<?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 Pedram on Medium]]></title>
        <description><![CDATA[Stories by Pedram on Medium]]></description>
        <link>https://medium.com/@7pedram?source=rss-1acb6289f136------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*2aDcDviiUp8SixYHvNZ0sw.jpeg</url>
            <title>Stories by Pedram on Medium</title>
            <link>https://medium.com/@7pedram?source=rss-1acb6289f136------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 15:54:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@7pedram/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[Stop Asking AI Chatbots for Yes/No Answers: Meet Google’s Decision 1 and Jev]]></title>
            <link>https://medium.com/@7pedram/stop-asking-ai-chatbots-for-yes-no-answers-meet-googles-decision-1-and-jev-530952fbef7d?source=rss-1acb6289f136------2</link>
            <guid isPermaLink="false">https://medium.com/p/530952fbef7d</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Pedram]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 15:29:07 GMT</pubDate>
            <atom:updated>2026-11-19T15:29:07.459Z</atom:updated>
            <content:encoded><![CDATA[<h4>Google just released Gemini Decision 1, joining TypeSafe AI’s Jev in a new wave of models built to decide, not chat. Here is how they work in plain English, with real project examples.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7C9VDy2W_42G6vlm6Wk_eQ.png" /><figcaption><em>Gemini Decisions API official launch visual. Source: Google Keyword Blog (blog.google)</em></figcaption></figure><p>Imagine hiring a bestselling novelist just to sort your morning mail into two piles: <strong>Bills</strong> and <strong>Junk</strong>.</p><p>They will do the job. But before putting an envelope in a pile, they might write a three-paragraph speech explaining <em>why</em> it looks like a bill, charge you by the word, and occasionally invent a third pile called <em>“Philosophical Letters”</em> that breaks your filing cabinet.</p><p>Funny as that sounds, that is how most teams have been using AI for the last three years.</p><p>Whenever an app needs a simple judgment call (like <em>“Is this comment polite?”</em> or <em>“Which department should get this ticket?”</em>), we send the whole thing to a giant AI chatbot and beg it to reply with a single word.</p><p>Now, that is finally changing.</p><h3>A New Wave: Google Decision 1 and TypeSafe AI’s Jev</h3><p>On <strong>November 11, 2026</strong>, Google released a brand-new model called <strong>Gemini Decision 1</strong> (through its new Gemini Decisions API). Unlike regular Gemini models, it does not write a single word of text. It is built for one thing only: making fast, structured decisions.</p><p>And Google is not alone here.</p><p>We have been watching this trend build momentum since <strong>TypeSafe AI</strong> launched <strong>Jev</strong> back on <strong>September 15, 2026</strong> (followed by a big <strong>Jev 1.14</strong> update on <strong>October 14, 2026</strong>). TypeSafe calls Jev a <strong>&quot;System One&quot; model</strong>, named after the fast, automatic gut-instinct thinking your brain does before slow, heavy reasoning kicks in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vww7R52gGNmhNoWdjM1Ztg.png" /><figcaption><em>TypeSafe AI’s System One &amp; Jev launch artwork. Source: TypeSafe AI Blog (typesafe.ai)</em></figcaption></figure><p>Now that both a well-funded startup (TypeSafe AI) and a tech giant (Google) are building in this exact same lane, it is clear that <strong>decision models</strong> are becoming a core part of modern apps.</p><p>Wait: if they don&#39;t write text, what do they actually do? Let&#39;s break it down simply.</p><h3>What Are Decision Models, in Plain English?</h3><p>A normal AI model (like standard ChatGPT, Claude, or Gemini) is built to talk. It writes answers one word at a time.</p><p>Even if you only want a Yes or No, a chatbot still has to &quot;think out loud&quot; in words. Sometimes it replies &quot;Yes.&quot; Sometimes it replies &quot;Sure, based on the context provided, the answer is yes!&quot;, and suddenly your app crashes because your code only expected a single word.</p><p><strong>Decision models like Decision 1 and Jev skip the writing completely.</strong></p><p>You give them a piece of content, ask a clear closed-ended question, and in a fraction of a second (~100 to 250 milliseconds), they return two things:</p><ol><li><strong>The pick</strong> (your answer)</li><li><strong>A confidence score</strong> from 0 to 1 showing how sure the model is (for example, 0.94 means 94% sure).</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PiaX2PdaiegDiIVRW7bwCA.png" /><figcaption><em>How Gemini Decision 1 processes inputs and returns answers with confidence scores in one pass. Source: Google Keyword Blog (blog.google)</em></figcaption></figure><blockquote><strong>The simplest way to remember it:</strong> Chatbots are built to <em>write</em>. Decision models are built to <em>choose</em>.</blockquote><h3>The 3 Questions You Can Ask Them (With Easy Examples)</h3><p>You don&#39;t need fancy prompts to use Decision 1 or Jev. Both models are built around <strong>three basic question types</strong> (Jev 1.14 also adds an experimental fourth one for drawing boxes around objects in images):</p><h4>1. Yes or No (called noul)</h4><p>Use this when you want a simple true/false check.</p><ul><li><strong>Example question:</strong> <em>&quot;Is this customer email asking for a refund?&quot;</em></li><li><strong>What you get back:</strong> Yes (with a 0.97 confidence score).</li></ul><h4>2. Multiple Choice (called choice)</h4><p>Use this when the model needs to pick <strong>one winner</strong> from a list of up to 255 options.</p><ul><li><strong>Example question:</strong> <em>&quot;Which team should handle this support ticket: Billing, Technical Support, or Sales?&quot;</em></li><li><strong>What you get back:</strong> Billing (plus the exact chance for every option on your list).</li></ul><h4>3. Quick Rating (called score)</h4><p>Use this when you want to grade something on an ordered scale (like 1 to 5 stars).</p><ul><li><strong>Example question:</strong> <em>&quot;Rate how urgent this bug report is from 1 (can wait) to 5 (site is down).&quot;</em></li><li><strong>What you get back:</strong> 4.3 (an average score weighted by how likely each level is).</li></ul><p>Best of all? You can ask <strong>dozens or even hundreds of these questions at the exact same time</strong> about the same piece of content in a single request (Decision 1 lets you batch up to 1,000 questions in one call).</p><h3>How to Use Them Effectively in Real Projects</h3><p>Think of a decision model as a <strong>smart, fast front door</strong> sitting in front of your main app or your bigger AI models.</p><p>Here is a real-world example of how that looks in practice.</p><h4>Real Project Example: The &quot;Smart Bouncer&quot; for Customer Support</h4><p>Suppose you run an app that gets <strong>20,000 customer messages a day</strong>, some with screenshots, voice notes, or PDF receipts attached.</p><p>If you send all 20,000 messages to a large reasoning AI model, your cloud bill explodes and users wait several seconds for every reply.</p><p>Instead, you put <strong>Gemini Decision 1</strong> or <strong>Jev</strong> at the front door:</p><ol><li><strong>Instant Check (~150 ms):</strong> A message arrives. In one quick call, the decision model answers three questions at once:<br>• <em>Is this message safe and not spam?</em> (Yes/No)<br>• <em>Which folder does it belong in?</em> (Multiple Choice: Refund, Bug, Account, General)<br>• <em>How frustrated does the user sound?</em> (Score: 1 to 5)</li><li><strong>Fast Path for Easy Cases:</strong> If the model is <strong>95% sure</strong> it&#39;s a standard password help request, your normal code sends the help link right away. No big chatbot needed.</li><li><strong>Smart Escalation for Hard Cases:</strong> If the model is only <strong>55% sure,</strong> or scores the customer&#39;s frustration at 4.8 / 5 , your code forwards that ticket to a stronger reasoning model or a real human teammate.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xGwXhYkjwi432vqUyNDBbg.png" /><figcaption><em>Example workflow using Jev: small parallel questions (Bool, Score, Choice) guiding regular app logic. Source: TypeSafe AI Blog (typesafe.ai)</em></figcaption></figure><h4>Other Everyday Places Where Decision Models Shine</h4><ul><li><strong>Traffic Routing:</strong> Google actually uses Decision 1 inside <strong>Gemini Notebook</strong> to decide which AI model a user&#39;s question needs, cutting response times on everyday questions by up to 40%.</li><li><strong>Content Moderation &amp; Safety Guardrails:</strong> Checking user uploads or AI tool calls for unsafe content in a fifth of a second before anything goes live.</li><li><strong>Picking Agent Tools:</strong> Helping an AI agent quickly choose which tool to click next without burning 500 words of &quot;thinking&quot; tokens first.</li><li><strong>Sorting Resumes, Leads, or Documents at Scale:</strong> Scoring thousands of records for a tiny fraction of a cent.</li></ul><h3>Google Decision 1 vs. TypeSafe AI Jev: What’s the Difference?</h3><p>While both models solve the same core problem, they have different strengths right now:</p><p><strong>Google Gemini Decision 1</strong></p><ul><li><strong>Release Date:</strong> Nov 11, 2026</li><li><strong>What Inputs It Reads:</strong> Text, JSON, Images, <strong>Audio, Video, and PDFs</strong></li><li><strong>Speed:</strong> ~150–250 ms</li><li><strong>Input Price:</strong> <strong>$0.10</strong> per 1M input tokens (plus a Free Tier)</li><li><strong>Output Price:</strong> <strong>$0 (Free)</strong></li><li><strong>Parallel Questions:</strong> Up to 1,000 questions per call</li></ul><p><strong>TypeSafe AI Jev (</strong><strong>jev-1.14.0)</strong></p><ul><li><strong>Release Date:</strong> Sept 15, 2026 (Jev 1.14 on Oct 14, 2026)</li><li><strong>What Inputs It Reads:</strong> <strong>Text &amp; JSON</strong> (Images are currently experimental)</li><li><strong>Speed:</strong> ~100 ms typical (70–500 ms range)</li><li><strong>Input Price:</strong> <strong>$0.042</strong> per 1M input tokens</li><li><strong>Output Price:</strong> <strong>$0 (Free)</strong></li><li><strong>Parallel Questions:</strong> Dozens or hundreds (up to context token limit)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LXXlHVHqQG1hDaX32RVz9g.png" /><figcaption>Google’s published benchmark and cost comparison between Gemini <em>Decision 1, TypeSafe Jev 1.14, and peers. Source: Google Keyword Blog (blog.google)</em></figcaption></figure><h3>The Honest Pros and Cons</h3><p>Before you rip out all your existing prompts, it helps to look at both the wins and the weak spots.</p><h4>The Pros (Why Builders Love Them)</h4><ul><li><strong>Blazing Fast:</strong> Answers come back in ~100 to 250 milliseconds, which is fast enough to run on every button click without the user noticing a lag.</li><li><strong>Way Cheaper:</strong> With regular chatbots, output tokens (the words the AI writes) are the most expensive part. Decision models generate <strong>zero output tokens,</strong> so you only pay a tiny fee for the input you send in ($0.042 per million tokens on Jev; $0.10 on Decision 1).</li><li><strong>Zero Formatting Headaches:</strong> Because they never write free-form text, you never have to fix broken JSON or weird extra sentences in your code.</li><li><strong>Built-in Honesty (Confidence Scores):</strong> Every answer comes with a number telling you whether the model is confident or guessing.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UwRShLFy1Y2qG61HAWeK-w.png" /><figcaption><em>Cost vs. accuracy comparison across four automated workflows. Source: TypeSafe AI Blog (typesafe.ai)</em></figcaption></figure><h4>The Cons (What to Watch Out For)</h4><ul><li><strong>No Explanations:</strong> They will give you the answer, but they cannot write a sentence explaining <em>why</em> they picked it.</li><li><strong>Bad at Math and Dates:</strong> Both Google and TypeSafe warn that these models are weak at counting, doing math, or comparing dates.</li><li><strong>Media Limits on Jev:</strong> Jev is great with text and JSON, and its image support is still marked experimental (&quot;Danger Zone&quot;), with no native audio or video yet. Decision 1 handles audio, video, images, and PDFs out of the box.</li><li><strong>Not Always the Top Accuracy Leader:</strong> While Google&#39;s own chart puts Decision 1 ahead on 5 of 7 tests, independent testing from <em>Artificial Analysis</em> ranked Decision 1 mid-pack (#15 of 25 decision models) and noted higher per-task costs on their test suite, while <em>MindStudio</em> found it struggled to read emotion in voice clips. Always test on your own data first!</li></ul><h3>5 Simple Best Practices for Your Projects</h3><p>As TypeSafe AI likes to say: <strong>&quot;The model makes judgments; your code makes decisions.&quot;</strong></p><p>If you plug Decision 1 or Jev into an app this week, follow these five simple rules:</p><ol><li><strong>Keep Each Question Small and Focused:</strong> Instead of asking one giant question with three conditions (<em>&quot;Is this email polite, urgent, and about billing?&quot;</em>), split it into three separate questions in the same call.</li><li><strong>Let Normal Code Handle Math and Dates:</strong> Never ask a decision model <em>&quot;Was this order placed less than 30 days ago?&quot;</em> Check days &lt; 30 in one line of regular code, and only ask the AI to judge the customer&#39;s message.</li><li><strong>Trim the Noise:</strong> Don&#39;t dump 50 pages of unrelated data into the input just because you can. Extra noise lowers accuracy (what TypeSafe calls <em>&quot;context rot&quot;</em> ). Send only what matters for the question.</li><li><strong>Use the Confidence Score as a Safety Net:</strong> Pick a cutoff score for your app. For example: above 0.85, act automatically; below 0.85, pass the item to a bigger reasoning AI or a human.</li><li><strong>Give an &quot;Other / None of the Above&quot; Escape Hatch:</strong> In real life, user inputs are messy. When using Multiple Choice, either include an other option or add a quick Yes/No check alongside it so the model isn&#39;t forced to guess when none of the choices fit.</li></ol><h3>Frequently Asked Questions (FAQ)</h3><p><strong>Can Gemini Decision 1 or Jev replace ChatGPT or regular Gemini?</strong></p><p>No, and they aren&#39;t meant to. Whenever you need an AI to write an email, summarize a report, write code, or explain an idea, you still need a generative chatbot. Decision models only handle the quick sorting, grading, and routing steps in between.</p><p><strong>Which one should I pick: Google Decision 1 or Jev?</strong></p><p>Pick <strong>Jev</strong> if your data is mostly text or JSON and you want the lowest possible cost per million tokens ($0.042/1M) and super snappy ~100 ms checks. Pick <strong>Gemini Decision 1</strong> if your project needs to judge images, audio, video, or PDFs directly, or if you want a free tier inside Google AI Studio to experiment with first.</p><p><strong>Are there newer versions coming soon?</strong></p><p>TypeSafe AI raised an $870M Series A in October 2026 and teased more models ahead, while Google noted that Google Cloud (Vertex AI) support for Decision 1 is &quot;coming soon,&quot; though neither company has announced a specific release date for their next update yet.</p><h3>Final Thought</h3><p>For the past few years, we treated every problem in AI like a conversation.</p><p>Models like <strong>Google’s Gemini Decision 1</strong> and <strong>TypeSafe AI’s Jev</strong> remind us that a huge part of software isn&#39;t about writing essays. It&#39;s about making thousands of small, fast, reliable judgment calls behind the scenes.</p><p>Have you tried swapping a bulky classifier prompt for a decision model yet? Drop a comment below. I’d love to hear what you’re building with them!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=530952fbef7d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why We Stopped Blogging on Our Own Website (And What We Did Instead)]]></title>
            <link>https://medium.com/@7pedram/why-we-stopped-blogging-on-our-own-website-and-what-we-did-instead-d4bcb58ebd52?source=rss-1acb6289f136------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4bcb58ebd52</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[reddit]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[medium]]></category>
            <dc:creator><![CDATA[Pedram]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 14:06:18 GMT</pubDate>
            <atom:updated>2026-07-30T14:06:18.615Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QS2V-9zL1XQZi0ymCdEGAg.png" /></figure><p>People are writing blog posts for business websites that almost nobody read.</p><p>Not because the writing is bad. Not because the topics were wrong. The content is solid, well-researched, sometimes even useful. But it sits on domains with no authority, buried somewhere on page four of Google, collecting dust next to a cookie consent banner and a stock photo of people shaking hands.</p><p>Recently I pulled up Google Trends data for “Medium” and “Reddit” over the past 10+ years. Both platforms hit their all-time peak search interest in April 2026. Not a small bump. An absolute spike, the highest point on the entire chart since 2004.</p><p>That was the moment I stopped treating our website blog as the centre of our content strategy. Here is exactly what we changed, why it works, and what I think most small businesses are getting wrong right now.</p><h3>The Problem Nobody Talks About</h3><p>If you run a small or medium-sized business, chances are someone told you at some point that you need a blog on your website. “It’s good for SEO,” they said. “Google rewards fresh content.”</p><p>That advice made sense in 2016.</p><p>It does not make sense in 2026.</p><p>Here is what actually happens when you publish an article on a website with low domain authority: Google crawls it eventually, maybe indexes it, and ranks it somewhere behind Reddit threads, Medium articles, established news outlets, and now AI-generated summaries that pull from all of the above. Your carefully written 1,200-word piece about “5 Tips for Hiring in Brighton” sits on page three. Nobody scrolls to page three.</p><p>Meanwhile, the same topic posted as a discussion on Reddit with 40 upvotes and 15 comments? That shows up in the AI Overview at the top of the results page. Google’s algorithm reads those upvotes as community validation. It reads those comments as engagement signals. It treats that thread as more trustworthy than your blog post, because real people interacted with it.</p><p>This is not speculation. About 68% of Google searches now end without the user clicking on anything at all. The AI Overview answers their question directly, and the sources it cites are overwhelmingly platforms like Reddit, Medium, Wikipedia, and established publications.</p><p>So the question becomes: why are we still publishing where nobody can find us?</p><h3>The Playbook: Medium First, Website Second</h3><p>The switch is simpler than most people think, and it costs nothing.</p><p><strong>Step one:</strong> Create a Medium publication for your business. Not a personal account, a publication. This gives you a branded space, the ability to add multiple writers later, and access to Medium’s built-in audience of readers who are already browsing by topic.</p><p><strong>Step two:</strong> Publish your primary content on Medium first. Every article, every guide, every case study. Medium’s domain authority is massive. When you publish there, Google indexes your content faster and ranks it higher than it would on your own site. I have seen articles indexed within hours on Medium that took weeks to appear from a standalone domain.</p><p><strong>Step three:</strong> Embed your Medium RSS feed back onto your website’s blog page. Most website builders support this natively, and there are plugins for WordPress that handle it in minutes. Your website visitors still see a blog section full of fresh content. They click through to Medium to read the full post. You lose nothing on your main site, but you gain everything in search visibility.</p><p>This is not about abandoning your website. Your site still matters for conversions, for your portfolio, for your contact page. But your blog content, the stuff that is supposed to attract new visitors through search, needs to live where search engines are actually looking.</p><h3>Real-World Case Study: Brighton Part Timer on Reddit</h3><p>I want to talk about something specific we have been doing, because theory is easy and execution is where most advice falls apart.</p><p>We run <strong>Brighton Part Timer</strong>, a local resource for part-time and flexible work in Brighton. For a few weeks, we tried the traditional route: publish job roundups and hiring insights on the website, share the link on social media, hope for traffic.</p><p>Traffic was flat. Engagement was... The content was fine, but it existed in a vacuum.</p><p>So we changed the approach completely. Instead of writing generic job advice articles, we started publishing <strong>monthly local hiring reports directly on Brighton’s Reddit community</strong>. Raw numbers. Real data. Actual trends pulled from local listings.</p><p>Things like: how many new part-time roles appeared this month compared to last. Which sectors are hiring and which are shrinking. What the average hourly rate looks like across hospitality, retail, and remote work. Whether weekend-only roles are increasing or declining.</p><p>No fluff. No “top 10 tips.” Just the numbers, formatted clearly, posted where locals actually spend their time.</p><p>The response was immediate, and it surprised me.</p><p>People started commenting with their own experiences. “Yeah, I noticed hospitality wages going up around the Lanes.” “The retail drop makes sense, three shops on Western Road closed last month.” The threads turned into genuine local conversations, with people validating, questioning, and adding context to the data we posted.</p><p>Those Reddit threads now rank on the first page of Google for multiple Brighton-specific job searches. Some of them get cited in AI Overviews when someone searches for things like “part time work Brighton 2026” or “Brighton hiring trends.”</p><p>We did not pay for that visibility. We did not build backlinks. We posted useful, specific data in a community that cared about it, and the search algorithms did the rest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QHlsVda9zY_q_7KhM0cTlA.png" /></figure><h3>Why AI Engines React to Data the Way the Human Brain Reacts to Sugar</h3><p>This is the part that took me a while to wrap my head around, and I think it is the most important thing to understand about where search is heading.</p><p>Large language models and AI search engines are drowning in generic content. There are millions of blog posts that say roughly the same thing in roughly the same way. “How to find a part-time job.” “Tips for improving your CV.” “What employers look for in 2026.” The AI has seen all of it, thousands of times over, reworded slightly each time.</p><p>But original data? Specific local numbers? Monthly reports with real figures? That is rare. And AI engines treat it like the human brain treats sugar: an immediate, high-priority signal that this content is worth extracting, citing, and surfacing to users.</p><p>Here is why:</p><p><strong>Original data cannot be faked easily.</strong> When you post that Brighton hospitality wages grew 4.2% month-over-month across 85 tracked listings, that is a specific, verifiable claim. The AI recognises it as high-value grounding material, something it can cite with confidence rather than paraphrasing from yet another generic advice post.</p><p><strong>Numbers are easy for AI to extract.</strong> LLMs are built on retrieval systems that parse and structure information. A thread full of percentages, comparisons, and trend data is computationally easy to turn into a clean AI Overview answer. A 2,000-word opinion piece with no data points? The model has to work much harder to figure out what, if anything, is actually useful to cite.</p><p><strong>Community engagement validates the data.</strong> When a Reddit post with original numbers gets upvoted and discussed, the algorithm reads those signals as proof that the information is trusted by the community. Upvotes function like peer review for the AI. A post with 80 upvotes and 25 comments full of local residents confirming the trends carries more weight than a corporate blog post with zero engagement.</p><p><strong>Freshness compounds over time.</strong> When you publish monthly reports, you create a pattern. The AI learns that your source produces regular, updated information. Over time, it starts treating your content as a reliable recurring source rather than a one-off. This is why consistency matters more than volume. Twelve monthly reports posted over a year carry more algorithmic weight than 50 generic articles posted whenever someone had time to write one.</p><p>Think of it this way: every AI-generated search answer needs to be grounded in something real. The model needs to point at a source and say, “this is where I got this information.” If your content is specific, data-rich, community-validated, and regularly updated, you become that source. If your content reads like everything else on the internet, the AI has no reason to pick you over the ten thousand other pages saying the same thing.</p><h3>What This Means for Paid Ads Too</h3><p>I should mention this because I know people are going to ask.</p><p>If you are running Google Ads for your business, the economics are shifting under your feet. Costs per click keep climbing while click-through rates are declining, because the AI Overview now sits above both organic results and paid ads for a growing number of queries. Users get their answer from the AI summary and never scroll down to your ad.</p><p>This does not mean paid ads are dead. They still work well for high-intent commercial queries where someone is ready to buy. But for informational and consideration-stage searches, the ones where you are trying to build awareness and trust, organic visibility through community platforms is becoming a better investment than throwing money at broad-match keywords.</p><p>The businesses that will do well over the next few years are the ones that figure out how to be the source the AI cites, not the ones paying to appear below an AI answer that already gave the user what they needed.</p><h3>How to Start This Week</h3><p>If any of this resonates, here is what I would actually do, in order, starting this week:</p><p><strong>Pick one data point you already have.</strong> Every business tracks something. Sales trends, customer questions, local market changes, seasonal patterns. You do not need to build a fancy dashboard. You need one interesting number and the context around it.</p><p><strong>Write it up plainly and post it where your audience is.</strong> If your audience is local, find the relevant subreddit or community group. If your audience is professional, post on Medium or LinkedIn. Do not overthink the formatting. A clear headline, the key numbers, and your honest interpretation of what they mean.</p><p><strong>Do it again next month.</strong> Consistency is what separates a one-off post from a trusted source. The first report might get modest engagement. The third or fourth one is when people start expecting it, referencing it, and sharing it.</p><p><strong>Connect it back to your website.</strong> Embed your Medium feed, link to your Reddit posts from your site, build a “Reports” section that aggregates everything. Your website becomes a hub, not the primary distribution channel.</p><p>The rules of search have changed. The winners are not the people publishing the most content on their own domains. They are the people feeding search engines real human data, in real communities, on a regular schedule.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4bcb58ebd52" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The question has moved on from “which model is best”]]></title>
            <link>https://medium.com/@7pedram/the-question-has-moved-on-from-which-model-is-best-91a205f4871c?source=rss-1acb6289f136------2</link>
            <guid isPermaLink="false">https://medium.com/p/91a205f4871c</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[deepseek]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[moonshots]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Pedram]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 16:37:22 GMT</pubDate>
            <atom:updated>2026-07-09T16:37:22.166Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2WkiTpvt3b-E-E14ZkPMIA.jpeg" /></figure><p>I spent some time going through the Stanford HAI 2026 AI Index Report this week.</p><p>One thing I didn’t expect personally:<br>The performance gap between the best closed-source and best open-weight AI models <strong>actually widened!</strong> It went from 0.5% in August 2024 to 3.3% by March 2026. Closed labs kept pushing on hard reasoning tasks, the kind where even PhD experts struggle, and open models haven’t fully caught up there.</p><p>Lets talk about other things in this report:<br>Open-weight models now reach close to frontier performance at release. And the open-source community closes the remaining distance within a few months through fine-tuning and distillation. For most production work, that last few percent <strong>doesn’t justify the cost difference really</strong>, especially when API pricing scales linearly with every token and self-hosting doesn’t.</p><p>The part I found most interesting was the transparency data. Stanford’s Foundation Model Transparency Index dropped from 58 to 40. You’d assume open-weight models would score well here. They don’t. Several major open-weight developers scored poorly on disclosing training data and compute.</p><p>The other shift the report captures: Chinese labs like Alibaba, DeepSeek, and Moonshot AI are now producing some of the strongest open-weight models globally. Two years ago that wasn’t the case.</p><p>My read on where this lands practically:</p><blockquote>If you need data sovereignty, cost control at scale, or domain-specific fine-tuning, open models are the obvious pick today. If you need the absolute ceiling on complex multi-step reasoning or want managed infrastructure with zero ops burden, closed models still earn their price.</blockquote><p><strong>The question has moved on from “which model is best” to “best for what, at what cost.”</strong></p><p>Sources I referenced: Stanford HAI AI Index Report 2026 (<a href="http://aiindex.stanford.edu/">aiindex.stanford.edu</a>), Stanford Foundation Model Transparency Index, LMSYS Chatbot Arena (<a href="http://lmarena.ai/">lmarena.ai</a>), Artificial Analysis (<a href="http://artificialanalysis.ai/">artificialanalysis.ai</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=91a205f4871c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Scaling multi-agent systems]]></title>
            <link>https://medium.com/@7pedram/scaling-multi-agent-systems-3ab2c2d2fd92?source=rss-1acb6289f136------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ab2c2d2fd92</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Pedram]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 16:27:23 GMT</pubDate>
            <atom:updated>2026-07-09T16:27:23.287Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hSMGGoxHtiP9tuUmTPX-eg.jpeg" /></figure><p>Scaling multi-agent systems is mostly an architecture problem.<br>Not a prompting one.</p><p>Once you have real orchestration running in production, the bottleneck tends to be the same thing every time: the main agent accumulating context until it starts dropping instructions. Slowly at first… then completely.</p><p>The fix we keep coming back to is transient subagents with hard context boundaries! A planning agent that never touches raw tool output. Worker agents that spin up, complete a bounded task, and return a structured result. Read-only research agents that never write to the workspace. Separate execution sandboxes for anything that involves running code or modifying files.<br>The overhead is smaller than you’d expect.</p><p>And the reliability difference is significant, particularly when you enforce strict input/output schemas between agents. That single decision probably cuts debugging time in half, in my experience at least.<br>The orchestrator only ever sees clean structured updates. It never drowns in noise.</p><p>If you’re already running multi-agent workflows, I’m curious how are you handling state management between your planning and execution layers?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ab2c2d2fd92" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I Switched from Outlier to Alignerr: A Generalist Writer’s Honest Review (2026)]]></title>
            <link>https://medium.com/@7pedram/why-i-switched-from-outlier-to-alignerr-a-generalist-writers-honest-review-2026-3f6a7a571fc4?source=rss-1acb6289f136------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f6a7a571fc4</guid>
            <category><![CDATA[money]]></category>
            <category><![CDATA[freelancing]]></category>
            <category><![CDATA[side-hustle]]></category>
            <category><![CDATA[remote-work]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Pedram]]></dc:creator>
            <pubDate>Wed, 31 Dec 2025 01:44:04 GMT</pubDate>
            <atom:updated>2026-01-01T10:12:56.810Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jPcloMkuZRplGTUW_qLdzg@2x.jpeg" /></figure><p>If you have been looking for remote work or side hustles recently, you have probably heard the mixed reviews about the big AI training platforms. Like many of you, I spent time working on Outlier, but after dealing with the frustration of empty queues and unstable projects, I started looking for alternatives.</p><p>For the last couple of months, I have been working with a platform called Alignerr (by Labelbox), specifically as a Generalist Writer in English. It has been a refreshing change, and since I know many people are looking for legitimate options right now, I wanted to share my honest experience.</p><p>What the work actually looks like</p><p>The role is fully remote and asynchronous, which basically means you can work whenever you want without a boss breathing down your neck.</p><p>I work on the Generalist Writer project, and my day-to-day isn’t just mindlessly clicking buttons. It involves genuinely interesting tasks like educating the AI by reviewing text for clarity, helping the model reason through complex arguments, and “Red Teaming.” Red Teaming is actually the fun part; you try to break the AI or trick it into revealing biases and inaccuracies so the developers can fix them.</p><p>The Pay and Why I Prefer It to Outlier</p><p>Let’s talk about the money, because that is why we are all here. The pay consistency on Alignerr has been the biggest upgrade for me. They pay well, and most importantly, they pay on time. I haven’t had to chase support for missing hours or delayed transfers, which was a constant headache on other platforms.</p><p>The stated pay range for this role is roughly $25 to $100 per hour. Obviously, hitting the top end of that range depends on the complexity of the specific task, but generally speaking, I have found the rates to be significantly better than what I was earning before.</p><p>Beyond the money, the work feels more respectable. You are using your brain to fix structure, tone, and logic, rather than just labeling images. I have enjoyed it enough that I’ve even started upskilling to apply for other specialized roles on their job board.</p><p>How to Get Started</p><p>The onboarding process is pretty straightforward. You will go through a screening with their AI interviewer, Zara, followed by some skill assessments. It works on a rolling basis, so once you pass, you simply wait for a project match.</p><p>If you are looking to get away from the instability of other sites or just want a side gig that actually pays what it promises, I highly recommend giving it a shot.</p><p>I used a referral link when I first applied, and I felt like it helped speed up my onboarding process (though I can’t promise it works the same for everyone!). If you want to use mine to potentially flag your application in their system, here is the link:</p><p>(((It looks like my referral link has officially reached its maximum usage limit because so many of you signed up. If you are still interested in joining, please drop me a direct message (DM). I will double-check if I can generate a fresh link or get a new one, and if I do, I will send it to you personally so you don’t miss out.)))</p><p>Transparency note: That is a referral link. I might earn a small bonus if you successfully onboard, but using it helps support me and hopefully helps you get noticed by the system faster.</p><p>If you have any questions about the testing or the workflow, feel free to leave a comment. I am happy to answer anything that doesn’t violate my NDA.</p><p>Hope this helps you find something stable for the New Year!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3f6a7a571fc4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>