<?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 AIQ on Medium]]></title>
        <description><![CDATA[Stories by AIQ on Medium]]></description>
        <link>https://medium.com/@Artificialintelligencequotient?source=rss-f9d65f147a25------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xPdZ_QVKHKYqSzEun3L5NA.png</url>
            <title>Stories by AIQ on Medium</title>
            <link>https://medium.com/@Artificialintelligencequotient?source=rss-f9d65f147a25------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 12:03:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Artificialintelligencequotient/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[What Happens When the AI You Depend On Isn’t Yours to Keep?]]></title>
            <link>https://medium.com/@Artificialintelligencequotient/what-happens-when-the-ai-you-depend-on-isnt-yours-to-keep-19684c7062b9?source=rss-f9d65f147a25------2</link>
            <guid isPermaLink="false">https://medium.com/p/19684c7062b9</guid>
            <category><![CDATA[fable-5]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude-fable-5]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <dc:creator><![CDATA[AIQ]]></dc:creator>
            <pubDate>Tue, 23 Jun 2026 21:31:11 GMT</pubDate>
            <atom:updated>2026-06-23T21:31:11.762Z</atom:updated>
            <content:encoded><![CDATA[<p>When the Fable 5 news landed, my first reaction was not “oh no, the model is down.” It was “wait, a government did that?”</p><p>Quick background if you missed it: Anthropic launched Fable 5 and Mythos 5 on June 9th. Three days later, on June 12th, the U.S. Department of Commerce issued an export-control directive. Anthropic couldn’t figure out how to segment its users by nationality fast enough, so it did the only thing it could do: It turned the models off…for everyone… globally…within hours.</p><p>I want to sit with that for a second. Not because of the specific politics (that’s a separate conversation), but because of what it revealed about how most businesses are currently built on top of AI.</p><p>Snyk, the developer security company, wrote a piece about it that I couldn’t stop thinking about. They called it a demonstration of “model-layer dependence,” which is a technical phrase that deserves a plain English translation. Here’s how I think about it: if your business has a workflow, a product, or a customer-facing feature that runs through a single AI provider’s specific model, you have just handed a critical piece of your operations to a third party you do not control. That third party can be shut down not by a bug or an outage or anything they did wrong, but by forces neither you nor they have any say in, like a government directive tied to a geopolitical dispute involving a Korean telecom company with “supposed China ties,” which is per the AI Daily Brief what apparently accelerated this particular ban.</p><p>Businesses that had built on Fable 5 didn’t get a grace period, they got hours.</p><p>The part that made me stop was the irony buried in Snyk’s reporting. The “jailbreak” the government cited as justification? It was asking the model to read a specific codebase and fix any software flaws. IBM X-Force researchers had simultaneously been complaining that Fable 5 was too paranoid, rejecting requests that were “tangentially cyber related” even for legitimate defensive security work. The thing that got it banned was basically its most useful feature, depending on who you asked.</p><p>What I took away from all of this has less to do with Fable specifically and more to do with a mental model I think every business leader should have right now. The AI Daily Brief put it well: the teams that handled this week without chaos weren’t the ones who found the best single replacement model. They were the ones who had already built for redundancy, with multiple models and smart routing. So when Fable went dark, instead of a fire drill, it was a configuration update.</p><p>This is not a technical problem. It’s the same logic as not sourcing everything from a single supplier or not running your whole company on a single cloud provider. The AI era is introducing a new layer of dependency that most businesses haven’t thought through yet. You can own your data, control your infrastructure, and still have a single government directive remove the engine from your product overnight.</p><p>I keep thinking about all the companies that are right now in the process of building AI deeply into something customer-facing. Are any of them asking what happens if the model goes away?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19684c7062b9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[China’s chips are better than the headlines say]]></title>
            <link>https://medium.com/@Artificialintelligencequotient/chinas-chips-are-better-than-the-headlines-say-ac6e34dbb4d8?source=rss-f9d65f147a25------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac6e34dbb4d8</guid>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[us-china-chip-war]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AIQ]]></dc:creator>
            <pubDate>Tue, 16 Jun 2026 00:46:00 GMT</pubDate>
            <atom:updated>2026-06-16T00:46:00.662Z</atom:updated>
            <content:encoded><![CDATA[<p>The story everyone tells about US chip export controls goes like this: cut off China’s access to the most advanced chipmaking equipment and you slow down their AI ambitions. I’ve heard that version across nearly every podcast and newsletter I follow for the past two years. It’s become almost axiomatic.</p><p>This week, analyst Dylan Patel at SemiAnalysis published something that put a dent in it.</p><p>SemiAnalysis just built a physical chip teardown lab in Oregon called STEEL, specifically to reverse-engineer and analyze the world’s most important semiconductors. Their first public report dropped this week, and their target was the chip inside Huawei’s latest smartphone, made by China’s SMIC foundry.</p><p>Their finding: SMIC’s newest manufacturing process has a minimum metal pitch (basically, how tightly you can pack the wiring inside a chip) of 32.5 nanometers (roughly 3,000 times thinner than a human hair, if that helps). Intel’s latest chips on their much-anticipated 18A process measure 36nm. SMIC is about 10% tighter.</p><p>That number surprised me. Here’s why it should surprise you too.</p><p>For the past several years, the US government has operated on a specific bet: if you block China from accessing the most advanced chipmaking equipment, you block China from making advanced chips. The key piece of equipment is an EUV lithography machine, made by a Dutch company called ASML, which uses extreme ultraviolet light to etch incredibly fine circuit patterns onto silicon. The US pressured the Netherlands to stop exporting EUV machines to China. No EUV, no advanced chips. Theory closed.</p><p>What SMIC did instead was use older machines they already had, called DUV, and run multiple exposure passes on the same chip layer to achieve finer resolution. Think of it like tracing the same outline twice to make the line darker. It’s slower, more labor-intensive, and less power-efficient than EUV. But on the density metric that matters most, they got there.</p><p>To be clear about what “getting there” does and doesn’t mean: SMIC’s chips are not a straight replacement for TSMC’s most advanced production. The SemiAnalysis teardown notes that SMIC achieves comparable density but pays for it in complexity, efficiency, and manufacturing precision. Their yields are lower, their chips run hotter, and scaling this process is harder. The headline metric is close. The full picture is more complicated.</p><p>But here’s what stops me from writing this off as a footnote: China’s AI company Zhipu already trained a 744 billion parameter model entirely on Huawei chips made by SMIC. Real AI work, at serious scale, is happening on domestically produced silicon right now. That’s the part that changes the conversation.</p><p>The US Bureau of Industry and Security revised its export control stance in January 2026, moving away from automatic denial toward case-by-case review. That policy shift might reflect what the SemiAnalysis teardown is now showing in hard data: the original theory about what the controls would prevent has run into the reality of what they actually prevented.</p><p>What I find genuinely interesting about this, separate from the geopolitics, is the dynamic it illustrates. Constraints tend to produce different kinds of innovation than abundance does. SMIC did not get access to the best tools. They made something competitive anyway. That story keeps showing up in technology, and it rarely ends where anyone expected.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac6e34dbb4d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can new AI startups survive?]]></title>
            <link>https://medium.com/@Artificialintelligencequotient/can-new-ai-startups-survive-22066aae15dd?source=rss-f9d65f147a25------2</link>
            <guid isPermaLink="false">https://medium.com/p/22066aae15dd</guid>
            <dc:creator><![CDATA[AIQ]]></dc:creator>
            <pubDate>Wed, 24 Apr 2024 21:01:17 GMT</pubDate>
            <atom:updated>2024-04-25T02:51:00.282Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1W9jsLV-GoIUBrmcqroZUg.png" /></figure><h4>“The financial artillery that you need to create one of these [LLMs] is going to wall out anyone but the biggest players.” — Ezra Klein</h4><p>In the evolving world of artificial intelligence, competition is undergoing a cataclysmic shift. AI is no longer a quaint feature nestled in the toolkit of tech giants; it is the battleground on which the future of technology, the economy, and humanity as we know it, will be decided. With incumbents boasting simulation technology akin to sci-fi novels, the question arises across the industry: Can smaller AI startups survive?</p><p>At the heart of this puzzle is the incredible escalation in the cost and complexities of pioneering AI models. New, state-of-the-art models require a staggering amount of data, computing power and financial firepower. Dario Amodei’s insights into the ‘scaling laws’ I laid out in my <a href="https://medium.com/@Artificialintelligencequotient/a-data-drought-23e96f27e52e">earlier post</a> give way to the financial barriers facing any would-be competitor to industry juggernauts.</p><p>Consider the economics of the computing power necessary to even begin building a fresh model. Larger models necessitate more compute per iteration to outdo rivals. They need to run models for longer periods, injecting an increasing amount of data — and that number of GPU chips (the backbone of AI computation) multiplied by the time they use those chips, is essentially dollar value — as chips are commonly rented by the hour.</p><p>Training OpenAI’s GPT-4 for instance, used an estimated $78 million worth of compute, while Google’s Gemini Ultra cost a staggering $191 million for compute, <a href="https://aiindex.stanford.edu/wp-content/uploads/2024/04/HAI_AI-Index-Report-2024.pdf">reports</a> the AI Index Report 2024.</p><p>Dario <a href="https://podcasts.apple.com/us/podcast/what-if-dario-amodei-is-right-about-a-i/id1548604447?i=1000652234981">notes</a> that the models in training now cost closer to a billion dollars. “I think in 2025, 2026, we’ll get more towards $5 or $10 billion.”</p><p>The scale of data required and the sheer computational power needed to process it all often leads to costs spiraling.</p><p>Despite the daunting data scale and immense computational prowess needed, large corporations are increasingly eager to invest in AI startups recognizing their need for fresh innovations. Likewise, newer startups need the resources and capital that big tech has to offer. <a href="https://aiindex.stanford.edu/wp-content/uploads/2024/04/HAI_AI-Index-Report-2024.pdf">Funding</a> for AI startups in 2023 octupled from the year prior reaching $25.2 billion with the major tech players — namely Microsoft, Alphabet, Amazon, and Nvidia — both investing in <em>and</em> partnering with emerging AI players.</p><p>Sam Altman, OpenAI’s CEO, <a href="https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/">says</a> the era of creating bigger, faster models is winding down. “I think we’re at the end of an era where it’s going to be these, like, giant, giant models. We’ll make them better in other ways.” Instead future advancements in AI will spring from innovative, groundbreaking ideas.</p><p>This shift in focus raises questions about the future landscape of the tech industry. In the age of AI, will the same big tech companies remain the <em>same</em> big tech companies… only more formidable?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22066aae15dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A data drought: challenges of information scarcity in AI]]></title>
            <link>https://medium.com/@Artificialintelligencequotient/a-data-drought-23e96f27e52e?source=rss-f9d65f147a25------2</link>
            <guid isPermaLink="false">https://medium.com/p/23e96f27e52e</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[technology-news]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[AIQ]]></dc:creator>
            <pubDate>Tue, 16 Apr 2024 05:45:07 GMT</pubDate>
            <atom:updated>2024-04-25T02:52:41.777Z</atom:updated>
            <content:encoded><![CDATA[<p>Since its debut on November 20, 2022, the ChatGPT phenomenon marked a pivotal moment in the trajectory of artificial intelligence. AI surged forward at an unprecedented pace, competitors emerged and the capabilities of these models have defied expectations. But how fast are AI timelines and will innovation continue to charge forward at the same speed?</p><p>The visionary co-founder of Anthropic, Dario Amodei, developed a set of principles to predict and quantify this rate at which AI is advancing. These principles are the compass guiding us through the dizzying landscape of AI advancement. Simply put, his ‘scaling laws’ say the more compute power and data we pour into AI systems, the more they evolve, morphing into formidable entities capable of feats previously deemed unimaginable. The relationship is exponential. (<a href="https://podcasts.apple.com/us/podcast/what-if-dario-amodei-is-right-about-a-i/id1548604447?i=1000652234981"><em>Ezra Klein Show</em></a>)</p><h3><strong>How LLMs are trained</strong></h3><p>Large Language models (LLMs) like Chat GPT, Gemini and Claude are reared on a diet of high and (gasp!) low quality data. Think of the data as a pizza that’s been cut into slices, or <a href="https://media.hubspot.com/e3t/Ctc/UB+113/d2q-Vf04/VXjjbz2g_KJhN3MHdMGdDGVKW7HlLKX5cXp75N4Rsyfd4Zwp2W5BWr2F6lZ3lhVTtm0H6CgHl4W6f4z4N8Sw3RWN2BMYDVRX47dW4WSYJw2vNLCCW6XQN1N6mj3McW4k1N0t7QZWSpW1HctrF52qBrfW1bn9K91hrjMJW8BhLvw75NwsHW7mxbT43ps0WRW7HMFZ57GnvvVW7PBnbm6BfMdyN956GBKT6hhRN1bSF0m2Qqh-W1QRP5W1913RtW3Btr8m6wLCVhW920PjW5S1MbJW23pxCP8-p-RTW3bN2Xq48bBnGW3TPryv4-HRnCW4RGcrb4GGyvDN1bpHNgZC0XDW5BF_xW896nGsW6D5WpL4khVhqN3G6w7kZqtXmW16_4wY76ThZDW80qvWZ9j8jwcW4g8B7t54vbpGW2gNgWm5nX_HvW15GwdR6VL1bKW5j0ZdX4vgHD8VCK7xF1F5wlhW7C9v1W6Zw9f2V6szgs7xH7_mds1MqH04">tokens</a> as they’re called. Tokens can range from full words to their fragments or even single characters.</p><p>Where does this data come from? Data sets are gathered from a range of sources spanning from a company’s own proprietary data — think Google mining data from Google Search, Maps, etc. Data can be scraped from websites or publishing sources — think Reddit, Wikipedia or New York Times articles for example. Of course data sets can be purchased or can even be generated by LLMs themselves (this is called synthetic data — more on that later). French AI startup Mistral apparently feeds their models “reams of 19th-century French literature, which is no longer subject to copyright law” (<a href="https://www.nytimes.com/2024/04/12/business/artificial-intelligence-mistral-france-europe.html?smid=nytcore-ios-share&amp;referringSource=articleShare"><em>NYT</em></a>)</p><p>AI researcher Pablo Villalobos estimates that GPT-4 trained on up to 12 trillion tokens (think 12 trillion pieces of pizza, if we’re continuing with that analogy) <a href="https://www.wsj.com/tech/ai/ai-training-data-synthetic-openai-anthropic-9230f8d8?utm_campaign=mb&amp;utm_medium=newsletter&amp;_hsenc=p2ANqtz-__lsiTRY9dOi0nvAWChy8wD2a43fKjSXYrxDmGSRzj33Kl9HWZjBwVPrtt9nHil7pNqrftK3-V1PQH-dvGplCWTemSYg&amp;_hsmi=302600334&amp;utm_source=morning_brew">per</a> The Wall Street Journal. He estimates a GPT-5 would need 60 trillion to 100 trillion tokens.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*reT-nz1RrPRKnz-Yf7Ktbg.png" /></figure><h3>A data heist</h3><p>“Companies are racing to get as much data as they can from any sources they can get their hands on” says Kevin Roose said on <a href="https://podcasts.apple.com/us/podcast/hard-fork/id1528594034?i=1000652238000"><em>HardFork</em></a>.</p><p>The <a href="https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html"><em>New York Times</em></a><em> </em>reports OpenAI has been scraping YouTube videos — more than 1M hours of YouTube videos were feed into GPT4. They even built Whisperer which according to their <a href="https://openai.com/research/whisper">blog post</a>, “is an automatic speech recognition (ASR) system trained on 680,000 hours of multilingual and multitask supervised data collected from the web.” In other words, when they ran out of available English language data, they turned to transcribing text from videos through their own speech recognition tool.</p><p>Meta for one, has been struggling to keep up according to the <a href="https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html"><em>New York Times</em></a>, because they don’t have access to as much good, high quality data within their products to train their models.</p><p>The story <a href="https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html">reports</a>, “Ahmad Al-Dahle, Meta’s vice president of generative A.I., told executives that his team had used almost every available English-language book, essay, poem and news article on the internet to develop a model, according to recordings of internal meetings, which were shared by an employee. Meta could not match ChatGPT unless it got more data, Mr. Al-Dahle told colleagues.”</p><blockquote>It’s no overstatement to say the collective of LLMs today comprise cumulatively of nearly every piece of digital data available. The more data they consume, the more powerful and capable they become. But alas, there’s a plot twist. Data is becoming scarce!</blockquote><p>The shortage of quality data and the constant quest for data sources have put a speed bump on the AI advancement expressway — and fears there could be a further slowdown as experts concern an even greater data drought within 2 years.</p><p>This crunch has given birth to ingenious data acquisition strategies amongst AI startups. Meanwhile, the big fish in the corporate sea have turned to dubious methods to amass data, often keeping the sources under wraps.</p><h3><strong>So what is synthetic data?</strong></h3><p>Well, it’s like asking an LLM to cook its own meal. It can whip up data that simulates real-world data. LLMs can cook up a diverse platter of synthetic data, such as text, code, and even images. This technique comes in handy when real-world data is scarce, sensitive, or too pricey to collect and label.</p><p>Tech titan Microsoft, for instance, birthed and launched a family of small models called <a href="https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/">Phi</a>, who were spoon-fed on synthetic data produced by LLMs. Isn’t that adorable?</p><p>Air Street Press <a href="https://press.airstreet.com/p/data-acquisition-strategies-for-ai?utm_source=post-email-title&amp;publication_id=43676&amp;post_id=143261614&amp;utm_campaign=email-post-title&amp;isFreemail=true&amp;r=ir7sc&amp;triedRedirect=true&amp;utm_medium=email">presents</a> a reality check though stressing, “No matter how sophisticated, it can only ever create an approximation of a problem domain. Over-reliance on it risks the model overfitting to the characteristics present in the synthetic data generation process.”</p><p>In the grand scheme of AI’s development, the integration of synthetic data highlights the complex balancing act between innovation, speed and ethics, a challenge that echoes the scaling laws — where the pursuit of larger, more capable models necessitates careful consideration of the quality and diversity of the data that trains them, synthetic or not.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=23e96f27e52e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Like Rosetta Stone, but for AI news.]]></title>
            <link>https://medium.com/@Artificialintelligencequotient/like-rosetta-stone-but-for-ai-news-30b23f91c59a?source=rss-f9d65f147a25------2</link>
            <guid isPermaLink="false">https://medium.com/p/30b23f91c59a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[technology-news]]></category>
            <category><![CDATA[technology-trends]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AIQ]]></dc:creator>
            <pubDate>Tue, 16 Apr 2024 03:21:01 GMT</pubDate>
            <atom:updated>2024-04-16T05:47:18.341Z</atom:updated>
            <content:encoded><![CDATA[<p>Just like you, I’m hearing about AI everywhere. It’s truly omnipresent. Over the past year and a half, the every day capabilities of machine learning and artificial intelligence have grown from mere text predictors to catalysts of what might just be the most transformative movement in our lifetime. For better or for worse, the ride is already moving.</p><p>If I’m honest, I find this all incredibly thrilling. I’ve burrowed down the rabbit hole of AI, exploring its many corners with an insatiable curiosity. I love the theatrics of it all — the influential players, the dynamic power struggles, the fierce competition, and how the reigning champions are fortifying their strongholds.</p><p>Each week, my diet consists of a hefty portion of AI news and info, served mostly via podcasts. My mind is practically brimming with all things AI, so I figured, why not share what I’ve learned? Think of this series as your AI Rosetta Stone, transforming intricate concepts into manageable, easily-digestible pieces..</p><p>So without further ado, let’s dive into a series of posts where I’ll be distilling the latest AI trends I’ve been tracking. I sincerely hope you find it as captivating as I do and I’d love to hear your thoughts or topic suggestions.</p><p>And, just to keep things transparent — I use <a href="https://www.jasper.ai/">Jasper</a> to aid my writing. Because after all, isn’t it great when we humans and AI work together?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30b23f91c59a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>