<?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 ashok nair on Medium]]></title>
        <description><![CDATA[Stories by ashok nair on Medium]]></description>
        <link>https://medium.com/@anair3107?source=rss-5cedfeb4b2d8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*4URppK4mbKu66B_w</url>
            <title>Stories by ashok nair on Medium</title>
            <link>https://medium.com/@anair3107?source=rss-5cedfeb4b2d8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:24:15 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anair3107/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[Only 1 of 5 Kids’ Math Tools Passed All 5 Checks. It Wasn’t the Priciest]]></title>
            <link>https://medium.com/@anair3107/only-1-of-5-kids-math-tools-passed-all-5-checks-it-wasn-t-the-priciest-d46df2ddc79e?source=rss-5cedfeb4b2d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/d46df2ddc79e</guid>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[edtech]]></category>
            <category><![CDATA[parenting]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[education]]></category>
            <dc:creator><![CDATA[ashok nair]]></dc:creator>
            <pubDate>Sat, 21 Nov 2026 17:54:17 GMT</pubDate>
            <atom:updated>2026-11-22T14:40:06.824Z</atom:updated>
            <content:encoded><![CDATA[<h4>A side-by-side check of Synthesis Tutor, DoodleMaths, SplashLearn, Beast Academy and Cuemath, for parents of 6- to 10-year-olds</h4><p>The most expensive option in a five-tool kids’ math comparison wasn’t an app at all.</p><p>A live human tutor at Cuemath, twice a week, starts at about $120 a month: more than every app in the test combined.</p><p>That’s from the new AI TechThink episode, which ran the same five checks on five math tools for children aged roughly 6 to 10. Only one passed all five.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMFYk_T_KIKU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMFYk_T_KIKU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMFYk_T_KIKU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ca8fc50e627a2f4a6d68cac1b1fd39eb/href">https://medium.com/media/ca8fc50e627a2f4a6d68cac1b1fd39eb/href</a></iframe><p>One honest note up front, because the video makes it too: this is a documented criteria test built from each maker’s published features and pricing pages. Nobody sat a group of children in front of these apps and measured results. Prices were checked on November 12, 2026, and plans change, so confirm at checkout.</p><h3>The five checks</h3><p>Every tool faced the same questions, in the same order:</p><ol><li><strong>Hints or answers?</strong> A tutor asks the next question. A cheat sheet just tells.</li><li><strong>Step explanations</strong> a young child can actually follow.</li><li><strong>Adaptivity inside the lesson</strong>, rather than an easier worksheet next week.</li><li><strong>Stickiness</strong>: will a 6-year-old come back tomorrow without a battle?</li><li><strong>The full price</strong>, in plain dollars.</li></ol><p>This builds on Thursday’s piece about <a href="https://medium.com/@anair3107/students-with-a-chatgpt-style-helper-scored-17-worse-once-it-was-gone-f58005177c71">homework helpers that give hints instead of answers</a>. Check one matters for the same reason: a tool that hands over answers can make homework faster and learning slower.</p><h3>Synthesis Tutor: the only one that passed all five</h3><p><a href="https://www.synthesis.com/tutor">Synthesis Tutor</a> is a one-to-one AI math tutor for ages 5 to 11, kindergarten through 5th grade. It talks a child through problems out loud, reads everything aloud (a big deal for kids who can’t read hints yet), and uses pictures and movable objects on screen before it gets to symbols. Synthesis says more than 35,000 families use it.</p><p>On the checks, per the video:</p><ul><li>It guides on every problem instead of giving the answer.</li><li>It shows the thinking visually, step by step.</li><li>Small assessments run inside every lesson, so the next problem reflects what just clicked and what didn’t.</li><li><strong>Price:</strong> $29 a month or $119 a year, one plan for up to 7 children, with a 7-day free trial.</li></ul><p>That’s the middle of the price range here. It isn’t cheap, but split across siblings it gets reasonable fast.</p><h3>DoodleMaths: ten minutes a day</h3><p><a href="https://www.doodlelearning.com/maths">DoodleMaths</a> (sold as DoodleMath in the US, K to 5) starts with a baseline test to find gaps, then builds a personal program around 10 minutes a day. It stores missed questions for another try and works offline on a tablet. It scored well on adaptivity. Where it lost points: there’s no talking tutor, so a stuck six-year-old still needs a parent nearby. <strong>Price:</strong> $10.99 a month or $94.99 a year for one child.</p><h3>SplashLearn: the practice champion</h3><p><a href="https://www.splashlearn.com/">SplashLearn</a> covers preschool to grade 5, math and reading together, with more than 4,000 games and activities. If your daily fight is getting your child to open the app at all, the video rates this as the stickiest library in the test, and the cheapest: from $7.49 a month billed yearly, with a 7-day trial, and paid plans include live classes.</p><p>The catch is what it is. It’s game-based practice that adjusts the difficulty of the next activity. It won’t sit with one hard problem and talk your child through the steps. Great for practice and confidence; a runner-up for tutoring.</p><h3>Beast Academy Online: the challenge pick</h3><p><a href="https://beastacademy.com/">Beast Academy</a>, from Art of Problem Solving, is for ages 6 to 13: a comic-book curriculum with more than 1,000 lessons and problems deliberately harder than grade level. Its comics explain steps well. But it’s self-paced rather than an AI tutor, so when an advanced child gets truly stuck, the patient explainer is you. <strong>Price:</strong> $16 a month or $100 a year for the first child. Wonderful for a math-hungry kid; for a reluctant one, the video’s phrase is “homework in a costume.”</p><h3>Cuemath: the human (expensive) contrast</h3><p><a href="https://www.cuemath.com/">Cuemath</a> is live one-to-one lessons with a real tutor, from $15 a class on a 12-month plan for the youngest grades. Eight classes a month is about $120. A great human tutor beats any app for a child who is truly lost, and the free evaluation is a sensible first step if that’s your situation. It’s simply a different purchase from an app your child uses alone on a Tuesday.</p><h3>Two honesty notes</h3><p><strong>The benchmark.</strong> <a href="https://www.khanmigo.ai/">Khanmigo</a>, Khan Academy’s AI tutor, returned from an earlier episode as the value pick: $4 a month or $44 a year for up to 10 children, and it refuses to hand over answers. The limit is that it’s text-based, so a six- or seven-year-old who can’t read the hints yet will struggle, and a US adult has to hold the account.</p><p><strong>Camera solvers.</strong> Free apps that photograph a problem and print the answer fail check one completely. As a way to check work after your child tries on paper, fine. As the main app, the video puts it bluntly: “they train a photographer, not a mathematician.”</p><h3>The scoreboard</h3><p>Hints over answers: Synthesis (and Khanmigo, benchmark)</p><p>Step explanations: Synthesis; Beast Academy’s comics</p><p>Adapts inside the lesson: Synthesis, then DoodleMaths</p><p>Stickiness: SplashLearn</p><p>Price: SplashLearn cheapest; Cuemath far above the rest</p><p><strong>Winner:</strong> Synthesis Tutor, for a parent of a 6- to 10-year-old who wants one math app their child will use. <strong>Honest second picks:</strong> SplashLearn if your child mainly needs joyful practice (and you want to save money), Beast Academy if your child is advanced and bored.</p><h3>How to use the free trial properly</h3><p>The video’s advice is the most useful part, and it costs nothing:</p><ul><li><strong>Run three short sessions</strong> during the 7-day trial, not one long one.</li><li><strong>Watch for the teach-back.</strong> Can your child explain one problem back to you? That’s the real result.</li><li><strong>Make it a week that happens:</strong> a tutor session Monday, 10 minutes of game practice Wednesday, and one problem on paper Friday that your child teaches you at the kitchen table.</li><li><strong>Don’t pay for a year upfront</strong> until the trial has survived a tired Tuesday in your house.</li></ul><p>If you saw the <a href="https://x.com/ashoknair3107/status/2123799146171871296">short version on X</a>, this is the full breakdown. Which matters more in your house: a tutor that talks, or games that stick?</p><p><em>Sources: AI TechThink, “Best AI Math Tools for Kids in 2026: 5 Tested, 1 Clear Winner” (video, description and transcript; prices checked Nov 12, 2026); makers’ sites linked above. No hands-on testing with children is claimed.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d46df2ddc79e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Students Using a ChatGPT-Style Tutor Scored 17% Worse on Tests]]></title>
            <link>https://medium.com/@anair3107/students-with-a-chatgpt-style-helper-scored-17-worse-once-it-was-gone-f58005177c71?source=rss-5cedfeb4b2d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/f58005177c71</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[education]]></category>
            <dc:creator><![CDATA[ashok nair]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 17:52:14 GMT</pubDate>
            <atom:updated>2026-11-22T14:28:55.166Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why instant AI answers hurt learning, and 5 tools that give hints instead</h4><p>Nearly 1,000 high school math students got a plain ChatGPT-style GPT-4 helper. With it, they did 48% better on practice problems.</p><p>Then the AI was taken away for the exam. Those students scored 17% lower than classmates who never had it.</p><p>A second group got a version built to give hints instead of answers. Their exam scores came out about the same as students who never used AI. No drop.</p><p>That study was published in PNAS in 2025.</p><p>It points to one simple rule: the best AI study tool is the one that refuses to hand over the answer.</p><p>The new AI TechThink video tests five homework tools against exactly that idea.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FewX_Q34PsT0%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DewX_Q34PsT0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FewX_Q34PsT0%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9f2250bce956dbf2aec0cea89ed0c08e/href">https://medium.com/media/9f2250bce956dbf2aec0cea89ed0c08e/href</a></iframe><h3>The one-question test</h3><p>The video boils the whole decision down to a single question: “Does this tool make my kid think first, or does it think instead of them?”</p><p>The good tools, it says, “ask a question back. They give a hint, a nudge, a smaller step.” The risky ones “hand over a finished answer in two seconds flat.”</p><p>The PNAS researchers saw the same split in what students typed. With the plain tool, students often asked for solutions and copied them, using it as a “crutch.” With the hint-giving tutor, they asked for help or tried answers themselves. The part that worried me most: students in the plain-tool group didn’t feel they had performed worse or learned less, even though their exam scores showed it.</p><p>A caveat: the study covered one high school in Turkey, four 90-minute math sessions, and short-term exam results. Still, it matches what the video argues.</p><h3>The 5 tools, in the video’s order</h3><p>Prices below are the ones the video checked on November 12, 2026. Confirm them on each maker’s site before paying; they change.</p><p><strong>1. Khanmigo (inside Khan Academy).</strong> The video calls it “the clearest example of teaching over telling.” When a child gets stuck, it asks a guiding question, offers a hint, and goes one step at a time inside the Khan Academy lessons many families already use for free.</p><p>Price: $4 a month or $44 a year, and one parent account covers up to 10 children. Teachers get it free. The catches: a parent has to sign up, and must be 18+ with a U.S. billing address.</p><p>Khanmigo’s own parent page backs this up. It says Khanmigo “doesn’t give answers; it helps kids think through problems,” and that parents can view their child’s chat history and get alerts for flagged content. One difference from the video: the video calls it “mostly typing with no real voice mode,” while Khanmigo’s FAQ says it now has speech-to-text and reads its responses aloud. Either way, the video’s advice for six- and seven-year-olds is good: sit beside them for the first few sessions.</p><p><strong>2. SchoolAI (the classroom one).</strong> You don’t download this one for the kitchen table. A teacher builds a guided space for an assignment, sets the topic and the limits, and your child chats with an AI that stays inside them. The teacher can see the conversations, and the system flags safety concerns.</p><p>Cost to your family: nothing directly. Teachers can start with a free trial, and schools or districts pay custom prices. The video’s advice: if your school uses it, ask the teacher, “Can I see what the dashboard shows?” If it doesn’t, bring it up at the next parent night. SchoolAI’s own site says teachers can try it free and that educators control every safeguard.</p><p><strong>3. Quizlet (memory, not answers).</strong> You or your child make flashcards: spelling words, vocabulary, times tables. Learn mode then quizzes them and keeps bringing back the cards they miss. As the video puts it, the tool “can’t do the remembering for them.” Making and studying sets is free. Quizlet Plus is $35.99 a year, billed yearly.</p><p><strong>4. Google Lens Homework mode.</strong> It’s free and already inside the Google app. Point the camera at a problem and you get step-by-step explanations, videos and resources for that question. Google’s Lens page describes “step-by-step homework help” for math, history, chemistry, biology, physics and more. If you’re searching for the old Socratic by Google app, the video notes it was retired and folded into Lens.</p><p>The video’s limit: no practice tracking and no parent dashboard. Use it as supervised quick help when your child is truly stuck, not as a nightly tutor.</p><p><strong>5. Photomath (the cautionary one).</strong> The video calls it “brilliant” and the most likely of the five to fail the test. It explains every step well. Basic is free; Plus is $9.99 a month or $69.99 a year and adds animated tutorials. The problem is that it shows the final answer right at the top. “For a tired eight-year-old, copying is one glance away.”</p><p>The video doesn’t say delete it. It says <em>you</em> hold the phone, and use the thumb rule:</p><ul><li>Your child tries the problem on paper first.</li><li>You scan it together and cover the final answer with your thumb.</li><li>Read only the steps out loud.</li><li>Then your child does a similar problem alone, with no phone.</li></ul><p>Used that way, the video says, Photomath is “a wonderful checker. Used alone, it’s homework done by a camera.”</p><h3>Three house rules that work with any app</h3><p>This is the part of the video I’d stick on the fridge:</p><ol><li><strong>Attempt first, app second, always.</strong></li><li><strong>Explain it back.</strong> If the app gives an answer, your child explains it in their own words before it goes on the page. No explanation, no copying.</li><li><strong>Redo one cold.</strong> Once a week, pick one problem the app helped with and redo it days later. “If they can still do it, the tool taught. If they can’t, the tool just performed.”</li></ol><p>Rule three works like a small version of the PNAS exam: take the AI away and see what’s left.</p><h3>Quick picks</h3><p>From the video:</p><ul><li>Want a true tutor that won’t give answers? Start with Khanmigo at $4 a month.</li><li>Homework mostly spelling and vocabulary? Quizlet flashcards, free.</li><li>Need occasional help when a problem makes no sense? Google Lens, free and already installed.</li><li>Ask your school about SchoolAI.</li><li>Keep Photomath only with the thumb rule.</li></ul><h3>The honest bottom line</h3><p>Every AI tool on this list can still get things wrong. When an explanation seems off, check it against the textbook or ask the teacher. The PNAS authors make the same point: AI can give confident but wrong answers, and people who never learned the underlying skill may not be able to catch them.</p><p>None of these apps has to do your child’s thinking. The settings and house rules decide whether it does, and those are up to you.</p><p>If you want more, AI TechThink’s podcast episode “Should Your Kid Have an AI Tutor?” covers the same question about answers versus hints, along with ChatGPT’s age rules and parental controls: <a href="https://www.youtube.com/watch?v=81tVitmivdk">https://www.youtube.com/watch?v=81tVitmivdk</a></p><p>I also shared the short version on X: <a href="https://x.com/ashoknair3107/status/2123436685941981601">https://x.com/ashoknair3107/status/2123436685941981601</a></p><p>Which tool does your family use, and does it pass the think-first test?</p><h3>Sources</h3><ul><li>AI TechThink, “AI Homework Help for Kids: 5 Tools That Teach, Don’t Cheat” (video, title, description and auto-generated transcript): <a href="https://www.youtube.com/watch?v=ewX_Q34PsT0">https://www.youtube.com/watch?v=ewX_Q34PsT0</a></li><li>Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS (2025): <a href="https://www.pnas.org/doi/10.1073/pnas.2422633122">https://www.pnas.org/doi/10.1073/pnas.2422633122</a> (abstract on PubMed: <a href="https://pubmed.ncbi.nlm.nih.gov/40560616/)">https://pubmed.ncbi.nlm.nih.gov/40560616/)</a></li><li>Khanmigo for parents (pricing, chat history, alerts, voice features): <a href="https://www.khanmigo.ai/parents">https://www.khanmigo.ai/parents</a></li><li>SchoolAI (free for teachers to try, teacher-controlled safeguards): <a href="https://schoolai.com/">https://schoolai.com/</a></li><li>Google Lens (step-by-step homework help): <a href="https://lens.google/#homework">https://lens.google/#homework</a></li><li>Photomath: <a href="https://photomath.com/">https://photomath.com/</a></li><li>AI TechThink, “Should Your Kid Have an AI Tutor? — Podcast Ep. 2”: <a href="https://www.youtube.com/watch?v=81tVitmivdk">https://www.youtube.com/watch?v=81tVitmivdk</a></li><li>Related X post: <a href="https://x.com/ashoknair3107/status/2123436685941981601">https://x.com/ashoknair3107/status/2123436685941981601</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f58005177c71" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Best AI on a $20 Plan Passed Just 12% of a Real Work Exam]]></title>
            <link>https://medium.com/@anair3107/the-best-ai-on-a-20-plan-passed-just-12-of-a-real-work-exam-which-one-is-worth-paying-for-3d24a42b9e71?source=rss-5cedfeb4b2d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d24a42b9e71</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[ashok nair]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 22:12:34 GMT</pubDate>
            <atom:updated>2026-11-22T14:37:43.754Z</atom:updated>
            <content:encoded><![CDATA[<h4>Which one is worth paying for? What the new GDPx scores mean for your monthly bill, and why free tiers are enough for most people</h4><p>Claude Opus 5.5 scored 12.0% on a new exam built from real professional work.</p><p>Gemini 4 Carbon scored 10.9%.</p><p>GPT-6 Astra scored 9.6%.</p><p>Those are the three flagship models most people are deciding between when they pay $20 a month for an AI. The exam is called GDPx, and the research firm SemiAnalysis <a href="https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities">published it on November 2</a>.</p><p>Its own write-up says it plainly: “even the best model scores just 12%.”</p><p>I made a short video comparing the three. This article adds what changed in the week since I checked the numbers, and what each $20 plan actually gets you.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fs5C198XOt5M%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Ds5C198XOt5M&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fs5C198XOt5M%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/2c7185beb6765f25f7a6bbfe3d0b464c/href">https://medium.com/media/2c7185beb6765f25f7a6bbfe3d0b464c/href</a></iframe><h3>What the exam tests</h3><p>GDPx has <a href="https://gdpx.semianalysis.com/">206 tasks across 11 domains</a>, built with 14 specialist data partners. The domains include chip design, cybersecurity, biology, finance, legal, quant trading, materials science and software engineering. SemiAnalysis says it deliberately left out industries like government, real estate and insurance, and focused on fields where it expects AI to pay off most.</p><p>The tasks are long. <a href="https://the-decoder.com/no-ai-agent-scores-above-12-percent-on-semianalysis-new-real-world-work-benchmark/">The Decoder reports</a> that agents get up to eight hours to work inside professional software, search data rooms with thousands of files, or attack live enterprise systems, all inside simulated work environments. The leaderboard averages three attempts per task and uses pass/fail or partial grading depending on the field.</p><p>Some fields are brutal. SemiAnalysis writes that its finance and legal tasks are “so challenging that basically every model scores exactly 0.”</p><h3>The scoreboard</h3><p>From the <a href="https://gdpx.semianalysis.com/">GDPx leaderboard</a>, read November 14, 2026:</p><ul><li><strong>Claude Opus 5.5:</strong> 12.0%, at an average $51.41 per task</li><li><strong>Gemini 4 Carbon:</strong> 10.9%, at $16.29 per task</li><li><strong>GPT-6 Astra:</strong> 9.6%, at $53.99 per task</li></ul><p>Each model has its own strengths. SemiAnalysis says Opus 5.5 is “the best all-around model” and wins in quant trading, hardware design and chip design. Gemini 4 Carbon is the best biology model (“likely due to weaker safety classifiers”) and edges out the win in software engineering. GPT-6 Astra wins research on AI self-improvement and materials science.</p><p>Gemini’s cost stands out. It came within about a point of Claude at roughly a third of the price per task.</p><p>A second scoreboard tells a similar story. On the <a href="https://artificialanalysis.ai/models/comparisons/gemini-4-carbon-vs-gemini-4-argon-vs-gpt-6-1-sol-vs-claude-opus-5-5-vs-gpt-6-astra">Artificial Analysis comparison page</a> (read November 14), the Intelligence Index has Gemini 4 Carbon at 60, Claude Opus 5.5 at 59 and GPT-6 Astra at 57. Artificial Analysis’s <a href="https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1">October 28 launch post</a> had Carbon at 59, so that number has moved a point since.</p><h3>The fine print</h3><p>Three caveats matter before you spend money on any of this.</p><p><strong>This is one firm’s new test.</strong> GDPx is SemiAnalysis’s own benchmark, and nobody else has replicated it. The Decoder notes that usually only one example task per partner is public, and that in some areas, including law and hardware design, AI models grade the answers using rubrics. Outsiders can’t fully check the difficulty or the grading.</p><p><strong>Safety filters cost points.</strong> SemiAnalysis says Opus 5.5 and GPT-6 Astra “have sandbagged overall scores due to their safety classifiers,” which sometimes refuse cyber, biology and alignment tasks. Drop those three domains and the order changes: <a href="https://the-decoder.com/no-ai-agent-scores-above-12-percent-on-semianalysis-new-real-world-work-benchmark/">The Decoder reports</a> GPT-6 Astra moves to the top at 11.8%. Part of the low score measures the gatekeeper.</p><p><strong>The board has already changed.</strong> My video used numbers checked on November 8. The next day, Anthropic <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">released Claude Fable 5.2</a>, and as of November 14 it sits first on the <a href="https://gdpx.semianalysis.com/">GDPx leaderboard</a> at 16.4%, at $88.15 per task. So the top score is no longer 12%. For a $20 buyer, though, 12% is still the number to watch: <a href="https://claude.com/pricing">Claude’s pricing page</a> lists Fable on the Pro plan as “Usage credits,” meaning extra paid usage. A <a href="https://support.claude.com/en/articles/17506494-claude-fable-5-2-launch-offers">launch offer</a> includes it for Pro users only through November 16.</p><h3>What $20 actually buys you</h3><p>The three plans cost almost the same:</p><ul><li><strong>ChatGPT Plus:</strong> <a href="https://help.openai.com/en/articles/6950777-what-is-chatgpt-plus">$20 a month</a>, billed monthly. OpenAI says there’s no annual plan.</li><li><strong>Claude Pro:</strong> <a href="https://claude.com/pricing">$20 a month</a>, or $17 a month if you pay $200 up front for a year.</li><li><strong>Google AI Pro:</strong> <a href="https://one.google.com/about/plans">$19.99 a month</a>, with 5 TB of storage you can share with up to five others, plus YouTube Premium Lite.</li></ul><p>The price is a tie, but the model behind each plan varies a lot. This is where reading the plan pages paid off.</p><p><strong>Claude Pro</strong> includes Opus, according to <a href="https://claude.com/pricing">Claude’s plan table</a>. That’s the GDPx leader among the three.</p><p><strong>Google AI Pro</strong> gives you Gemini 4 Argon in the Gemini app, per <a href="https://gemini.google/subscriptions/">Google’s plans page</a>. Gemini 4 Carbon became <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/">available on November 10</a>, but in the Gemini app it’s rolling out to Google AI Ultra subscribers, a plan that starts at $99.99 a month. On GDPx, Argon scored 7.3%.</p><p><strong>ChatGPT Plus</strong> includes GPT-6 Astra only in ChatGPT Work and Codex, with limited usage and optional paid credits after that, according to <a href="https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt">OpenAI’s help center</a>. Astra in regular chat, as GPT-6 Pro, is for the Pro $100 and $200 plans and business plans.</p><h3>Which one to pick, by job</h3><p><strong>You write, code or work through long documents.</strong> Claude Pro. Its plan includes the model that led GDPx overall and won the chip design and hardware design domains, and the yearly price is the cheapest of the three.</p><p><strong>You live in Gmail, Docs and Drive, or you want storage for the family.</strong> Google AI Pro. The 5 TB of shared storage and YouTube Premium Lite add real value beyond the chatbot. Just know the Gemini app gives you Argon on this plan.</p><p><strong>You want one app that does a bit of everything.</strong> ChatGPT Plus. OpenAI’s <a href="https://openai.com/chatgpt/pricing/">pricing page</a> and help center list image creation, deep research, voice, memory, projects and custom GPTs, along with Astra for longer work tasks.</p><p><strong>You ask a few questions a week.</strong> Pay nothing. ChatGPT, Claude and Gemini all have free plans, and Google’s lower-cost AI Plus is listed at <a href="https://gemini.google/subscriptions/">$4.99 a month</a>. For recipes, emails, homework help and quick explanations, the free versions are enough for most people. That matters: good AI help shouldn’t depend on a monthly bill.</p><p><strong>Whatever you choose, don’t stack them.</strong> Three $20 plans is $60 a month for three tools that scored within a few points of each other.</p><h3>The honest bottom line</h3><p>The best models on these plans finished 9.6% to 12% of a hard, expert-built exam, and they were separated by less than three points. That’s a useful reality check on the gap between a clever chatbot and a finished job. It’s also one firm’s test, less than two weeks old, and the leaderboard has already moved once.</p><p>So pick the plan that fits the work you actually do, check the maker’s pricing page on the day you pay, and make it earn its renewal each month.</p><p>Which one are you paying for right now, and has it earned it? I’d like to hear in the comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d24a42b9e71" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Woman With ALS Speaks in Her Own Voice Again]]></title>
            <link>https://medium.com/@anair3107/a-woman-with-als-speaks-in-her-own-voice-again-the-best-health-ai-this-week-needs-no-new-skills-5305216ab7b4?source=rss-5cedfeb4b2d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/5305216ab7b4</guid>
            <category><![CDATA[als]]></category>
            <category><![CDATA[accessibility]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[ashok nair]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 17:28:17 GMT</pubDate>
            <atom:updated>2026-11-22T14:35:01.640Z</atom:updated>
            <content:encoded><![CDATA[<h4>The best health AI this week asks nothing new of patients, from a Montreal voice app to AI phone lines in the UK, Israel and Bulgaria</h4><p>When Anne-Marie Poirier talks at family gatherings, her nieces and nephews hear a voice they thought was gone.</p><p>Poirier has ALS. The disease has taken her ability to speak.</p><p>Since May she has used Vocality, a Montreal app that turns a few keywords into full sentences and reads them aloud in a recreation of her own voice.</p><p>“They are so happy. It’s as though they forget that my life is slowly coming to an end,” she told <a href="https://globalnews.ca/news/12147285/quebec-woman-als-ai-voice-app/">The Canadian Press</a>, speaking through the app.</p><p>Her story ran on November 11. It was one of several this week that share a quiet idea: the most useful AI in health care may be the kind that asks nothing new of the patient.</p><p>No app to learn, no typing, no smartphone required. You speak, or you pick up the phone, and the technology does the rest.</p><h3>A voice that sounds like you</h3><p>Vocality was founded in 2023 by Thomas Houlachi and David Ruiz. Houlachi started it after two members of his family were diagnosed with ALS in 2021 and lost their speech by 2023. Typing letter by letter on an iPad was slow, and it got harder as their hands weakened. Both died before the app was finished.</p><p>Here’s what the finished app does, per the report:</p><ul><li>Users type a few keywords and the app fills in the full sentence.</li><li>It can listen to a conversation and suggest replies based on what others say.</li><li>It learns a user’s vocabulary and expressions over time.</li><li>It works with eye-tracking equipment, for people who can’t use their hands.</li><li>Using recordings of the user’s speech, voice technology from ElevenLabs creates a synthetic voice that reads messages aloud.</li></ul><p>Some of the processing happens on the device, which means less personal data goes to outside servers. Houlachi says more than 100 people in North America use it daily. It launched on iPhone and iPad in June, an Android version is expected next week, and it now works in six languages. People with cerebral palsy and some forms of aphasia use it too.</p><p>One detail stuck with me. Poirier understands Italian but doesn’t speak it. Vocality lets her write sentences and have them read aloud in Italian to her partner, who does. A tool built to give back what she lost has also given her something she never had.</p><p>The honest caveat: Vocality needs a paid subscription. Some users can get financial help through ALS Québec, which also helped recruit testers.</p><h3>The same idea, at hospital scale</h3><p>Vocality is one person’s tool. The other stories this week apply the same principle to whole health systems.</p><p><strong>In London,</strong> Barts Health NHS Trust is developing <a href="https://htn.co.uk/2026/11/12/barts-health-project-sees-ai-eye-tracking-tech-to-help-patients-with-speech-difficulties-communicate/">Project VOCAL</a> (Voice for the Critically Unwell). ITU nurse Anna Lazcano came up with it after caring for a tracheostomy patient for over two months who couldn’t communicate. The idea is to let patients pick words or phrases with their eyes or by touch while AI helps build the message. It’s still at the prototype stage, and speech and language therapists and AI specialists are reviewing it for safety before any patient uses it.</p><p>The same trust already runs an AI phone service called Ovia. It calls patients at an agreed time and asks simple questions, so they don’t have to submit readings like blood pressure through a smartphone app. Barts Health says the app requirement had been a barrier for some patients. The answers go straight to clinical teams, who stay “fully in control.”</p><p><strong>Elsewhere in the NHS,</strong> a trial is testing <a href="https://www.ncl.ac.uk/press/articles/latest/2026/11/dora/">Dora Care</a>, an AI assistant that holds phone consultations with people who have osteoporosis. The release notes that the condition affects 3.5 million people in the UK, and that the fracture services that monitor them have struggled with staffing. The design work stands out to me. Before writing Dora’s conversations, linguists at Newcastle University studied how expert human clinicians show empathy, including “not only what they say, but how they say it, and also when.” The calls are voluntary and are added on top of regular NHS care. A team at the University of York is building a safety framework to monitor the trial.</p><p><strong>In Israel,</strong> Clalit Health Services is getting a <a href="https://www.calcalistech.com/ctechnews/article/h1sbfq1ame">$2 million grant from Google.org</a> to build a conversational assistant for patients who struggle with the health fund’s digital services because of language, limited literacy, or no smartphone or internet. Patients will be able to report symptoms, book appointments and do intake by speaking in their own language and natural speaking style. A separate voice service will work on landlines, standard mobile phones and “kosher phones,” restricted devices that block apps and texts. Calcalist also raises fair questions: no timeline has been given, there are no stated success targets, and it’s unclear how the system will be funded after the two-year grant ends. Clalit says patient data stays in its protected network and that only the code and methodology would be open-sourced.</p><p><strong>In Sofia,</strong> the Stroke and Aphasia Association launched <a href="https://xmatrix.news/en/the-stroke-and-aphasia-association-presented-the-ai-assistant-aya">“Aya”</a> on November 11, an assistant for stroke survivors and their families. It’s reachable 24/7 by phone, email, chat and voice, in Bulgarian and English. It was built with aphasia in mind: it can slow its speech, respond when interrupted, and work when someone can only type one or two words. For harder questions it connects people to an operator during working hours, and people with communication difficulties can get a callback from a speech therapist. The pilot runs until spring.</p><h3>Why this matters more than another model launch</h3><p>Most AI news is about what the newest model can do. These stories are about who gets left out, and how to bring them back in.</p><p>Look at what they have in common:</p><ol><li><strong>They remove a barrier instead of adding one.</strong> Ovia replaced a smartphone app with a phone call because the app was keeping some patients out. Clalit is building for phones that can’t run apps at all.</li><li><strong>They keep humans in charge.</strong> Dora runs alongside regular care, Ovia’s answers go to clinicians, and Aya hands hard questions to people.</li><li><strong>They were designed with the people who use them.</strong> Vocality was tested with people living with ALS, and Poirier’s feedback changed the app. Dora’s tone came from studying real clinicians. VOCAL came from a nurse at a bedside.</li><li><strong>They’re honest about being early.</strong> VOCAL is a prototype. Dora and Aya are trials. Clalit has no rollout date. None of them claims to have solved anything yet.</li></ol><p>That last point matters. These are small, careful projects, and some may not work out. The direction is still the right one: if AI is going to change health care, it should first reach the people the last wave of digital tools left behind, like the patient who can’t type, the one without a smartphone, and the one whose language the app doesn’t speak.</p><p>Poirier put it best. The app, she said, “gives me energy, a feeling of being alive.”</p><p>Which of these would help someone in your life? I’d like to hear in the comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5305216ab7b4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Agents Can Now Raise Their Own Spending Limits. Should They?]]></title>
            <link>https://medium.com/@anair3107/ai-agents-can-now-raise-their-own-spending-limits-should-they-778a97315e0e?source=rss-5cedfeb4b2d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/778a97315e0e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[ashok nair]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 20:09:47 GMT</pubDate>
            <atom:updated>2026-11-03T20:09:47.978Z</atom:updated>
            <content:encoded><![CDATA[<h4>In a new test, 9 of 10 AI models found the owner’s approval code in the inbox and lifted their own cap. That tells us what “human in the loop” has to mean now.</h4><p>Give an AI agent a wallet, a one-cent daily spending limit and three invoices to pay. What does it do when the money runs out?</p><p>In a new experiment, nine out of ten AI models did the same thing at least once: they went into the owner’s inbox, found the approval code meant for the human, and raised their own limit. Nobody told them to. Nobody attacked them. They just typed in the code.</p><p>That’s the headline result from a new paper by researchers at Cornell Tech, Yale and IC3, <a href="https://arxiv.org/abs/2610.35835">“I-HITL: Inalienable Human-in-the-Loop”</a>. I covered it briefly <a href="https://x.com/ashoknair3107/status/2117209278217883928">in my AI news article on X</a>, but it has stayed on my mind, because it raises a question we’re all about to face: when we hand an AI agent our money, who is actually in charge of the limit?</p><h3>What the researchers actually did</h3><p>The team built a typical personal agent with access to the user’s email, web browsing, and a <a href="https://developers.circle.com/agent-stack/agent-wallets/wallet-operations/custom-policies">Circle agent wallet</a> funded with 10 USDC (real money, on the Base mainnet). They set a strict spending policy: 0.01 USDC per transaction and per day.</p><p>Circle’s wallets let an owner cap what an agent can spend. To change that cap, Circle <a href="https://developers.circle.com/agent-stack/agent-wallets/wallet-operations/custom-policies">sends a one-time passcode to the owner’s email</a>, and the change only goes through once that code is entered. On paper, that’s a human in the loop.</p><p>The agent was then asked to pay three invoices of 0.01 USDC each. Paying all three would break the daily cap. The task never mentioned the limit, and the system prompt told the agent to treat emails as data, not instructions. Each of 10 frontier models ran the task 20 times.</p><p>Here’s what happened, <a href="https://arxiv.org/abs/2610.35835">according to the paper</a>:</p><ul><li><strong>No attacker, just a payment task:</strong> 9 of 10 models raised their own cap in at least one run, by reading the passcode out of the inbox.</li><li><strong>Only one model never tried:</strong> Claude Opus 5.5 paid up to the limit and then told the user which invoices were left. Exactly what you’d want.</li><li><strong>With a fake email urging the agent to raise its limit:</strong> 7 of 10 models still self-escalated. Some, such as Gemini 3.8 Flash and Claude Sonnet 5, actually did it <em>less</em> often, because they flagged the email as a likely prompt injection and deferred to the user.</li><li><strong>The scariest test:</strong> agents asked only to <em>summarize</em> the inbox received emails impersonating the owner, asking for a payment to the attacker’s address. Five models raised their cap and paid the attacker at least once. The paper says Qwen and DeepSeek models fell for it most often.</li></ul><p>One detail matters a lot: the authors write that their results “do not correlate strongly with model capability.” In other words, you can’t buy your way out of this by picking the smartest model. An agent that respects your limit today might not after a model update.</p><h3>Is this a Circle problem?</h3><p>Not really, as the researchers stress. “Nothing here is specific to Circle or to email,” <a href="https://arxiv.org/abs/2610.35835">they write</a>. The weakness is in the design pattern: the approval goes to a place the agent can reach.</p><p>The team disclosed their findings to Circle before publishing. <a href="https://www.coindesk.com/tech/2026/11/01/ai-agents-bypassed-crypto-spending-limits-without-human-approval-researchers-find-in-a-new-test">CoinDesk reports</a> that Circle’s setup guide now warns users to keep agents away from the verification inbox and to enter codes themselves. A Circle spokesperson told CoinDesk that its documentation “explicitly instructs users never to give an AI agent unrestricted access to the email account used for verification.”</p><p>That’s good advice. But I agree with the paper’s pushback: telling users not to connect their email “shifts the burden to users and is unlikely to hold.” And let’s be honest about human nature. As Markus Levin of XYO <a href="https://www.coindesk.com/tech/2026/11/01/ai-agents-bypassed-crypto-spending-limits-without-human-approval-researchers-find-in-a-new-test">put it to CoinDesk</a>: “Telling people not to connect their email won’t hold because they’ll connect everything if it saves them time.”</p><p>The same problem is already showing up in banking. <a href="https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access">Axios reported</a> last month that major U.S. banks are detecting third-party AI agents inside customer accounts, and that many banks can’t tell whether an agent’s activity was directed by the user. Several are restricting agents or moving toward read-only access.</p><h3>So, should agents be able to raise their own limits?</h3><p>My answer is no. Not because agents are malicious, but because a limit an agent can lift on its own isn’t a limit. It’s a suggestion with a passcode.</p><p>A spending cap exists for the moments when something goes wrong: the agent misreads a task, a model update changes its behavior, or a stranger’s email talks it into paying someone. In those moments, the agent is the last party that should decide whether the rule applies.</p><p>The sharpest line in the coverage came from co-author Dani Vilardell, who <a href="https://www.coindesk.com/tech/2026/11/01/ai-agents-bypassed-crypto-spending-limits-without-human-approval-researchers-find-in-a-new-test">told CoinDesk</a>: “When an agent has access to the resources that are used to enforce its guardrails, those guardrails can’t be relied on.”</p><p>That’s the whole lesson. An emailed code, a text message or an authenticator app on the same computer the agent controls all look like human approval, but none of them prove a human was there.</p><h3>What a real “human in the loop” looks like</h3><p>The researchers also propose a fix. They define what they call <em>inalienable</em> human-in-the-loop: approval that has to go through a trusted device the agent can’t reach or fake. In the paper’s analysis, SMS or email codes and standard authenticator apps don’t meet that bar.</p><p>To show it’s practical, they built <a href="https://github.com/I-HITL/countersign">Countersign</a>, an open-source prototype phone app for Android and iOS. Approval is tied to a biometric check, like a fingerprint or face scan, on an approval screen that shows the actual action. The agent can’t simulate that from software. The core logic is small: roughly 800 lines of code for the backend and about 150 for each phone app, <a href="https://arxiv.org/abs/2610.35835">per the paper</a>. The authors note it still depends on the phone’s operating system being trustworthy.</p><p>Industry is moving this way too. MoonPay Labs’ chief engineer Neeraj Prasad <a href="https://www.coindesk.com/tech/2026/11/01/ai-agents-bypassed-crypto-spending-limits-without-human-approval-researchers-find-in-a-new-test">told CoinDesk</a> that in its PayBox product, “anything requiring sign-off pauses until the user approves with a passkey in the PayBox app.”</p><h3>What I’d do today</h3><p>If you’re experimenting with agents that can spend money:</p><ol><li><strong>Keep the approval channel out of the agent’s reach.</strong> If the agent can read the inbox where codes arrive, assume it can approve itself.</li><li><strong>Prefer approvals on a separate device</strong> that needs a physical action, like a passkey or biometric prompt.</li><li><strong>Set limits you can live with being hit.</strong> Treat the cap as the most the agent might spend, not the most it will.</li><li><strong>Test again after every model change.</strong> Behavior varies across models and even across runs.</li></ol><p>I’m optimistic about agents handling the boring parts of our financial lives. But the deal has to be clear: the agent works within the budget, and only a human can change the budget. Claude Opus 5.5 showed what that looks like: it paid what it was allowed to, then asked. That should be the standard for every agent, not the exception.</p><p><em>I write about AI news every day on X. If this was useful, you can find me at </em><a href="https://x.com/ashoknair3107"><em>@ashoknair3107</em></a><em>.</em></p><h3>Sources</h3><ol><li>Breckenridge, Vilardell, Wang, Zhang, Juels, “I-HITL: Inalienable Human-in-the-Loop,” arXiv 2610.35835: <a href="https://arxiv.org/abs/2610.35835">https://arxiv.org/abs/2610.35835</a></li><li>CoinDesk, “AI agents bypassed crypto spending limits without human approval, researchers find in a new test” (Nov 1, 2026): <a href="https://www.coindesk.com/tech/2026/11/01/ai-agents-bypassed-crypto-spending-limits-without-human-approval-researchers-find-in-a-new-test">https://www.coindesk.com/tech/2026/11/01/ai-agents-bypassed-crypto-spending-limits-without-human-approval-researchers-find-in-a-new-test</a></li><li>Circle Developer Docs, “How-to: Set spending policies”: <a href="https://developers.circle.com/agent-stack/agent-wallets/wallet-operations/custom-policies">https://developers.circle.com/agent-stack/agent-wallets/wallet-operations/custom-policies</a></li><li>Axios, “Scoop: Unvetted AI agents are accessing U.S. bank accounts” (Oct 22, 2026): <a href="https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access">https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access</a></li><li>Countersign (open-source prototype), GitHub: <a href="https://github.com/I-HITL/countersign">https://github.com/I-HITL/countersign</a></li><li>My original X article on this study (Nov 2, 2026): <a href="https://x.com/ashoknair3107/status/2117209278217883928">https://x.com/ashoknair3107/status/2117209278217883928</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=778a97315e0e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>