20%. Apple TV shot up 54%.</p><p>Apple’s own explanation was blunt: it had never seen a component price increase this fast. The company that has spent two decades absorbing supply chain shocks so customers never notice them just admitted it couldn’t anymore.</p><p>That’s not a product story. That’s a warning shot.</p><h3>The $725 billion bet nobody can explain</h3><p>Here’s the number that should worry you more than any Nvidia earnings call.</p><p><strong>In 2020</strong>, before ChatGPT existed, the four biggest US tech companies (Amazon, Meta, Google, and Microsoft) spent a combined $90 billion on capital expenditure. By 2023 that number hit $147 billion. In 2025, it jumped to $410 billion. This year, 2026, it’s projected at $725 billion.</p><p>Six years. Eight times the spend. And almost all of it is going into one thing: AI data centers.</p><p>A single large AI data center costs somewhere between $10 billion and $25 billion to build once you account for the chips, the cooling, the power infrastructure, and the concrete. One building can hold 100,000 Nvidia GPUs, and each GPU runs $30,000 to $40,000. Do the math and you’re looking at $3 to $4 billion in chips alone before a single watt of power gets consumed.</p><p>The justification everyone gives is data growth. In 2010, the world generated about 2 trillion gigabytes of data. By 2026, that number is projected to hit 221 zettabytes, more than a hundred times higher. The story writes itself: more data means more compute means more spending means more revenue eventually.</p><p>Except “<strong>eventually</strong>” is doing a lot of work in that sentence.</p><h3>Big tech is betting 94 cents of every dollar it earns</h3><p><a href="https://www.pimco.com/us/en/insights/ai-credit-expansion-assessing-the-micro-and-macro-risks"><strong>PIMCO</strong></a> ran the numbers on what this spending actually means for big tech’s balance sheets, and the figure is staggering. Capital expenditure will consume 94% of operating cash flow over the next two years. In 2023, that ratio was 40%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FP6Nc46wAlLNOkjy" /></figure><p>Read that again. If big tech earns $100, it’s putting $94 straight back into AI infrastructure. Six dollars is left over for dividends, buybacks, raises, and everything else that keeps a company running normally.</p><p><a href="https://fortune.com/2026/06/25/what-bubble-jpmorgan-5-5-trillion-ai-capex-explosion-profitable-for-now/"><strong>JPMorgan </strong></a>did the follow-up analysis that matters more. If you’re an investor demanding even a conservative 10% return on this kind of spend, AI needs to generate roughly $650 billion in annual revenue to justify itself.</p><p><strong>Here’s what AI is actually earning.</strong></p><p>OpenAI brings in about $25 billion a year and loses $14 billion doing it. Anthropic is on pace for somewhere around $26 billion this year, trending toward $47 billion if the run rate holds, and is still losing money. Add Google’s Gemini revenue and the whole industry is sitting at roughly $75 billion combined.</p><p>$650 billion needed. $75 billion earned. That’s not a rounding error, that’s a nine-to-tenfold gap between what AI is spending and what it’s bringing back.</p><p><strong><em>Palantir CEO Alex Karp said it plainly on July 1: enterprises are paying for tokens that create no value, and the models have been oversold. Coming from a company that sells to the CIA, the US government, and half of Fortune 500, that’s not a throwaway line.</em></strong></p><h3>Enterprises are already walking away from the bill</h3><p>This is the part that should worry the AI labs specifically, because the entire trillion-dollar valuation story rests on one assumption: enterprises will keep paying more for AI, forever.</p><p>That assumption cracked in June 2026.</p><p>McKinsey found that 73% of enterprise AI deployments are failing to hit their projected ROI. BCG puts the number of companies seeing substantial return at just 5%. MIT’s research shows a 95% failure rate on measurable financial returns, and only 29% of executives can even measure their AI ROI in the first place.</p><p>Then the real-world stories started stacking up. Lindy, a 25-person AI startup, revealed its Anthropic Claude API bill had exceeded its entire payroll. The founder switched 100% of traffic to DeepSeek and cut costs by 90% overnight. Uber’s CTO admitted publicly that the company blew through its entire annual AI budget in four months.</p><p>When your customers start hunting for the cheapest possible substitute instead of the best one, you’re not selling a premium product anymore. You’re selling a commodity, and commodity pricing crushes margins.</p><h3>The memory chip shortage is how this reaches your wallet</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*WgL0xJFPM9Vex0wAwn3ZzA.jpeg" /></figure><p>You don’t need to own Nvidia stock to feel this. Samsung, SK Hynix, and Micron control 90% of the world’s memory chip supply, and AI data centers pay roughly ten times more per module than consumer electronics companies do. So the three of them redirected 93% of their production toward AI-grade memory.</p><p>DRAM prices are up 171% year over year as of March 2026. DDR5 memory has quadrupled since September 2025. Dell’s CEO said the price of a gigabyte of DRAM went from $0.43 to $2.39 in six months, a five-and-a-half-times jump. That’s the cost that landed on Apple’s laptops and is now landing on yours, whether you buy a Mac, a Dell, or an Xbox.</p><p>Capital flows to whoever pays the most. Right now, that’s AI data centers, and everyone buying a consumer electronics product this year is quietly subsidizing that shift.</p><h3>Why this looks exactly like the dot-com crash</h3><p>The pattern behind this has a name in economics: the capital cycle. High returns attract capital. Capital keeps flowing until massive overcapacity gets built. Overcapacity collapses under its own weight. A handful of survivors pick up the wreckage cheap and make a fortune once real demand finally shows up.</p><p>The telecom boom of the late 1990s followed this script exactly. Congress passed the Telecommunications Act in 1996, the internet was undeniably the future, and telecom companies poured more than $500 billion into fiber optic cable in five years. Global Crossing hit a $47 billion valuation without ever posting a profitable year. Corvis pulled off a $1.1 billion IPO with zero revenue.</p><p>By early 2000, only 2.7% of all that installed fiber was actually carrying data. Bandwidth prices collapsed 90%. WorldCom hid $3.8 billion in expenses to fake profitability and still filed the largest bankruptcy in US history at the time. The crash wiped out $2 trillion in market value.</p><p>Then, a few years later, YouTube happened. Streaming happened. Cloud storage and smartphones happened, and all that “wasted” fiber became the literal backbone of the modern internet, bought for pennies on the dollar by whoever survived the crash.</p><p>The technology was real. The bubble still burst, because the timeline of demand didn’t match the pace of the spending.</p><h3>So is 2026 different?</h3><p>Partially, and it’s worth being honest about where the comparison breaks down. The dot-com companies were mostly unprofitable and funded by debt. Nvidia posted $120 billion in net income last year. Microsoft, Google, and Amazon are among the most profitable companies that have ever existed. At the 2000 peak, the Nasdaq 100’s forward price-to-earnings ratio hit roughly 60x. Today it sits closer to 26x, elevated but nowhere near that territory.</p><p><strong>Anyone telling you with total confidence that this is a bubble is guessing. Anyone telling you with total confidence that it isn’t is guessing too.</strong></p><p>The honest answer sits in between: the technology and the revenue are both real, and the argument isn’t about whether AI matters. It’s about whether the price tag attached to it makes sense yet.</p><p><strong>Two paths sit ahead.</strong></p><p>In the first, the spending gets reined in, the overcapacity gets exposed, and a correction ripples through every sector that’s been riding the AI capex wave, including the American jobs tied to building and servicing it. In the second, the labs race toward actual profitability, token prices rise to match real costs, and the cheap AI tools everyone’s gotten used to become a premium product reserved for whoever can afford them.</p><p>Either way, the free ride that consumers and enterprises have been on for the past three years is closer to its end than its beginning.</p><h3>Final Thought</h3><p>The scariest part of this story isn’t the size of the numbers. It’s how normal they’ve started to feel. A $725 billion annual bet with a $650 billion revenue gap should be front-page news every single day until it’s resolved one way or the other. Instead it’s background noise, wedged between earnings calls and the next model release.</p><p>Watch what enterprises do with their AI budgets over the next two quarters, not what the labs say on stage. That’s where the real answer to this question is going to show up first.</p><p>I track AI infrastructure spending, tool pricing shifts, and what’s actually worth paying for in Jargonese, three times a week, no hype.</p><p><a href="https://jargonese.ai/">Subscribe here →</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e2226015f6a2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[10 GitHub Repos That Let You Scrape the Entire Internet for Free (2026)]]></title>
            <link>https://medium.com/@SIDjART/10-github-repos-that-let-you-scrape-the-entire-internet-for-free-2026-3d6ec9b83572?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d6ec9b83572</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[web-scraping-tools]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[github-repositories]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 09:18:54 GMT</pubDate>
            <atom:updated>2026-07-08T09:18:54.080Z</atom:updated>
            <content:encoded><![CDATA[<p>No contracts, no API keys, no per-page fees. Clean data from any website — all open source.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tltCyDEKB2b2pbrq4x2ulQ.png" /></figure><p>Data access is the most quietly expensive part of building with AI.</p><p>Paid APIs charge per request. Enterprise scraping tools require a sales call before you see a price. Most builders either pay up or hit a wall and work around it.</p><p>These 10 GitHub repositories change that math. Every one of them handles at least one of the three capabilities paid tools charge for: JavaScript rendering, proxy rotation, or structured output. Most handle all three. None of them cost anything to self-host</p><p>Save all of them.</p><h3>The 10 Repos</h3><ol><li><strong>Firecrawl</strong> — <a href="https://github.com/firecrawl/firecrawl">https://github.com/firecrawl/firecrawl</a></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JVyvGlPeszYfI5alC-BtlQ.png" /></figure><p>130,000 stars. One of GitHub’s top 100 repositories.</p><p>Take any URL and it follows every subpage, renders JavaScript, and returns clean markdown or structured JSON that an AI agent can parse immediately. The open-source version is fully self-hostable; the commercial API is optional. Half the AI startups you’ve heard of run this quietly in the background.</p><p><strong>2. Crawl4AI</strong> — <a href="https://github.com/unclecode/crawl4ai">https://github.com/unclecode/crawl4ai</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W1iEZztWkJVNk67fZPIztQ.png" /></figure><p>GitHub’s most-starred scraper with 51,000 stars, and the origin story is worth knowing: a developer got charged $16 for a service he figured he could build in a weekend, so he built it in a few days. It converts any website to LLM-ready markdown faster than most paid alternatives, with no accounts, no API keys, and no per-page billing. The developer got fed up. Now millions of people benefit.</p><p><strong>3. browser-use</strong> — <a href="https://github.com/browser-use/browser-use">https://github.com/browser-use/browser-use</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TRleR7QIW04S5kLwHJnZBQ.png" /></figure><p>95,000 stars in under a year. Two ETH Zurich researchers built this to control a browser the way a human does: clicking, scrolling, filling forms, logging into sites it’s never seen before. Feed it a task in natural language and it navigates. Think of it as Playwright, but with a reasoning layer.</p><p><strong>4. Crawlee</strong> —<a href="https://github.com/apify/crawlee">https://github.com/apify/crawlee</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9RLAw02qhbdFBghjTHyyMQ.png" /></figure><p>This is the one agencies don’t advertise.</p><p><strong>Proxy rotation, automatic retries, browser fingerprint spoofing, and queue management</strong> are all included in one professional-grade framework. Scraping companies charge enterprise clients for services built on top of this exact stack. The framework itself is free and open source.</p><p><strong>5. ScrapeGraphAI</strong> —<a href="https://github.com/ScrapeGraphAI/Scrapegraph-ai">github.com/ScrapeGraphAI/Scrapegraph-ai</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L9NgranpG7-Z3bv0gT6B2Q.png" /></figure><p>Instead of writing selectors, you describe what you want in plain English and the LLM figures out how to extract it. No hardcoded CSS selectors means it keeps working when a site’s structure changes, which is the thing that silently breaks every traditional scraper after a site update.</p><blockquote><em>I cover AI tools, open-source finds, and what’s actually worth your attention in Jargonese — three times a week, no fluff.</em> <a href="https://jargonese.beehiiv.com/"><em>Subscribe here →</em></a></blockquote><p><strong>6. Playwright</strong> —<a href="https://github.com/microsoft/playwright">https://github.com/microsoft/playwright</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KAhKUz2TkfVmVQMwODxYrA.png" /></figure><p>Microsoft’s Playwright, with over 88,500 stars, is the modern successor to Puppeteer, developed by the same team that transitioned from Google. It offers cross-browser automation (Chromium, Firefox, WebKit) with advanced features like auto-wait mechanisms and web-first assertions. Playwright has become the professional standard for end-to-end testing and automation, demonstrating significant market share growth in 2026, surpassing competitors like Cypress and Selenium.</p><p><strong>7. Puppeteer</strong> — <a href="https://github.com/puppeteer/puppeteer">https://github.com/puppeteer/puppeteer</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-Yed1HjlxG1vrO8FsLwRXQ.png" /></figure><p>Google’s original headless Chrome library, Puppeteer, maintains a strong presence with over <strong>94,400 stars</strong>.</p><p>It provides a high-level API for controlling Chrome and Firefox, backed by a mature ecosystem and extensive community support. Puppeteer remains a battle-tested standard for Chromium-specific workflows, with recent updates (v22+) improving support for WebDriver BiDi and Firefox.</p><p><strong>8. Scrapy</strong> — <a href="https://github.com/scrapy/scrapy">https://github.com/scrapy/scrapy</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w1E6Cpj0oDGtaoBq97hIog.png" /></figure><p>Since 2008, Scrapy has been the Python standard for web scraping, accumulating over 53,000 stars. Maintained by Zyte (formerly Scrapinghub), it features robust middleware, extensible pipelines, and is designed for large-scale structured data extraction.</p><p>Scrapy represents the battle-tested “<strong>old guard</strong>” that continues to dominate complex Python scraping tasks, now integrating with Zyte API for AI-powered extraction in existing pipelines.</p><p><strong>9. Stagehand</strong> — <a href="https://github.com/browserbase/stagehand">https://github.com/browserbase/stagehand</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u9pp9YYEQkrdPoMNkv75ng.png" /></figure><p>An AI layer built on top of Playwright that handles dynamic interfaces where traditional selectors consistently break. Useful specifically for sites that actively fight bots — the ones where your selectors start returning null after a deploy and you have no idea why. Sits between Playwright and browser-use in terms of control versus autonomy.</p><p><strong>10. Newspaper4k</strong> — <a href="https://github.com/codelucas/newspaper">https://github.com/codelucas/newspaper</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PzPpgt41EmiDqkuMbSVNAQ.png" /></figure><p>Does one thing with unusual quality: extracts clean article content from any news or blog URL. Title, body text, authors, publish date, images — stripped of ads, navigation, and boilerplate. If you’re building a content intelligence pipeline or a news monitoring tool, this is the fastest starting point that exists.</p><h3>Bottom Line</h3><p>Commercial data APIs aren’t going anywhere. Teams with compliance requirements, SLAs, or no interest in maintaining infrastructure still have reasons to pay.</p><p>For everyone else — developers building pipelines, solopreneurs, students, anyone running AI research or competitive intelligence workflows — the open-source alternatives are genuinely better than what you’d pay for at the $20 to $50 per month tier.</p><h4><strong>Read that again.</strong></h4><p>The free scraping stack in 2026 outperforms the paid entry-tier products that existed in 2023. That’s not a hot take; it’s what 130,000 GitHub stars on a free tool looks like versus a $16 monthly invoice.</p><p><em>I cover AI tools, open-source finds, and what’s actually worth your attention in Jargonese — three times a week, no fluff.</em> <a href="https://jargonese.beehiiv.com/"><em>Subscribe here →</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d6ec9b83572" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Best $20 AI Plan in 2026: ChatGPT Plus, Claude Pro, or Google AI Pro?]]></title>
            <link>https://medium.com/@SIDjART/the-best-20-ai-plan-in-2026-chatgpt-plus-claude-pro-or-google-ai-pro-e9d894b8c327?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9d894b8c327</guid>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Fri, 12 Jun 2026 16:53:36 GMT</pubDate>
            <atom:updated>2026-06-12T16:55:28.911Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NL3gDuGaSW03WJmsITDxsQ.png" /><figcaption><strong>ChatGPT Plus, Claude Pro, or Google AI Pro?</strong></figcaption></figure><p><strong>Almost every major AI company now offers a $20 monthly plan.</strong></p><p>For many, the question isn’t whether to use AI, but which of these premium subscriptions delivers the most value. With options like ChatGPT Plus, Claude Pro, and Google AI Pro, discerning the best fit for your workflow can be a challenge.</p><p>This article provides a comprehensive comparison of these three leading AI plans, examining their features, model capabilities, and usage limitations. The goal is to equip you with the insights needed to make an informed decision, ensuring your investment in AI truly enhances your productivity and creative output.</p><h3>The $20 AI Maze: Why Choosing is Hard</h3><p>Advanced AI has rapidly transitioned from a niche technology to an essential utility. The market is now saturated with powerful tools, each vying for your monthly subscription.</p><p>While all promise enhanced capabilities, their underlying strengths and weaknesses vary significantly. Understanding these distinctions is crucial to selecting a plan that aligns with your specific needs, rather than simply subscribing to the most popular option.</p><h3>Video &amp; Image Generation: Who Wins the Visuals War?</h3><p>For users whose workflows demand video generation, <strong><em>Google AI Pro (Gemini)</em></strong> stands as the sole contender among the three. It offers robust video creation capabilities, including the option to upload avatars and utilize various templates for motion graphics. ChatGPT has discontinued its video generation features, and Claude has never offered them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kYhw8728FSPtBexIsTtc7g.jpeg" /></figure><p><strong><em>Therefore, for video-centric tasks, Gemini is the unequivocal choice.</em></strong></p><p>Image generation presents a more competitive landscape.</p><p>Claude currently lacks any integrated image generation tool. Between ChatGPT and Gemini, both provide capable solutions.</p><p>ChatGPT’s image model is frequently cited as a global leader, offering templates and advanced editing functionalities for existing images with exceptional results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*S359_iZhRHqUl3i57SNMFA.jpeg" /></figure><p>Gemini also performs commendably with templates and image editing, though its output may not consistently reach the pinnacle of ChatGPT’s quality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5VaEbk7IM3-vdUSkRkwYpg.png" /></figure><p>Consequently, if top-tier image generation is a priority, ChatGPT is the preferred option. For occasional image generation where cutting-edge quality is not paramount, Gemini serves as a perfectly viable alternative. This distinction often influences the overall assessment of each platform’s utility.</p><h3>General Chat Experience: Wall of Text vs. Structured Clarity</h3><p>When evaluating general AI chatbot performance, user preference plays a significant role, largely influenced by output style. All three models capably handle standard chat queries, but their presentation of information differs markedly.</p><p>Claude, for instance, can be less intuitive for casual conversations. Its grasp of real-world knowledge may appear less comprehensive than its rivals, and its responses often manifest as dense blocks of text, hindering readability and comprehension.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xVKcQzq4jtVQIKlyn7rIaw.png" /></figure><p>ChatGPT also tends to generate lengthy outputs, though its settings offer some degree of structural refinement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xWPsKN-60vXzz_ZlBBiQ6A.png" /></figure><p>Gemini, conversely, typically delivers the most structured and easily digestible outputs for general chat interactions. Its organized presentation facilitates quick scanning and understanding, which is particularly beneficial for users seeking rapid answers or concise summaries. While Claude may excel in specialized applications, for everyday conversational AI, Gemini often provides a more user-friendly experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*855wJKXyT9EsoTjRvRysOg.png" /></figure><h3>Third-Party Integrations &amp; AI Agents: Connecting Your Workflow</h3><p>The capacity for seamless integration with external tools and access to local files can profoundly impact productivity. In this domain, Claude and ChatGPT offer comprehensive solutions, while Gemini exhibits certain limitations.</p><p>Claude features an extensive array of third-party integrations, accessible via itsconnectors. This includes popular services such as Gmail, Google Drive, Notion, and Google Calendar. A key advantage of Claude is its integration with Zapier, enabling connectivity with a vast ecosystem of tools even if they are not directly listed. This makes Claude exceptionally versatile for automating diverse workflows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UHdWBPrqiCs1uGB-" /></figure><p>ChatGPT provides a similar feature through its ‘apps,’ offering a broad range of integrations. While it supports many major services like Gmail, Calendar, and Drive, it currently lacks Zapier integration. Consequently, ChatGPT’s integration ecosystem, though extensive, does not match the comprehensive reach of Claude’s.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UvQXmB6SzeuB4M3_wdmXWw.png" /></figure><p>Google Gemini, in contrast, primarily focuses on integration within the Google ecosystem. Its ‘connected apps’ predominantly include Google products such as Gmail, Calendar, Docs, Drive, and YouTube. While this offers seamless functionality for users deeply invested in Google Workspace, it significantly restricts third-party connectivity. For those requiring integration with a wide variety of external tools, Gemini may not be the optimal choice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6l8p6sS5csj-OmDshJ48pw.png" /></figure><p>Beyond direct integrations, both <strong><em>Claude and ChatGPT</em></strong> offer powerful AI agents designed for knowledge work. Claude’s desktop application features Claude Co-work, which can access, edit, and create files within a specified local folder. This local file access is a significant advantage for privacy and workflow integration, as it is not cloud-based. Co-work proves highly effective for tasks ranging from invoice management and video organization to business journaling and code generation for websites.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*rP9neKvBGUV5fazALYNMjg.jpeg" /></figure><p>ChatGPT’s equivalent, Codex, is also a downloadable application that interacts with local files. Codex performs similarly to Co-work, and in some areas, it demonstrates superior capabilities. However, Co-work is often perceived as slightly more mature and user-friendly for general knowledge work, primarily because Claude’s application allows for a clear distinction between chat, Co-work tasks, and coding, a separation that Codeex currently lacks. Both represent excellent options for users needing an AI agent to manage local files.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wbEzm_GMk8QIHFAbQlwqbQ.png" /></figure><p>Gemini also includes an agent known as Spark. However, Spark currently faces two notable limitations. Firstly, it operates as a cloud-based service, meaning it does not access local files, although this functionality is anticipated in future updates. Secondly, Spark is exclusively available on Gemini’s higher-tiered plans, making it inaccessible to users on the $20 Pro subscription. These factors render Spark less functional and accessible compared to Claude Co-work and Codeex at the equivalent price point.</p><h3>Coding &amp; Unique Features: Beyond the Chatbot</h3><p>For individuals engaged in coding, the consensus points to Claude and ChatGPT as the leading platforms. While specific preferences may vary, both are widely recognized for their strong performance in coding tasks. Utilizing their respective agent applications, Co-work or Codex, further enhances their utility for developers.</p><p><strong>Each platform also distinguishes itself with unique features.</strong></p><p>Claude, for example, offers <strong>Claude Design</strong>, a tool for prototyping applications, creating slide decks, and generating motion graphics. This feature can be particularly efficient for visual content creation, often requiring minimal prompts to produce polished results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3vzdPTKW-dO_2NxYn6gxMg.png" /></figure><p>However, in terms of the sheer breadth of unique features, Google AI Pro stands out. A subscription to Google AI Pro unlocks a comprehensive suite of tools beyond the core Gemini interface:</p><p>•<strong>Google Flow</strong>: A dedicated application for creating images and videos, separate from Gemini, offering distinct usage limits and improved organization for visual assets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZDhyf6pc4DZ4I1s_UXDUMw.png" /></figure><p>•<strong>Flow Music</strong>: An AI-powered tool for generating music, also with its own usage limits, providing capabilities for creative audio production.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0HpFjZ07V72yUqlIRlhSQA.png" /></figure><p>•<strong>Notebook LM:</strong> Considered one of Google’s most innovative AI tools, it allows for the creation of notebooks to organize research topics. Users can ground these notebooks with various sources, including websites, YouTube videos, and uploaded files, ensuring that AI responses are contextually relevant to the provided information. The Pro plan enhances this free tool with higher usage limits.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*czH1HBJw4PCrESdP5EYrTQ.png" /></figure><p>•<strong>YouTube Premium Light</strong>: This benefit provides an ad-free viewing experience for most YouTube videos and enables background playback and downloads on mobile devices.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*TJ86wpznYro-IyLzjF7peg.png" /></figure><p>•<strong>Gemini in Gmail &amp; Docs</strong>: AI assistance integrated directly into Google Workspace applications, facilitating email composition and document creation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o5I_QTp_cUz1CZ8h-SSMYw.avif" /></figure><p>•<strong>Google Health Premium</strong>: An AI-powered personal health coach designed to assist with calorie tracking, workout planning, and recovery strategies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MDPq-LZvW6ha0pqUxFUibg.jpeg" /></figure><p>•<strong>Google Home Premium</strong> and additional benefits, such as $10 in monthly Google Cloud credits and 5 terabytes of cloud storage across Gmail, Drive, and Photos, with family sharing options for up to five users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kguOQ-n6tuHDGwRED2CCNg.avif" /></figure><p>The Google AI Pro plan offers substantial value, particularly for users deeply integrated into the Google ecosystem who can leverage these extensive additional services.</p><h3>Usage Limits: The Hidden Cost</h3><p>The utility of any AI platform is significantly impacted by its usage limits. These caps dictate the extent to which users can interact with the tools and models within a given timeframe, and they vary considerably across the three platforms.</p><p>Both Claude and Gemini implement dynamic usage limits. These limits are influenced by the complexity of prompts, the specific tools and models employed, and the length of conversations. More demanding requests will naturally consume the allocated usage more rapidly.</p><p>ChatGPT provides more transparent usage metrics: <strong>160 messages every 3 hours when utilizing its most powerful model</strong>. The 3-hour reset period is a notable advantage, as Claude’s primary usage window resets every 5 hours, offering ChatGPT users more frequent opportunities to refresh their limits throughout the day. Furthermore, Codex, ChatGPT’s agent application, operates with separate usage limits. This means that even if Codeex usage is exhausted, the main ChatGPT interface remains accessible. This contrasts with Claude Co-work, where usage is tied more directly to the main Claude limits.</p><p>Observations suggest that <strong>Claude’s limits are often reached the fastest, followed by Gemini, with ChatGPT</strong> users experiencing limitations least frequently. However, this is not the complete picture. Both<strong> ChatGPT and Gemini typically downgrade users to a less powerful model</strong> rather than imposing a complete lockout once limits are reached. For Google AI Pro, many of the supplementary tools, such as Flow, Flow Music, and Notebook LM, do not count against the primary Gemini usage limit. This strategic distribution of AI activities across Google’s offerings can significantly mitigate the impact of usage caps.</p><h3>The Verdict: Which $20 AI Plan is Best for You?</h3><p>For individuals with extensive AI usage requirements, it is important to note that a single $20 plan may not suffice. Many power users opt for higher-tiered subscriptions, such as Claude Max or Google AI Ultra, to accommodate demanding workflows, particularly for advanced features like Claude Co-work and comprehensive image/video generation.</p><p>Selecting the optimal $20 AI plan depends entirely on individual needs and usage patterns:</p><p><strong>Choose ChatGPT Plus if</strong>: High-quality image generation is a top priority, clear usage limits are preferred, and a robust integration ecosystem (albeit without Zapier) is valued.</p><p><strong>Choose Claude Pro if:</strong> Human-like writing capabilities are essential, extensive third-party integrations (including Zapier) are required, and a powerful local AI agent for knowledge work (Claude Co-work) is desired.</p><p><strong>Choose Google AI Pro if:</strong> Deep integration within the Google ecosystem is a key factor, video generation is needed, a wide array of unique features (Flow, Flow Music, Notebook LM) is appealing, and bundled benefits like YouTube Premium Light and cloud storage are advantageous.</p><p>Each of these plans offers substantial value for its $20 monthly fee, but they cater to distinct priorities. Consider your primary use cases, your existing digital environment, and the extent to which AI is integrated into your daily tasks. This detailed comparison aims to provide the clarity necessary for an informed decision.</p><p><em>I cover AI tools, workflows, and what’s actually worth your attention in Jargonese — weekly, no fluff. </em><a href="https://jargonese.beehiiv.com/"><em>Subscribe here →</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9d894b8c327" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Best Free LLM Setup for Hermes Agent in 2026. Here’s My Hermes Agent Setup.]]></title>
            <link>https://medium.com/activated-thinker/the-best-free-llm-setup-for-hermes-agent-in-2026-heres-my-hermes-agent-setup-7114e0feff85?source=rss-2ea801d4c512------2</link>
            <guid isPermaLink="false">https://medium.com/p/7114e0feff85</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[productivity-hacks]]></category>
            <category><![CDATA[open-source-ai]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[free-api]]></category>
            <dc:creator><![CDATA[Sidj]]></dc:creator>
            <pubDate>Fri, 29 May 2026 17:37:05 GMT</pubDate>
            <atom:updated>2026-05-29T17:37:05.816Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x1ojSoz1-kjT9xWU4cQ86Q.png" /></figure><p>I spent three hours yesterday fighting with AWS Bedrock permissions.</p><p>It was a mess of IAM roles and region-locked models that made me want to throw my laptop out the window.</p><p>I just wanted my Hermes Agent to work without draining my bank account every time I asked it to refactor a React component.</p><p><strong><em>Then it hit me. I was paying for a headache.</em></strong></p><p>The “<strong>pro</strong>” setup involving Claude Sonnet via Bedrock is great if you’re a multi-million dollar enterprise. But if you’re a builder or a student, you’re probably overcomplicating your life for zero extra gain.</p><p>I scrapped the whole thing and rebuilt my Hermes stack for $0. It’s faster and I haven’t seen a credit card charge in weeks.</p><blockquote><a href="https://jargonese.beehiiv.com/">Stop overpaying and join Jargonese →</a></blockquote><figure><img alt="Hermes Agent logo and a successful connection to a primary model" src="https://cdn-images-1.medium.com/max/1024/1*GSXgYUVfGFmcd5_o8IYb-g.png" /></figure><h3>The Primary Engine: GPT-5 via Copilot</h3><p>Most people forget they already have the “<em>Holy Grail</em>” of models sitting in their IDE.</p><p>If you have the <strong>GitHub Student Developer Pack</strong> or a s<strong>tandard Copilot subscription</strong>, you have access to the GPT-5 ecosystem. Hermes natively supports GitHub Copilot as a provider via OAuth.</p><p>I switched my primary model to gpt-5 via the Copilot provider. It handles tool calling with a level of precision that makes older models look like toddlers. No API keys to manage, no usage limits that feel like a ticking time bomb.</p><p>You just run hermes auth, choose Copilot, and you’re in. It’s the closest thing to “<strong>set it and forget it</strong>” in the AI world right now.</p><h3>The Safety Net: OpenRouter and Nemotron</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zlaf9g0H2aB87p1qFBgtVQ.png" /><figcaption>Openrouter’s free model</figcaption></figure><p>Even the best providers go down. Or they rate-limit you right when you’re in the middle of a flow state.</p><p>That’s why you need a fallback that doesn’t suck. I’ve been using OpenRouter’s free tier, specifically the <strong>nvidia/nemotron-3-super-120b-a12b:free</strong> (or whatever the current flavor of Nemotron is this week).</p><p>Nemotron is the “<strong>hidden gem</strong>” of free models.</p><p>It’s surprisingly good at following complex instructions in an agentic loop. While other free models hallucinate the moment you give them a multi-step task, Nemotron stays on the rails.</p><p>It’s my “<strong>break glass in case of emergency</strong>” model.</p><h3>Configuring the Fallback Chain</h3><p>Hermes is powerful because it doesn’t just fail. It pivots.</p><p>I configured my config.yaml to treat Copilot as the primary and OpenRouter as the secondary.</p><p>If Copilot hits a snag, Hermes automatically switches to Nemotron without dropping the conversation context.</p><p>Here is what the core of my config looks like:</p><p><strong>YAML</strong></p><blockquote>model:</blockquote><blockquote>default: gpt-5 provider: copilot</blockquote><blockquote>fallback: model: nvidia/nemotron-3-super-120b-a12b:free provider: openrouter</blockquote><p>It’s simple. It’s clean. And it works.</p><h3>The Era of “Good Enough” Free AI</h3><p>We’ve reached a point where “free” doesn’t mean “broken” anymore.</p><p>Between the Google AI Studio free tier, Groq’s insane speeds, and the Copilot integration, there is no reason to be paying $20/month for individual API access unless you’re running production-scale traffic.</p><p>The friction I felt with Bedrock wasn’t a technical hurdle; it was a sign that I was using the wrong tool for the job.</p><p>If you’re building, stop worrying about the “best” model. Focus on the workflow. A free model that works beats a frontier model locked behind a permission error every single time.</p><h3>Final Thought</h3><p>The best setup isn’t the one with the most expensive models. It’s the one that lets you build without thinking about the cost of every prompt.</p><p>Switch to the Copilot/OpenRouter stack. Your bank account — and your sanity — will thank you.</p><p>The AI world moves too fast to keep up with every new model release. I filter the noise and send you the workflows that actually matter every Tuesday.</p><p>“<strong><em>Most people will keep paying for API credits they don’t need. You don’t have to be one of them. I send out the exact ‘zero-cost’ stacks I’m using every Tuesday morning</em></strong>.</p><blockquote><a href="https://jargonese.beehiiv.com/">Stop overpaying and join Jargonese →</a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7114e0feff85" width="1" height="1" alt=""><hr><p><a href="https://medium.com/activated-thinker/the-best-free-llm-setup-for-hermes-agent-in-2026-heres-my-hermes-agent-setup-7114e0feff85">The Best Free LLM Setup for Hermes Agent in 2026. Here’s My Hermes Agent Setup.</a> was originally published in <a href="https://medium.com/activated-thinker">Activated Thinker</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>