<?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 Prateek Upadhyay on Medium]]></title>
        <description><![CDATA[Stories by Prateek Upadhyay on Medium]]></description>
        <link>https://medium.com/@Deeplearning_AI?source=rss-58db1debbd67------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*oz0Qx5LzYQ60N3X5_bZ-GA.jpeg</url>
            <title>Stories by Prateek Upadhyay on Medium</title>
            <link>https://medium.com/@Deeplearning_AI?source=rss-58db1debbd67------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 23:08:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Deeplearning_AI/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[AI Has Left the Chat]]></title>
            <link>https://medium.com/@Deeplearning_AI/ai-has-left-the-chat-9ea987496b3b?source=rss-58db1debbd67------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ea987496b3b</guid>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Prateek Upadhyay]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 09:39:55 GMT</pubDate>
            <atom:updated>2026-11-13T09:39:55.861Z</atom:updated>
            <content:encoded><![CDATA[<h4>Robots, robotaxis and driverless trucks are no longer demos. Here is 2026 in a 3-minute read.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vt9XvZnD9lO_gjoMY9P4gA.png" /></figure><p>For two years, AI lived inside a chat box.</p><p><strong>In 2026, it stepped outside.</strong> It walks factory floors, drives strangers home and hauls freight down highways.</p><p>This is <strong>Physical AI</strong>: AI that doesn’t just talk. It moves.</p><h3>1. Robots are getting real jobs</h3><p><strong>19,100 humanoid robots shipped in the first half of 2026. That’s up 272% in one year.</strong></p><ul><li><strong>Chinese companies shipped 97%</strong> of them. Two firms, AgiBot and Unitree, made about 3 out of every 4.</li><li><strong>7 in 10 went to work</strong> in factories, warehouses and logistics, not to labs or stage shows.</li><li>Experts expect about <strong>60,000 this year</strong> and <strong>500,000 by 2030</strong>.</li></ul><p><strong>Takeaway:</strong> the numbers are small today, but this is the steep part of the curve.</p><h3>2. Robotaxis: the driver’s seat is empty</h3><p><strong>Waymo’s cars drive about 3.9 million miles a week with nobody behind the wheel.</strong></p><p>That’s <strong>more than 150 trips around the Earth. Every week.</strong></p><ul><li>About <strong>half a million rides a week</strong>, in 10+ US cities.</li><li>More than <strong>270 million</strong> driverless miles so far.</li><li>Tesla, Baidu and others are racing to catch up.</li></ul><p><strong>Takeaway:</strong> in some cities, a robotaxi is no longer a novelty. It’s just a ride.</p><h3>3. Trucks: freight goes driverless</h3><ul><li>Aurora now runs driverless trucks on <strong>10 routes</strong> across the southern US.</li><li>Nearly <strong>440,000 driverless miles</strong> by mid-2026.</li><li>Goal: <strong>200+ driverless trucks</strong> on the road by the end of the year.</li></ul><p>A truck that doesn’t need sleep can make a 1,000-mile trip <strong>up to 40% faster</strong> (Aurora’s estimate).</p><p><strong>Takeaway:</strong> the first big business for self-driving may be boxes, not people.</p><h3>4. China is changing how cars think</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Z1fMMWQAm1wJKDy5hRP-w.png" /></figure><p><strong>Old way: a rulebook.</strong> Engineers write thousands of “if this, then that” rules. <em>If the light is red, stop.</em> A situation nobody wrote a rule for? The car gets stuck.</p><p><strong>New way: end-to-end AI.</strong> One big model learns from millions of hours of real driving, like a person gaining experience. Camera in, steering and braking out.</p><ul><li><strong>7 in 10 new cars</strong> in China already come with driver assist.</li><li>Huawei calls its 2026 system a move from “rule-driven” to an <strong>AI agent</strong>. Xpeng and many others are making the same switch.</li><li>China’s first mandatory safety standard for more advanced self-driving starts in <strong>July 2027</strong>.</li></ul><p><strong>Takeaway:</strong> China’s cars are going from following a rulebook to <strong>learning from experience</strong>.</p><h3>What it means for you</h3><ul><li><strong>Jobs will change first in driving, warehouses and delivery.</strong></li><li><strong>Safety is the real test.</strong> Trust will be won one mile at a time.</li><li><strong>It’s a US vs China race.</strong> The US leads robotaxis. China leads robots and scale.</li></ul><h3>The bottom line</h3><p><strong>First AI learned to talk. Now it’s learning to walk, drive and deliver.</strong></p><p>The question is no longer <em>if</em> machines will share our roads and workplaces. It’s <strong>how fast.</strong></p><p><em>Which would you trust first: a robotaxi, a robot coworker or a driverless truck? Tell me in the comments.</em> 👇</p><p><em>Sources: Smart Analytics Global via </em><a href="https://techrepublic.com/article/news-humanoid-robot-shipments-chinese-vendors"><em>TechRepublic</em></a><em> (Aug 2026); Waymo safety data via </em><a href="https://electrek.co/2026/10/27/musk-robotaxi-victory-waymo-austin-miles/"><em>Electrek</em></a><em> (Oct 2026); </em><a href="https://finance.yahoo.com/sectors/technology/articles/waymo-scales-500k-weekly-rides-123118125.html"><em>Yahoo Finance</em></a><em> (May 2026); Aurora via </em><a href="https://gcn.com/aurora-launches-second-generation-driverless-trucks/20501/"><em>GCN</em></a><em> (Aug 2026); MIIT via </em><a href="https://eu.36kr.com/en/p/3915288958539657"><em>36Kr</em></a><em>; </em><a href="https://autonews.gasgoo.com/articles/market-industry/after-gpt-6-do-autonomous-driving-companies-have-only-one-path-left-2117964492666982401"><em>Gasgoo</em></a><em>. All images are original graphics by the author.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ea987496b3b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM vs Jev vs CDM: The Writer, the Judge and the Clerk]]></title>
            <link>https://medium.com/@Deeplearning_AI/llm-vs-jev-vs-cdm-the-writer-the-judge-and-the-clerk-5a45094cf061?source=rss-58db1debbd67------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a45094cf061</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[cdm]]></category>
            <category><![CDATA[llm-cost-optimization]]></category>
            <category><![CDATA[jevs]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Prateek Upadhyay]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 06:04:54 GMT</pubDate>
            <atom:updated>2026-10-22T06:04:54.435Z</atom:updated>
            <content:encoded><![CDATA[<h4>A visual, no-hype guide to the new “decision layer” of AI — what each model really does, when to use which, and how to wire all three into one agent (with code)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ComfUu2wD2gkOc62WJqUpQ.png" /></figure><p>It is 2:07 a.m. and Priya’s phone buzzes.</p><p>Priya (a composite character — the details are illustrative, the pain is real) runs the support platform at a mid-sized e-commerce company in Pune. Every incoming ticket goes to a frontier LLM with a carefully engineered prompt: <em>read this ticket, return JSON with team, priority and refund_eligible.</em></p><p>It works. Mostly. Tonight, the model returned &quot;team&quot;: &quot;tech_support&quot;. The allowed values are billing, technical and sales. A downstream switch statement fell through, 3,000 tickets piled into a dead queue, and the on-call engineer — Priya — is awake.</p><p>When she does the post-mortem the next morning, three numbers jump out:</p><ul><li>Each triage call takes <strong>2–4 seconds</strong>, for an answer that is essentially <em>three words</em>.</li><li>Triage is <strong>the single biggest line on her LLM bill</strong>, bigger than the customer-facing chatbot.</li><li>The model never once said <em>“I’m not sure.”</em> It was confidently wrong.</li></ul><p>If that story feels familiar, you’re the audience for this article. Because between <strong>15 September and 17 October 2026</strong>, the AI industry shipped a wave of models built for exactly Priya’s problem — models that <strong>don’t write</strong>, they <strong>decide</strong>. Two names led the conversation: <strong>Jev</strong> (TypeSafe AI’s “System One” model) and <strong>CDMs</strong> (ScaleDown’s “Constrained Decoding Models”).</p><p>So: are they LLM killers? Rebranded classifiers? Hype? Let’s find out — with pictures.</p><h3>TL;DR (for the busy reader)</h3><blockquote><strong><em>LLM = the Writer.</em></strong><em> Generates any string, token by token. Unbeatable for open-ended work (chat, code, plans). Slow and pricey when all you need is a label — and it can invent labels.</em></blockquote><blockquote><strong><em>Jev = the Judge.</em></strong><em> You send a </em>state<em> plus several </em>typed questions<em> (choice / score / yes-no). One request returns a probability distribution for every question, in parallel. Vendor-reported: 70–500 ms, $0.042 per million input tokens, output free. Can’t generate text — by design.</em></blockquote><blockquote><strong><em>CDM = the Clerk.</em></strong><em> Small task-specific models (Classify, Extract, Summarize, Compress) whose answers are restricted to </em>labels you define<em> or </em>spans of your input<em>. 1M-token, multimodal input. Vendor-reported: $0.05 per million input tokens, output free.</em></blockquote><blockquote><strong><em>The real lesson:</em></strong><em> stop sending </em>decisions<em> to a </em>writer<em>. Use the LLM for the hard, creative part and cheap specialised models for the high-volume judgments around it — “cheap by default, frontier on exception.”</em></blockquote><blockquote><strong><em>The caveat:</em></strong><em> almost every speed/cost multiplier you’ll see (200×, 444×, 467×) is a </em>vendor<em> benchmark. Test on your own data.</em></blockquote><h3>The 60-second version: one restaurant, three jobs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*csUM1yJhEh-bMCTnAyc6eQ.png" /></figure><p><em>Figure 1 — The best kitchens use all three. Nobody asks the head chef to seat guests.</em></p><p>Think of a busy restaurant.</p><ul><li><strong>The head chef (LLM)</strong> can cook anything, including dishes nobody ordered. Brilliant, creative — and you pay for every minute of their time.</li><li><strong>The maître d’ (Jev)</strong> glances at a guest and instantly makes <em>several</em> calls at once: which table, is this a VIP, are they in a hurry. They’ll tell you how sure they are. And they will never seat you at a table that doesn’t exist.</li><li><strong>The prep station (CDM)</strong> chops, sorts and labels — but only what’s already on the counter. One job per station, done very fast.</li></ul><p>For two years, most AI products have asked the head chef to also seat guests and peel every onion. That’s the “decision tax”, and it’s what this new generation of models is trying to abolish.</p><p>Now let’s open the kitchen doors.</p><h3>Part 1 — The Writer: what an LLM actually does</h3><p>A large language model is, at its core, a <em>next-token predictor</em>. Give it text, it computes a probability distribution over its whole vocabulary (often 100k+ tokens), picks one, appends it, and repeats. A 200-token answer is 200 sequential steps. That <strong>autoregressive</strong> loop is the source of the LLM’s superpower — it can produce <em>any</em> string: a poem, a SQL query, a legal memo, a plan for a ten-step agent task.</p><p>It’s also the source of its three weaknesses when you embed it inside software:</p><ol><li><strong>Latency scales with output.</strong> Even a tiny answer pays for the prompt plus a token-by-token loop — and with reasoning models, a hidden chain of thought on top.</li><li><strong>The output space is unbounded.</strong> Ask for one of three labels and the model <em>can</em> still return a fourth — tech_support, billing?, or a friendly &quot;Sure! Here&#39;s the JSON…&quot;.</li><li><strong>Confidence is not built in.</strong> You can ask the model to “rate its confidence from 0 to 1”, but that number is just more generated text. TypeSafe’s launch post puts it bluntly: even when prompted for a confidence estimate, models “tend to be overconfident and inconsistent”.</li></ol><p>The industry’s patch for problem #2 has been <strong>constrained decoding</strong> (a.k.a. structured outputs): at each step, mask out every token that would break a JSON schema or grammar. Open-source libraries like <strong>Outlines</strong> and <strong>XGrammar</strong>, and serving engines like <strong>vLLM</strong>, <strong>SGLang</strong>, <strong>TensorRT-LLM</strong> and <strong>llama.cpp</strong>, all support it today.</p><p>It works remarkably well — for <em>structure</em>. A September 2026 arXiv study by Akash Chavan tested five small open models (0.6B–4B) on 14 structured-output tasks: constrained decoding pushed schema validity from <strong>78.6–92.9 % to 100 %</strong>. But content accuracy did not follow; the paper’s memorable line is that constrained decoding’s “reach ends exactly where schema conformance ends.”</p><blockquote><strong><em>Key insight:</em></strong><em> Valid JSON ≠ correct JSON. Constrained decoding fixes the </em>shape<em> of the answer, not the </em>judgment<em> inside it — and it still pays the autoregressive latency bill.</em></blockquote><p>That is the gap Jev and CDMs aim at.</p><h3>Part 2 — The Judge: Jev, the “System One” model</h3><h4>The origin story</h4><p>On <strong>15 September 2026</strong>, San Francisco–based <strong>TypeSafe AI</strong> came out of two years in stealth with <strong>Jev</strong> and a <strong>US$40 million seed round led by DCVC</strong>. Its CEO, <strong>Diogo Almeida</strong>, spent roughly four years at OpenAI working on RLHF, InstructGPT, ChatGPT and GPT-4. His launch post opens with a question that has nagged many of us:</p><blockquote>“Models have been superhuman at chat for years, so where is all the automation?”</blockquote><p>His answer: chat models are optimised for what <em>human raters</em> prefer — fluent prose — not for decisions software can depend on. So TypeSafe built a different stack: a new model architecture, a <strong>parallel sampler</strong>, and a training method it calls <strong>Reinforcement Learning for Calibrated Decisions (RLCD)</strong>.</p><p>Two fun naming details:</p><ul><li><strong>“System One”</strong> comes from Daniel Kahneman’s <em>Thinking, Fast and Slow</em> — fast, intuitive judgment, as opposed to slow, deliberate “System 2” reasoning.</li><li><strong>“Jev”</strong> honours <strong>William Stanley Jevons</strong>, whose <em>Jevons paradox</em> says that making a resource cheaper increases total consumption. TypeSafe’s bet: make machine judgment 100× cheaper and you’ll use it 1,000× more.</li></ul><h4>How Jev works (the part that matters for builders)</h4><p>TypeSafe’s own one-liner is the best summary: <em>unstructured state in, typed probabilistic decisions out.</em></p><p>You send:</p><ul><li>a <strong>state</strong> — a string, a JSON object, or an array of strings (a ticket, a game frame described as JSON, a chat history), and</li><li>one or more <strong>questions</strong>, each of one of three types (TypeSafe calls them <em>primitives</em>):</li><li><strong>Choice</strong> — What it asks: “Which of these options fits?” (unordered) · What comes back: choice, probabilities, confidence</li><li><strong>Score</strong> — What it asks: “Where does this fall on this ordered rubric?” · What comes back: score (probability-weighted average), probabilities, confidence</li><li><strong>Noul</strong> — What it asks: “Is this statement true?” · What comes back: noul, a probability between 0 and 1</li></ul><p>Every question is evaluated <strong>in parallel and in isolation</strong> against the same state, in a single request. The docs note that adding questions “barely changes the response time”, and because each question is isolated, there’s no “context rot” between them. The flip side: questions can’t depend on each other — if Q2 needs Q1’s answer, you make two calls.</p><p>Other specs worth knowing (from the docs and public reporting): up to <strong>64k tokens</strong> per input (32k for state plus the longest question); choice cardinality up to <strong>255</strong> options; text-only state; pricing of <strong>$0.042 per million input tokens</strong> with <strong>output free</strong>.</p><p>What is <em>not</em> public: parameter count, architecture details, weights, or a technical paper. TypeSafe describes it as transformer-based; outside observers have speculated it may sit on an open-weight LLM or a BERT-like encoder. Keep that in mind when you read benchmark claims.</p><h4>Show me the code</h4><p>Here’s TypeSafe’s quick-start pattern in Python — one ticket, three questions, one call:</p><pre># pip install typesafe-sdk      (Python 3.10+)<br># export TYPESAFE_API_KEY=&quot;...&quot;<br>from typesafe_sdk import Choice, Noul, Score, TypeSafeClient<br> <br>ticket = (&quot;Our API integration started returning 500 errors on every request &quot;<br>          &quot;about 20 minutes ago, and we can&#39;t process any customer orders.&quot;)<br> <br>with TypeSafeClient(model=&quot;jev-latest&quot;) as client:   # reads TYPESAFE_API_KEY<br>    r = client.system_one(<br>        state=ticket,<br>        questions={<br>            &quot;department&quot;: Choice(<br>                instructions=&quot;Which team should handle this&quot;,<br>                criteria={<br>                    &quot;billing&quot;:   &quot;Payment or subscription issues&quot;,<br>                    &quot;technical&quot;: &quot;Bugs or integration problems&quot;,<br>                    &quot;sales&quot;:     &quot;Pricing or account questions&quot;,<br>                },<br>            ),<br>            &quot;frustration&quot;: Score(<br>                instructions=&quot;How frustrated the customer appears&quot;,<br>                criteria=[&quot;Calm, just stating facts&quot;,<br>                          &quot;Frustrated but civil&quot;,<br>                          &quot;Very angry, strong language&quot;],<br>            ),<br>            &quot;is_urgent&quot;: Noul(<br>                instructions=&quot;The message conveys urgency or time-sensitivity&quot;,<br>            ),<br>        },<br>    )<br> <br>print(r.choices[&quot;department&quot;].choice)    # technical<br>print(r.scores[&quot;frustration&quot;].score)     # 0.67  (position on the 0–2 rubric)<br>print(r.nouls[&quot;is_urgent&quot;].noul)         # 0.99  (probability)</pre><p>The raw API is a single POST https://api.typesafe.ai/v1/systemone. In TypeSafe&#39;s example response, department comes back with &quot;confidence&quot;: 1.0 and frustration with &quot;confidence&quot;: 0.5 — a built-in signal that the second judgment is shakier. Notice what&#39;s missing: no prompt engineering for JSON, no parser, no retry loop. In TypeScript, the SDK even infers the literal type &quot;billing&quot; | &quot;technical&quot; | &quot;sales&quot;.</p><blockquote><strong><em>Why “can’t hallucinate” is half true.</em></strong><em> TypeSafe says Jev “can’t hallucinate” in the sense that it can’t return a value outside the type you declared — a schema violation is impossible by construction. It can absolutely still pick the </em>wrong<em> option. Commentators, including Mehul Gupta’s popular explainer on Medium, make the same point. Translate the claim as: </em>no type errors, plus an honest probability<em> — which is still a big deal.</em></blockquote><h4>The demos that made developers grin</h4><p>Jev’s launch went viral less for its tables than for its toys: a <strong>Doom</strong>-playing bot (TypeSafe’s engineer worried that 10 queries per second would cost about <strong>$7/hour</strong>; the team thought that was cheap), a <strong>Wikiracing</strong> agent choosing among hundreds of links per hop, and community-built Minecraft, Subway-Surfers-style and drone-simulation bots. MindStudio’s write-up reports a two-minute Minecraft session costing about <strong>1 cent</strong>. Fireship’s 21 September video — <em>“An ex-OpenAI researcher just deleted language from the LLM…”</em> — carried the idea well beyond the AI-research bubble.</p><p>Two honest notes from the sources: the Doom demo fed Jev a <em>text/JSON description</em> of the game, not pixels; and every demo ran in simulators, not on real robots.</p><h3>Part 3 — The Clerk: CDMs, the Constrained Decoding Models</h3><h4>The launch</h4><p>On <strong>13 October 2026</strong>, <strong>ScaleDown</strong> (CEO Neal Patel) launched what it called <strong>Constrained Decoding Models</strong>, with day-one integrations from <strong>LangChain</strong> (langchain-scaledown plus LangSmith LLM Gateway support), <strong>OpenRouter</strong>, <strong>Vercel</strong> and <strong>LiteLLM</strong>.</p><p>The idea: instead of one general model that can write any string, ship <strong>four single-task models</strong>, each trained for one job and allowed to return only what that job permits:</p><ul><li><strong>Classify</strong> — Job: Score text against labels <em>you</em> define (each with a yes/no “rubric”) · Output: softmax probability per label + optional one-sentence reason</li><li><strong>Extract</strong> — Job: Pull fields you describe in plain language · Output: exact spans with <strong>character offsets</strong>, confidence and context</li><li><strong>Compress</strong> — Job: Cut context down to what’s relevant to a query · Output: the relevant parts of the input</li><li><strong>Summarize</strong> — Job: Condense long text · Output: a short abstractive summary</li></ul><p>All accept up to <strong>1M tokens</strong>; Classify, Extract and Summarize also accept <strong>images and scanned PDFs</strong>. ScaleDown lists pricing at <strong>$0.05 per million input tokens</strong> with output free (OpenRouter ran a $0.04 launch price for Extract and Summarize).</p><p>The headline claim — <strong>“200× faster and 467× cheaper than frontier models”</strong> on these narrow tasks — is ScaleDown’s own benchmark. The per-task numbers, against the average of three OpenAI models, are more modest: Summarize <strong>41× cheaper, 2.7× faster</strong>; Compress <strong>138× cheaper, 1.8× faster</strong>.</p><h4>Show me the code</h4><p>The raw Classify API is refreshingly simple:</p><pre>import os, requests<br> <br>resp = requests.post(<br>    &quot;https://api.scaledown.xyz/classify&quot;,<br>    headers={&quot;x-api-key&quot;: os.environ[&quot;SCALEDOWN_API_KEY&quot;]},<br>    json={<br>        &quot;text&quot;: &quot;I was charged twice for my subscription and need a refund immediately.&quot;,<br>        &quot;labels&quot;: [<br>            {&quot;name&quot;: &quot;billing&quot;,   &quot;rubric&quot;: &quot;Is this about a charge, refund or payment problem?&quot;},<br>            {&quot;name&quot;: &quot;technical&quot;, &quot;rubric&quot;: &quot;Is this about a bug or something not working?&quot;},<br>            {&quot;name&quot;: &quot;account&quot;,   &quot;rubric&quot;: &quot;Is this about login, password or account settings?&quot;},<br>            {&quot;name&quot;: &quot;other&quot;,     &quot;rubric&quot;: &quot;Does this fit none of the other categories?&quot;},<br>        ],<br>        &quot;reasoning&quot;: True,   # adds a one-sentence explanation (billed extra)<br>    },<br>).json()<br> <br>print(resp[&quot;top_label&quot;], resp[&quot;scores&quot;])   # billing {&#39;billing&#39;: 0.92, ...}<br>print(resp.get(&quot;reasoning&quot;))</pre><p>Two production tips straight from ScaleDown’s docs: Classify <strong>always returns a winner</strong>, so add a catch-all label like other; and apply your own confidence threshold on scores. There&#39;s also a multi_label mode (independent probabilities, several tags per input) and a chunks batch mode where the shared label &quot;profile&quot; is billed once.</p><h4>The skeptic’s corner (and why it matters)</h4><p>CDMs drew sharper pushback than Jev. An ibl.ai analysis titled <em>“Constrained Decoding Models Are Not a New Architecture”</em> makes three points worth knowing:</p><ol><li><strong>Constrained decoding is a decoding-time technique</strong>, available as a library on models you already run — not a new architecture.</li><li><strong>Not all CDMs skip generation.</strong> By ScaleDown’s own docs, Classify is a scoring head and Extract points at spans of the input — but <strong>Summarize is abstractive</strong>, writing “in the model’s own words.” So “skips token generation entirely” is true for two of the three launch models, not all of them.</li><li><strong>Open weights are close behind on some tasks.</strong> ScaleDown’s own head-to-head page shows it leading Fastino’s Apache-2.0 <strong>GLiNER2.5</strong> by <strong>+15.5 points on classification</strong> but only <strong>+0.2 points on extraction</strong> — and GLiNER runs locally, for free.</li></ol><p>To ScaleDown’s credit, it publishes its methodology (public datasets like CLINC-OOS, CUAD, QMSum and FinanceBench) and its harness on GitHub (scalebench) — more transparency than the category norm.</p><blockquote><strong><em>Key insight:</em></strong><em> “CDM” is best read as a </em>product category<em> — narrow, metered, task-specific endpoints — rather than a new neural architecture. That’s not a criticism; it’s a buying decision.</em></blockquote><h3>Head-to-head: what’s actually different?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uB2gTy2DimVbiPpezQEvHQ.png" /></figure><p><em>Figure 2 — Same ticket, three very different computation shapes.</em></p><p>The deepest difference isn’t speed or price. It’s the <strong>output space</strong> — the set of things each model is <em>allowed</em> to say.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LP8o0NKdIeCnEtYPNzwBOQ.png" /></figure><p><em>Figure 3 — What a model is allowed to say decides what can go wrong.</em></p><ul><li><strong>Think of it as</strong> — LLM: The Writer · Jev (System One): The Judge · CDM (ScaleDown): The Clerk</li><li><strong>Input</strong> — LLM: Prompt / messages · Jev (System One): <em>State</em> + typed questions · CDM (ScaleDown): Text, docs, images + labels / field descriptions</li><li><strong>Output</strong> — LLM: Any string · Jev (System One): Probability distribution per question · CDM (ScaleDown): A label from your list, or spans of your input (Summarize: a short summary)</li><li><strong>Computation</strong> — LLM: Sequential, token by token · Jev (System One): One request, all questions in parallel · CDM (ScaleDown): Task-specific; output restricted to the task</li><li><strong>Confidence</strong> — LLM: Not native · Jev (System One): Built-in, calibrated (RLCD) · CDM (ScaleDown): Label probabilities; per-span confidence</li><li><strong>Context</strong> — LLM: Model-dependent · Jev (System One): 64k per input (32k state + question), text-only · CDM (ScaleDown): Up to 1M tokens, multimodal</li><li><strong>Price (input / output)</strong> — LLM: ~$0.20–$10 per 1M / output ~5× more* · Jev (System One): $0.042 per 1M / free · CDM (ScaleDown): $0.05 per 1M / free</li><li><strong>Latency (vendor)</strong> — LLM: 3–329 s end-to-end* · Jev (System One): 70–500 ms · CDM (ScaleDown): 2.2×–13.9× faster than frontier lineups</li><li><strong>Open weights?</strong> — LLM: Many · Jev (System One): No · CDM (ScaleDown): No (licensed self-hosted containers exist)</li><li><strong>Best for</strong> — LLM: Chat, code, writing, planning · Jev (System One): Routing, scoring, guardrails, game/agent loops · CDM (ScaleDown): Extraction, context compaction, document triage</li><li><strong>Not for</strong> — LLM: High-volume micro-decisions · Jev (System One): Writing anything · CDM (ScaleDown): Open-ended reasoning</li></ul><p><em>*Ranges as published in TypeSafe’s launch comparison. All figures vendor-reported, Sep–Oct 2026.</em></p><h3>The economics — and the fine print</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nQL_t_xg5q4WrSY1HqX6dQ.png" /></figure><p><em>Figure 4 — The numbers, as the vendors publish them.</em></p><p>On a log scale, the gap is dramatic: one to two orders of magnitude on price, and from seconds to milliseconds on latency. For Priya’s 40,000 tickets a day, even a conservative 20× saving changes the conversation from “can we afford AI triage?” to “why aren’t we running it on <em>every</em> message?” — Jevons paradox in action.</p><p>Now the fine print — because the best bloggers tell you what the vendor didn’t put in bold:</p><ul><li><strong>TypeSafe’s</strong> 193.6× faster / 444.6× cheaper headline comes from four workflows written by its own team; TypeSafe itself says these are “on the higher end of real world gains” and that “some bias could exist.” Reference answers were the average of two frontier models, not human labels.</li><li><strong>ScaleDown’s</strong> 200×/467× is its own benchmark on narrow tasks; ibl.ai also flagged an arithmetic oddity in one headline average.</li><li><strong>Latency figures depend on hardware and geography</strong> (TypeSafe notes its evals ran from laptops on the US West Coast).</li><li><strong>Independent, neutral benchmarks don’t exist yet.</strong> A community “Decision Index” has emerged, but early rankings were already revised within days.</li></ul><p><strong>Rule of thumb:</strong> treat every multiplier as a hypothesis, and run a 500-example bake-off on your own traffic before you migrate anything.</p><h3>Zoom out: this isn’t two products, it’s a new layer</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b5sWxqeurQYk_Vkq3daP5Q.png" /></figure><p><em>Figure 5 — In about five weeks, “decide” split off from “generate”.</em></p><p>Jev didn’t stay alone for long. Within weeks came Convai’s <strong>Laya</strong> and Fastino’s <strong>GLiNER2.5-Decide</strong> (both Apache 2.0 encoders), Nace AI’s <strong>Drex 1.5</strong> (open weights, restrictive licence), the vLLM Semantic Router team’s <strong>Decision 2.0</strong> family (0.6B–27B, Apache 2.0), <strong>OpenAI’s Decisions API</strong> (public beta on 6 October, $0.10 per million input tokens, output free), <strong>Microsoft-Decision-1</strong>, ScaleDown’s CDMs, and <strong>Perplexity’s pplx-decider</strong> models.</p><p>Researchers moved just as fast:</p><ul><li><strong><em>Jev in the Wild</em></strong> (arXiv, Sep 2026) counted <strong>2,170 public GitHub repositories</strong> using Jev by 22 September — one week after launch (1,865 of them brand-new) — mostly for attribute judgment (77 %), scoring/ranking (52 %) and action selection (31 %), with ~70 % using it for two or more purposes.</li><li><strong><em>LLM-as-Jev</em></strong> (arXiv 2610.02076, Li &amp; Wagle) argues that ordinary LLMs <em>already</em> contain this capability: with a training-free readout over option probabilities, frozen Qwen models behave as strong decision models — and can even use “thinking mode” for a slower “Thinking Jev.”</li></ul><p>When one startup says “the decision layer should be separate”, that’s positioning. When a dozen ship in five weeks and OpenAI follows, that’s architecture.</p><h3>Putting it together: one agent, three kinds of model</h3><p>Here’s the pattern the LangChain team has been promoting across its Jev and ScaleDown posts — “cheap by default, frontier on exception”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CD6sVkOihFnjzxYCu5Aqag.png" /></figure><p><em>Figure 6 — The LLM thinks; the small models decide and tidy up.</em></p><p><strong>1. Route to the right model (Jev).</strong> A cheap lookup shouldn’t burn your most expensive model:</p><pre># pip install &quot;langchain-typesafe[experimental]&quot;<br>from langchain.agents import create_agent<br>from langchain_typesafe.experimental.middleware import ModelChoice, ModelRouterMiddleware<br> <br>router = ModelRouterMiddleware(<br>    choices={<br>        &quot;fast&quot;:     ModelChoice(model=&quot;openai:luna&quot;,<br>                                criteria=&quot;Direct lookups, extraction, and localized changes.&quot;),<br>        &quot;powerful&quot;: ModelChoice(model=&quot;openai:sol&quot;,<br>                                criteria=&quot;Architecture and high-stakes decisions.&quot;),<br>    },<br>    instructions=&quot;Choose the least costly model that can complete the task.&quot;,<br>)<br>agent = create_agent(&quot;openai:gpt-5.6-luna&quot;, middleware=[router])</pre><p><strong>2. Guard dangerous tool calls (Jev).</strong> The “auto mode” safety classifiers that coding harnesses keep behind closed doors, now as a one-liner:</p><pre>from langchain_typesafe.experimental.middleware import AutoModeMiddleware<br>agent = create_agent(&quot;openai:gpt-5.6-luna&quot;,<br>                     middleware=[AutoModeMiddleware(tools=[&quot;bash&quot;])])</pre><p><strong>3. Keep context lean (CDMs).</strong> Summarise old history, compress fat tool results, and extract exact fields:</p><pre># uv pip install -U langchain-scaledown<br>from deepagents import create_deep_agent<br>from langchain_scaledown import (ScaledownSummarizationMiddleware,<br>                                 ScaledownToolCompressionMiddleware)<br> <br>agent = create_deep_agent(<br>    model=&quot;openai:gpt-6-sol&quot;,<br>    tools=[search_docs, fetch_url],<br>    middleware=[<br>        ScaledownSummarizationMiddleware(&quot;openai:gpt-6-sol&quot;,<br>                                         trigger=(&quot;fraction&quot;, 0.8),<br>                                         keep=(&quot;fraction&quot;, 0.1)),<br>        ScaledownToolCompressionMiddleware(tools=[&quot;fetch_url&quot;]),<br>    ],<br>)</pre><p>In LangChain’s own trace, compression cut a documentation page from <strong>18,402 to 3,110 characters</strong> before the main model saw it — and if a ScaleDown call fails, the middleware leaves history untouched so the agent keeps running.</p><p><em>Model names above are as they appear in LangChain’s Sep–Oct 2026 posts; substitute whatever models you use. Snippets are adapted from the official docs — check the linked repos for the current API.</em></p><h3>The superpower nobody talks about: confidence-based routing</h3><p>The quiet revolution isn’t speed. It’s that a calibrated probability lets you turn AI into a plain if statement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q3Y1pwYk0q7I4UC291mZgQ.png" /></figure><p><em>Figure 7 — The System 1 → System 2 hand-off, in code.</em></p><pre>def triage(ticket: str):<br>    ans = jev_department(ticket)            # wraps the Jev call shown earlier<br>    if ans.confidence &gt;= 0.90:<br>        return route(ans.choice)            # System 1: fast, cheap, automatic<br>    if ans.confidence &gt;= 0.60:<br>        return llm_triage_with_reasoning(ticket)   # System 2: think it through<br>    return human_queue.put(ticket)          # neither model is sure → a person decides</pre><p>This is exactly what Priya was missing at 2:07 a.m. Not a smarter model — a model that knows when it doesn’t know. (Thresholds here are illustrative; tune them on labelled samples of your own traffic.)</p><h3>Build-it-yourself: a mini “Jev-style” decision function with an open LLM</h3><p>Want the idea without a vendor? Inspired by the <em>LLM-as-Jev</em> paper, here’s a self-hosted, ~30-line sketch: present the options as letters and read the <strong>next-token probabilities</strong> for those letters in <strong>one forward pass</strong>. No generation, no parsing — and the answer can’t fall outside your options.</p><pre># pip install torch transformers<br>import torch<br>from transformers import AutoTokenizer, AutoModelForCausalLM<br> <br>MODEL = &quot;Qwen/Qwen3-0.6B&quot;   # any small instruct model works; bigger = better judgment<br>tok = AutoTokenizer.from_pretrained(MODEL)<br>lm = AutoModelForCausalLM.from_pretrained(MODEL, torch_dtype=torch.float32).eval()<br> <br>@torch.no_grad()<br>def choice(state: str, question: str, options: list[str]) -&gt; dict:<br>    letters = [chr(ord(&quot;A&quot;) + i) for i in range(len(options))]<br>    menu = &quot;\n&quot;.join(f&quot;{l}. {o}&quot; for l, o in zip(letters, options))<br>    prompt = (f&quot;State:\n{state}\n\nQuestion: {question}\n{menu}\n&quot;<br>              f&quot;Reply with a single letter.\nAnswer:&quot;)<br>    ids = tok(prompt, return_tensors=&quot;pt&quot;).input_ids<br>    next_logits = lm(ids).logits[0, -1]                       # ONE forward pass<br>    letter_ids = [tok.encode(&quot; &quot; + l, add_special_tokens=False)[0] for l in letters]<br>    probs = torch.softmax(next_logits[letter_ids], dim=0)     # renormalise over options only<br>    k, p_max = len(options), probs.max().item()<br>    return {<br>        &quot;choice&quot;: options[int(probs.argmax())],<br>        &quot;probabilities&quot;: dict(zip(options, [round(p, 3) for p in probs.tolist()])),<br>        # simple heuristic: 0 = uniform guess, 1 = certain (not TypeSafe&#39;s formula)<br>        &quot;confidence&quot;: round((p_max - 1 / k) / (1 - 1 / k), 3),<br>    }<br> <br>print(choice(&quot;My card was charged twice for order A-104.&quot;,<br>             &quot;Which team should handle this ticket?&quot;,<br>             [&quot;billing&quot;, &quot;technical&quot;, &quot;sales&quot;]))</pre><p>Caveats, honestly stated: raw LLM probabilities are <strong>not calibrated</strong> the way RLCD-trained models claim to be (check with a reliability plot or Brier score on your data), and letter-position bias is real (shuffle options and average). For production self-hosting, look at the Apache-2.0 decision encoders (GLiNER2.5-Decide, Laya, Decision 2.0) or constrained decoding in vLLM/SGLang via Outlines or XGrammar.</p><h3>So… which one should I use?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i3DkTAkTp_pYjd11ihNO6A.png" /></figure><p><em>Figure 8 — A 30-second decision guide.</em></p><p>In one breath:</p><ul><li><strong>Writing anything</strong> (answers, code, emails, plans) → <strong>LLM</strong>.</li><li><strong>Many small judgments about one state, in milliseconds, with honest probabilities</strong> (route, score, guard, act) → <strong>Jev</strong> or another decision model (e.g. OpenAI’s Decisions API).</li><li><strong>Copying exact fields, compacting huge or visual inputs, single-label triage over long docs</strong> → <strong>CDM-style</strong> task models.</li><li><strong>Data can’t leave your network</strong> → open-weight decision encoders, or your own LLM plus constrained decoding.</li></ul><h3>Five myths, busted</h3><ol><li><strong>“Jev and CDMs will replace LLMs.”</strong> No. Both vendors — and LangChain — explicitly say to keep a general LLM for open-ended generation. They replace the <em>decision tax</em>, not the writer.</li><li><strong>“Can’t hallucinate = always right.”</strong> No type errors ≠ no wrong answers. You still need evaluation on your own data.</li><li><strong>“Structured outputs already solved this.”</strong> They solved <em>shape</em>, not judgment, latency or calibration (see the Chavan paper).</li><li><strong>“200× faster everywhere.”</strong> These are vendor multipliers on vendor-chosen tasks. Expect smaller gains on yours — still potentially large.</li><li><strong>“CDM is a brand-new neural architecture.”</strong> It’s a product family built on a well-known technique. Judge it on accuracy, price, context length and deployment options.</li></ol><h3>Your 30-minute getting-started checklist</h3><ol><li><strong>Inventory your LLM calls.</strong> Tag each as <em>generate</em>, <em>decide</em> or <em>extract</em>. In most agent stacks, the last two dominate call volume.</li><li><strong>Pick 2–3 decision points</strong> (e.g., ticket routing, tool-call safety, “is the agent looping?”).</li><li><strong>Write atomic questions.</strong> TypeSafe’s docs advise one well-scoped “gut-check” per question; combine them in code, not in a mega-prompt.</li><li><strong>Build a 300–500 example golden set</strong> from real traffic, labelled by humans.</li><li><strong>Bake off</strong> your current LLM vs Jev vs ScaleDown vs an open-weight option — measure accuracy, p95 latency, cost per 1,000 calls and calibration.</li><li><strong>Ship with thresholds</strong> and a human fallback. Log every probability.</li><li><strong>Re-test monthly.</strong> This category changed weekly in its first five weeks.</li></ol><h3>Closing: Jevons was right</h3><p>In 1865, William Stanley Jevons noticed that better steam engines didn’t reduce Britain’s coal use — they exploded it, because coal-powered work became cheap enough to do everywhere.</p><p>Priya’s team didn’t cut their AI usage when triage got cheap. They put a judgment on <em>every</em> message, every tool call, every agent step. That’s the real story of LLM vs Jev vs CDM: not a cage match, but a division of labour. The <strong>Writer</strong> gets to focus on what only it can do. The <strong>Judge</strong> and the <strong>Clerk</strong> handle the thousand small decisions around it — fast, cheap, and honest about their doubts.</p><p>The question for your stack is no longer <em>“which model is smartest?”</em> It’s <em>“which decisions am I still paying a writer to make?”</em></p><p><em>If this helped, give it a clap (or fifty), share it with the engineer who got paged last night, and tell me in the comments which decision you’d move first.</em></p><h3>References &amp; further reading</h3><p><strong>Official launches and docs</strong></p><ul><li>TypeSafe AI — <a href="https://typesafe.ai/blog/introducing-system-one-models-and-jev">Introducing System One Models &amp; Jev</a> (15 Sep 2026)</li><li>TypeSafe docs — <a href="https://docs.typesafe.ai/introduction">Introduction</a> · <a href="https://docs.typesafe.ai/introduction/quickstart">Quick start</a> · <a href="https://docs.typesafe.ai/primitives">Primitives</a> · <a href="https://docs.typesafe.ai/models">Models</a></li><li>TypeSafe workflow evals — <a href="https://evals.typesafe.ai/">evals.typesafe.ai</a></li><li>ScaleDown — <a href="https://scaledown.ai/">scaledown.ai</a> · <a href="https://docs.scaledown.ai/quickstart">Quickstart</a> · <a href="https://docs.scaledown.ai/api-reference/classify">Classify API</a> · <a href="https://docs.scaledown.ai/benchmarks/providers">Benchmarks</a> · <a href="https://docs.scaledown.ai/benchmarks/methodology">Methodology</a></li><li>LangChain — <a href="https://www.langchain.com/blog/building-a-harness-with-jev">Building a Harness with Jev</a> · <a href="https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents">Building Long-Running Agents with ScaleDown and Deep Agents</a></li><li>LangChain integration docs — <a href="https://docs.langchain.com/oss/python/integrations/providers/typesafe">TypeSafe</a> · <a href="https://docs.langchain.com/oss/python/integrations/middleware/scaledown">ScaleDown middleware</a></li></ul><p><strong>GitHub</strong></p><ul><li><a href="https://github.com/typesafe-ai/skills">typesafe-ai/skills</a> — TypeSafe agent skill</li><li><a href="https://github.com/typesafe-ai/system-one-adapter-python">typesafe-ai/system-one-adapter-python</a> — wrapper used to get structured decisions from LLMs in TypeSafe’s evals</li><li><a href="https://github.com/scaledown-team/langchain-scaledown">scaledown-team/langchain-scaledown</a> — LangChain middleware (MIT)</li><li><a href="https://github.com/scaledown-team/scalebench">scaledown-team/ScaleBench</a> — ScaleDown benchmark harness</li><li><a href="https://github.com/fastino-ai/GLiNER2">fastino-ai/GLiNER2</a> — open-weight extraction / decision models</li><li><a href="https://github.com/CruiseDevice/small-llm-structured-benchmark">CruiseDevice/small-llm-structured-benchmark</a> — code for the constrained-decoding study</li><li><a href="https://github.com/allebee/jevk5">allebee/jevk5</a> — independent open-weight Jev-style reproduction (not affiliated with TypeSafe)</li></ul><p><strong>Papers</strong></p><ul><li>Chavan — <a href="https://arxiv.org/abs/2609.23742">Constrained Decoding Eliminates Structural Failures in Small LLMs but Reveals a Scale-Dependent Semantic Gap</a> (arXiv 2609.23742)</li><li>Li &amp; Wagle — <a href="https://arxiv.org/abs/2610.02076">LLM-as-Jev: LLMs Are Already Jev-Style Decision Models</a> (arXiv 2610.02076)</li><li>Ling, Xue &amp; Ye — <a href="https://arxiv.org/html/2609.30216v1">Jev in the Wild</a> (arXiv 2609.30216)</li></ul><p><strong>Explainers, videos and critiques worth your time</strong></p><ul><li>Fireship — <a href="https://www.youtube.com/watch?v=TbkUKCm3CHQ">An ex-OpenAI researcher just deleted language from the LLM…</a> (YouTube, 21 Sep 2026)</li><li>Mehul Gupta, <em>Data Science in Your Pocket</em> — <a href="https://medium.com/data-science-in-your-pocket/what-is-jev-ai-b8294d980001">What is Jev AI?</a> (Medium)</li><li>MindStudio — <a href="https://www.mindstudio.ai/blog/jev-system-one-model-launch">Jev Explained: Typesafe AI’s Non-Autoregressive System-1 Model</a></li><li>Matt Crabtree, DataCamp — <a href="https://www.datacamp.com/blog/system-one-models-jev">Jev: TypeSafe’s System One Model Explained</a></li><li>The New Stack — <a href="https://thenewstack.io/typesafe-jev-system-one/">launch coverage of TypeSafe’s Jev</a></li><li>ibl.ai — <a href="https://ibl.ai/blog/constrained-decoding-models-not-a-new-architecture">Constrained Decoding Models Are Not a New Architecture</a> · <a href="https://ibl.ai/blog/decision-layer-models-nine-days-own-the-router">Six Decision Models in Three Weeks</a></li><li>AICoder — <a href="https://aicoder.com/news/news-20261013-scaledown-constrained-decoding-models-langchain-deep-agents">ScaleDown ships Constrained Decoding Models</a></li><li>Wikipedia — <a href="https://en.wikipedia.org/wiki/Jev_(AI_model)">Jev (AI model)</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a45094cf061" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI in Education Portal : 6 ways AI is reshaping Education Industry]]></title>
            <link>https://medium.com/@Deeplearning_AI/ai-in-education-portal-c3c383933da5?source=rss-58db1debbd67------2</link>
            <guid isPermaLink="false">https://medium.com/p/c3c383933da5</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[elearning]]></category>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Prateek Upadhyay]]></dc:creator>
            <pubDate>Sat, 11 Jun 2022 14:36:59 GMT</pubDate>
            <atom:updated>2022-06-11T14:45:35.084Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Why use AI in eLearning</strong></p><p>The traditional corporate training model presents some challenges,including :</p><p>● <strong>Too much content and way too long</strong> : Most traditional learning programs still are made of long-form modules. And creating an hour of such training content requires way too many hours. This overwhelms people very easily!</p><p>● <strong>Lack of personalized experiences : </strong>Because content creation is time-consuming, eLearning courses are often too generic and do not adapt to the specific needs of each employee.</p><p>● <strong>Don’t really track the program’s effectiveness : </strong>Calculating the ROI of traditional training involves time-consuming data collection and entry processes.</p><p>● <strong>It doesn’t match the digital workforce’s existing needs and expectations : </strong>Training is no longer an initiative that happens once or twice a year but is an ongoing process. Today’s workforce (including mostly millennials and Gen Z) prefer to self-manage their learning experiences, according to a LinkedIn Learning report.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/651/1*xomlHgG_YslzZ9JQK6hpWw.png" /></figure><p><strong>Six Ways AI will Shape eLearning</strong></p><p>1. Real-Time Questioning</p><p>2. Natural Language Processing</p><p>3. Improve Accessibility</p><p>4. Personalize Learning</p><p>5. Gamification</p><p>6. Automatic Grading</p><ol><li><strong>Real-Time Questioning</strong></li></ol><p>● One of the major issues an individual comes across when learning is the inability to clear their doubts as soon as they pop up in their minds.</p><p>● At times it might just be the ‘fear of looking unintelligent’ or just that the trainer isn’t available.</p><p>● But it is highly vital that learners ask quick/immediate questions,because if not,this will only complicate the learning process even more.</p><p>● By integrating<strong> AI </strong>into our <strong>eLearning courses</strong>,the necessity to walk up to the trainer or look up the internet every time a slight confusion/doubt arises in the midst of learning can be avoided.</p><p>● All one would have to do is ask the AI engine and receive the appropriate answer.</p><p>2. <strong>Natural Language Processing</strong></p><p>● How many times have we wished we could talk to our instructors in our native language ? or communicate with the eLearning platform in the language we prefer ?</p><p>● <strong>AI </strong>has indeed made our wishes true by integrating our eLearning courses with an AI assistant with whom we can ask queries and clarify doubts in the language we prefer.</p><p>● This will not just make learning a lot easier but also save an ample amount of time and make <strong>learning even more engaging and efficient.</strong></p><p>3. <strong>Improve Accessibility</strong></p><p>● <strong>AI </strong>can play a great role in training employees with disabilities too and can be used to convert “spoken language into captions and transcripts” thus helping the hearing-impaired comprehend the subject easily.</p><p>4. <strong>Personalize Learning</strong></p><p>● Delivering the same form of content to every single learner is a very common approach in corporate training.It can indeed be quite overwhelming for a trainer to cater to and meet the needs of every learner.But this one-size-fits-all approach is definitely not appropriate as everybody has their own style and pace of learning.</p><p>● <strong>AI</strong> can be used to track the previous performance of an individual and use that data to alter the current learning material,thus providing a <strong>personalized learning experience.</strong></p><p>● Within an eLearning course,AI can track learner’s progress,which will help identify the areas where each learner lacks proficiency thus altering the material accordingly.</p><p>5. <strong>Gamification</strong></p><p>● Gamification is an essential approach used in eLearning to engage and motivate learners towards acquiring new knowledge.</p><p>● With <strong>AI</strong>, planning tactical games for online training material is feasible.</p><p>● It helps to process large amounts of data ,which is then used to determine learners’ behavior and to provide them with updates about their learning progress.</p><p>6. <strong>Automatic Grading</strong></p><p>● By integrating <strong>AI</strong>,we can also leverage powerful automation tools that facilitate grading.</p><p>● Integration of the solution can hasten the process of checking and grading the learners’ papers in different languages.It frees the time of the teachers for <strong>preparing materials for online courses</strong>,<strong>communication with the learners</strong> etc</p><p><strong>References:</strong></p><p>1. <a href="https://blog.commlabindia.com/elearning-design/artificial-intelligence-elearning">https://blog.commlabindia.com/elearning-design/artificial-intelligence-elearning</a></p><p>2. <a href="https://www.shiftelearning.com/blog/artificial-intelligence-elearning#:~:text=According%20to%20eLearning%20Industry%20%22An,making%2C%20and%20much%20more.%22">https://www.shiftelearning.com/blog/artificial-intelligence-elearning#:~:text=According%20to%20eLearning%20Industry%20%22An,making%2C%20and%20much%20more.%22</a></p><p>3. <a href="https://elearningindustry.com/artificial-intelligence-based-platform-impact-future-elearning">https://elearningindustry.com/artificial-intelligence-based-platform-impact-future-elearning</a></p><p>4. <a href="https://becominghuman.ai/the-role-of-artificial-intelligence-in-e-learning-41ac88ee3e8d">https://becominghuman.ai/the-role-of-artificial-intelligence-in-e-learning-41ac88ee3e8d</a></p><p>5. <a href="https://www.tituslearning.com/artificial-intelligence-elearning/">https://www.tituslearning.com/artificial-intelligence-elearning/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c3c383933da5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Autoencoders :Complete Guide]]></title>
            <link>https://medium.com/@Deeplearning_AI/autoencoders-complete-guide-81892446ae21?source=rss-58db1debbd67------2</link>
            <guid isPermaLink="false">https://medium.com/p/81892446ae21</guid>
            <category><![CDATA[autoencoder]]></category>
            <category><![CDATA[variational-autoencoder]]></category>
            <category><![CDATA[convolutional-autoencoder]]></category>
            <category><![CDATA[denoising-autoencoder]]></category>
            <category><![CDATA[anomaly-detection]]></category>
            <dc:creator><![CDATA[Prateek Upadhyay]]></dc:creator>
            <pubDate>Sun, 22 May 2022 13:59:35 GMT</pubDate>
            <atom:updated>2022-05-22T14:07:06.076Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Autoencoders : Complete Guide</strong></h3><p><strong>Autoencoders</strong> are a specific type of feedforward neural networks where the input is the same as the output .They compress the input into lower dimensional code and then reconstruct the output from this representation. The code is a compact “summary” or “compression” of the input also called the latent-space representation.</p><ul><li>Unsupervised Dimensionality reduction algorithm</li><li>Autoencoders are considered to be self-supervised because they generate their own labels from the training data</li><li>An Autoencoder learns to compress the data while minimizing the reconstruction error.</li><li>Autoencoders are lossy compression where the output will not be the same as the input but a close degraded representation.</li><li>Alternatives are gzip</li></ul><p>An autoencoder consists of 3 components: <strong>encoder</strong>,<strong>code</strong> and <strong>decoder</strong>.</p><p>The encoder compresses the input and produces the code,the decoder then reconstructs the input using this code.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*1VYs4bQYjPhlc_E8" /><figcaption>Autoencoders: Image Reconstruction</figcaption></figure><p>To build an autoencoder we need 3 things:</p><ul><li>Encoding method (based on ANN)</li><li>Decoding method (based on ANN)</li><li>Loss function (to compare the output with the target) (either MSE or cross-entropy)</li></ul><p><strong>Autoencoder Architecture</strong></p><ul><li>Encoder and decoder are fully-connected feedforward neural networks</li><li>Code (latent space) is a single layer of an ANN with the dimensionality of our choice.The number of nodes in the code layer (code size) is a hyperparameter that we set before training the autoencoder.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ir1OALYayM5mSxC3" /><figcaption>Autoencoders : Latent space</figcaption></figure><p><strong>Hyperparameters :</strong></p><ul><li><strong>Code Size : </strong>Represents the number of nodes in the middle layer where less number of nodes results in more compression.</li><li><strong>Number of Layers: </strong>Represents the number of layers in encoder and decoder network</li><li><strong>Number of nodes per layer: </strong>The autoencoder network in the above figure is a stacked autoencoder where the number of nodes per layer decreases with each subsequent layer of the encoder ,and increases back in the decoder where the decoder is symmetric to the encoder.</li><li><strong>Loss Function: </strong>If the input values are in the range of [0,1] then we typically use cross entropy ,otherwise we use Mean Squared Error (MSE)</li></ul><h3><strong>Use Cases</strong></h3><ul><li><strong>Image Reconstruction</strong></li><li><strong>Image Denoising</strong></li><li><strong>Anomaly Detection on ECG Signals</strong></li></ul><p><strong>Following section will get you through code implementation of each of the use cases mentioned above.</strong></p><h3><strong>Implementation : Image Reconstruction</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/372/0*prZPf6RyEu_nBk3o" /><figcaption>Use case : Image Reconstruction</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/372/0*jzLYmbvtAXHiqcOk" /><figcaption>Use case : Image Reconstruction</figcaption></figure><p>For below implementation,we will be using <a href="http://vis-www.cs.umass.edu/lfw/">LFW dataset</a> (Labeled faces in the wild dataset) which is a public benchmark for face verification also known as pair matching.</p><p><strong>Loading and visualization of dataset</strong></p><pre>import matplotlib.pyplot as plt<br>def show_image(x):<br> plt.imshow(np.clip(x + 0.5, 0, 1))<br>show_image(X[6])</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/251/0*Nt1T_uJokFWUmK-_" /><figcaption>Input data</figcaption></figure><p><strong>Implementing the Autoencoder</strong></p><p>The autoencoder architecture consists of an encoder,decoder based on ANN and a code size (latent_dim = 64 ) which can be tweaked as per the requirement.Logically,lower the code_size is,the more the image will compress,but lesser the features will be preserved where the reproduced image will be that much different from the original.</p><p><strong>Encoder Layer</strong></p><p>The input shape taken in the encoder is (32x32x3) which is flattened into a 1D array using Flatten() layer before being passed to the Dense() layer.The last layer in the encoder is the Dense() layer which tries to find the optimal parameters that achieves the best output ,the output size of which is set to the code_size or latent_dim.</p><p><strong>Decoder Layer</strong></p><p>The decoder takes the encoded input and tries to reconstruct it in the form of a row ,which is then stacked into a (32x32x3) matrix through the Dense layer.The final Reshape() layer will reshape it into an image.</p><pre>latent_dim = 64<br>img_shape = X.shape[1:]<br>class Autoencoder(Model):<br> def __init__(self,latent_dim):<br> super(Autoencoder,self).__init__()<br> self.latent_dim = latent_dim<br> self.encoder = tf.keras.Sequential([<br> layers.Flatten(),<br> layers.Dense(latent_dim,activation=’relu’),</pre><pre> ])<br> self.decoder = tf.keras.Sequential([<br> layers.Dense(np.prod(img_shape),activation=’sigmoid’),<br> layers.Reshape(img_shape) <br> ])<br> def call(self,x):<br> encoded=self.encoder(x)<br> decoded=self.decoder(encoded)<br> return decoded<br>autoencoder=Autoencoder(latent_dim)</pre><p><strong>Model training</strong></p><pre>history = autoencoder.fit(x_train,x_train,epochs=20,shuffle=True,<br> validation_data=(x_test,x_test))</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/0*5jvTbXahYF1z_84Y" /></figure><p><strong>Model Result Visualization</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/368/0*nbKOCDRmtTB-EgVJ" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/368/0*vzU-rQHvaPWvt8RI" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/368/0*R6QjQspchnvU0FTc" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/368/0*Sj7WEPumnjx_igIs" /></figure><p><strong>The entire code implementation can be found </strong><a href="https://github.com/prateekupadhyay1997/Autoencoders.git"><strong>here</strong></a></p><h3><strong>Implementation : Image Denoising</strong></h3><p>An autoencoder can also be trained to remove noise from images. In the following section, you will create a noisy version of the Fashion MNIST dataset by applying random noise to each image. You will then train an autoencoder using the noisy image as input, and the original image as the target.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mwDR6u_uV7a8Icg-" /><figcaption>Use case : Image Denosing</figcaption></figure><p><strong>Loading and Visualization of dataset</strong></p><pre>n=10<br>plt.figure(figsize=(20,2))<br>for i in range(n):<br> ax=plt.subplot(1,n,i+1)<br> plt.title(“original+noise”)<br> plt.imshow(tf.squeeze(x_test_noisy[i]))<br> plt.gray()<br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*toLozom6FeSazVwu" /><figcaption>Input Data</figcaption></figure><p><strong>Implementing the Autoencoder for Image Denoising</strong></p><pre>class Denoise(Model):<br> def __init__(self):<br> super(Denoise,self).__init__()<br> self.encoder = tf.keras.Sequential([<br> layers.Input(shape=(28,28,1)),<br> layers.Conv2D(16,(3,3),activation=’relu’,padding=’same’,strides=2),<br> layers.Conv2D(8,(3,3),activation=’relu’,padding=’same’,strides=2)])<br> self.decoder = tf.keras.Sequential([<br> layers.Conv2DTranspose(8,kernel_size=3,strides=2,padding=’same’),<br> layers.Conv2DTranspose(16,kernel_size=3,strides=2,padding=’same’),<br> layers.Conv2D(1,kernel_size=(3,3),activation=’sigmoid’,padding=’same’)])<br> def call(self,x):<br> encoded=self.encoder(x)<br> decoded=self.decoder(encoded)<br> return decoded<br>autoencoder = Denoise()</pre><p><strong>Autoencoder Model Training</strong></p><pre>autoencoder.fit(x_train_noisy,x_train,<br> epochs=10,<br> shuffle=True,<br> validation_data=(x_test_noisy,x_test))</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*C6C0zS4IVd23-MNT" /></figure><p><strong>Model result visualization</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*C5K_vMVQH8nsywgK" /><figcaption>Model Result</figcaption></figure><p><strong>The entire code implementation for Image denosing can be found </strong><a href="https://github.com/prateekupadhyay1997/Autoencoders.git"><strong>here</strong></a></p><h3><strong>Implementation : Anomaly Detection on ECG Dataset</strong></h3><p>In this example, you will train an autoencoder to detect anomalies on the <a href="http://www.timeseriesclassification.com/description.php?Dataset=ECG5000">ECG5000 dataset</a>. This dataset contains 5,000 <a href="https://en.wikipedia.org/wiki/Electrocardiography">Electrocardiograms</a>, each with 140 data points. You will use a simplified version of the dataset, where each example has been labeled either 0 (corresponding to an abnormal rhythm), or 1 (corresponding to a normal rhythm). You are interested in identifying the abnormal rhythms.</p><p>Note: This is a labeled dataset, so you could phrase this as a supervised learning problem. The goal of this example is to illustrate anomaly detection concepts you can apply to larger datasets, where you do not have labels available (for example, if you had many thousands of normal rhythms, and only a small number of abnormal rhythms).</p><p>How will you detect anomalies using an autoencoder? Recall that an autoencoder is trained to minimize reconstruction error. You will train an autoencoder on the normal rhythms only, then use it to reconstruct all the data. Our hypothesis is that the abnormal rhythms will have higher reconstruction error. You will then classify a rhythm as an anomaly if the reconstruction error surpasses a fixed threshold.</p><p><strong>Loading and visualization of dataset</strong></p><pre>plt.grid()<br>plt.plot(np.arange(140),normal_train_data[0])<br>plt.title(“A Normal ECG”)<br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/370/0*2gAaDTxAbGIoBjiR" /></figure><pre>plt.grid()<br>plt.plot(np.arange(140),anomalous_train_data[0])<br>plt.title(“An Anomalous ECG”)<br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/370/0*IbgLt1BTyQBeD6gr" /></figure><p><strong>Implementing the autoencoder for Anomaly Detection</strong></p><pre>class AnomalyDetector(Model):<br> def __init__(self):<br> super(AnomalyDetector,self).__init__()<br> self.encoder = tf.keras.Sequential([<br> layers.Dense(32,activation=”relu”),<br> layers.Dense(16,activation=”relu”),<br> layers.Dense(8,activation=”relu”)<br> ])<br> self.decoder = tf.keras.Sequential([<br> layers.Dense(16,activation=”relu”),<br> layers.Dense(32,activation=”relu”),<br> layers.Dense(140,activation=”sigmoid”)<br> ])<br> def call(self,x):<br> encoded = self.encoder(x)<br> decoded = self.decoder(encoded)<br> return decoded<br>autoencoder = AnomalyDetector()</pre><p><strong>Model Training</strong></p><pre>history = autoencoder.fit(normal_train_data,normal_train_data,<br> epochs=20,<br> batch_size=512,<br> validation_data = (test_data,test_data),<br> shuffle=True)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/662/0*z6q7Tf3zd9vg1_2p" /></figure><p><strong>Model Performance Visualization</strong></p><pre>encoded_data = autoencoder.encoder(normal_test_data).numpy()<br>decoded_data = autoencoder.decoder(encoded_data).numpy()</pre><pre>plt.plot(normal_test_data[0],’b’)<br>plt.plot(decoded_data[0],’r’)<br>plt.fill_between(np.arange(140),decoded_data[0],normal_test_data[0],color=’lightcoral’)<br>plt.legend(labels = [“Input”,”Reconstruction”,”Error”])<br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/370/0*B64C5Wz-G6tG1I4e" /><figcaption>Model Result</figcaption></figure><p><strong>The entire code implementation for anomaly detection can be found </strong><a href="https://github.com/prateekupadhyay1997/Autoencoders.git"><strong>here</strong></a></p><p>If you have any feedback feel free to reach out to me on <a href="https://twitter.com/Prateek13268676">twitter</a>.</p><p><strong>References :</strong></p><ol><li><a href="https://towardsdatascience.com/applied-deep-learning-part-3-autoencoders-1c083af4d798">https://towardsdatascience.com/applied-deep-learning-part-3-autoencoders-1c083af4d798</a></li><li><a href="https://www.tensorflow.org/tutorials/generative/autoencoder">https://www.tensorflow.org/tutorials/generative/autoencoder</a></li><li><a href="https://www.deeplearningbook.org/">https://www.deeplearningbook.org/</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=81892446ae21" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Universal Usability]]></title>
            <link>https://medium.com/@Deeplearning_AI/universal-usability-89ea8b93882b?source=rss-58db1debbd67------2</link>
            <guid isPermaLink="false">https://medium.com/p/89ea8b93882b</guid>
            <category><![CDATA[user-interface]]></category>
            <category><![CDATA[universal-usability]]></category>
            <category><![CDATA[human-comuter-interaction]]></category>
            <dc:creator><![CDATA[Prateek Upadhyay]]></dc:creator>
            <pubDate>Fri, 09 Oct 2020 13:40:23 GMT</pubDate>
            <atom:updated>2020-10-09T13:40:23.563Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>— “A Comprehensive Design Tactic for Web Usability”</strong></h3><h3><strong>What is Usability?</strong></h3><p>Understanding in Layman’s terms, we can break down the word ‘usability’ into two root words namely ‘use’ and ‘ability’. So, heading by its literal meaning, we can say that usability takes into account the very cognition of the target user while using any particular web interface.</p><p>To be more vivid, we can explain usability as the extent of the efficiency with which software and such other computing applications are used to accomplish the expected motives efficiently and adequately. The level of complexity that is entailed while using a particular interface is estimated by its usability.</p><p>Here, we can use the term <strong>‘Universal Usability’</strong> to depict the enlarged horizon of users in global perspective.</p><p>The word ‘universal’ signifies it’s affinity to bring into inclusion all kinds of netizens. So, Universal usability attributes to the idea of information and communication assistance that are suitable for users belonging to different hierarchy and age groups.</p><p>The concept of ‘universal usability’ was first introduced by <strong><em>Ben Shneiderman, </em></strong>a computer scientist at the ‘Human Computer Interaction Lab’ at the University of Maryland, College Park.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*mLqXa-Dg_egv8xr2DR-lYg.jpeg" /><figcaption><em>Fig 1. Image of Ben Shneiderman</em></figcaption></figure><p>He closely associated the concept of ‘Universal usability’ and ‘Design for all’ to denote the usage, access and convenience level of majority of the households enforcing them as successful users of information and communication services at par with regular basis.</p><h3><strong>Advantages of Universal usability</strong></h3><p>Universal usability is primly obligated to ensure that the professional as well as academics is accessible for learners with disability. As we can see that approachability is the most quintessential part of inclusive Education today. Designing convenient physical spaces, customised e- content and learning contexts demands for additional thought and outlining in every stage of implementation. Universal usability is a virtue for a greater expanse of mass coming under the spectrum of differentiated learning styles, choices and potential.</p><p>Here, we can give emphasis on <strong>“<em>Digital Assistive Technology”</em></strong></p><p>As we know that technology today has become the bone marrow for the netizens. In case of the specially-abled learners too, technology has opened its Pandora’s Box. Digital assistive technology is one such boon in the arena of Universal usability.</p><p>Assistive technology basically means usage of any device, equipment, software or programs in order to assist the people with specific liabilities. They can overcome the challenges and get their work accomplished with the help of the advanced technologies. Some prominent examples of assistive technology widely used all over the world are text-to-text speech and word prediction etc.</p><p>Despite of all the pros, we also come across few hurdles in the implementation of Universal usability, let us spread some light over the cons that we need to focus on -</p><ol><li><strong>Rifts in user’ awareness</strong></li></ol><p>The very first step to overcome this would be to make the users familiar with the information and communication products and services, so that they can manually relate it with their level of understanding.</p><p><strong>2. Heterogeneity of users</strong></p><p>One of the major pros of universal usability is that it welcomes users with diverse skills sets, expertise, belonging to different generations, users with liabilities and impairments, cultural variations of people who represent the fluctuating literacy level of the nation. These disparities will apparently help the designers to customise and make improvements to deal with the glitches.</p><p><strong>3. Paradigm shift through advanced technology</strong></p><p>Universal usability basically establishes the access and equity between wide ranges of applications. The privilege of the netizens being able to adopt anything from the web and convert it into its desirable form of content is also noteworthy. Especially in today’s world of creators and influencers, it’s very much evident how the right display of content changes the life of multitude of viewers and vice versa.</p><p>Assistive technology is like a virtue for the section of people who work with challenges constantly struggling with their learning incapability. It helps them to gain confidence, and communicate in a better way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QIUExl6VxeDXROpEVdjV1Q.jpeg" /><figcaption><em>Fig 2. Understanding Assistive Technology: How Do People with Mobility Disabilities Use Technology?</em></figcaption></figure><h3><strong>Universal usability — From an Indian perspective</strong></h3><p>In India, Universal usability is still striving to attain its rightful space in the software engineering process. There are some institutions where they have a very well established space. The participation of usability practitioners are also estimated in the entire project budgeting and project outlining process but there are very few institutions where it is done so religiously. People are still trying to learn usability through articles published on websites, by attending workshops here and there and they are trying to exercise it, learn it on the field, or on the job.</p><p>With the passage of time, this very schema has evolved in such a way that it affects the other domains and successfully brings about positive changes. Here we can cite the example of education system, communication, logistics, e-governance etc. Keeping in mind the population explosion and the majority of it coming under the rural periphery where access and connectivity is yet another challenge to surpass universal usability in India is yet to gain momentum.</p><h3><strong>Conclusion</strong></h3><p>After a detailed analysis, I can conclude that Universal usability is suitably used in conjunction with user-centred design, a process by which an output is devised according to the requirements and blueprints of users. Within the last several years, the usability of Web sites has become a hot issue for Web developers. Many significant web sites engage usability engineers to assure that they have an easy-to-use, friendly site that provides a convincing customer experience. Apart from that, advancement of technology in every sphere will bring about noticeable changes and promote it even beyond imagination.</p><h3><strong>References</strong></h3><ul><li><a href="https://www.umiacs.umd.edu/people/ben">https://www.umiacs.umd.edu/people/ben</a></li><li><a href="http://web.mit.edu/16.459/www/UnivUsab.pdf">http://web.mit.edu/16.459/www/UnivUsab.pdf</a></li><li><a href="http://universalusability.com/">http://universalusability.com/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=89ea8b93882b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>