<?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 Zaza Zakaria on Medium]]></title>
        <description><![CDATA[Stories by Zaza Zakaria on Medium]]></description>
        <link>https://medium.com/@zazazakaria18?source=rss-ef40c0d2a1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*IwMsTtd9QdvwOQLDHKbsqg.jpeg</url>
            <title>Stories by Zaza Zakaria on Medium</title>
            <link>https://medium.com/@zazazakaria18?source=rss-ef40c0d2a1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:49:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zazazakaria18/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[ Stop Building Python Projects Like It’s 2009! Modernize Your Workflow in 2024 ]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://blog.devgenius.io/stop-building-python-projects-like-its-2009-modernize-your-workflow-in-2024-2c5905101c7d?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/795/1*903aKlVJKwLZV2D-j20BwQ.png" width="795"></a></p><p class="medium-feed-snippet">Modern practices like type hints, Pydantic, and Ruff will make your code cleaner, faster, and easier to debug. &#x1F40D;&#x1F4A1;</p><p class="medium-feed-link"><a href="https://blog.devgenius.io/stop-building-python-projects-like-its-2009-modernize-your-workflow-in-2024-2c5905101c7d?source=rss-ef40c0d2a1------2">Continue reading on Dev Genius »</a></p></div>]]></description>
            <link>https://blog.devgenius.io/stop-building-python-projects-like-its-2009-modernize-your-workflow-in-2024-2c5905101c7d?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c5905101c7d</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[clean-code]]></category>
            <category><![CDATA[software]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Tue, 11 Mar 2025 09:10:44 GMT</pubDate>
            <atom:updated>2025-03-11T18:51:11.554Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[February 2025: The Month AI Took Over Your Life (And You Didn’t Even Notice)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/aimonks/february-2025-the-month-ai-took-over-your-life-and-you-didnt-even-notice-cd94d18a9cc3?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/994/1*FgkZeB6PesyWZJbgDiHSXQ.png" width="994"></a></p><p class="medium-feed-snippet">Feb 2025: AI agents invade daily life, billions pour into data hubs, and copyright chaos looms. The future&#x2019;s here&#x200A;&#x2014;&#x200A;did you notice?</p><p class="medium-feed-link"><a href="https://medium.com/aimonks/february-2025-the-month-ai-took-over-your-life-and-you-didnt-even-notice-cd94d18a9cc3?source=rss-ef40c0d2a1------2">Continue reading on  . »</a></p></div>]]></description>
            <link>https://medium.com/aimonks/february-2025-the-month-ai-took-over-your-life-and-you-didnt-even-notice-cd94d18a9cc3?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd94d18a9cc3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Fri, 07 Mar 2025 08:29:43 GMT</pubDate>
            <atom:updated>2025-03-07T10:50:27.117Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Docker in 2025: 10x Things You MUST Know — — NOW !]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://blog.devgenius.io/docker-in-2025-10x-things-you-must-know-now-70936844a544?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/1428/1*h9WLfBqYAgilSNrXjb6yzQ.png" width="1428"></a></p><p class="medium-feed-snippet">Docker has become the cornerstone of modern application deployment, yet many developers still wrestle with manual configurations and&#x2026;</p><p class="medium-feed-link"><a href="https://blog.devgenius.io/docker-in-2025-10x-things-you-must-know-now-70936844a544?source=rss-ef40c0d2a1------2">Continue reading on Dev Genius »</a></p></div>]]></description>
            <link>https://blog.devgenius.io/docker-in-2025-10x-things-you-must-know-now-70936844a544?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/70936844a544</guid>
            <category><![CDATA[docker]]></category>
            <category><![CDATA[containers]]></category>
            <category><![CDATA[dockerfiles]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[developer]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Fri, 28 Feb 2025 13:41:44 GMT</pubDate>
            <atom:updated>2025-03-01T15:55:07.792Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[1 Million Tokens, Ultra-Low Cost: Why You Should Love Gemini Flash ]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/aimonks/gemini-apis-is-all-you-need-cheap-powerfull-and-1m-token-lenght-da095adc2f73?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/1430/1*b4efJxVoLbG5svupCp7jPQ.png" width="1430"></a></p><p class="medium-feed-snippet">In the fast-evolving AI landscape, developers need high-performance tools without breaking the bank.</p><p class="medium-feed-link"><a href="https://medium.com/aimonks/gemini-apis-is-all-you-need-cheap-powerfull-and-1m-token-lenght-da095adc2f73?source=rss-ef40c0d2a1------2">Continue reading on  . »</a></p></div>]]></description>
            <link>https://medium.com/aimonks/gemini-apis-is-all-you-need-cheap-powerfull-and-1m-token-lenght-da095adc2f73?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/da095adc2f73</guid>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Mon, 24 Feb 2025 08:47:50 GMT</pubDate>
            <atom:updated>2025-02-24T12:11:38.900Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Power of Generative AI in Software Engineering Education]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/aimonks/the-power-of-generative-ai-in-software-engineering-education-3ee9cbffaf7b?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/813/1*bITzgIHDwGFRX3icwhhNTA.png" width="813"></a></p><p class="medium-feed-link"><a href="https://medium.com/aimonks/the-power-of-generative-ai-in-software-engineering-education-3ee9cbffaf7b?source=rss-ef40c0d2a1------2">Continue reading on  . »</a></p></div>]]></description>
            <link>https://medium.com/aimonks/the-power-of-generative-ai-in-software-engineering-education-3ee9cbffaf7b?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ee9cbffaf7b</guid>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Wed, 12 Feb 2025 10:28:22 GMT</pubDate>
            <atom:updated>2025-02-16T23:07:12.753Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Hosting DeepSeek R1 on AWS: The Game-Changing Cost & Scale ]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/aimonks/hosting-deepseek-r1-on-aws-the-game-changing-cost-scale-38db12f21f2a?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/2048/1*Y3yPCAdgLUXvYGf6DNDRXw.jpeg" width="2048"></a></p><p class="medium-feed-snippet">DeepSeek R1: Open-source LLM hosted on a single AWS instance, serving 2240+ users at a fraction of the cost. better than chatgpt &amp; cheaper</p><p class="medium-feed-link"><a href="https://medium.com/aimonks/hosting-deepseek-r1-on-aws-the-game-changing-cost-scale-38db12f21f2a?source=rss-ef40c0d2a1------2">Continue reading on  . »</a></p></div>]]></description>
            <link>https://medium.com/aimonks/hosting-deepseek-r1-on-aws-the-game-changing-cost-scale-38db12f21f2a?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/38db12f21f2a</guid>
            <category><![CDATA[deepseek]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Fri, 07 Feb 2025 08:51:19 GMT</pubDate>
            <atom:updated>2025-02-16T23:05:34.485Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[January 2025: The Month AI Rewrote the Future]]></title>
            <link>https://medium.com/aimonks/january-2025-the-month-ai-rewrote-the-future-bdcc3458d394?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/bdcc3458d394</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deepseek]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Tue, 04 Feb 2025 17:10:01 GMT</pubDate>
            <atom:updated>2025-02-06T16:30:49.053Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*hWgQgImkqPXX1_uqK_hxJg.png" /></figure><p>If the launch of ChatGPT in 2022 marked the dawn of generative AI, <strong>January 2025 will be remembered as the moment the world fully embraced its seismic potential</strong>. This wasn’t just another month of incremental progress — it was a whirlwind of breakthroughs, billion-dollar bets, and paradigm shifts that redefined what AI can achieve. Let’s unpack why this 31-day span may go down as the most pivotal chapter in AI history.</p><h4>checkout the article for free <a href="https://www.linkedin.com/feed/update/urn:li:activity:7292582429993697280">here</a></h4><h3>1. The Funding Frenzy: Capital Floods the AI Ecosystem</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*gFkiYQlRAxlWg8qd.jpg" /><figcaption><a href="https://imgflip.com/memegenerator/Shut-Up-And-Take-My-Money-Fry">Shut Up And Take My Money</a></figcaption></figure><p>The month began with a financial earthquake. <strong>OpenAI</strong>, already a titan, secured a staggering <strong>$40 billion funding round</strong>, catapulting its valuation to <strong>$340 billion</strong> — a figure rivaling the GDP of small nations. But the real story wasn’t just the number; it was the intent. Insiders revealed that this capital will fuel OpenAI’s pivot toward <strong>AI-as-a-service infrastructure</strong>, positioning it as the AWS of artificial intelligence.</p><p>Meanwhile, <strong>Meta</strong> dropped jaws with a <strong>$65 billion commitment</strong> to AI hardware. Their plan? Deploy <strong>1.3 million next-gen GPUs</strong> (including Blackwell architecture chips) and build a sprawling data center in Louisiana to power <strong>Llama 4</strong>, their flagship multimodal model. Analysts noted this move isn’t just about competing with OpenAI — it’s about controlling the physical backbone of AI computation.</p><p>Then came <strong>Reliance Industries</strong>, India’s conglomerate giant, announcing a <strong>$20–30 billion investment</strong> to construct the world’s largest AI data center. With a <strong>3-gigawatt capacity</strong> (enough to power 3 million homes) and 100% renewable energy, this project underscores a critical trend: <strong>AI’s future is being built in the Global South</strong>, far beyond Silicon Valley.</p><h3>2. The Open-Source Revolution: Democratizing Superintelligence</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/545/0*DWmopWpE2joAHghI" /><figcaption><a href="https://www.ndtv.com/offbeat/deepseeks-explosive-success-triggers-hilarious-meme-fest-check-the-best-ones-7576719">DeepSeek’s Explosive Success Triggers Hilarious Meme Fest.</a></figcaption></figure><p>January shattered the myth that only tech giants can play in the AI big leagues. China’s <strong>DeepSeek-R1</strong> stole headlines by releasing a <strong>free, open-source model</strong> that operates at <strong>95% lower cost</strong> than rivals. Its secret? A novel <strong>sparse MoE architecture</strong> that dynamically allocates computational resources, allowing it to outperform GPT-4o on reasoning tasks while costing pennies per query. Within 72 hours, it became the <strong>#1 free app on iOS</strong> globally — a wake-up call for proprietary model vendors.</p><p>But DeepSeek wasn’t alone. <strong>Mistral AI</strong>, Europe’s open-source darling, launched <strong>Small 3 24B</strong>, a lean model that punches far above its weight. Benchmarks showed it matching Meta’s <strong>Llama 3.3 70B</strong> on coding tasks while using 40% less energy — a sustainability win. Then came <strong>Ai2’s Tulu 3 405B</strong>, which achieved state-of-the-art results on the <strong>Humanity’s Last Exam</strong> benchmark, solving complex physics and philosophy problems that stumped earlier models.</p><p>Even Alibaba threw its hat in the ring with <strong>Qwen 2.5-Max</strong>, optimized for Asian languages and business workflows. The message was clear: <strong>The age of centralized AI monopolies is over</strong>.</p><h3>3. AI Agents: From Labs to Living Rooms</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/0*GhzP8YlyIBMGUafI" /><figcaption><a href="https://www.youtube.com/watch?v=H2YqOBrpKQU">OpenAI’s New OPERATOR Agent Surprises Everyone!</a></figcaption></figure><p>While models grabbed headlines, <strong>AI agents quietly became the month’s unsung heroes</strong>. OpenAI’s <strong>Operator</strong>, now in research preview, demonstrated startling autonomy. In one demo, it booked a multi-city flight, negotiated a hotel discount via email, and ordered groceries — all within a single chat session. The key innovation? A <strong>“chain-of-thought” memory system</strong> that lets the agent backtrack and self-correct errors mid-task.</p><p>Enterprises got their own tools. <strong>Microsoft AutoGen v0.4</strong> introduced <strong>event-driven agent teams</strong>, where specialized AIs (e.g., a legal bot, a finance bot) collaborate asynchronously on projects. Early adopters like Accenture reported slashing contract review times by 70%. Not to be outdone, <strong>Amazon Bedrock’s Multi-Agent Collaboration (MAC)</strong> framework enabled agents to handle everything from travel planning (Delta Airlines) to drug discovery (Moderna).</p><p>The sleeper hit? <strong>Snowflake’s SwiftKV</strong>, a storage optimization technique that reduced LLM inference costs by <strong>75%</strong> — critical for scaling real-time AI analytics.</p><h3>4. Governments and Giants: The New AI Power Brokers</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*emHPH7wXTTtGhoIk.jpg" /><figcaption><a href="https://www.google.com/url?sa=i&amp;url=https%3A%2F%2Fwww.bbc.com%2Fnews%2Farticles%2Fcy4m84d2xz2o&amp;psig=AOvVaw3paaGlRL1HdD1A6HtYGG8C&amp;ust=1738775221197000&amp;source=images&amp;cd=vfe&amp;opi=89978449&amp;ved=0CBcQjhxqFwoTCNi8wczAqosDFQAAAAAdAAAAABAE">Stargate: Tech giants announce AI plan worth up to $500bn</a></figcaption></figure><p>Politics and AI collided spectacularly. On January 21, U.S. President Donald Trump unveiled <strong>Project Stargate</strong>, a <strong>$500 billion public-private partnership</strong> to build “AI infrastructure for the 21st century.” Backed by OpenAI, SoftBank, Oracle, and UAE’s MGX, the initiative aims to create a global network of data centers, quantum computing hubs, and AI talent pipelines. Critics called it a geopolitical power play; supporters hailed it as the modern-day Apollo Program.</p><p>Meanwhile, <strong>ChatGPT Gov</strong> launched as a secure, FedRAMP-certified tool for U.S. agencies. Early use cases included automating veterans’ benefit claims and optimizing emergency response logistics during winter storms. Across the Atlantic, the EU fast-tracked its <strong>AI Infrastructure Act</strong>, mandating that 50% of AI compute power reside within member states by 2030.</p><h3>5. Market Tremors: Open-Source vs. Proprietary</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*7mvPCXnouNDmrhE8.png" /><figcaption><a href="https://www.indmoney.com/blog/us-stocks/what-is-deepseek-how-it-hit-nvidia-broadcom-tech-stocks">DeepSeek hits Nvidia, Broadcom and Global Tech Stocks | INDmoney</a></figcaption></figure><p>The financial fallout was immediate. <strong>NVIDIA’s stock plummeted 17%</strong> as investors questioned the sustainability of its hardware-centric model amid the open-source boom. “Why pay $30,000 for an H100 GPU cluster when DeepSeek-R1 runs on commodity hardware?” argued Morgan Stanley in a viral research note.</p><p>Yet the bigger shift was cultural. The release of <strong>Humanity’s Last Exam</strong> — a 3,000-question benchmark spanning neurobiology to constitutional law — symbolized a growing focus on <strong>AI’s qualitative leap over human cognition</strong>. When Tulu 3 405B scored 89% versus the human average of 72%, it sparked urgent debates: <em>Are we benchmarking AI… or is AI benchmarking us?</em></p><h3>The Road Ahead: What January 2025 Teaches Us</h3><p>This month wasn’t just a burst of innovation — it was a <strong>blueprint for AI’s next decade</strong>:</p><ol><li><strong>Open-source will dominate</strong>, but with a twist: hybrid models where free base models are monetized via premium APIs.</li><li><strong>AI agents will become ambient</strong>, embedded in everything from Excel macros to smart refrigerators.</li><li><strong>Nation-states are the new VCs</strong>, with AI infrastructure now a pillar of geopolitical strategy.</li></ol><p>But challenges loom. Can we sustainably power 3-gigawatt data centers? Will agentic AI exacerbate job displacement? And who governs projects like Stargate?</p><p><strong>Join the Conversation</strong></p><p>we’re dissecting these trends daily. Dive deeper with us:</p><ul><li><strong>Deep Dive Series</strong>: How sparse MoE architectures work (and why they’re game-changers).</li><li><strong>Interviews</strong>: Stargate Project architects reveal their 2030 roadmap.</li><li><strong>Ethics Roundtables</strong>: Should open-source models have safety “kill switches”?</li></ul><p><em>Comment below with topics you want explored — or share your take on January’s AI big bang. The future is a team sport. 🚀</em></p><blockquote>if you like the article, drop 50 claps for it <em>🚀🚀🚀</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bdcc3458d394" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimonks/january-2025-the-month-ai-rewrote-the-future-bdcc3458d394">January 2025: The Month AI Rewrote the Future</a> was originally published in <a href="https://medium.com/aimonks">𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Minimalist BI Infrastructure with PostgreSQL, FDW, and Superset]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://blog.devgenius.io/building-a-minimalist-bi-infrastructure-with-postgresql-fdw-and-superset-c92e5241f767?source=rss-ef40c0d2a1------2"><img src="https://cdn-images-1.medium.com/max/810/1*oAVxlfOSadGTzfEe9gos3Q.png" width="810"></a></p><p class="medium-feed-snippet">Overview</p><p class="medium-feed-link"><a href="https://blog.devgenius.io/building-a-minimalist-bi-infrastructure-with-postgresql-fdw-and-superset-c92e5241f767?source=rss-ef40c0d2a1------2">Continue reading on Dev Genius »</a></p></div>]]></description>
            <link>https://blog.devgenius.io/building-a-minimalist-bi-infrastructure-with-postgresql-fdw-and-superset-c92e5241f767?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/c92e5241f767</guid>
            <category><![CDATA[data-warehouse]]></category>
            <category><![CDATA[superset]]></category>
            <category><![CDATA[postgresql]]></category>
            <category><![CDATA[finops]]></category>
            <category><![CDATA[business-intelligence]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Tue, 21 Jan 2025 16:00:30 GMT</pubDate>
            <atom:updated>2025-02-14T15:44:05.377Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
        </item>
        <item>
            <title><![CDATA[Deep Dive into Large Language Model Parameters: A Comprehensive Guide]]></title>
            <link>https://medium.com/@zazazakaria18/deep-dive-into-large-language-model-parameters-a-comprehensive-guide-c11f866b33b4?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/c11f866b33b4</guid>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Thu, 27 Jun 2024 17:31:37 GMT</pubDate>
            <atom:updated>2024-06-27T17:31:37.416Z</atom:updated>
            <content:encoded><![CDATA[<p>If you’ve been working with Large Language Models (LLMs) like GPT-4 recently, you’re likely familiar with the array of parameters available at your fingertips. Most often, we find ourselves adjusting the temperature setting, yet the purpose and impact of the other parameters can remain somewhat nebulous. This lack of clarity can lead to uncertainty when fine-tuning these models to achieve the desired output.</p><p>Today, we’re going to embark on an exploratory journey into six key parameters of LLMs. This deep dive aims not just to elucidate what each parameter does, but also to reveal how they profoundly influence the behavior and output of the models. By demystifying these parameters, we hope to empower you with the knowledge and confidence needed to fine-tune your LLMs with peace of mind, ensuring that your next interaction with these AI marvels is both fruitful and enlightening. Let’s dive in!</p><h3>Temperature Parameter: Balancing Randomness and Coherence</h3><h4>TL;DR Summary:</h4><p>The Temperature parameter in LLMs controls the level of randomness in the model’s output. A lower temperature yields more predictable and coherent responses, while a higher temperature encourages creativity and diverse responses.</p><h4>Theoretical Explanation:</h4><p>LLMs operate by assigning probabilities to each potential next token in their vocabulary. For instance, given a sentence like “The cat in the…”, the model might assign probabilities such as “hat: 80%, house: 5%, basket: 4%, best: 4%…”. The sum of these probabilities equals 100%, ensuring a complete distribution across the model’s extensive vocabulary.</p><p>Temperature modifies these probabilities before the model selects the next token. Here’s how:</p><ul><li><strong>Temperature = 1</strong>: This default setting maintains the original probability distribution. The randomness inherent in the model’s training is preserved, providing outputs that are varied yet aligned with the training data.</li><li><strong>Temperature &lt; 1:</strong> This setting skews the distribution towards more likely tokens. For instance, “hat” might jump to a 95% likelihood, reducing the model’s unpredictability and increasing its adherence to likely or common responses.</li><li><strong>Temperature = 0:</strong> This extreme creates deterministic output. The most probable token (“hat” at 100%) is always chosen, eliminating randomness entirely.</li><li><strong>Temperature &gt; 1:</strong> This increases randomness, boosting the probabilities of less likely tokens. It expands the model’s creative reach, allowing for more unusual or novel outputs.</li></ul><h4><strong>Examples for Each Temperature Setting</strong></h4><h4><strong>Low Temperature (e.g., 0.5):</strong></h4><p><strong>Context:</strong> Generating a formal email response.</p><p><strong>Input:</strong> “Thank you for your inquiry about our services.”</p><p><strong>Output with Low Temperature:</strong> “We appreciate your interest and are happy to provide more detailed information. Please let us know your specific requirements.”</p><p><strong>Analysis:</strong> The response is coherent, professional, and closely aligned with the input context.</p><h4>Default Temperature (1.0):</h4><ul><li><strong>Context:</strong> Writing a travel blog entry.</li><li><strong>Input: </strong>“My journey to Spain was filled with unexpected adventures.”</li><li><strong>Output with Default Temperature: </strong>“In Barcelona, I stumbled upon a hidden tapas bar, and in Granada, I joined a spontaneous flamenco dance gathering.”</li><li><strong>Analysis: </strong>The output is engaging, varied, yet plausible, reflecting a balance of creativity and coherence.</li></ul><h4>High Temperature (e.g., 1.5):</h4><ul><li><strong>Context: </strong>Brainstorming for a science fiction story.</li><li><strong>Input: </strong>“In a world where time travel is possible…”</li><li><strong>Output with High Temperature: </strong>“I found myself attending Shakespeare’s first play, then dining at a futuristic underwater city, and later joining a rebellion on Mars.”</li><li><strong>Analysis:</strong> The response is imaginative and diverse, exploring far-fetched yet intriguing scenarios.</li></ul><h4>Extreme Temperature (0 or close to 0):</h4><ul><li><strong>Context:</strong> Generating code based on specific instructions.</li><li><strong>Input: </strong>“Create a Python function to calculate Fibonacci numbers.”</li><li><strong>Output with Extreme Low Temperature:</strong></li></ul><p>“def fibonacci(n): a, b = 0, 1 for _ in range(n): a, b = b, a + b return a.”</p><ul><li><strong>Analysis:</strong> The output is precise, deterministic, and adheres strictly to the input command, showcasing the utility of a low-temperature setting for technical tasks.</li></ul><h4>Extreme Temperature (e.g., 2 or above):</h4><ul><li><strong>Context:</strong> Creating abstract poetry.</li><li><strong>Input: </strong>“Write a poem about the essence of dreams.”</li><li><strong>Output with Extreme High Temperature: </strong>“In slumber’s weave, time dances backward, stars whisper secrets to the moon, and the sun sings lullabies to galaxies unborn.”</li><li><strong>Analysis:</strong> The output is highly imaginative and abstract, showcasing the model’s capacity for creative and unconventional language use.</li></ul><p>In conclusion, temperature is a versatile tool in an engineer’s toolkit when working with LLMs. By adjusting this parameter, one can steer the AI’s output from highly structured and predictable to wildly imaginative and unexpected, depending on the task at hand.</p><h3>Frequency Penalty Parameter: repeat or no-repeat</h3><h4>TL;DR Summary</h4><p>The Frequency Penalty parameter plays a pivotal role in controlling the repetition of words and phrases in generated text. By adjusting this parameter, users can influence the model to either repeat or avoid certain words, enhancing the uniqueness or emphasis in the text.</p><h4>Theoretical Explanation</h4><p>The Frequency Penalty works by adjusting the likelihood of a word’s reappearance based on its previous occurrences in the text. This parameter operates on a simple principle: the more frequently a word has appeared, the greater the penalty applied to its future selection.</p><ul><li><strong>Mechanism:</strong> When the model generates text, each token (word or piece of a word) has a certain probability of being chosen next. The Frequency Penalty modifies these probabilities. If a token has already appeared, its chances of being selected again are reduced proportionally to the number of times it has been used.</li><li><strong>Impact on Probability Distribution:</strong> This results in a dynamic probability distribution where frequently used words become less likely to be chosen as the text progresses. The penalty is cumulative, meaning the more a word is used, the greater the reduction in its probability.</li></ul><h4>Example: Low vs. High Frequency Penalty</h4><h4>Low Frequency Penalty Setting:</h4><ul><li><strong>Example Prompt:</strong> “The future of technology is…”</li><li><strong>Generated Text:</strong> “The future of technology is technology-driven. Technology will revolutionize technology sectors.”</li><li><strong>Characteristics:</strong> Here, the word ‘technology’ is repeated often, emphasizing the subject but potentially leading to redundant or monotonous text.</li></ul><h4>High Frequency Penalty Setting:</h4><ul><li><strong>Example Prompt: </strong>“The future of technology is…”</li><li><strong>Generated Text: </strong>“The future of technology lies in innovation, artificial intelligence, and sustainable development.”</li><li><strong>Characteristics:</strong> The diverse range of terms avoids repetition, making the text more engaging and varied.</li></ul><p>By understanding and adjusting the Frequency Penalty, you can effectively control the level of repetition in the generated text, enabling a balance between emphasis and diversity. This parameter is crucial for tasks where either reinforcement of concepts or exploration of diverse ideas is needed.</p><h3>Top-P (Nucleus Sampling) Parameter</h3><h4>TL;DR Summary</h4><p>The Top-P parameter, also known as Nucleus Sampling, is a critical setting in LLMs that controls the breadth of the AI’s thought process. By adjusting Top-P, we influence how wide or narrow the AI’s focus is when selecting the next word in a sentence.</p><h4>Theoretical Explanation</h4><p>Top-P sampling operates on a simple yet powerful principle: instead of considering all possible next words equally, it focuses on a subset (the “nucleus”) that cumulatively makes up a certain percentage (P) of the probability mass. This subset represents the most probable words the model might choose next.</p><ul><li><strong>Lower Top-P Value:</strong> Here, the model is limited to a smaller, more probable set of words. This results in more predictable and safe outputs, as the model sticks closely to the most likely word choices.</li><li><strong>Higher Top-P Value:</strong> This setting expands the range of words considered, including those less likely but potentially more creative or contextually interesting. It allows the model to explore a wider array of word choices, leading to more diverse and potentially unexpected outputs.</li></ul><h4>Practical Examples</h4><h4>Low Top-P Example:</h4><ul><li><strong>Prompt: </strong>“The future of transportation is…”</li><li><strong>Low Top-P Output:</strong> “…likely to be dominated by electric cars and high-speed trains.”</li><li><strong>Analysis: </strong>Here, the output is safe and predictable, sticking closely to current trends and common expectations.</li></ul><h4>High Top-P Example:</h4><ul><li><strong>Prompt: </strong>“The future of transportation is…”</li><li><strong>High Top-P Output:</strong> “…an exciting realm of flying bicycles, teleportation portals, and solar-powered hovercrafts.”</li><li><strong>Analysis: </strong>In this case, the model ventures into more imaginative and less conventional ideas, reflecting a broader range of possibilities.</li></ul><p>Understanding and manipulating the Top-P parameter allows for a tailored approach in generating text. Whether you’re seeking conventional, safe responses or aiming for more adventurous, creative outputs, adjusting the Top-P value is a crucial tool in your LLM toolkit. This parameter’s adjustment can significantly impact the flavor and direction of the AI’s generated content, making it an indispensable aspect of fine-tuning LLMs for specific applications.</p><h3>Presence Penalty Parameter</h3><h4>TL;DR Summary</h4><p>The Presence is essential for managing topic diversity in generated text. A lower setting leads to more repetition and focus on the initial topic, while a higher setting encourages the exploration of new topics, reducing redundancy.</p><h3>Theoretical Explanation</h3><p>At its essence, the Presence Penalty impacts the probability of reusing words or phrases already present in the text. Each time a token (word or phrase) appears, its future probability decreases if the Presence Penalty is high. This mechanism ensures that the AI avoids overusing certain terms, thereby nudging it towards a wider range of topics and ideas. It’s a dynamic balance between repetition and novelty, governed by how heavily the model penalizes the recurrence of tokens.</p><h4>Practical Examples</h4><p>Let’s consider the prompt: “The best sport is.” How does varying the Presence Penalty impact the AI’s response?</p><ul><li><strong>Presence Penalty = 0.0:</strong> “The best sport is soccer, soccer, soccer.” Here, the AI latches onto the word ‘soccer,’ leading to a monotonous and repetitive text.</li><li><strong>Presence Penalty = 0.5:</strong> “The best sport is soccer, basketball, and tennis.” This setting strikes a balance, maintaining coherence while introducing variety.</li><li><strong>Presence Penalty = 1.0:</strong> “The best sport is soccer, rugby, and chess.” At this level, the AI ventures into a wider array of sports, displaying a high level of diversity and unpredictability in the response.</li></ul><h4>Application Tips</h4><p>When to adjust the Presence Penalty? It depends on your objective:</p><ul><li>For Focused, Repetitive Content: A lower Presence Penalty is suitable if you aim to emphasize a specific theme or concept, ensuring that the AI remains anchored to the initial topic.</li><li>For Diverse, Explorative Content: A higher Presence Penalty is ideal for generating content that spans a wide range of ideas, preventing the AI from dwelling too long on a single topic.</li></ul><p>The key lies in striking the right balance. Setting the Presence Penalty too high or too low can skew the relevance and coherence of the generated text. Mindful adjustment of this parameter can significantly enhance the quality and relevance of the AI’s output, aligning it more closely with your specific content goals.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c11f866b33b4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The LLM Scalability Laws are Broken Again: Llama 2 Falls to the Newcomer, Mistral 7B]]></title>
            <link>https://medium.com/@zazazakaria18/the-llm-scalability-laws-are-broken-again-llama-2-falls-to-the-newcomer-mistral-7b-dffe07ee3daf?source=rss-ef40c0d2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/dffe07ee3daf</guid>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[llama-2]]></category>
            <category><![CDATA[hugging-face]]></category>
            <dc:creator><![CDATA[Zaza Zakaria]]></dc:creator>
            <pubDate>Wed, 27 Sep 2023 19:02:27 GMT</pubDate>
            <atom:updated>2023-10-01T13:12:28.749Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y7eR2aFWTyXgx88y9M46Dw.jpeg" /></figure><p>In the relentless race of open-source language models, the benchmarks of excellence are constantly being reset. Once, breaking the scalability laws was synonymous with outshining behemoths like Meta/Chinchilla 70B, a prestigious title Llama 2 clinched with pride. But in a twist of events, Llama 2, the erstwhile disruptor, now finds itself outpaced by a novel force: Mistral 7B.</p><h3>The Evolving Landscape of Open-Source Models</h3><p>The accelerating competition in the open-source domain is more than a testament to sheer innovation; it&#39;s emblematic of a continuous cycle of reinvention. As Mistral 7B rises to prominence, it&#39;s clear that this landscape is shifting. With its stellar performance metrics, versatility across tasks, and the allure of open-source adaptability, Mistral 7B encapsulates the fervor of this dynamic ecosystem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*19c3O_UfpR6yj5X94xs42g.png" /><figcaption>Mistral vs LLama 2 familly benchmark</figcaption></figure><h3>Mistral 7B’s Notable Triumphs</h3><ol><li><strong>Size and Efficiency:</strong> Despite its 7.3B parameters, Mistral 7B consistently overshadows Llama 2’s 13B model across the board. Its prowess even challenges Llama 1’s massive 34B model in several benchmarks.</li><li><strong>Technical Ingenuity:</strong>** The integration of Grouped-query attention (GQA) assures faster inference, while the Sliding Window Attention (SWA) is designed to manage longer sequences with reduced computational demands.</li><li><strong>Balanced Mastery: </strong>Whether it’s code or English tasks, Mistral 7B achieves an enviable equilibrium, mirroring the performance metrics of CodeLlama 7B.</li><li><strong>Open Source &amp; Versatility:</strong> Beyond its performance, Mistral 7B’s adaptability is impressive. With compatibility spanning platforms like AWS, GCP, and Azure, and its availability on HuggingFace, it truly is a model for the masses.</li></ol><h3>Mistral 7B: A Prelude to the Future of AI &amp; LLM</h3><p>Beyond benchmarks and technicalities, Mistral 7B paints a promising picture for the future of AI and Large Language Models (LLMs). Its emergence and subsequent dominance hint at the pace at which the AI community is advancing. If a 7B model can pose such stiff competition today, the prospect of similar models outpacing giants like GPT-4 in the foreseeable future becomes ever more plausible.</p><h3>In Conclusion</h3><p>The narrative of AI’s evolution is being redefined with each groundbreaking release. Mistral 7B’s ascendancy in this narrative is emblematic of the community’s unyielding drive. As we gaze ahead, the horizon gleams with potential. For skeptics who once doubted the capabilities of smaller models in the shadow of giants like GPT-4, Mistral 7B serves as a compelling counter-narrative, signaling a future rifle with possibilities.</p><p>Link to the original blog by mistral https://mistral.ai/news/announcing-mistral-7b/</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dffe07ee3daf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>