<?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 Krishna Sankar on Medium]]></title>
        <description><![CDATA[Stories by Krishna Sankar on Medium]]></description>
        <link>https://medium.com/@ksankar?source=rss-578c7297699c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*UWMX53cnUvh6Fo7F.jpg</url>
            <title>Stories by Krishna Sankar on Medium</title>
            <link>https://medium.com/@ksankar?source=rss-578c7297699c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 07:03:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ksankar/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[Germany — It’s Economic Influence in the new World Order]]></title>
            <link>https://ksankar.medium.com/germany-its-economic-influence-in-the-new-world-order-160cb8ed5bfa?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/160cb8ed5bfa</guid>
            <category><![CDATA[germany]]></category>
            <category><![CDATA[ukraine]]></category>
            <category><![CDATA[european-union]]></category>
            <category><![CDATA[elections]]></category>
            <category><![CDATA[economic-visualization]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Sun, 23 Feb 2025 20:15:33 GMT</pubDate>
            <atom:updated>2025-02-23T20:15:33.679Z</atom:updated>
            <content:encoded><![CDATA[<h3>Germany — It’s Economic Influence in the new World Order</h3><p>The Guardian has an excellent set of videography visualizing the economic stats of Germany.</p><blockquote><strong>As a Data Scientist, this level of storytelling brings tears of joy — numbers, trends, and history woven into a visual masterpiece</strong></blockquote><p>I can only show a teaser — trust me, no summary can do justice; you got to see it yourself [<a href="https://www.theguardian.com/world/ng-interactive/2025/feb/19/the-rise-of-the-far-right-global-trade-and-ukraines-future-why-the-german-election-matters-visual-explainer">Here</a>].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fPuyQR_xZ0g5MFbSIoxEeQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=160cb8ed5bfa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Warren Buffett’s Trillion-Dollar Advice for Uncle Sam]]></title>
            <link>https://ksankar.medium.com/warren-buffetts-trillion-dollar-advice-for-uncle-sam-92043eca6306?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/92043eca6306</guid>
            <category><![CDATA[warren-buffett]]></category>
            <category><![CDATA[berkshire-hathaway]]></category>
            <category><![CDATA[wisdom-financial]]></category>
            <category><![CDATA[investment]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Sun, 23 Feb 2025 17:42:37 GMT</pubDate>
            <atom:updated>2025-02-23T21:15:43.976Z</atom:updated>
            <content:encoded><![CDATA[<h3>Warren Buffett’s Trillion-Dollar Advice for Uncle Sam 💰🇺🇸</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DnU2GsRAaScEI4iZSkaVQA.jpeg" /></figure><h3>A Master Class in Economic Wisdom</h3><p>Poetic, compassionate, and razor-sharp — Warren Buffett delivers words of wisdom to the U.S. administration in Berkshire Hathaway’s Annual Report [<a href="https://berkshirehathaway.com/2024ar/2024ar.pdf"><em>Here</em></a>]. Of course, for the fans of the <strong><em>Oracle of Omaha</em></strong>, these insights are more than just financial advice — they are <strong><em>a masterclass in economic foresight and common sense.</em></strong></p><p>Buffett wraps up his letter with a direct message to the U.S. government:</p><blockquote>“<strong>So thank you, Uncle Sam</strong>”, <em>writes Buffett</em>, “<strong>Someday your nieces and nephews at Berkshire hope to send you even larger payments than we did in 2024. Spend it wisely.</strong></blockquote><blockquote><strong>Take care of the many who, for no fault of their own, get the short straws in life. They deserve better.</strong></blockquote><blockquote><strong>And never forget that we need you to maintain a stable currency and that result requires both wisdom and vigilance on your part.</strong>”</blockquote><blockquote>“<strong>I thought honestly in a very subtle way that was a powerful message</strong>,” said CFRA Research analyst Cathy Seifert.</blockquote><p>Despite <strong><em>capitalism’s imperfections</em></strong>, Buffett remains bullish on <strong><em>America’s economic resilience</em></strong>, crediting it for the country’s <strong><em>extraordinary progress over 235 years</em></strong>.</p><p>Buffett writes <strong><em>“From a base of only four million people — and despite a brutal internal war early on, pitting one American against another — America changed the world in the blink of a celestial eye.</em></strong>”</p><h3><strong>Billion-Dollar Advice? Make That Trillion-Dollar Advice!</strong> 💵</h3><p>As of today, BRK-A has a Market Cap of <strong>$1T</strong> (yep, T and in Trillion!) My initial title for this blog was “<strong><em>Billion Dollar Advice</em></strong>”, but honestly “<strong><em>Trillion Dollar Advice</em></strong>” seems far more fitting !!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZWsmZFuTnnzj8iWYcCmI0Q.png" /></figure><h3><strong>Buffett’s Key Insights: A Distilled Dose of Wisdom</strong> 📖</h3><p>To help you navigate through the 150 pages (13 page Chair’s letter + various 10-Ks, A-1 and A-2), here’s a <strong><em>top 10 curated summary</em></strong> of Buffett’s wisdom !!</p><h3>🔟 Final Thoughts: Buffett’s Legacy Lives On:</h3><p>Buffett’s <strong>2024 letter</strong> is packed with <strong>wisdom, humor, and timeless investment lessons</strong>. His message remains clear:</p><blockquote>💡 <strong>Acknowledge mistakes. Stay patient. Invest for the long haul<br>💰 Avoid risky spending. Hoard cash until great opportunities arise<br>🇺🇸 America’s best days are still ahead — if managed wisely</strong></blockquote><blockquote><strong>🏫 Never look at where a candidate has gone to school. You will see them from all walks of life</strong></blockquote><p>The Special Letters — from Charlie (<a href="https://berkshirehathaway.com/SpecialLetters/CTM%20past%20present%20future%202014.pdf"><strong><em>Vice Chairman’s Thoughts — Past and Future</em></strong></a>) and from Warren (<a href="https://berkshirehathaway.com/SpecialLetters/WEB%20past%20present%20future%202014.pdf"><strong><em>Berkshire — Past, Present and Future</em></strong></a>) are also a good additional read !</p><h3>1️⃣ Introduction: The Purpose of the Annual Report</h3><p>Warren Buffett opens the letter by emphasizing that Berkshire Hathaway’s annual report isn’t just a legal obligation but <strong><em>a means of transparent communication with shareholders</em></strong>.</p><ul><li>His philosophy is to provide <strong><em>the type of insight he would expect if the roles were reversed </em></strong>— if shareholders were the CEO and Buffett was a passive investor trusting them with his savings.</li></ul><h3><strong>2️⃣ Mistakes — Yes, We Make Them at Berkshire</strong></h3><p>A beautiful section, you should read it ! I will give couple of quotes</p><ul><li>Buffett candidly discusses his history of mistakes, particularly in <strong>capital allocation and personnel decisions</strong>. He acknowledges that errors in evaluating businesses and managers have been inevitable, but the most important lesson, as his longtime partner Charlie Munger taught him, is to take swift corrective action rather than “thumb-sucking” and delaying necessary decisions.</li></ul><blockquote><em>Buffett writes</em> “<strong>During the 2019–23 period, I have used the words “mistake” or “error” 16 times in my letters to you. Many other huge companies have<em> never</em> used either word over that span. Amazon, I should acknowledge, made some brutally candid observations in its 2021 letter. Elsewhere, it has generally been happy talk and pictures</strong>”</blockquote><p>Buffett warns against <strong><em>the illusion of managerial perfection</em></strong>:</p><blockquote>He adds candidly “<strong>That taboo (</strong>“mistake” or “wrong” were forbidden words at board meetings or analyst calls<strong>), implying managerial perfection, always made me nervous (</strong>though, at times, there could be legal issues that make limited discussion advisable. We live in a very litigious society<strong>)</strong>”</blockquote><p><strong>Key Lesson? <em>Don’t sugarcoat mistakes. Own them. Fix them. Move forward.</em></strong></p><h3>3️⃣ Berkshire’s Record-Breaking Tax Payments 💰</h3><p>In a stunning milestone, Berkshire Hathaway became <strong>the largest corporate taxpayer in U.S. history</strong>, paying <strong>$26.8 billion in federal income taxes in 2024</strong> — about <strong>5% of total U.S. corporate tax receipts</strong>.</p><blockquote>To illustrate the magnitude of Berkshire’s tax contributions, Buffett explains, with humor and pride “<strong>If Berkshire had sent the Treasury a $1 million check <em>every 20 minutes</em> throughout all of 2024 — visualize <em>366 days and nights </em>because 2024 was a leap year — we still would have owed the federal government a significant sum at yearend. Indeed, it would be well into January before the Treasury would tell us that we could take a short breather, get some sleep, and prepare for our 2025 tax payments !</strong>”</blockquote><p><strong>Why So Much?</strong> <em>Berkshire </em><strong><em>reinvests profits instead of paying dividends</em></strong><em>, compounding its tax burden over time. </em><strong><em>Lesson?</em></strong><em> Compounding isn’t just for investments — it applies to taxes too !</em></p><h3>4️⃣ Will You Be Getting the <em>60 Years of BRK</em> Book? 📚</h3><p>Buffett invites shareholders to Berkshire’s annual meeting on May 3rd, 2025, promising engaging discussions and a showcase of Berkshire-owned brands. This year, a special edition book, <strong>“60 Years of Berkshire Hathaway”</strong>, will be available.</p><blockquote>I was able to get a copy of the previous edition - “<strong>Berkshire Hathaway Inc.: Celebrating 50 Years of a Profitable Partnership</strong>” ! I am keeping an eye on the book shops checking for the availability of the 60 Years book — probably will be in the market after May 3rd.</blockquote><blockquote>If you’re a Buffett fan, this <strong>one’s a must-read !</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_w16CPwqBDB3oC1t9ov06A.png" /></figure><h3>5️⃣ Cash Position — Waiting for the right opportunity ?</h3><p>Berkshire now holds <strong><em>$334 billion cash</em></strong>, double the <strong><em>$168 billion</em></strong> cash Berkshire <strong><em>held a year ago</em></strong>.</p><blockquote>With its massive cash reserves, Berkshire could acquire nearly any US company outside of the biggest giants.</blockquote><ul><li>Market experts suggest that Buffett’s reluctance to deploy cash <strong><em>reflects high valuations rather than a negative market outlook </em></strong>[1]</li><li>He writes “<strong><em>Paper money can see its value evaporate if fiscal folly prevails. In some countries, this reckless practice has become habitual, and, in our country’s short history, the U.S. has come close to the edge!</em></strong>”</li><li>Buffett himself addressed this issue [5] during last annual general meeting (when they had $182B in cash !) “<em>We’d love to spend it, but we won’t spend it unless we think we’re doing something that has very little risk and can make us a lot of money.</em>” He added, “<strong><em>We only swing at pitches we like. It isn’t like I’ve got a hunger strike or something like that going on. It’s just that things aren’t attractive</em></strong><em>.</em>”[3]</li><li>Another factor contributing to the cash buildup is the <strong><em>temporary halt in stock buybacks</em> </strong>in the third quarter. The company maintains that <strong><em>buybacks only occur when Buffett believes the stock is trading below its intrinsic value</em></strong>. [5]</li><li>But there are other voices as well — Investor Bill Smead of Smead Capital Management said Buffett’s actions show he’s actually “<strong><em>bearish as hell but won’t admit it</em></strong>”. Buffett likely thinks <strong><em>the stock market is terribly expensive.</em></strong></li></ul><h3>6️⃣ Berkshire’s Expanding Footprint in Japan 🇯🇵</h3><p>Berkshire has been steadily increasing its investments in <strong>five major Japanese trading companies (Itochu, Marubeni, Mitsubishi, Mitsui, Sumitomo)</strong>. Initially purchased in 2019 because of their low valuations, Buffett now sees them as long-term holdings.</p><ul><li>These companies align with Berkshire’s investment philosophy i.e., <strong>Dividend growth, Share buybacks at sensible prices and Conservative executive compensation.</strong></li></ul><blockquote>Berkshire’s Japanese investments had a market value of <strong>$23.5 billion </strong>at year-end, significantly above their <strong>$13.8 billion cost basis</strong>.</blockquote><p>Berkshire’s holding has been keeping below the 10% — as they approach the loimit the five companies have relaxed the celing and so Berkshire would increase it’s holding.</p><blockquote>Buffett highlights the company’s use of yen-denominated debt to hedge currency risk.</blockquote><h3>7️⃣ CEO Succession: Greg Abel Takes the Helm</h3><p>Buffett, now 94, acknowledges that his time leading Berkshire Hathaway is nearing an end, and he reaffirms that <strong><em>Greg Abel will succeed him as CEO</em></strong>. He expresses confidence that Abel shares Berkshire’s values and commitment to honest reporting. He warns against the dangers of companies fooling their shareholders, which can lead to management deceiving itself.</p><blockquote>Buffet of CEO selections: “<strong>I never look at where a candidate has gone to school. Never!</strong>”</blockquote><h3>8️⃣ Investment Strategy: Controlled vs. Marketable Equities</h3><p>Buffett discusses Berkshire’s <strong>two primary investment approaches viz. Controlling entire businesses</strong> (e.g., GEICO, BNSF, Berkshire Hathaway Energy) and <strong>Owning minority stakes in high-quality public companies</strong> (e.g., Apple, American Express, Coca-Cola, Moody’s).</p><ul><li>While <strong>direct ownership offers more control</strong>, Buffett acknowledges that great companies are rarely available for outright purchase. Instead, <strong><em>Berkshire takes advantage of Wall Street’s occasional mispricing of publicly traded businesses</em></strong>. (<strong><em>hint, hint, hint … if you see Buffett buying shares, it is an indication of an arbitrage opportunity 💱</em></strong>)</li></ul><h3>9️⃣ Insurance: Berkshire’s Underrated Cash Machine</h3><ul><li><strong>Property-Casualty (P/C) Insurance</strong> remains <strong>Berkshire’s financial backbone</strong>, generating <strong>$32 billion in after-tax underwriting profits</strong> over the last <strong>20 years</strong>.</li><li>Buffett warns about <strong>mispricing risk</strong>:</li></ul><blockquote><em>“</em><strong><em>With long-tail insurance, a company may report large but fictitious profits for years — even decades — before realizing it’s been losing money all along.</em></strong><em>”</em></blockquote><p>His <strong><em>biggest concern? Climate change.</em></strong></p><blockquote><strong>“Convective storms are increasing, and climate change may have been announcing its arrival.”</strong></blockquote><p>Berkshire <strong><em>prices policies conservatively</em></strong>, ensuring they <strong><em>never chase market share at the expense of profitability</em></strong>.</p><h3>Well, That’s all Folks !</h3><blockquote><strong>Did this blog inspire you to dive into Buffett’s work ? Let me know your thoughts! 🚀</strong></blockquote><h3>References</h3><ol><li><a href="https://www.business-standard.com/world-news/why-warren-buffett-s-cash-pile-is-growing-and-what-it-means-for-markets-125022200530_1.html">https://www.business-standard.com/world-news/why-warren-buffett-s-cash-pile-is-growing-and-what-it-means-for-markets-125022200530_1.html</a></li><li><a href="https://nypost.com/2025/02/22/us-news/warren-buffett-urges-trump-to-spend-government-funds-wisely/">https://nypost.com/2025/02/22/us-news/warren-buffett-urges-trump-to-spend-government-funds-wisely/</a></li><li><a href="https://apnews.com/article/warren-buffett-berkshire-hathaway-annual-shareholders-letter-89d882488c22faf86c684993022c1428">https://apnews.com/article/warren-buffett-berkshire-hathaway-annual-shareholders-letter-89d882488c22faf86c684993022c1428</a></li><li><a href="https://www.aol.com/finance/only-swing-pitches-warren-buffett-110100744.html">https://www.aol.com/finance/only-swing-pitches-warren-buffett-110100744.html</a></li><li><a href="https://www.business-standard.com/world-news/why-warren-buffett-s-cash-pile-is-growing-and-what-it-means-for-markets-125022200530_1.html">https://www.business-standard.com/world-news/why-warren-buffett-s-cash-pile-is-growing-and-what-it-means-for-markets-125022200530_1.html</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=92043eca6306" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ChatGPT, no reason can do, hmm ? Even at 4o!]]></title>
            <link>https://ksankar.medium.com/chatgpt-no-reason-can-do-hmm-even-at-4o-838d84a91b36?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/838d84a91b36</guid>
            <category><![CDATA[hallucinations]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Sat, 18 May 2024 02:06:55 GMT</pubDate>
            <atom:updated>2024-05-18T17:46:04.351Z</atom:updated>
            <content:encoded><![CDATA[<h3>ChatGPT, reason there is not, hmm? Even at 4o!</h3><p>As Yoda would say it …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q8JhCqmkc_Oy0bZAxPkpjw.png" /></figure><p>Enough said — the following interaction w/ ChatGPT says it all.</p><p>The problem is that ChatGPT doesn’t really reason. There is a lesson for us on the nature of LLMs — powerful, yet very naive. Still stochastic Parrots ! My blog [<a href="https://ksankar.medium.com/chatgpt-the-smooth-talking-stochastic-parrot-d655e6c0aae7"><strong><em>Here</em></strong></a>]</p><p>Is it hallucination ? An error ? Lack of training ? No reasoning capability ? You be the judge …</p><p>This is a fundamental characteristic of the LLM — not something we can patch by external mechanisms</p><blockquote><strong>See the interaction below (5.17.24 GPT-4o/Plus Plan) — I had to try it myself !</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*edY7V-NjHnMws_23KvaoUQ.png" /></figure><h3>References:</h3><ol><li><a href="https://amistrongeryet.substack.com/p/gpt-4-capabilities">https://amistrongeryet.substack.com/p/gpt-4-capabilities</a></li><li>Based on Rich Heimann’s LinkedIn Post <a href="https://www.linkedin.com/posts/rheimann_there-is-a-popular-puzzle-called-thewolf-activity-7196958934023237635-qr6D?utm_source=share&amp;utm_medium=member_desktop">https://www.linkedin.com/posts/rheimann_there-is-a-popular-puzzle-called-thewolf-activity-7196958934023237635-qr6D</a></li><li>That refer’s to Gary Marcus’ post</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*uW-zfaEbOxQVdeET" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=838d84a91b36" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Llama 3 Trivia : For the Enquiring Minds]]></title>
            <link>https://ksankar.medium.com/llama-3-trivia-for-the-enquiring-minds-a51c38ac90f9?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a51c38ac90f9</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[llama-3]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[meta]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Sun, 21 Apr 2024 20:51:09 GMT</pubDate>
            <atom:updated>2024-04-21T20:51:09.291Z</atom:updated>
            <content:encoded><![CDATA[<h3>Llama 3 Trivia : For the Enquiring Minds</h3><p>Meet Llama 3, Meta’s latest leap into the world of Generative AI — this new foundation model is setting the stage for some exciting developments.</p><blockquote>Wondering what makes Llama 3 tick and it’s consequences ? Let me dive into the intuitive side of things and unpack what this could mean for the future of AI.</blockquote><p>For those of you hungry for the hardcore technical details, check out these links: [<a href="https://llama.meta.com/llama3/"><em>Here</em></a>], [<a href="https://ai.meta.com/blog/meta-llama-3/"><em>Here</em></a>], and GitHub [<a href="https://github.com/meta-llama/llama3"><em>Here</em></a>]. Let’s get into it!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x0qK5lDammCRSYUC" /></figure><h3><strong>1 — Correlation or Causation ? — Nvidia falls 10% as Llama 3 hits the market</strong></h3><p>May be both, but rest assured Llama 3 does make developing grounds up foundation models a luxury for most of the companies.</p><ul><li>The word on the street is that <strong><em>Meta spent anywhere between $100M — $1B, over 7.7M GPU hours on (reportedly) 600,000 H100s.</em></strong></li><li><em>Please add comments as and when you see more reliable numbers. The </em><a href="https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md"><em>model card</em></a><em> has good details.</em></li></ul><blockquote><strong>And, free to use till and commercial license required only if you have more than 700M monthly active users (I think almost all of us fall in that group of plebeians </strong>🙃<strong>) If you don’t believe me, please check their </strong><a href="https://github.com/meta-llama/llama3/blob/main/LICENSE"><strong>license</strong></a><strong> !!</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Dm5Y3xP1vzp7UqlLpSe5g.png" /></figure><h3><strong>2 — But, Meta still can’t get more R-E-S-P-E-C-T</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZIwBfSLphHb927HIm2UCCg.png" /></figure><blockquote>“Basically they leveled the playing field in AI while keeping their edge on the social media market “ — Heard at Twitter or LinkedIn</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SIvdfKNKK6DXCNhd3w6fvQ.png" /></figure><h3><strong>3 — Llama 3 does disrupt the market , even though it is still an evolution and not a revolution</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pXp4Co4DAZp-vTAKM7m6ew.png" /></figure><h3><strong>4 — Data is everything !</strong></h3><p>I think the major innovation in Llama 3 is the way they use data</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8v8-rInpWGyveXzXvU7Z5A.png" /></figure><p>Llama 3 is only 300 lines of code (of course, over millions of PyTorch). You just need something robust that you can pour trillions of tokens of high quality training data into (15T+ for Llama 3). (<em>Ref : Nice write up by Tobias Zwingmann in LinkedIn [</em><a href="https://www.linkedin.com/posts/tobias-zwingmann_meta-facebook-just-spent-over-100000000-activity-7187500623704076288-_vbG"><em>Here</em></a><em>]</em>)</p><ul><li>Lots of interesting data layer tweaks — Better tokenizer, Massive pre-training corpus, Filtering pre training data, Overtraining/Post training data quality, curating the data mix (for best performance) and performing multiple rounds of QA on human annotations.</li><li>The evolution in architecture helped - but attention to data quality and the data pipeline are the main innovation</li><li>Their data filtering pipelines include heuristic filters, NSFW filters, semantic deduplication approaches, and text classifiers to predict data quality. And this was done mainly by Llama 2 !</li></ul><h3><strong>5 — OpenAI &amp; Google will catchup</strong></h3><p>Not sure if this prediction will come true — it will be self evident by the time you read this blog tomorrow !</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_cIWGRbMWGOwb_2GQWdmDg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MAozzeJgwgBlFx5Q" /></figure><blockquote>I think this picture is a good footnote for this blog !</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a51c38ac90f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 7 from (502 pages of) Stanford’s 2024 AI Index Report]]></title>
            <link>https://ksankar.medium.com/top-7-from-502-pages-of-stanfords-2024-ai-index-report-662031a632f5?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/662031a632f5</guid>
            <category><![CDATA[stanford-university]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Tue, 16 Apr 2024 00:44:42 GMT</pubDate>
            <atom:updated>2024-04-16T02:17:08.428Z</atom:updated>
            <content:encoded><![CDATA[<p>Interesting report. <a href="https://aiindex.stanford.edu/report/">https://aiindex.stanford.edu/report/</a></p><blockquote>It is only 502 pages long. But my blog is much shorter 😇</blockquote><blockquote><strong>And, am sure you will be relieved to hear that no Generative AI LLMs were harmed during the creation of the blog …</strong></blockquote><blockquote>This is from a quick review — I plan to go over one more time (at least) for a deep dive …</blockquote><p><strong>1 AI beats humans on some tasks, but not on all</strong></p><ul><li><em>Nothing new. The Google AGI levels is of relevance here.</em></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Glfh0TROp5-57JOd2NVPqg.png" /></figure><p><strong>2 Frontier models get way more expensive</strong></p><ul><li><em>According to AI Index estimates, the training costs of state-of-the-art AI models have reached unprecedented levels. For example, </em><strong><em>OpenAI’s</em></strong><em> </em><strong><em>GPT-4</em></strong><em> used an estimated </em><strong><em>$78M </em></strong><em>worth of compute to train, while </em><strong><em>Google’s Gemini</em></strong><em> Ultra costs </em><strong><em>$191M</em></strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yfqmDM8Z9oglWB99o1SJcg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b8awAWH4QvuSY38BQTszNA.png" /><figcaption>Training Cost of Foundation Models</figcaption></figure><p><strong>3 Robust and standardized evaluations for LLM responsibility are seriously lacking</strong></p><ul><li><em>New research from the AI Index reveals a significant lack of standardization in responsible AI reporting. Leading developers, including OpenAI, Google, and Anthropic, primarily test their models against different responsible AI benchmarks. This practice complicates efforts to systematically compare the risks and limitations of top AI models</em></li><li><em>MACHIAVELLI Benchmark — an interesting framework for assessing traits such as deception, morality, and power-seeking in sophisticated AI systems ! — Interesting read</em></li></ul><p><strong>4 People across the globe are more cognizant of AI’s potential impact — and more nervous</strong></p><ul><li>Keep this in mind as we develop Generative AI systems</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m-hD5OOrXTqy8KzByyMxrA.png" /></figure><p><strong>5 Thanks to LLMs, robots have become more flexible (and interactive)</strong></p><ul><li><em>Robots become very interactive - Beyond their improved robotic capabilities, these models can ask questions, which marks a significant step toward robots that can interact more effectively with the real world</em></li></ul><p><strong>6 More technical research in agentic AI</strong></p><ul><li><em>Yep, LAMs (Large Agent Models) will rule …</em></li></ul><p><strong>7 Political deepfakes are easy to generate and difficult to detect</strong></p><blockquote><strong>Good, I have exactly 7 ! What caught your attention ?</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=662031a632f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Generative AI needs deeper Organizational Surgery]]></title>
            <link>https://ksankar.medium.com/generative-ai-needs-deeper-organizational-surgery-349986066c6a?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/349986066c6a</guid>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[mckinsey]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Wed, 06 Mar 2024 08:06:35 GMT</pubDate>
            <atom:updated>2024-03-07T01:47:17.745Z</atom:updated>
            <content:encoded><![CDATA[<h3>Effective Generative AI &amp; Digital Transformation needs deeper Organizational Surgery — McKinsey</h3><p>McKinsey has a thought-provoking narrative that should resonate deeply with those seeking to leverage the potential of Generative AI. Let me summarize the five ideas (and a couple of quotes) that stood out for me.</p><blockquote><strong>I am not a huge fan of McKinsey, but found some kindred souls on this one !</strong></blockquote><p>You should definitely read the blog [<a href="https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/a-generative-ai-reset-rewiring-to-turn-potential-into-value-in-2024"><strong><em>here</em></strong></a>], alongside additional insightful readings — a book [<a href="https://www.mckinsey.com/featured-insights/mckinsey-on-books/rewired"><strong><em>here</em></strong></a>] and the “Rewired In Action Case Collection” [<a href="https://www.mckinsey.com/capabilities/mckinsey-digital/how-we-help-clients/rewired-in-action"><strong><em>here</em></strong></a>].</p><blockquote>I have started reading the Rewired book — plan to write a blog summarizing the ideas w.r.t <strong>Digital Transformation &amp; Generative AI - A dynamic duo !</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cltmNmgLZb4JbGi8a8bsrA.jpeg" /></figure><h4><strong>1. Generative AI has enormous potential, … but the payoff may only come when companies do deeper organizational surgery on their business</strong></h4><ul><li>As I had written “<strong><em>Hallucination as a Feature not a Bug</em></strong>” [<a href="https://bit.ly/hallucination-is-a-feature"><strong><em>Here</em></strong></a>], only by employing Generative AI <strong><em>natively</em></strong> can we extract it’s value.</li></ul><blockquote><strong>Opportunistic applications (where Generative AI is used now) are not an optimal sweet spot for Generative AI</strong></blockquote><blockquote><strong>While “faster &amp; cheaper and with less people” is worth pursuing, the focus should be on “doing things better to create a competitive advantage” — which Generative AI can do, once these fantastic beasts are tamed properly !!</strong></blockquote><ul><li>For example, “<em>Customer service is a commodity capability, not part of the core business for most companies. While Generative AI might help with productivity in such cases, it won’t create a competitive advantage</em>.”</li></ul><h4>2. Launching pilots is (relatively) easy; getting pilots to scale &amp; create meaningful value is hard because they require a broad set of changes to the way work actually gets done</h4><ul><li>This another important point. Developing a prototype is easy because the foundation models do a lot of the work — so we can quickly show interesting results.</li><li>But deploying them in production requires lot more effort — Guardrails, Red teaming, Observability, Risk Evaluation, Governance and Compliance are all things that take time and thought. <strong><em>They can’t be solved by prompt engineering or RAG !</em></strong></li></ul><blockquote>For my new readers I have blogs on the guardrails [<a href="https://bit.ly/llm-guardrails"><strong>here</strong></a>] and Red Teaming [<a href="https://bit.ly/ai-red-teaming"><strong>here</strong></a>]</blockquote><ul><li>Organizations need to set up the technology architecture to scale — composability, orthogonal extensibility are all important architectural pragmas</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mb5ruPI5NA4S_PrRextEPg.png" /></figure><h4>3. Up-skill the talent you have but be clear about the Generative AI-specific skills you need</h4><ul><li>But they aren’t the normal coders or data scientists. <em>The best Generative AI technical talent has design skills to uncover where to focus solutions, contextual understanding to ensure the most relevant and high-quality answers are generated, collaboration skills to work well with knowledge experts and strong forensic skills to figure out causes of breakdowns</em></li></ul><h4>4. Form a centralized team to establish standards that enable responsible scaling</h4><ul><li>I am a big proponent of a centralized cross-functional Generative AI WG that is in the middle of all the activities (see figure below)</li></ul><blockquote><strong>To facilitate responsible scaling, this centralized cross-functional Generative AI Working Group (WG) emerges as a strategic necessity. This team serves as the linchpin, developing protocols and standards to support scalable deployment while mitigating risks effectively.</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rWc0pSYe4iroj16JtQl8Ig.png" /></figure><h4>5. Keep in mind that the domain is still very nascent</h4><p>As Sam Altman mentions in his interview with Gates [<a href="https://bit.ly/gates-altman"><strong><em>here</em></strong></a>], the current models are at their stupidest and we are on a very steep improvement curve.</p><p><em>“In some ways, this article is premature — so much is changing that we’ll likely have a profoundly different understanding of Generative AI and its capabilities in a year’s time”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*q7ctRmtJLuAeU83j.png" /></figure><p><strong>Finally, couple of interesting quotes heard at the Authors talk video about the genesis of the book</strong> [<a href="https://www.mckinsey.com/featured-insights/mckinsey-on-books/author-talks-what-is-the-key-to-unlocking-digital-transformation"><strong><em>here</em></strong></a>]</p><ul><li>“<em>They’re (Generative AI) amazing. They pass the Arthur C. Clarke test of technology that’s indistinguishable from magic</em>” — Rodney Zemmel</li><li>“<strong><em>One use case never does the job</em></strong><em>. It is a family of many that needs to be strong together to actually create meaningful impact and change experiences</em>” — Eric Lamarre</li><li>“<em>There is a difference </em><strong><em>doing digital vs. being digital </em></strong><em>— an old IT culture vs. the new digital culture</em>” — Rodney Zemmel</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=349986066c6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Unreasonable Effectiveness of Google Gemini 1.5 — Alea iacta est]]></title>
            <link>https://ksankar.medium.com/the-unreasonable-effectiveness-of-google-gemini-1-5-alea-iacta-est-945aa9b4f4f1?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/945aa9b4f4f1</guid>
            <category><![CDATA[google-gemini-ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[generative-ai]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Mon, 19 Feb 2024 05:40:09 GMT</pubDate>
            <atom:updated>2024-02-20T03:19:41.759Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Unreasonable Effectiveness of Google Gemini 1.5 — Alea iacta est</h3><p>This is going to be a short blog on the <strong><em>disruptively impressive capabilities</em></strong> of the <a href="https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/#gemini-15">Google Gemini 1.5 preview</a>.</p><p>Google has unveiled a private preview edition of the Gemini 1.5 Pro — “… <strong><em>starting today, a limited group of developers and enterprise customers can try it with a context window of up to 1 million tokens …”</em></strong></p><h3>What does it mean ?</h3><p>A huge jump in the token size, meaning it can process much larger amounts of data compared to other LLMs/chatbots as shown in the table below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fMUsxUrvTC-6-cWlhZ_00g.jpeg" /></figure><blockquote>Token size is the currency in NLP — Tokens can be entire parts or subsections of words, images, videos, audio or code. The bigger a model’s context window, the more information it can take in and process in a given prompt — making its output more consistent, relevant and useful.</blockquote><blockquote>With a token size of 32,768, gpt-4–32k can intake around 25,000 words. Gemini 1.5 can ingest 32 times — 1 hour of video, 11 hours of audio, codebases with over 30,000 lines of code or over 700,000 words (1,400 pages!!). In Google’s research, they’ve successfully tested up to 10 million tokens !</blockquote><p>1M + is definitely impressive — granted, it may lead to substantial challenges such as extensive hallucinations and unintended consequences; nevertheless as the saying goes, “<strong><em>Alea iacta est</em></strong>” (the die is cast) — a barrier broken, the bar has been risen !</p><p>Gemini 1.5 not only handles vast volumes of text but processes them with remarkable accuracy !</p><blockquote>“<strong>For text, Gemini 1.5 Pro achieves 100 percent recall up to 530k tokens, 99.7 percent up to 1M tokens, and 99.2 percent accuracy up to 10M tokens</strong>” — Results shared by Rowan Cheung from early access</blockquote><p>While I don’t have access to the preview, <a href="https://twitter.com/rowancheung">Rowan Cheung</a> has and he spent this Saturday working on this and write an excellent blog in twitter [<a href="https://twitter.com/rowancheung/status/1759280384930459941"><strong><em>Here</em></strong></a>] ! Rowan’s firsthand experience, as detailed in his excellent Twitter blog, provides valuable insights into Gemini 1.5’s capabilities.</p><p>Of course, I am on the waiting list !!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I35WFHmKWzhbwyvOoUEJiQ.jpeg" /></figure><h3>The Larger Picture</h3><ul><li>As Rowan writes and replies from others, “<strong><em>… so many LLM capabilities just unlocked …. so many more apps will be built …</em></strong>” — that is the key. Suddenly opens up to a host of previously unknown capabilities.</li><li>“<strong><em>This is huge progress. Guess we will have to wait to see what brings GTP-5 to fight this</em></strong>” — hence my note Alea iacta est. This has raised the bar !</li><li>“<strong><em>… Felt surreal to watch Gemini ingest multiple transcripts and spit out accurate details in less than a minute</em></strong>”</li></ul><h3>The Architecture</h3><ul><li>Am going to slide-in a quick peek into some of the interesting technical details</li><li>Gemini 1.5 is not a single model, but a MoE ie Mixture of Experts, with conditional computation. “<strong><em>While a traditional Transformer functions as one large neural network, MoE models are divided into smaller “expert” neural networks</em></strong>.<strong><em>Depending on the type of input given, MoE models learn to selectively activate only the most relevant expert pathways (using the learned routing function) in its neural network. This specialization massively enhances the model’s efficiency</em></strong>.”</li><li>BTW, the MoE architecture has been the core of the (very successful) Mistral family of models from the French company</li><li>The model is natively multimodal and supports interleaving of data from different modalities, it can support a mix of audio, visual, text, and code inputs in the same input sequence</li><li>Gemini 1.5 Pro is trained on multiple 4096-chip pods of Google’s TPUv4 accelerators, distributed across multiple datacenters, and on a variety of multimodal and multilingual data.</li></ul><h3>Will it replace RAG (Retrieval Augmentation)?</h3><ul><li>An obvious question — <a href="https://www.linkedin.com/in/alexander-ratner-038ba239/overlay/about-this-profile/">Alexander Ratner</a> has an interesting answer</li><li>Long context models — like Gemini 1.5 — will definitely eat up a lot of simpler use cases + pre-production development (that is very common nowadays)</li><li>But, RAG still wins from a cost, latency, and scale perspective. Even more durably: a RAG approach is modular. <em>So for more complex, scaled, and/or production settings- RAG likely here to stay</em></li></ul><h3>Rowan has tried six interesting capabilities</h3><ol><li><strong>Breaking down + understanding a long video</strong> — he uploaded the entire NBA dunk contest from last night and asked which dunk had the highest score. Gemini 1.5 was incredibly able to find the specific perfect 50 dunk and details from just its long context video understanding!</li><li><strong>Understanding and comparing full movie transcripts from Interstellar and Ad Astra </strong>— Gemini 1.5 was able to understand, compare, and contrast entire transcripts from both movies to help me decide which movie to watch.</li><li><strong>Watching, understanding, and distinguishing if an OpenAI Sora video is AI or not</strong> — Gemini 1.5 highlighted the famous Sora cat video and highlighted key factors of why it could be AI-generated. He was surprised by how in-depth it went on the response. — encouraging. <strong><em>&lt;ks&gt; I always knew that we need LLMs to understand /guardrail LLMs. This is very good. I had mentioned this in my blogs [</em></strong><a href="https://bit.ly/agi-turing-test"><strong><em>Here</em></strong></a><strong><em>] &amp; [</em></strong><a href="https://bit.ly/llm-guardrails"><strong><em>Here</em></strong></a><strong><em>]&lt;/ks&gt;</em></strong></li><li><strong>Finding, understanding, and explaining a small figure in a long paper </strong>— Gemini 1.5 was able to extract “table 8” from the Gemini 1.5 Pro paper by DeepMind and explain what the table meant. <strong><em>&lt;ks&gt; Excellent capability for financial applications, operations optimization et al &lt;/ks&gt;</em></strong></li></ol><blockquote>I will leave you with an urge to try out the Gemini 1.5 and all it’s wonders! The more interesting exploration is to comprehend and internalize the speed in which the LLM capabilities are evolving … as Sam Altman mentions in his interview with Bill Gates [<a href="https://bit.ly/gates-altman">Here</a>], <strong>we will be on this steep improvement curve for the next 5+ years …</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r1Tc4dkkjkXn3cxU.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=945aa9b4f4f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hallucination is a feature not a bug !]]></title>
            <link>https://ksankar.medium.com/hallucination-is-a-feature-not-a-bug-910d12a8fbab?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/910d12a8fbab</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[generative-ai-guardrails]]></category>
            <category><![CDATA[hallucinations]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Tue, 06 Feb 2024 05:00:20 GMT</pubDate>
            <atom:updated>2024-03-05T00:17:31.570Z</atom:updated>
            <content:encoded><![CDATA[<h3>Hallucination is a feature not a bug !</h3><p>You may have heard this sentiment expressed about Generative AI and ChatGPT in numerous contexts. While it may sound like a catchphrase, this statement holds significant relevance and carries various implications…</p><h3>Hallucination &amp; Generative AI</h3><ul><li>In the realm of artificial intelligence, the term ‘hallucination’ often conjures images of distorted perceptions or false sensory experiences. However, within the context of Generative AI, hallucination takes on a unique and fascinating dimension.</li><li>It refers to the phenomenon where AI models generate outputs that diverge from the underlying reality captured in the training data.</li><li>While this may initially seem like a flaw in the model’s performance, a closer examination reveals that hallucination is not merely a bug to be fixed but rather a feature inherent to the generative process.</li></ul><blockquote>In this blog, we delve into the intriguing world of generative AI hallucination, exploring ̶i̶t̶s̶ ̶c̶a̶u̶s̶e̶s̶,̶ its implications, and creative potential. F̶r̶o̶m̶ ̶d̶i̶s̶t̶o̶r̶t̶e̶d̶ ̶i̶m̶a̶g̶e̶s̶ ̶t̶o̶ ̶n̶o̶n̶s̶e̶n̶s̶i̶c̶a̶l̶ ̶t̶e̶x̶t̶,̶ ̶w̶e̶ ̶u̶n̶c̶o̶v̶e̶r̶ ̶t̶h̶e̶ ̶c̶o̶m̶p̶l̶e̶x̶i̶t̶i̶e̶s̶ ̶o̶f̶ ̶h̶a̶l̶l̶u̶c̶i̶n̶a̶t̶i̶o̶n̶ ̶a̶n̶d̶ ̶i̶t̶s̶ ̶r̶o̶l̶e̶ ̶i̶n̶ ̶s̶h̶a̶p̶i̶n̶g̶ ̶t̶h̶e̶ ̶f̶u̶t̶u̶r̶e̶ ̶o̶f̶ ̶A̶I̶-̶d̶r̶i̶v̶e̶n̶ ̶c̶r̶e̶a̶t̶i̶v̶i̶t̶y̶ ̶a̶n̶d̶ ̶i̶n̶n̶o̶v̶a̶t̶i̶o̶n̶.̶</blockquote><p>Note : I have a little bit more to say about the above section, at the end !</p><h3>Currently Generative AI is used for opportunistic applications, not the ultimate applications</h3><ul><li>An interesting <a href="https://www.oneusefulthing.org/p/on-holding-back-the-strange-ai-tide">blog</a> aptly named “<em>On holding back the strange AI tide</em>” and <a href="https://www.linkedin.com/posts/emollick_on-holding-back-the-strange-ai-tide-activity-7160411295416995840-imaS">LinkedIn post</a> by <a href="https://substack.com/@oneusefulthing">Prof.Ethan Mollick</a> is very insightful.</li><li>Ethan says and I quote “<em>Companies keep picking the same few safe* projects to focus on for AI: customer service, support copilots, sales, and “talk to my documents.” All of these are existing IT solutions that can be given an AI gloss</em>”</li></ul><h3>And, that is challenging primarily because that is not an optimal sweet spot for Generative AI</h3><ul><li>It works, works better than the tools we have now — but to apply Generative AI for these applications, we have to do good amount of scaffolding ! “.. <em>trying to stop AI from being weird</em>” !</li><li>I have written about the Guardrails and other intricacies of Generative AI [<a href="https://bit.ly/llm-guardrails"><strong><em>here</em></strong></a>] and [<a href="https://bit.ly/ai-red-teaming"><strong><em>here</em></strong></a>]. Initial <em>initial slides [</em><a href="https://github.com/xsankar/AI-Red-Teaming/blob/main/S61261-compressed.pdf"><strong><em>here</em></strong></a><em>] for m</em>y talk at Nvidia GTC <em>[</em><a href="https://www.nvidia.com/gtc/session-catalog/?tab.allsessions=1700692987788001F1cG&amp;search=guardrails#/session/1692674753904001nZeG"><strong><em>here</em></strong></a><em>]</em></li><li>“.<em>. these approaches often highlight the weak points of AI more than technologists expect. No matter how much RAG or fine-tuning you do, your LLM system will hallucinate details. Thus areas where accuracy is expected, like talking to internal documents or interacting with customers, can be more challenging than people think. And external-facing systems are vulnerable to prompt injection</em>” — Prof. Ethan Mollick</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XITvgLUtLtoazMu2ZthXqg.jpeg" /></figure><h3>Generative AI is best used natively (including it’s hallucination) than simple automation</h3><ul><li>“<em>.. even worse, the substantial benefits of AI are going to be greatly reduced by trying to pretend it is just like previous waves of technology..</em>”</li><li>“.. <em>But LLMs are inherently weird and powerful in ways that don’t fit into typical IT solutions, or even our conventional approaches to organizational capabilities and boundaries. </em><strong><em>Make sure you are devoting at least some effort to doing much stranger and more powerful experiments, because AIs are only getting better from here</em></strong>” — Prof. Ethan Mollick</li></ul><h3>These are the stupidest the models will ever be !</h3><ul><li>Said Sam Altman during an interview with Bill Gates [<a href="https://bit.ly/gates-altman"><strong><em>here</em></strong></a>]</li><li>We are on a very steep improvement curve and will stay there for next 5–10 years, driving the cost of intelligence to nearly zero !</li></ul><blockquote>Which means they will get more creative and will hallucinate more — but the good kind of hallucinations !</blockquote><p><strong><em>In short, definitely use generative AI wherever you get performance improvements (with appropriate guardrails), also look for futuristic, non-linear, disruptive opportunities</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_Yr4EvBHvdBrAWVc.png" /></figure><h3>Our work systems are not built for AI, so we will need to rebuild them — as Agents : License to Operate !</h3><ul><li>This is an interesting and frightening reality. We shouldn’t apply Generative AI just to automate isolated components within existing work flow — such as customer service (internal or external) or the automation of operations. <strong><em>We should reengineer our processes using Generative AI ! That is Operations Optimization in this new world !!</em></strong></li><li>Agents and License to Operate ! The other day I saw an interesting company <a href="https://sema4.ai/">sema4.ai</a> — LinkedIn announcement [<a href="https://www.linkedin.com/pulse/sema4ai-raises-305-million-bring-open-source-powered-ai-mission-critical-i8wsc/"><strong><em>here</em></strong></a>] and Mayfield [<a href="https://www.mayfield.com/partnering-with-sema4-to-accelerate-the-agent-economy/"><strong><em>here</em></strong></a>].</li><li>Their goal “.. <em>build category-defining intelligent agents that connect context to action, changing how work gets done in the enterprise</em>”</li></ul><blockquote>I have much more to discuss about agents and the essential skills required for achieving Artificial General Intelligence (AGI) — perhaps a topic for another day. You can find my preliminary thoughts on AGI in the rough-cut blog [<a href="https://bit.ly/agi-turing-test"><strong><em>here</em></strong></a>]</blockquote><h3>Cognitive Plumbing !</h3><blockquote><a href="https://www.mayfield.com/">Mayfield</a> has an interesting diagram (below). Slightly self serving, highlighting their investments, still useful — food for thought !</blockquote><p>[Update] A recent blog by Navin talks about Cognition As A Service (CaaS) [<a href="https://www.mayfield.com/cognition-as-a-service-caas-rise-of-assistants-digital-coworkers/"><strong><em>here</em></strong></a>].</p><ul><li>That makes sense. Of course, it is different from rule based automation. In fact it need not be automation at all — meaning humans can be in the loop as needed by the workflow; even escalation would be part of it.</li><li>I do have couple of disagreements with <a href="https://www.mayfield.com/team/navin-chaddha/">Navin Chaddha</a>’s CaaS blog — the travel example is an agent and doesn’t have cognition per se; same with the fraud alert — it is a rule driven agent; we can send forms and block transactions now. The Digital Co-worker, Digital Twin et al are CaaS. The key is to let humans do what they can do best and machines do their best.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*Z_wCcTDfBGZQ-Ot7" /></figure><h3>Epilogue</h3><p>You might be wondering why I had some sentences in s̶t̶r̶i̶k̶e̶ ̶t̶h̶r̶o̶u̶g̶h̶ it is because that initial section was written by ChatGPT v3.5 (rest is mine) ! <em>The </em>s̶t̶r̶i̶k̶e̶ ̶t̶h̶r̶o̶u̶g̶h̶<em> is my edit - it doesn’t know what I plan to write under the title;</em> <em>the formatting is mine too, of course</em>.</p><blockquote><strong>What a better way to start a blog on Hallucinating Generative AI than by enlisting a Hallucinating Generative AI to compose the introduction !</strong></blockquote><p>And, it did an excellent job — without knowing anything about Hallucination or that it is Hallucinating, it effectively argues for the significance of Hallucination !! Take another look at the opening section “Hallucination &amp; Generative AI”</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=910d12a8fbab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Role of AI Red Teaming in Cybersecurity — a.k.a. What the heck is this AI Red Teaming?]]></title>
            <link>https://ksankar.medium.com/the-role-of-ai-red-teaming-in-cybersecurity-a-k-a-what-the-heck-is-this-ai-red-teaming-62b8bda73300?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/62b8bda73300</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[red-teaming]]></category>
            <category><![CDATA[ai-red-teaming]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Fri, 02 Feb 2024 18:33:49 GMT</pubDate>
            <atom:updated>2024-04-06T04:07:21.072Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Role of AI Red Teaming in Cybersecurity — a.k.a. What the heck is this AI Red Teaming?</h3><p>Ever since NIST coined the term AI Red Teaming, questions have been raised, given that Red Teaming has a certain connotation in traditional cybersecurity.</p><blockquote><strong>In this blog post, let me share the current wisdom, backed by references (at the end) !</strong></blockquote><ul><li>Update: [3.23.24] Very informative blog “<em>What’s the Difference Between Traditional Red-Teaming and AI Red-Teaming?</em>” [<a href="https://www.cranium.ai/traditional_vs_ai_red_teaming/"><strong><em>Here</em></strong></a>] — A must read, if you are here … Of course, you are here 😏 …</li><li>Update : [4.5.24] Found two insightful talks on AI Red Teaming viz. <strong>AI Red Teaming LLM: Past, Present, and Future</strong> [<a href="https://www.youtube.com/watch?v=e4GQNN5EG4E"><strong><em>here</em></strong></a>] and <strong>Red Teaming Large Language Models</strong> [<a href="https://www.youtube.com/watch?v=QMS19i2BUjM"><strong><em>here</em></strong></a>]</li></ul><h4>Sidebar</h4><blockquote>On a different note, please join me at Nvidia GTC in March 2024 for a talk on Guardrails [<a href="https://www.nvidia.com/gtc/session-catalog/?tab.allsessions=1700692987788001F1cG&amp;search=guardrails#/session/1692674753904001nZeG">here</a>]! It’s a set of insights packed into a quick session (initial slides [<a href="https://github.com/xsankar/AI-Red-Teaming/blob/main/S61261-compressed.pdf">here</a>] — with practice I could cover in 16 min) — though truth be told, I might need a solid 90 minutes to truly do justice to this topic. Who knows, maybe writing a book is the next logical step… (Stay tuned! 😉)</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nS70ZHjQ-4mPNphOS2CO_g.png" /></figure><h4>Back to the Main Feature …</h4><p>An excellent example of AI Red Teaming, is the GRT (Generative AI Red Teaming) at Def Con 2023. I’ve shared a detailed review of the process and results in my blog [<a href="https://bit.ly/3P3NYFP"><strong><em>here</em></strong></a>]. Very interesting — the results are illuminating.</p><blockquote><strong>Please note that these results are dated, and significant progress has been made since then. Nonetheless, the Def Con 2023 GRT continue to offer valuable insights into the methodology and the essence of AI Red Teaming.</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q7Y0ZsAVkB7vXTufONO_oQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*USKqtQEo4bwEvqCH8nwgAQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XfHBvw5KFWUb95CDmYpe_Q.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qV6AmE10ByU-WV7-qgUDlg.jpeg" /></figure><h3>Summary</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ao_e5sIdg2TrpHXR7dWbRw.png" /></figure><h3>Pragmatics : Part 1 — How to construct the AI Red Teaming Metrics, Datasets, Prompt Banks &amp; Benchmarks</h3><ul><li>AI Red Teaming Tests are composed of Prompt Banks, developed against a knowledge set/graph, and plausible contextual responses. Literally reminds us of the Voigt-Kampff tests from the Blade Runner [<a href="https://ksankar.medium.com/explainable-chatgpt-sentience-reasoning-voight-kampff-prompts-5879edf04e80"><strong><em>here</em></strong></a>]</li><li>Unlike traditional evaluation methods, responses are not judged based on exact matches but on implicature metrics. Language Models (LLMs) don’t guarantee identical responses every time, or semantically equivalent prompts</li><li>And, it is not just finding holes but also understanding the limitations and suitability</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/930/1*AXQh9im8N-1zbhJxmMp-bA.png" /></figure><ul><li>Tests follow a progressive nature, where a response could lead to another prompt deeper in the knowledge graph on the same topic. The prompt banks are topic-specific, with numerous available metrics and benchmarks (refer to my GitHub repository [Here]).</li><li>Evaluation is context-specific. For instance, the LLM Evaluation Triangle illustrates the relationships between the query, response, and context, along with the corresponding evaluation metrics.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*80y9NeiMaBBIDfo_iAHlFA.png" /></figure><h3>Pragmatics : Part 2 — How to run the AI Red Teaming Tests</h3><ul><li>Interestingly, the process of AI Red Teaming is the same as traditional Red Teaming. The enterprise practices around capturing the results, GRC (Governance, risk and Compliance) processes around incidents and so forth haven’t changed.</li><li>Some caveats — Generative AI systems probably are more dynamic — meaning documents for RAG need to be updated frequently, underlying foundation models would need updates, new benchmarks would be developed et al. All due to the nescent nature of the domain.</li><li>In short, the AI Red Teaming development and execution would be separate from traditional Red Teaming but the evaluation and GRC responses should flow through the traditional Red Teaming channels</li></ul><h3>Pragmatics : Part 3 — AI Red Teaming Metrics, Datasets, Prompt Banks &amp; Benchmarks</h3><p>This is a large topic — more and more benchmarks and datasets are being researched — in fact, just at the NeurIPS 2023 we saw a huge set of very innovative benchmarks and datasets — see my GitHub repository Awesome NeurIPS 2023 [<a href="https://github.com/xsankar/Awesome-NeurIPS-2023?tab=readme-ov-file#benchmarks"><strong><em>here</em></strong></a>]</p><p>I am collecting and updating papers and ideas related to the AI Red Teaming practices, benchmarks et al in my GitHub [<a href="https://github.com/xsankar/AI-Red-Teaming"><strong><em>here</em></strong></a>]. I plan to spend some time on this in the next few weeks — please bookmark the GitHub and visit in a few weeks.</p><blockquote>AI Red Teaming is not easy and it will stay this way for a long time — Sam Altman’s interview with Bill Gates is very illuminating on the trajectory of the new world of Generative AI. My blog [<a href="https://bit.ly/gates-altman"><strong><em>here</em></strong></a>]</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FudRlpjj1ZsLq3ON.png" /></figure><h3>TTD (Things To Do)</h3><ul><li>Summarize AI Red Teaming work that NIST has started</li></ul><blockquote>I have created the GitHub repository <a href="https://github.com/xsankar/Awesome-NIST"><strong><em>Awesome-NIST</em></strong></a> to collect all NIST relevant papers, best practices et al from NIST and others. But more focus on Generative AI stuff. Inspired by the awesome-* trend in the GitHub</blockquote><ul><li>The <strong><em>NIST AISIC</em></strong> [<a href="https://www.nist.gov/artificial-intelligence/artificial-intelligence-safety-institute"><strong><em>here</em></strong></a>] is about to start and it has AI Red Teaming work streams</li><li>I am yet to summarize the readings from the <strong><em>NIST Secure Software Development Framework for Generative AI and for Dual Use Foundation Models Virtual Workshop</em></strong> [<a href="https://www.nist.gov/news-events/events/nist-secure-software-development-framework-generative-ai-and-dual-use-foundation"><strong><em>here</em></strong></a>]</li></ul><blockquote>Most probably that will be my next blog</blockquote><ul><li>Haven’t seen the results of the <strong><em>NIST AI Public Working Groups</em></strong> [<a href="https://airc.nist.gov/generative_ai_wg"><strong><em>here</em></strong></a>]. The Pre-deployment testing work stream is all about AI Red Teaming</li></ul><blockquote>In short, lot to read and write about, in this space !!</blockquote><h3>References</h3><ol><li><a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2023.pdf">https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2023.pdf</a></li><li>Github <a href="https://github.com/xsankar/Awesome-Awesome-LLM">https://github.com/xsankar/Awesome-Awesome-LLM</a></li><li><a href="https://www.applause.com/blog/red-teaming-for-testing-generative-ai">https://www.applause.com/blog/red-teaming-for-testing-generative-ai</a></li><li><a href="https://truera.com/ai-quality-education/generative-ai-rags/how-to-prevent-llms-from-hallucinating/">https://truera.com/ai-quality-education/generative-ai-rags/how-to-prevent-llms-from-hallucinating/</a></li><li><a href="https://truera.com/ai-quality-education/generative-ai-rags/what-are-the-different-ways-in-which-llms-hallucinate/">https://truera.com/ai-quality-education/generative-ai-rags/what-are-the-different-ways-in-which-llms-hallucinate</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62b8bda73300" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gates Interviews Sam Altman !!]]></title>
            <link>https://ksankar.medium.com/gates-interviews-sam-altman-c4b584578765?source=rss-578c7297699c------2</link>
            <guid isPermaLink="false">https://medium.com/p/c4b584578765</guid>
            <category><![CDATA[gpt]]></category>
            <category><![CDATA[sam-altman]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[bill-gates]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Krishna Sankar]]></dc:creator>
            <pubDate>Tue, 16 Jan 2024 01:49:58 GMT</pubDate>
            <atom:updated>2024-01-17T14:29:06.006Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Gates Interviews Sam Altman !!</strong></h3><p>The title tells it all ! An interesting interview with Altman by Gates. <em>Not sure who is more famous — should I say </em><strong><em>Bill</em></strong><em> Gates interviews Altman or Gates Interviews </em><strong><em>Sam</em></strong><em> Altman ;o)</em></p><p>Anyway the podcast is [<a href="https://www.gatesnotes.com/Podcast?podcast=s2e0"><em>here</em></a>] or [<a href="https://www.youtube.com/watch?v=PkXELH6Y2lM"><em>here</em></a>]. Trust me, it’s a must-listen! — But, before you hit play, catch a sneak peek at my blog for a quick summary …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-_2vKWE23mwlrpwyotX2Rg.png" /></figure><blockquote>There is a summary by Anthropic’s Claud 2 [<a href="https://twitter.com/IntuitMachine/status/1746278269165404164">Here</a>]. Please take a look after reading my blog and share your thoughts on whether I did a better job!</blockquote><p><strong>1st Impressions</strong></p><blockquote>Just beautiful — it is very important for all of us to tune in to this podcast and think about it. Lots of very important concepts and future trajectories !</blockquote><ul><li>The podcast was taped (probably around Nov 10th ’23 or so) a little before the whole OpenAI drama — where Sam Altman was fired/rehired and then the board was shuffled ! My chronicles of the episode — I see AGI everywhere !! [<a href="https://ksankar.medium.com/youre-fired-agi-to-sam-altman-c0b2d1e7429c?source=user_profile---------8----------------------------"><strong><em>Here</em></strong></a>]</li><li>Bill called Altman and Sam says “<em>the team has never felt more productive or more optimistic or bette</em>r”, it is “<em>a real moment of growing up</em>” for them and “<em>are ready for the challenges</em>” in front of them !</li><li>This interview has it’s place among the insightful discussions along with a) Fireside Chat w/ Ilya Sutskever (Chief Scientist Open AI) by Jensen Huang (CEO, Nvidia) (<a href="https://www.youtube.com/watch?v=-yquJiNKlAE"><strong><em>Here</em></strong></a>) ; b) An excellent interview of Ilya Sutskever (Chief Scientist Open AI) by Dwarkesh Patel (<a href="https://www.youtube.com/watch?v=Yf1o0TQzry8"><strong><em>Here</em></strong></a>). I write about AGI touching upon these blogs <strong>[</strong><a href="https://bit.ly/agi-turing-test"><strong><em>Here</em></strong></a>]</li><li><strong>Trivia :</strong> Gates asks his guests to bring their favorite music. Sam brought <em>The New Four Seasons — Vivaldi Recomposed by Max Richter, performed by Chineke! Orchestra, played on period instruments. </em>“<strong><em>Beautiful, Soaring and Optimistic</em></strong>” Naturally, I found the music at Amazon and promptly ordered <a href="https://www.amazon.com/gp/product/B09SVYS8WF"><em>this</em></a> and <a href="https://www.amazon.com/gp/product/B00IERJQXS/"><em>this</em></a> + the original Nigel Kennedy 25th Anniversary <a href="https://www.amazon.com/Vivaldi-Four-Seasons-25th-Anniversary/dp/B00LJWUP9E"><em>here</em></a> !</li></ul><blockquote><strong>Commentary : I bet you wouldn’t see the above in a GPT-created summary; Anthropic’s Claud definitely is not going to order Vivaldi from Amazon — at least not yet ! And when it starts ordering on it’s own, be afraid — very afraid !</strong></blockquote><p><strong>2 In a Blink: Executive Snapshot for the ADHD Minds (<em>In pictures !</em>)</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y9ZkR8zgcsVOaMTsZnRYrA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3zKPHdc8pSblXuMomyUVSw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WVg8tcfCssLwaxIE4RXA_A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AtOUyoyrjfNNfu7xrbdC6g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_T0Mi2EVu4sgqXNTU3bIwQ.png" /></figure><blockquote><strong>AGI is OpenAI’s target and they are committed. I had written about navigating the inevitable encounter with AGI and what makes AGI A<em>G</em>I [</strong><a href="https://bit.ly/agi-turing-test"><strong>here</strong></a>]</blockquote><h4>On Societal Impact</h4><ul><li>The best problems we should work on is applying AI to help humans to get along with each other ! <strong><em>Bill is troubled a lot by polarization and war — rightfully so.</em></strong></li><li>Sam believes that the technology will surprise us on the upside —on how much positive it can do. We just have to see — <strong><em>The only way out is through !!</em></strong></li></ul><blockquote>Trivia : Most frequently used mobile app — <strong>Altman : Slack; Gates — Outlook </strong>(“<em>old style e-mail guy</em>”)</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GEh-vnYMfscj9-vNf8MwHQ.jpeg" /></figure><p>This interview throws some light on what Sam was doing on his world tour with the leaders of various countries ! He was socializing an IAEA for huge LLMs !! Testing methodology similar to how we handle nuclear bombs !!</p><ul><li>If we are right, modern AI technologies will impact society, geopolitical balance of power and so many other things</li><li>Altman was socializing with the world leaders, a global regulatory body (because they have global impact) that looks at these super powerful systems (not like GPT-4, but 100,000 or million times the compute power of that) — a global body something like the IAEA !</li><li>From a global cooperation perspective, asking for a slowdown would be hard. “<em>Do whatever you want, but any computer cluster above a certain extremely high-power threshold — may be 5 in the world, has to submit to the equivalent of international weapons inspectors. </em><strong><em>The model has to be made available for safety audit, pass some tests during training and before deployment</em></strong>”</li></ul><blockquote><strong>This can help us with the biggest tier of risks — “There are still going to be things that are going to go wrong with much smaller-scale systems, in some cases, probably pretty badly wrong”</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bo0ch-LTO-ff-AEiJDo43A.jpeg" /></figure><ul><li>“<em>Both annoying, motivating and fun</em>”. Pushes them to be better and do things faster</li><li>OpenAI is very confident in their approach … to AGI !</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5MXPggI3OfVkrxfwlILZCA.jpeg" /></figure><ul><li>They started robots too early, so had to put the project on hold.</li><li>It wasn’t helping them make progress with the difficult parts of the ML research and … also realized that they need intelligence and cognition 1st and then adapt it to physicality</li><li>They have started investing in robotics and “<em>really exciting new (robotics) platforms are being built</em>”</li><li>The consensus prediction (10 years ago) was that the impact is going to be blue-collar 1st, white-collar work 2nd, creativity — may be never or certainly last as that was magic &amp; human</li><li>But now it has gone exactly the other direction !</li></ul><blockquote>One reason is the creativity, <strong>the hallucination (a feature not a bug) </strong>enabled the machine to discover new things !</blockquote><ul><li>Certainly a case of, you got to follow where technology goes — we have preconceptions, but sometimes science doesn’t want to go that way</li></ul><p>Couple of backward pointers to my referred blogs (for those who came in late) and other interviews:</p><ul><li>Navigating the inevitable encounter with AGI [<a href="https://bit.ly/agi-turing-test"><strong><em>Here</em></strong></a>]</li><li>Fireside Chat w/ Ilya Sutskever (Chief Scientist Open AI) by Jensen Huang (CEO, Nvidia) (<a href="https://www.youtube.com/watch?v=-yquJiNKlAE"><strong><em>Here</em></strong></a>)</li><li>An excellent interview of Ilya Sutskever (Chief Scientist Open AI) by Dwarkesh Patel (<a href="https://www.youtube.com/watch?v=Yf1o0TQzry8"><strong><em>Here</em></strong></a>)</li></ul><blockquote><strong>And, my GitHub repositories collection of Papers, Benchmarks and Frameworks on LLM Evaluation, Red Teaming, NeurIPS2023 et al</strong> [<a href="https://github.com/xsankar/Awesome-Awesome-LLM"><strong>Here</strong></a>]</blockquote><p><strong>Rumors</strong></p><p>Of course, rumors are rumors and can’t be fully correct, otherwise they wouldn’t be rumors !! Source : <a href="https://www.youtube.com/watch?v=JuYLCfb7CK8"><em>The AIGrid</em></a> and <a href="https://www.standard.co.uk/news/tech/chatgpt-5-release-date-details-openai-chatbot-b1130369.html"><em>The Standard</em></a></p><ul><li>OpenAI has developed a powerful model names “Arrakis” a.k.a GPT-5 (code named after one my favorite books The Dune, Of course)</li><li>Multimodal, has near human capabilities + priorities from above (Roadmap)</li><li>Hallucinates much less than GPT-4</li></ul><blockquote>Not sure if the last is progress, as I really think <strong>Hallucination is a feature, not a bug and the best value we get from GPT models is leveraging their hallucinations not muzzling them — of course, based on the context<em>.</em></strong></blockquote><ul><li>Trained with more than 50% synthetic data (This is interesting)</li><li>Has agent capabilities (Good)</li><li>Scheduled for 2024 (duh !)</li></ul><h4>Finally</h4><p>I leave you with a presentation from the past (1979, IBM) …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/812/1*X74-U22zNTLbY__x8KJDzw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c4b584578765" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>