<?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 Abhishek Maheshwarappa on Medium]]></title>
        <description><![CDATA[Stories by Abhishek Maheshwarappa on Medium]]></description>
        <link>https://medium.com/@abhishek-maheshwarappa?source=rss-cf986f69de56------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*O0HQdbCO3q3K99jjQzKwPQ.png</url>
            <title>Stories by Abhishek Maheshwarappa on Medium</title>
            <link>https://medium.com/@abhishek-maheshwarappa?source=rss-cf986f69de56------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 11:42:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@abhishek-maheshwarappa/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[The Genie Code: Transforming Data Application Development]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/the-genie-code-transforming-data-application-development-b2fc862c5445?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/b2fc862c5445</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Sat, 14 Mar 2026 21:51:39 GMT</pubDate>
            <atom:updated>2026-03-14T21:51:39.884Z</atom:updated>
            <content:encoded><![CDATA[<h2>Introduction</h2><p>In today’s fast-paced tech environment, efficiency and collaboration are essential for success. Enter <a href="https://www.databricks.com/blog/introducing-genie-code">Genie Code</a> a groundbreaking framework designed to revolutionize data application development by seamlessly integrating data processing, machine learning, and application deployment. This article introduces Genie Code, examining its features, benefits, and potential limitations. Whether you’re a developer, architect, manager, or leader in the tech industry, understanding Genie Code could be crucial for refining data science workflows and boosting productivity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*87gTx1MtBudHm4VZ" /><figcaption>Credits — Databricks</figcaption></figure><h2>What is Genie Code?</h2><p>Genie Code is an innovative framework by Databricks that simplifies the complexities of data application development. It integrates essential processes such as data engineering and machine learning into a cohesive platform, streamlining workflows and fostering teamwork among data teams. In an era that values efficiency and agility, Genie Code is a powerful tool for developers and organizations striving to stay competitive.</p><h2>Key Features of Genie Code</h2><h3>Unified Development Environment</h3><p>Genie Code offers a unified development environment that integrates data processing, machine learning, and application development. This platform facilitates smooth transitions between exploring data, training models, and deploying applications, significantly reducing the friction of traditional workflows. By centralizing these processes, Genie Code enhances productivity and maintains consistency throughout the development lifecycle.</p><h3>Automated Code Generation</h3><p>A standout feature of Genie Code is its automated code generation capability. By creating boilerplate code automatically, developers can focus on higher-level logic and functionality. This shift not only saves time but also allows developers to engage in more innovative and strategic project aspects, ultimately driving meaningful outcomes.</p><h3>Collaboration Tools</h3><p>Effective teamwork is critical in data projects, and Genie Code excels by providing collaboration tools. Shared notebooks and version control features enable real-time teamwork, ensuring all members are aligned and informed. This is particularly beneficial in dynamic environments where rapid changes and iterations are common.</p><h3>Integration with Existing Tools</h3><p>Genie Code’s compatibility with popular data science and machine learning libraries enhances its flexibility, allowing users to integrate seamlessly with existing tools and platforms. This interoperability streamlines data ingestion and processing, enabling teams to leverage their existing infrastructure while adopting Genie Code.</p><h3>Emphasis on Best Practices</h3><p>Promoting best practices is a core tenet of Genie Code. It encourages adherence to coding standards and provides built-in tools for testing and validation, ensuring developed applications are efficient, robust, and maintainable. This focus on quality minimizes errors and optimizes performance, leading to more reliable outcomes.</p><h2>Target Audience and Stakeholders</h2><p>Genie Code is designed for a broad audience, including developers, architects, managers, and tech industry leaders. Its comprehensive features offer value to anyone involved in data application development, making it a versatile tool for enhancing productivity and innovation.</p><h2>Potential Drawbacks and Considerations</h2><p>While Genie Code offers many advantages, it’s important to consider potential limitations. For instance, the learning curve associated with adopting a new framework might pose challenges, especially for teams accustomed to different tools. Additionally, as with any integrated platform, there might be concerns about vendor lock-in or limitations in customization. Evaluating these factors can help teams make informed decisions about adopting Genie Code.</p><h2>Conclusion</h2><p>Genie Code represents a significant advancement in data application development, offering a unified platform for integrating data processing, machine learning, and application deployment. Its emphasis on efficiency, collaboration, and best practices makes it a valuable asset for tech professionals seeking to enhance their workflows. While considering its potential drawbacks, Genie Code’s comprehensive features and benefits make it a compelling choice for those looking to stay ahead in the dynamic tech landscape.</p><p>Stay on the cutting-edge of AI! Follow me on Medium, connect on LinkedIn, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons!</p><p>LinkedIn — <a href="https://www.linkedin.com/in/abhishek-g-m/">https://www.linkedin.com/in/abhishek-g-m/</a> <br>X — <a href="https://x.com/AbhigmKar">https://x.com/AbhigmKar</a> <br>Medium — <a href="https://medium.com/@abhishek-maheshwarappa">https://medium.com/@abhishek-maheshwarappa</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b2fc862c5445" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What’s New from the 2025 Data + AI Summit Keynote]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/whats-new-from-the-2025-data-ai-summit-keynote-9081fae9a54f?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/9081fae9a54f</guid>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-intelligence]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Thu, 19 Jun 2025 03:55:04 GMT</pubDate>
            <atom:updated>2025-06-19T03:55:04.136Z</atom:updated>
            <content:encoded><![CDATA[<p>I went to <a href="https://www.databricks.com/">Databricks</a> Data + AI Summit in San Francisco, and the energy on stage was contagious. The keynote rolled out a slate of product releases that bring us closer to a single, end-to-end Data Intelligence Platform. Below, I have distilled every major announcement from transactional databases in the lakehouse to visual ETL designers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mjwX351PWpaKdwA6.png" /></figure><h4>1. Lakebase: Postgres for the Lakehouse</h4><p><strong>What it is.</strong> A fully managed, serverless Postgres–compatible engine built on your Delta Lake. Lakebase delivers sub-10 ms OLTP performance, seamless branching, point in time recovery, and native integration with Unity Catalog managed tables.<br> <strong>Why it matters.</strong> You no longer need separate operational databases. Build your AI-driven apps on a single platform with consistent governance, IAM, and audit across both OLTP and analytics workloads.</p><h4>2. Databricks Apps: Your Data + AI App Platform</h4><p><strong>What it is.</strong> A framework for deploying secure web applications and AI agents inside Databricks. Apps inherit platform SSO, RBAC, and compute isolation out of the box, and now support Python (Flask, FastAPI, Streamlit) and JavaScript (Node.js/React) stacks with 20+ pre-installed open-source packages.<br> <strong>Why it matters.</strong> Teams can prototype and ship data-centric applications — dashboards, internal tools, GenAI assistants — without standing up separate services or reinventing authentication.</p><h4>3. Agent Bricks: Low Code AI Agents</h4><p><strong>What it is.</strong> A framework to build, deploy, and manage production grade AI agents natively within Databricks. Agent Bricks orchestrate LLM calls, vector search, RAG chains, and custom tooling against your data, with built-in monitoring and governance.<br> <strong>Why it matters.</strong> It streamlines agent development: you write simple configuration or notebook code, and the platform handles scaling, retries, and observability with no external orchestration necessary.</p><h4>4. Spark Declarative Pipelines: Built In Workflow Definitions</h4><p><strong>What it is.</strong> An open source preview wextension to Apache Spark (now in Spark 4.0) that lets you define batch and streaming pipelines declaratively.<br> <strong>Why it matters.</strong> Declarative pipelines reduce boilerplate and simplify maintenance. Data engineers can version control a single pipeline definition, while the runtime handles dependency resolution and execution planning.</p><h4>5. Lakeflow: Code First Orchestration Service (GA)</h4><p><strong>What it is.</strong> A managed service for high-throughput data engineering tasks, offering programmatic APIs in Python, Scala, and SQL to define, schedule, and monitor pipelines.<br> <strong>Why it matters.</strong> You get enterprisegrade orchestration retry policies, lineage tracking, SLA alerts without managing your own scheduler or Kubernetes cluster.</p><h4>6. Lakeflow Designer: Visual ETL Canvas (Private Preview)</h4><p><strong>What it is.</strong> A drag &amp; drop interface for designing and deploying Lakeflow jobs, complete with schema inference, preview mode, parameterization, and lineage visualization.<br> <strong>Why it matters.</strong> Business analysts and citizen data engineers can build reliable ETL pipelines without writing code, while still benefiting from Lakeflow’s production grade runtime.</p><p><strong>Read more</strong><br><a href="https://www.databricks.com/blog/announcing-lakeflow-designer-no-code-etl">https://www.databricks.com/blog/announcing-lakeflow-designer-no-code-etl</a></p><h4>7. Unity Catalog Metrics: Certified Business Metrics</h4><p><strong>What it is.</strong> A governed semantic layer for defining measures and dimensions in YAML or via UI, exposed as first-class SQL objects. Metrics are consumable from Databricks SQL, notebooks, AI/BI Genie, and partner BI tools (Tableau, Sigma, Hex, DOMO, ThoughtSpot).<br> <strong>Why it matters.</strong> Eliminates metric drift by centralizing definitions your whole organization uses the same “single source of truth” for KPIs.</p><h4>8. Lakehouse Federation (Lakebridge) &amp; Iceberg Managed Tables</h4><p><strong>What they are.</strong></p><ul><li><strong>Lakehouse Federation (Lakebridge):</strong> Register and query external catalogs (BigQuery, Redshift, Snowflake, Iceberg) under Unity Catalog governance row or column-level access, tags, lineage.</li><li><strong>Iceberg Managed Tables (Public Preview):</strong> Manage Iceberg tables through Unity Catalog with full metadata, access controls, and audit.</li><li><strong>Why it matters.</strong> Creates a unified data mesh: hybrid workloads run seamlessly across systems, with consistent security and governance.</li></ul><h4>9. Native Gemini Support: Built In LLM Inference</h4><p><strong>What it is.</strong> Google’s Gemini models are now available natively in Databricks for inference, no external API calls or egress costs. You call the Foundation Model APIs directly against Gemini endpoints hosted inside your Databricks workspace.<br> <strong>Why it matters.</strong> Analysts and developers can run state of theart LLM inference with minimal latency and under your existing enterprise security and governance controls.</p><h4><strong>10. AI/BI Genie: Conversational Analytics (GA)</strong></h4><p><strong>What it is.</strong> A natural language interface for querying data &amp; dashboards &amp; ask questions in plain English, receive visualizations or SQL snippets, and iterate within the workspace.<br> <strong>Why it matters.</strong> Lowers the barrier for business users to self serve analytics, accelerates insight discovery, and reduces reliance on centralized BI teams.</p><h4>11. mlflow 3.0 &amp; AI Gateway</h4><p><strong>What they are.</strong></p><ul><li><strong>mlflow 3.0:</strong> A ground up redesign for GenAI, introduces LangGraph orchestration, agent aware tracking, high throughput model serving (250k QPS+), and deeper governance hooks.</li><li><strong>AI Gateway (GA):</strong> A secure, scalable edge-to-lakehouse API that proxies LLM and vector store calls from any client complete with rate limiting and audit logs.</li><li><strong>Why it matters.</strong> Delivers end 2 end observability and control over AI lifecycles from data ingestion through training, inference, and agent deployment.</li></ul><h4>12. Apache Spark 4.0 &amp; Real Time Mode</h4><p><strong>What it is.</strong> Contributions in Spark 4.0 include SQL UDFs, PIPE syntax, session variables, Plotly integration, polymorphic UDTFs, structured logging, and new streaming APIs. Real Time Mode enables fully declarative streaming.<br> <strong>Why it matters.</strong> Simplifies incremental pipelines to write pure SQL, eliminate orchestration glue, and get Spark scale performance in both batch and streaming contexts.</p><h4>13. Zerobus: Direct Write API to Unity Catalog</h4><p><strong>What it is.</strong> A new high throughput, low latency ingestion path that lets edge apps, IoT devices, or cloud services write data directly into Unity Catalog managed tables.<br> <strong>Why it matters.</strong> Bypasses complex ingestion layers for near-real-time analytics, personalization, and monitoring while preserving governance and audit.</p><h4><strong>14. Serverless GPU Compute (Beta)</strong></h4><p><strong>What it is.</strong> Fully managed A10 G (beta) and soon H100 GPUs for notebooks and jobs, with auto-scaling and no reservations.</p><p><strong>Why it matters.</strong> Train or run inference on GPUs without Ops overhead pay only for what you use.</p><h4>15. <strong>MCP Support (Model Context Protocol)</strong></h4><p><strong>What it is.</strong> First-class hosting of Anthropic’s MCP servers inside Databricks Apps, plus Playground integration.<br><strong>Why it matters.</strong> Lets you expose tools, knowledge, and retrieval endpoints to LLMs using a standard protocol no extra infra.</p><h4>References</h4><ol><li><a href="https://www.databricks.com/blog/announcing-lakebase-public-preview">https://www.databricks.com/blog/announcing-lakebase-public-preview</a></li><li><a href="https://www.databricks.com/blog/announcing-general-availability-databricks-lakeflow">https://www.databricks.com/blog/announcing-general-availability-databricks-lakeflow</a></li><li><a href="https://www.databricks.com/blog/mlflow-30-unified-ai-experimentation-observability-and-governance">https://www.databricks.com/blog/mlflow-30-unified-ai-experimentation-observability-and-governance</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9081fae9a54f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Illusion of AI Thinking: How Apple’s New Paper Changed the Way I See AGI]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/the-illusion-of-ai-thinking-how-apples-new-paper-changed-the-way-i-see-agi-71de6da15bd5?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/71de6da15bd5</guid>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[reasoning-model]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Tue, 17 Jun 2025 06:20:24 GMT</pubDate>
            <atom:updated>2025-06-17T06:20:24.466Z</atom:updated>
            <content:encoded><![CDATA[<p>I was in a deep conversation with my friend Amulya a doctoral candidate at BU(<a href="https://www.bu.edu/">Boston University</a>) discussing large language models and the future of AGI. Midway through, she dropped a link into our chat.</p><p>It was a recent <a href="https://www.apple.com/">Apple</a> research paper titled <strong>“</strong><a href="https://machinelearning.apple.com/research/illusion-of-thinking"><strong>The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.</strong></a><strong>”</strong> Published in June 2025, it was clear the moment I opened it wasn’t just another AI paper. It was something deeper. Something that challenged the very core of what we think LLMs can do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UsJ3h6CTuwTtzeh0iwe3pQ.png" /><figcaption>Image generated using Midjourney AI</figcaption></figure><h3>Not Just Another Benchmark</h3><p>We havee all seen pass or fail leaderboard screenshots, multiple choice scores, chain of thought traces, GSM(Grade School Math) type math reasoning sets. This paper is not that. Apple’s researchers shifted the focus entirely, instead of static accuracy, they measured the <strong>scalability of reasoning in modern LLMs</strong>.</p><h3>What Makes This Paper Different</h3><p>Instead of the usual benchmarks, Apple built something they call <strong>controllable puzzle environments</strong>. Here’s what sets this work apart:</p><p><strong>Controllable Puzzle Environments</strong><br> The team designed synthetic puzzle generators. These let them systematically increase the compositional complexity of problems without changing the underlying logic being tested. So, as the puzzles get harder, it’s not because the rules change but because the steps and reasoning required go up.</p><p><strong>Four Core Reasoning Domains</strong></p><ol><li>Symbolic manipulation</li><li>Mathematical and logical deduction</li><li>Planning and multi- step problem solving</li><li>Algorithmic pattern detection</li></ol><p><strong>Relative-Complexity Index</strong><br> Every puzzle instance is labeled with a single “relative complexity index,” so all results are plotted on the same scale no apples to oranges comparisons.</p><p><strong>Reasoning-Trace Analysis</strong><br> Instead of just looking at final answers, the researchers analyze every intermediate step. They study how the models’ reasoning processes degrade as complexity increases, not just when they get the answer wrong.</p><h3>What Did They Find? (And Why It Matters)</h3><p><strong>The Phase Transition in Model Performance:</strong></p><ul><li>All the major models like GPT4, Claude Opus, Gemini 1.5 Pro, Mistral handle low-complexity tasks well.</li><li>As soon as complexity increases by even a single step, performance doesn’t just decline gradually, it collapses.</li><li><strong>Example:</strong> In Boolean logic evaluation:</li></ul><ol><li>At depth 3, models perform with near-perfect accuracy.</li><li>At depth 4, accuracy drops by over half.</li><li>At depth 5, results are barely better than guessing.</li></ol><p><strong>This isn’t a slow fade. It’s a sharp “phase transition” one notch higher in complexity, and reasoning just falls apart.</strong></p><ul><li>The same pattern showed up across all domains.</li><li>This collapse was visible on performance curves, not just final scores.</li></ul><h3>Tools &amp; Scale Weren’t the Answer</h3><p><strong>Tool Use:</strong></p><ul><li>You’d think giving models calculators, code interpreters, or scratchpads would help. It didn’t.</li><li>Longer explanations? Yes.</li><li>Better accuracy? No.</li><li>Tool use polished their syntax, but didn’t fix scalable reasoning.</li></ul><p><strong>Model Size:</strong></p><ul><li>Apple tested smaller and the very largest available LLMs.</li><li>Larger models delayed the collapse by about one complexity level.</li><li>But the <strong>“cliff”</strong> still came just a bit later.</li><li>This points to a limit in architecture, not just scale or training data.</li></ul><h3>Why Most Benchmarks Miss This</h3><p><strong>Benchmarking Problem:</strong></p><ul><li>Most public benchmarks are clustered at low-complexity indices.</li><li>Models look great when tasks are easy.</li><li>The “illusion of thinking” is really an illusion of benchmark design.</li></ul><p><strong>What Apple Proposes:</strong></p><ul><li>Evaluate accuracy as a function of problem complexity — not as a single static number.</li><li>Study when and how performance collapses using reasoning-trace analysis.</li></ul><h3>The Illusion We’re Buying Into</h3><p>Let’s be clear:</p><ul><li>Fluent language isn’t the same as deep reasoning.</li><li>These models are brilliant at mimicking understanding until the reasoning bar is raised.</li><li>At that point, they revert to being sophisticated pattern matchers, not genuine problem solvers.</li><li>The Apple paper argues we’re confusing output fluency with scalable, generalizable intelligence.</li></ul><h3>What Do We Do Now? (The Authors’ Suggestions)</h3><p><strong>Hybrid Systems</strong></p><p>Combine LLM backbones with explicit symbolic modules, long-term memory, meta-reasoning layers, or planning agents.</p><p><strong>Complexity-Aware Evaluation</strong></p><p>Use benchmarks that highlight where and how performance collapses, not just a static pass or fail.</p><p><strong>Reasoning Transparency</strong></p><p>Make reasoning-trace analysis a standard part of model evaluation so we see not just answers, but how those answers break down.</p><h3>Why I Wrote This</h3><p>It wasn’t just the paper alone that compelled me to write. It was the conversation with Amulya the curiosity sparked by her sharing this deep dive into AI’s real limits. This research challenges what many of us have come to assume about LLMs and AGI.</p><p>If you’re following AI progress and wondering whether the hype reflects reality, take Apple’s findings seriously. Sometimes, all it takes is one thoughtful paper and one honest conversation to completely shift your perspective.</p><p>The future of AI depends on how well we understand these limits &amp; how boldly we innovate beyond them.</p><p><strong>References</strong></p><ul><li>Apple Machine Learning Research. (2025). <em>The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity</em>. Retrieved from <a href="https://machinelearning.apple.com/research/the-illusion-of-thinking">https://machinelearning.apple.com/research/the-illusion-of-thinking</a></li></ul><p>Stay on the cutting-edge of AI! Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71de6da15bd5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Grok 3 — The Age of Reasoning Agents]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/grok-3-the-age-of-reasoning-agents-980b9440b058?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/980b9440b058</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[grok-3]]></category>
            <category><![CDATA[agi]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Sat, 22 Feb 2025 17:43:40 GMT</pubDate>
            <atom:updated>2025-02-22T17:43:40.656Z</atom:updated>
            <content:encoded><![CDATA[<h3>Grok 3 — The Age of Reasoning Agents</h3><p>Elon Musk’s <a href="https://x.ai/">xAI</a> has recently unveiled Grok 3, its most advanced AI model to date, aiming to redefine the landscape of artificial intelligence.</p><p>This article delves into the inception of xAI and Grok, explores the groundbreaking features of Grok 3, highlights its standout attributes, compares it with leading AI models like Claude, GPT-4o, and DeepSeek R1, and discusses its projected impact for the year ahead.</p><h3><strong>Introduction to xAI and Grok</strong></h3><p>Founded by Elon Musk in July 2023, xAI’s mission is to “understand the universe” through the development of cutting-edge AI technologies.</p><p>The company’s flagship product, Grok, is an AI chatbot designed to provide insightful, unfiltered answers to a wide array of questions, drawing inspiration from the “Hitchhiker’s Guide to the Galaxy.” Grok integrates seamlessly with X (formerly Twitter), leveraging real-time data to enhance its responses.</p><h3><strong>What is Grok 3?</strong></h3><p>Launched in February 2025, Grok 3 represents a significant leap in AI capabilities. Trained on xAI’s Colossus supercluster, which boasts approximately 200,000 GPUs, Grok 3 utilized ten times the computational power of its predecessor, Grok 2. This extensive training has endowed Grok 3 with enhanced reasoning, mathematics, coding proficiency, and a vast repository of world knowledge.</p><p>Notably, Grok 3 introduces advanced reasoning capabilities, allowing it to process complex queries by breaking them down into manageable components and evaluating multiple solution pathways.</p><h3><strong>Standout Features of Grok 3</strong></h3><ul><li><strong>Advanced Reasoning:</strong> Grok 3 employs large-scale reinforcement learning to refine its problem-solving strategies, enabling it to think for extended periods, correct errors, and explore alternative solutions.</li><li><strong>Enhanced Computational Power:</strong> The model’s training on the Colossus supercluster with 10x the compute of previous models has significantly improved its performance across various tasks.</li><li><strong>User-Controlled Reasoning Modes:</strong> Users can engage “Think” mode for step-by-step reasoning or activate “Big Brain” mode for more complex problem-solving, utilizing additional computational resources.</li><li><strong>DeepSearch Integration:</strong> Grok 3 features DeepSearch, an AI-powered search engine that scans the internet and X to generate detailed summaries in response to user queries, enhancing its research capabilities.</li></ul><h3><strong>Comparison with Claude, GPT-4o, and DeepSeek R1</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F9-ujCeyY7GXueyobHqG8Q.png" /><figcaption>Source — X AI Grok</figcaption></figure><p>In benchmark tests, Grok 3 has demonstrated superior performance compared to its contemporaries:</p><ul><li><strong>Mathematics and Science:</strong> Grok 3 achieved a 93.3% score on the 2025 American Invitational Mathematics Examination (AIME) and 84.6% on graduate-level expert reasoning (GPQA), surpassing models like GPT-4o and Claude 3.5 Sonnet.</li><li><strong>Coding Proficiency:</strong> On the LiveCodeBench for code generation and problem-solving, Grok 3 attained a 79.4% score, indicating a higher accuracy and code quality than DeepSeek R1.</li><li><strong>Real-World Applications:</strong> Grok 3’s integration with DeepSearch offers users direct, clickable source links for every claim, ensuring transparency and credibility, a feature that sets it apart from other models.</li></ul><h3><strong>Projections for the Year Ahead</strong></h3><p>As AI technology continues to evolve, Grok 3 is poised to play a pivotal role in various sectors:</p><ul><li><strong>Enterprise Integration:</strong> xAI plans to offer Grok 3 through its enterprise API, enabling businesses to harness its advanced reasoning and data analysis capabilities for improved decision-making.</li><li><strong>Multimodal Enhancements:</strong> Upcoming features include a synthesized voice mode, allowing for more natural and interactive user experiences.</li><li><strong>Open-Source Initiatives:</strong> xAI intends to open-source Grok 2, fostering community-driven development and innovation in AI applications.</li></ul><p>Grok 3’s advancements signify a transformative shift in AI, offering enhanced reasoning and problem-solving capabilities that are set to influence various industries and applications in the coming year.</p><h3><strong>References</strong></h3><ul><li>xAI. (2025, February 19). <em>Grok 3 Beta — The Age of Reasoning Agents</em>. Retrieved from<a href="https://x.ai/?utm_source=chatgpt.com">X.ai</a></li></ul><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=980b9440b058" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI’s Deep Research: Leap Towards AGI]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/openais-deep-research-leap-towards-agi-4926e84097cd?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/4926e84097cd</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[agi]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Thu, 06 Feb 2025 02:21:21 GMT</pubDate>
            <atom:updated>2025-02-06T02:21:21.614Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>Artificial Intelligence (AI) continues to evolve, offering tools that enhance our ability to process and analyze vast amounts of information. OpenAI’s latest development, <a href="https://openai.com/index/introducing-deep-research/"><strong>Deep Research</strong></a>, represents a significant step in this direction, aiming to streamline complex research tasks through advanced AI capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jzGgW9EEPM7lbHDk2PDWkQ.png" /><figcaption>Source — OpenAI</figcaption></figure><p>What is Deep Research?</p><p>Deep Research is a new feature within ChatGPT that autonomously conducts multi-step research tasks on the internet. Leveraging a specialized version of OpenAI’s upcoming o3 model, it is optimized for web browsing and data analysis. Users provide a prompt, and Deep Research finds, analyzes, and synthesizes information from various online sources, including text, images, and PDFs, to create comprehensive reports. This process, which takes Deep Research only tens of minutes, would traditionally take a human many hours.</p><h3>Why Did OpenAI Develop Deep Research?</h3><p>OpenAI developed Deep Research to enhance the efficiency and depth of information gathering and analysis. By automating complex research tasks, it aims to assist professionals in fields such as finance, science, and engineering, providing them with comprehensive reports in a fraction of the time it would take manually. This development reflects OpenAI’s commitment to leveraging AI to augment human capabilities and streamline workflows.</p><h3>Key Features and Capabilities</h3><p>Deep Research offers several notable features:</p><ul><li><strong>Autonomous Multi-Step Research:</strong> It can independently plan and execute complex research tasks, navigating through multiple sources to gather relevant information.</li><li><strong>Comprehensive Analysis:</strong> The tool synthesizes data from various formats, including text, images, and PDFs, to generate well-rounded reports.</li><li><strong>Efficiency:</strong> Tasks that would take humans many hours are completed by Deep Research in tens of minutes, significantly enhancing productivity.</li><li><strong>Citations and Transparency:</strong> The generated reports include citations, allowing users to verify sources and ensuring transparency in the information provided.</li></ul><h3>Benchmarking AI: “Humanity’s Last Exam” and More</h3><p>To evaluate AI’s intelligence, Deep Research introduces <strong>new benchmarking frameworks</strong> aimed at testing advanced reasoning, decision-making, and adaptability.</p><h3>Key AI Benchmarks in Deep Research:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WSVqYjjH28BUy_4bVNw5Vw.png" /></figure><h3>How to Use Deep Research</h3><p>Deep Research is integrated into the ChatGPT interface. To utilize this feature:</p><ol><li><strong>Access ChatGPT:</strong> Ensure you have access to ChatGPT, with Deep Research currently available to Pro subscribers in the U.S.</li><li><strong>Input a Prompt:</strong> Enter a specific query or topic you wish to explore.</li><li><strong>Initiate Deep Research:</strong> The AI will autonomously conduct research, analyzing and synthesizing information from various online sources.</li><li><strong>Review the Report:</strong> Within tens of minutes, Deep Research will provide a comprehensive report complete with citations, allowing you to delve into the findings and verify sources as needed.</li></ol><h3>Availability</h3><p>As of now, Deep Research is available to <a href="https://openai.com/index/introducing-chatgpt-prox"><strong>ChatGPT Pro subscribers</strong></a> in the United States. OpenAI plans to expand access to other regions and user tiers in the future. Users can access this feature through the web version of ChatGPT, with plans to roll it out to mobile and desktop applications later in February.</p><h3>The Road Ahead: Enhancing AI-Driven Research</h3><p>OpenAI’s Deep Research marks a significant advancement in AI-driven analysis, offering a tool that can autonomously conduct complex research tasks efficiently. As AI continues to evolve, tools like Deep Research exemplify the potential to augment human capabilities, making information gathering and analysis more efficient and comprehensive.</p><h3>References</h3><ul><li><a href="https://openai.com/">https://openai.com/</a></li><li><a href="https://www.youtube.com/watch?v=YkCDVn3_wiw">https://www.youtube.com/watch?v=YkCDVn3_wiw</a></li></ul><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4926e84097cd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Running DeepSeek-R1 and Qwen2.5 Locally with Ollama]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/running-deepseek-r1-and-qwen2-5-vl-locally-with-ollama-6ed4e115f6bb?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ed4e115f6bb</guid>
            <category><![CDATA[qwen2-5]]></category>
            <category><![CDATA[deepseek-r1]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[alibaba]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Fri, 31 Jan 2025 12:48:32 GMT</pubDate>
            <atom:updated>2025-02-05T15:16:41.118Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>With the increasing popularity of open-source AI models, running powerful LLMs like <a href="https://medium.com/@abhishek-maheshwarappa/deepseek-r1-revolutionizing-reasoning-with-reinforcement-learning-and-distillation-24f9e1877627"><strong>DeepSeek-R1</strong></a> and <strong>Qwen2.5</strong> locally has become more accessible than ever. These models offer cutting-edge capabilities for natural language processing (DeepSeek-R1) and multimodal AI (, allowing users to leverage them for diverse applications like chatbots, summarization, code generation, and image-based reasoning.</p><p>To read more about DeepSeek, check out my other article on it — <a href="https://medium.com/@abhishek-maheshwarappa/deepseek-r1-revolutionizing-reasoning-with-reinforcement-learning-and-distillation-24f9e1877627">here</a></p><p><strong>Ollama</strong> is a lightweight and easy-to-use tool designed to run LLMs on local machines efficiently. It simplifies downloading, managing, and running models without the need for complex dependencies or cloud-based APIs.</p><h3>What is Ollama?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hpFiAn2ffM3UfRX7YtemcA.png" /><figcaption>Source — <a href="https://ollama.com/">https://ollama.com/</a></figcaption></figure><p><a href="https://ollama.com/">Ollama</a> is an open-source framework that enables users to run LLMs locally with minimal setup. It provides a seamless interface to download and execute various models, making it an excellent choice for developers, researchers, and AI enthusiasts who want to experiment with AI capabilities offline. Key features include:</p><ul><li>Simple installation and model management</li><li>Support for a variety of open-source models</li><li>Local execution with optimized performance</li><li>Interactive chat and inference capabilities</li></ul><h3>Uses of Ollama</h3><p>Ollama enables a variety of AI-powered applications, including:</p><ul><li><strong>Local AI Chatbots</strong> — Run interactive chat models without needing internet access.</li><li><strong>Text Summarization</strong> — Process large text inputs and generate concise summaries.</li><li><strong>Code Generation</strong> — Use models like DeepSeek-R1 to generate or complete code.</li><li><strong>Multimodal Processing</strong> — Utilize Qwen2.5-VL for image-based reasoning and vision-language tasks.</li><li><strong>AI-Assisted Writing</strong> — Enhance content creation with AI-powered assistance.</li></ul><h3>How to Install Ollama</h3><h3>Installation on macOS/Linux</h3><p>Open a terminal and run:</p><pre>curl -fsSL https://ollama.com/install.sh | sh</pre><h3>Installation on Windows</h3><p>Windows users need <strong>WSL 2 (Windows Subsystem for Linux)</strong>. Once WSL is installed, follow the macOS/Linux installation command in the WSL terminal.</p><p>After installation, verify that Ollama is working by running:</p><pre>ollama --version</pre><h3>How to Run Ollama</h3><p>Once installed, start the Ollama service by running:</p><pre>ollama serve</pre><p>This will allow you to interact with models locally.</p><h3>How to Pull Any Model</h3><p>To download and use a model, use the following command:</p><pre>ollama pull &lt;model-name&gt;</pre><p>For example, to pull <strong>DeepSeek-R1</strong>:</p><pre>ollama pull deepseek-r1</pre><p>To pull <strong>Qwen2.5</strong>:</p><pre>ollama pull qwen2.5</pre><h3>How to Chat with the Models</h3><p>Once a model is pulled, you can start a chat session:</p><pre>ollama run deepseek-r1</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TiYr2y4Vp5nj6V3pNpm2BQ.png" /><figcaption>Source — Deepseek.com</figcaption></figure><p>For a multimodal model like <strong>Qwen2.5</strong>, you can input both text and images:</p><pre>ollama run qwen2.5</pre><p>To exit the chat, use <strong>Ctrl+C</strong>.</p><h3>Useful Ollama Commands</h3><p>Here are some handy commands to interact with Ollama:</p><ul><li><strong>List available models</strong>:</li></ul><pre>ollama list</pre><ul><li><strong>Show details of a model</strong>:</li></ul><pre>ollama show deepseek-r1</pre><ul><li><strong>Remove a model</strong>:</li></ul><pre>ollama rm deepseek-r1</pre><ul><li><strong>Run a model with a custom prompt</strong>:</li></ul><pre>ollama run deepseek-r1 &quot;Summarize this article for me.&quot;</pre><h3>References and Citations</h3><ol><li><strong>DeepSeek-R1 Paper &amp; GitHub</strong>: <a href="https://github.com/deepseek-ai">https://github.com/deepseek-ai</a></li><li><strong>Qwen2.5 Model</strong>: <a href="https://huggingface.co/Qwen">https://huggingface.co/Qwen</a></li><li><strong>Ollama Documentation</strong>: <a href="https://ollama.com/">https://ollama.com</a></li></ol><p>By following this guide, you can easily run state-of-the-art AI models locally without requiring cloud-based resources, ensuring efficiency and privacy while exploring AI capabilities.</p><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ed4e115f6bb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Qwen 2.5-VL: Alibaba’s Latest Leap in Multimodal AI]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/qwen-2-5-vl-alibabas-latest-leap-in-multimodal-ai-10b46e5e6552?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/10b46e5e6552</guid>
            <category><![CDATA[qwen]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[alibaba]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Wed, 29 Jan 2025 22:41:16 GMT</pubDate>
            <atom:updated>2025-01-29T22:41:16.395Z</atom:updated>
            <content:encoded><![CDATA[<p>The world of AI is evolving at an unprecedented pace, and China is leading the charge with groundbreaking innovations. Hot on the heels of DeepSeek-R1, <a href="https://www.alibabacloud.com/en/solutions/generative-ai/qwen?_p_lc=1">Alibaba</a> has dropped Qwen2.5-VL, a state-of-the-art multimodal model that showcases just how rapidly AI is advancing. This release underscores how China is overtaking global players in AI innovation, delivering models that push the boundaries of what’s possible in natural language processing and vision-language understanding.</p><h3>The Rise of Qwen2.5-VL</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fFDVgMi7SCW_-HFi" /><figcaption>Source — Alibaba Qwen</figcaption></figure><p>Alibaba’s <a href="https://qwenlm.github.io/">Qwen</a> series has been making waves, and Qwen2.5-VL is no exception. Built on top of its previous iterations, this model is designed to handle vision-language tasks with remarkable efficiency. It integrates image and text processing capabilities, making it a powerful tool for applications in e-commerce, content creation, and even autonomous systems.</p><h3>What’s New in Qwen2.5-VL?</h3><ol><li><strong>Enhanced Multimodal Understanding</strong> — The model excels at interpreting images alongside text, making it a top contender in vision-language AI.</li><li><strong>Improved Response Quality</strong> — Alibaba has significantly optimized the model’s reasoning and response generation, producing more accurate and coherent answers.</li><li><strong>Faster Inference</strong> — Leveraging advanced architectures, Qwen2.5-VL is highly efficient, enabling real-time applications with lower latency.</li><li><strong>Bigger and Smarter</strong> — With an increased parameter count and enhanced training techniques, this model delivers superior performance compared to its predecessors.</li><li><strong>Open-Source Advantage</strong> — Qwen2.5-VL is <strong>open-source</strong>, making cutting-edge AI more accessible to researchers and developers worldwide.</li></ol><h3>Performance Benchmarks</h3><p>Qwen2.5-VL has been tested against industry-leading models, showing significant improvements in:</p><ul><li><strong>Image Captioning</strong></li><li><strong>Visual Question Answering (VQA)</strong></li><li><strong>Text-Image Generation</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oJJ7Z7-5_9heBFPp" /><figcaption>Source — Alibaba Qwen</figcaption></figure><h3>What Can Qwen2.5-VL Do?</h3><p>Here are some impressive capabilities of the model:</p><ul><li><strong>Describe complex images accurately</strong> — It can generate detailed descriptions of scenes, making it useful for accessibility tools.</li><li><strong>Answer questions about images</strong> — It can understand and respond to queries about the content of a given image.</li><li><strong>Generate captions for e-commerce</strong> — Automatically generate product descriptions and marketing content.</li><li><strong>Assist in creative design</strong> — Help generate visual ideas based on text prompts.</li><li><strong>Enhance AR and VR experiences</strong> — Power immersive AI-driven interactions.</li><li><strong>Parse and summarize documents</strong> — Extract key information from PDFs, contracts, and research papers, streamlining workflow automation.</li><li><strong>Read Chinese vertical text</strong> — Unlike many Western models, Qwen2.5-VL can process and interpret Chinese vertical reading formats, a crucial feature for digital archives and traditional content.</li><li><strong>Advanced handwriting recognition</strong> — Identify and transcribe handwritten notes, improving accessibility for digital note-taking applications.</li></ul><h3>The AI Race: How China is Leading</h3><p>The speed at which Alibaba released Qwen2.5-VL, right after the DeepSeek-R1 breakthrough, is a testament to how China is accelerating AI innovation. With companies like Alibaba, Baidu, and ByteDance pouring resources into LLMs and multimodal AI, China is positioning itself as a dominant force in the AI arms race. While Western firms focus on commercial AI, Chinese companies are aggressively iterating and improving models at a staggering pace.</p><h3>Why is China Overtaking the AI Space?</h3><ul><li><strong>Investment &amp; Infrastructure</strong> — Massive government and private sector funding in AI research.</li><li><strong>Talent Pool</strong> — A strong pipeline of AI researchers and engineers trained in cutting-edge technologies.</li><li><strong>Fast Iteration</strong> — Companies like Alibaba are not just developing AI — they are rapidly deploying and improving upon their models.</li><li><strong>Open-Source Culture</strong> — Unlike the walled-garden approach of some Western AI giants, China’s tech firms are making their AI tools widely accessible.</li></ul><h3>What’s Next?</h3><p>With Qwen2.5-VL setting new benchmarks in multimodal AI, the next wave of innovation will likely focus on real-world applications. Expect breakthroughs in AI-powered commerce, robotics, and generative content creation. As the competition heats up, the rapid evolution of models like Qwen2.5-VL will redefine how AI integrates into our daily lives.</p><p>For those looking to explore this model further, check out Alibaba’s official blog on Qwen2.5-VL <a href="https://qwenlm.github.io/blog/qwen2.5-vl/">here</a>.</p><p>China’s AI boom is not slowing down anytime soon. The question is — how will the rest of the world keep up?</p><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10b46e5e6552" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[DeepSeek-R1: Revolutionizing Reasoning with Reinforcement Learning and Distillation]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/deepseek-r1-revolutionizing-reasoning-with-reinforcement-learning-and-distillation-24f9e1877627?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/24f9e1877627</guid>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[deepseek-r1]]></category>
            <category><![CDATA[agi]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Sun, 26 Jan 2025 19:21:09 GMT</pubDate>
            <atom:updated>2025-01-26T19:36:58.369Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>In this article, we will explore the advancements and methodologies behind <a href="https://arxiv.org/pdf/2501.12948">DeepSeek-R1</a>, a cutting-edge approach to enhancing reasoning capabilities in Large Language Models (LLMs). It provide a comprehensive summary of the paper, highlighting key innovations, such as the use of reinforcement learning (RL) to incentivize reasoning and the distillation of capabilities into smaller models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O7oJgUICAgOnXW1R8hr0rA.png" /><figcaption>Source: DeepSeek AI</figcaption></figure><p>The quest to enhance reasoning capabilities in Large Language Models (LLMs) has seen significant progress with the advent of reinforcement learning (RL). In this article, we delve into <a href="https://www.deepseek.com/">DeepSeek</a>-Zero and DeepSeek-R1, groundbreaking approaches that comprise two distinct models:</p><ol><li><strong>DeepSeek-Zero</strong>: A pure RL-based model that achieves impressive reasoning capabilities without relying on supervised fine-tuning (SFT).</li><li><strong>DeepSeek-R1</strong>: A more refined version that integrates multi-stage training and cold-start data to improve readability and reasoning performance.</li></ol><h4>Novelty</h4><p>The uniqueness of these models lies in:</p><ul><li>Incentivizing reasoning purely through RL, eliminating the dependency on SFT.</li><li>Distilling reasoning capabilities from large models into smaller, efficient architectures.</li></ul><h3>Approach</h3><h4>Reasoning Capabilities Through Reinforcement Learning (RL)</h4><p>These models explore the potential of RL to enhance reasoning capabilities without using SFT as a cold start. The training pipeline transitions from DeepSeek-Zero, which focuses on self-evolution through RL, to DeepSeek-R1, which incorporates structured data for improved results.</p><h4>DeepSeek-Zero: Reinforcement Learning on the Base Model</h4><p><strong>Group Relative Policy Optimization (GRPO):<br></strong>GRPO is a computationally efficient RL technique that replaces the traditional critic model with a group score-based estimation, reducing computational overhead.</p><p><strong>Reward Modeling:</strong></p><p><strong>1. Accuracy Rewards:</strong> <br>Ensure correctness in deterministic tasks such as math and coding.<br>2. <strong>Format Rewards:</strong> <br>Enforce a structured reasoning process using and tags.</p><p><strong>Performance and Self-Evalution:</strong></p><p>Comparison of DeepSeek-R1-Zero and OpenAI o1 models on reasoning</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/681/1*_MQTxYxXlUutIEEcdBShBg.png" /><figcaption>Source: DeepSeek AI</figcaption></figure><p>DeepSeek-Zero demonstrates a steady improvement in reasoning benchmarks, with notable “aha moments” where the model autonomously refines its approach, achieving performance comparable to OpenAI’s o1 series.</p><h3>DeepSeek-R1: Reinforcement Learning with Cold Start</h3><h4>What is Cold Start?</h4><p>Cold start involves fine-tuning the base model(DeepSeek-V3-Base) with curated long Chain-of-Thought (CoT) data to stabilize RL training. This process ensures:</p><ul><li>Enhanced readability.</li><li>Structured responses with summaries at the end of each output.</li><li>To address the initial instability of RL training when starting with a raw, untuned model.</li><li>It creates a foundation for RL to build upon, enabling faster convergence and better performance on reasoning tasks.</li></ul><h4>Reasoning-Oriented Reinforcement Learning</h4><p>After the cold start, large-scale RL is applied to improve reasoning-intensive tasks like coding, math, and logic. To address language mixing, a language consistency reward aligns outputs with human preferences, enhancing coherence and readability.</p><h4>Data Usage</h4><p><strong>Reasoning Data:</strong></p><ul><li>This dataset includes prompts specifically designed for reasoning-intensive tasks, such as solving mathematical problems, logical reasoning, and structured problem-solving scenarios. The training relies on rule-based rewards to evaluate correctness, ensuring the model can handle well-defined problems with clear solutions.</li><li>Examples include tasks from standardized math and logic competitions like AIME or problem-solving platforms like Codeforces. The data helps the model excel in generating structured and logical outputs.</li></ul><p><strong>Non-Reasoning Data:</strong></p><ul><li>This dataset comprises prompts for tasks such as question-answering (QA), creative writing, and language translation. These tasks are included to diversify the model’s capabilities beyond reasoning and ensure its general applicability.</li><li>Non-reasoning data is drawn from diverse domains, including factual QA benchmarks, conversational tasks, and language-specific translations, helping the model handle a broader range of queries with fluency and coherence.</li></ul><h4>Distillation: Empower Small Models with Reasoning Capability</h4><p>DeepSeek-R1’s reasoning capabilities are distilled into smaller models like Qwen and Llama series, significantly enhancing their performance without requiring RL training. This approach democratizes access to advanced reasoning capabilities for research and industry applications.</p><h3>DeepSeek-R1 Evaluation</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kBVdcoZCIs3EjLHb6hvJXw.png" /></figure><p>DeepSeek-R1’s performance is benchmarked against industry-leading models, showcasing:</p><ul><li><strong>Reasoning Tasks:</strong> Achieving superior accuracy in benchmarks like AIME 2024 and MATH-500.</li><li><strong>General QA:</strong> Outperforming competitors like GPT-4o and Claude in creative writing and instruction-following tasks.</li><li><strong>Long-Context Understanding:</strong> Excelling in tasks requiring extended reasoning, such as AlpacaEval and ArenaHard.</li></ul><p>These results highlight the effectiveness of RL in improving reasoning and generalization across diverse tasks.</p><h3>Distillation vs. Reinforcement Learning</h3><h4>Distillation vs. RL</h4><p><strong>Advantages of Distillation:</strong></p><ul><li>Distillation achieves better performance for smaller models with less computational effort compared to RL.</li><li>DeepSeek-R1 distilled models outperform traditional RL-trained compact architectures, such as QwQ-32B.</li></ul><p><strong>Challenges with RL:</strong></p><ul><li>RL for smaller models is computationally intensive and may not yield results comparable to distillation.</li></ul><h4>Unsuccessful Attempts</h4><p>Some of the experiments conducted during the development process were unsuccessful, highlighting certain limitations and challenges. These include:</p><p><strong>Process Reward Models (PRM):</strong></p><ul><li>PRM guides the model’s reasoning process by evaluating intermediate steps (Lightman et al., 2023; Uesato et al., 2022).</li><li><strong>Challenges:</strong> Difficulty in defining fine-grain steps, reliance on manual annotations, and susceptibility to reward hacking hindered scalability (Gao et al., 2022).</li></ul><p><strong>Monte Carlo Tree Search (MCTS):</strong></p><ul><li>Inspired by AlphaGo (Silver et al., 2017a) and AlphaZero (Silver et al., 2017b), MCTS breaks problems into smaller parts to explore solutions systematically.</li><li><strong>Challenges:</strong> The exponential complexity of token generation and difficulty in training fine-grained value models led to suboptimal performance (Feng et al., 2024).</li></ul><p>These attempts provide valuable insights but highlight the limitations of certain techniques in reasoning-focused LLMs.</p><h3>Conclusion, Limitation, and Future Work</h3><h4>Conclusion</h4><p>DeepSeek-R1 exemplifies the potential of RL in advancing reasoning capabilities, achieving results comparable to state-of-the-art models like OpenAI-o1–1217. The distillation process further extends these capabilities to smaller, efficient models, making advanced reasoning more accessible.</p><h4>Limitations</h4><ul><li>Challenges in multi-turn interactions and complex role-playing tasks.</li><li>Language mixing issues when handling queries in non-English languages.</li><li>Sensitivity to prompts, where few-shot prompting degrades performance.</li></ul><h4>Future Work</h4><ul><li>Enhance prompt engineering to improve robustness.</li><li>Address language mixing by expanding the training dataset.</li><li>Incorporate asynchronous evaluations to improve efficiency in software engineering tasks.</li></ul><h3>References</h3><ul><li>DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning</li><li><a href="https://www.deepseek.com/">https://www.deepseek.com/</a></li></ul><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=24f9e1877627" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fine-Tuning DeepSeek LLM: Adapting Open-Source AI for Your Needs]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/fine-tuning-deepseek-llm-adapting-open-source-ai-for-your-needs-12a7e5572fa5?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/12a7e5572fa5</guid>
            <category><![CDATA[deepseek-v3]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[finetune-llm]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Wed, 22 Jan 2025 01:47:33 GMT</pubDate>
            <atom:updated>2025-01-22T01:47:33.905Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>DeepSeek LLM is a powerful open-source language model, but to maximize its potential for specific applications, fine-tuning is essential. In this guide, we’ll walk through the process of fine-tuning DeepSeek LLM using Supervised Fine-Tuning (SFT) with Hugging Face datasets, providing a step-by-step code walkthrough for training on a domain-specific dataset. We will also discuss the loss function used, why a subset of data was used, and how LoRA (Low-Rank Adaptation) enables memory-efficient fine-tuning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-eMCNFUD-60M3LpI.png" /><figcaption>Source: DeepSeek AI</figcaption></figure><p>For people who love to get hands dirty use this Google Colab</p><p><a href="https://colab.research.google.com/drive/1jyEhPnTf2aHAn0jttQ99M2fCaszoZBDU?usp=sharing">Google Colab</a></p><h3>Overview of Supervised Fine-Tuning (SFT)</h3><p>Supervised Fine-Tuning (SFT) is the process of further training a pre-trained model on a labeled dataset to specialize it for a particular task, such as customer support, medical Q&amp;A, or e-commerce recommendations.</p><h3>How Fine-Tuning Works</h3><p>Fine-tuning involves training the model on <strong>task-specific labeled data</strong>, where:</p><ul><li><strong>Input (X)</strong>: The text data given to the model.</li><li><strong>Target (Y)</strong>: The expected output based on labeled data (e.g., sentiment label, chatbot response, or summarized text).</li><li><strong>Loss Function</strong>: Measures how well the model’s predictions match the expected output. The most commonly used loss function for text generation is <strong>Cross-Entropy Loss</strong>.</li></ul><p>For example, when fine-tuning on the <strong>IMDB sentiment dataset</strong>:</p><ul><li><strong>Input (X)</strong>: A movie review like <em>“The movie had great visuals, but the plot was weak.”</em></li><li><strong>Target (Y)</strong>: The correct label, e.g., <strong>“Positive” or “Negative” sentiment</strong>.</li></ul><p>For <strong>text generation tasks</strong>, input could be a <strong>question</strong>, and the target could be the <strong>correct response</strong> generated by the model.</p><h3>Loss Function Used: Cross-Entropy Loss</h3><p>For language models, we use Cross-Entropy Loss, which calculates the difference between the predicted token distribution and the actual target distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/645/0*hE31UN7frZA7aP7e.png" /><figcaption>Credits — Latex to Image</figcaption></figure><p>The goal is to minimize this loss during training so that the model learns to generate more accurate text outputs.</p><h3>2. Why Use a Subset of Data?</h3><p>When fine-tuning large language models like DeepSeek LLM on <strong>resource-limited hardware</strong>, training on the full dataset (e.g., IMDB with 25,000 samples) can lead to excessive training time and GPU memory issues. To mitigate this, we:</p><ul><li><strong>Selected a subset</strong>: 500 samples for training and 100 for evaluation.</li><li><strong>Maintained representativeness</strong>: This subset retains enough diversity to achieve reasonable performance.</li></ul><p>Using a smaller dataset speeds up experimentation while demonstrating fine-tuning concepts effectively. For production-level fine-tuning, larger datasets should be used on more powerful infrastructure.</p><h3>3. Load the DeepSeek LLM</h3><p>Before fine-tuning, we need to load the <strong>DeepSeek LLM</strong> and prepare it for training.</p><h3>Install Required Libraries</h3><p>First, install the necessary dependencies:</p><pre>pip install -U torch transformers datasets accelerate peft bitsandbytes</pre><h3>Load the Model with 4-bit Quantization</h3><p>We use 4-bit quantization to make the large model compatible with limited GPU memory:</p><pre>from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig<br>from peft import LoraConfig, get_peft_model<br><br>model_name = &quot;deepseek-ai/deepseek-llm-7b-base&quot;<br># Configure 4-bit quantization<br>bnb_config = BitsAndBytesConfig(<br>    load_in_4bit=True,<br>    bnb_4bit_compute_dtype=torch.float16  # Use float16 for faster computation<br>)<br># Load tokenizer and model<br>tokenizer = AutoTokenizer.from_pretrained(model_name)<br>model = AutoModelForCausalLM.from_pretrained(<br>    model_name, <br>    quantization_config=bnb_config, <br>    device_map=&quot;auto&quot;<br>)<br># Apply LoRA for memory-efficient fine-tuning<br>lora_config = LoraConfig(<br>    r=8,  # Low-rank adaptation size<br>    lora_alpha=32,<br>    target_modules=[&quot;q_proj&quot;, &quot;v_proj&quot;],  # Apply LoRA to attention layers<br>    lora_dropout=0.05,<br>    bias=&quot;none&quot;<br>)<br>model = get_peft_model(model, lora_config)<br>model.print_trainable_parameters()<br>print(&quot;✅ DeepSeek LLM Loaded with LoRA and 4-bit Precision!&quot;)</pre><h3>4. Using Hugging Face Datasets for Training</h3><p>For fine-tuning, we need a high-quality dataset. Hugging Face provides access to various datasets:</p><h3>Select a Dataset</h3><p>For this example, let’s use the <strong>IMDB dataset</strong> for fine-tuning DeepSeek LLM on sentiment classification:</p><pre>from datasets import load_dataset<br><br># Load dataset<br>dataset = load_dataset(&quot;imdb&quot;)</pre><h3>Preprocess the Dataset</h3><p>Convert text into tokenized inputs for the model:</p><pre>def tokenize_function(examples):<br>    inputs = tokenizer(<br>        examples[&quot;text&quot;], <br>        truncation=True, <br>        padding=&quot;max_length&quot;, <br>        max_length=512<br>    )<br>    inputs[&quot;labels&quot;] = inputs[&quot;input_ids&quot;].copy()<br>    return inputs<br><br>tokenized_datasets = dataset.map(tokenize_function, batched=True)<br># Subset the dataset for faster experimentation<br>small_train_dataset = tokenized_datasets[&quot;train&quot;].shuffle(seed=42).select(range(500))<br>small_test_dataset = tokenized_datasets[&quot;test&quot;].shuffle(seed=42).select(range(100))<br># Print a sample tokenized entry<br>print(&quot;Tokenized Sample:&quot;)<br>print(small_train_dataset[0])</pre><h3>5. What is LoRA (Low-Rank Adaptation)?</h3><p><strong>LoRA (Low-Rank Adaptation)</strong> is a technique designed to make fine-tuning large models like DeepSeek LLM more memory-efficient by:</p><ul><li>Freezing the majority of the model’s weights.</li><li>Introducing low-rank trainable matrices in key layers (e.g., attention layers).</li></ul><p>This drastically reduces the number of trainable parameters while preserving the model’s performance. LoRA enables fine-tuning large language models on resource-constrained hardware (e.g., Colab GPUs).</p><h3>How LoRA Works</h3><ol><li>Decomposes parameter updates into low-rank matrices:<br>Where and are low-rank matrices.</li><li>Applies updates only to the decomposed matrices (e.g., attention projections).</li><li>Reduces memory and computational cost compared to full fine-tuning.</li></ol><h3>6. Code Walkthrough: Fine-Tuning DeepSeek LLM</h3><h3>Set Training Parameters</h3><pre>from transformers import TrainingArguments, Trainer<br><br>training_args = TrainingArguments(<br>    output_dir=&quot;./results&quot;,<br>    evaluation_strategy=&quot;epoch&quot;,<br>    learning_rate=3e-4,  # Lower learning rate for LoRA fine-tuning<br>    per_device_train_batch_size=1,  # Reduce batch size for memory efficiency<br>    gradient_accumulation_steps=8,  # Simulate larger batch size<br>    num_train_epochs=0.5,<br>    weight_decay=0.01,<br>    save_strategy=&quot;epoch&quot;,<br>    logging_dir=&quot;./logs&quot;,<br>    logging_steps=50,<br>    fp16=True,  # Mixed precision training<br>)</pre><h3>Initialize Trainer</h3><pre>trainer = Trainer(<br>    model=model,<br>    args=training_args,<br>    train_dataset=small_train_dataset,<br>    eval_dataset=small_test_dataset,<br>)<br>print(&quot;🚀 Trainer Initialized!&quot;)</pre><h3>Start Fine-Tuning</h3><pre>print(&quot;🚀 Starting Fine-Tuning...&quot;)<br>trainer.train()</pre><h3>Save the Fine-Tuned Model</h3><pre>trainer.save_model(&quot;./fine_tuned_deepseek&quot;)<br>tokenizer.save_pretrained(&quot;./fine_tuned_deepseek&quot;)<br>print(&quot;✅ Fine-Tuned Model Saved Successfully!&quot;)</pre><p>Checkout the google colab to get hands on practice</p><p><a href="https://colab.research.google.com/drive/1jyEhPnTf2aHAn0jttQ99M2fCaszoZBDU?usp=sharing">Google Colab</a></p><h3>7. Next Steps</h3><ul><li>Experiment with larger datasets for production-level training.</li><li>Explore more advanced LoRA configurations for efficient scaling.</li></ul><p>In the next article, we’ll explore how <strong>DeepSeek LLM</strong> can revolutionize <strong>e-commerce and retail</strong>. From personalizing product recommendations to generating engaging marketing content, we’ll dive into real-world use cases and practical examples. Learn how this open-source powerhouse can enhance customer experiences, optimize business operations, and drive growth in the competitive retail landscape.</p><p>Stay tuned for actionable insights and code walkthroughs to harness the potential of DeepSeek LLM in your e-commerce and retail projects! 🚀</p><h3>References for the Article:</h3><ol><li><strong>DeepSeek LLM:</strong>GitHub Repository: <a href="https://github.com/deepseek-ai/DeepSeek-LLM">DeepSeek LLM</a></li><li><strong>LoRA (Low-Rank Adaptation):</strong>Hu, Edward J., et al. “LoRA: Low-Rank Adaptation of Large Language Models.” <a href="https://arxiv.org/abs/2106.09685">arXiv preprint arXiv:2106.09685</a></li><li><strong>IMDB Dataset:</strong>Hugging Face Dataset: IMDB</li><li><strong>BitsAndBytes (Quantization):</strong>Official Repository: <a href="https://github.com/TimDettmers/bitsandbytes">BitsAndBytes GitHub</a></li></ol><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=12a7e5572fa5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Getting Started with DeepSeek LLM: A Hands-On Guide]]></title>
            <link>https://abhishek-maheshwarappa.medium.com/getting-started-with-deepseek-llm-a-hands-on-guide-0ddae3d6d667?source=rss-cf986f69de56------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ddae3d6d667</guid>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[deepseek-v3]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Abhishek Maheshwarappa]]></dc:creator>
            <pubDate>Wed, 15 Jan 2025 16:02:06 GMT</pubDate>
            <atom:updated>2025-01-15T16:02:06.965Z</atom:updated>
            <content:encoded><![CDATA[<p>· <a href="#7546">Introduction</a><br>· <a href="#31f8">Installing Dependencies</a><br>· <a href="#4262">Running DeepSeek LLM on Google Colab</a><br>· <a href="#f5b1">Loading DeepSeek LLM</a><br>· <a href="#935e">Running Inference with Basic Text Generation</a><br>· <a href="#9e6a">Experimenting with Different Prompts</a><br>· <a href="#823f">Summarization Example</a><br>· <a href="#3a78">Code Generation Example</a><br>· <a href="#79b4">Creative Writing Example</a><br>· <a href="#3c3b">Next Steps</a></p><h3>Introduction</h3><p><a href="https://github.com/deepseek-ai/DeepSeek-LLM">DeepSeek LLM</a> has emerged as a powerful open-source language model, offering researchers, developers, and businesses the flexibility to fine-tune and experiment without restrictions. In this guide, we will walk through the step-by-step process of setting up DeepSeek LLM, installing dependencies, running basic inference, and exploring different prompts to understand its capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Oj2ltVuG-qXZv703.png" /><figcaption>Source: DeepSeek AI</figcaption></figure><p>If you prefer hands-on learning over theory, jump straight into the code with Colab!</p><p><a href="https://colab.research.google.com/drive/1EhYQ4xMB2J-AcdsEfzW-5CdDyyBko6PF?usp=sharing">Google Colab</a></p><h3>1. Installing Dependencies</h3><p>To begin, ensure you have the following installed:</p><ul><li>Python 3.8+</li><li>PyTorch (CUDA-enabled for GPU acceleration)</li><li>Hugging Face Transformers library</li><li>accelerate for optimized inference</li><li>Google Colab or a local machine with an NVIDIA GPU</li></ul><p>Run the following command to install the necessary libraries:</p><pre>pip install torch transformers accelerate</pre><h3>2. Running DeepSeek LLM on Google Colab</h3><p>If you don’t have a powerful local GPU, you can run DeepSeek LLM on <a href="https://colab.research.google.com/"><strong>Google Colab</strong></a> with a free T4 GPU:</p><ol><li>Go to <a href="https://colab.research.google.com/">Google Colab</a>.</li><li>Select Runtime &gt; Change runtime type &gt; GPU.</li><li>Run the installation commands above to set up the environment.</li></ol><h3>3. Loading DeepSeek LLM</h3><p>DeepSeek LLM is available on <strong>Hugging Face</strong> and can be easily loaded with the transformers library.</p><pre>from transformers import AutoModelForCausalLM, AutoTokenizer<br>import torch</pre><pre>model_name = &quot;deepseek-ai/deepseek-llm-7b&quot;<br>tokenizer = AutoTokenizer.from_pretrained(model_name)<br>model = AutoModelForCausalLM.from_pretrained(model_name, device_map=&quot;cuda&quot;)</pre><h3>4. Running Inference with Basic Text Generation</h3><p>Once the model is loaded, you can generate text using a simple inference script:</p><pre>input_text = &quot;What are the key advancements in DeepSeek LLM?&quot;<br>inputs = tokenizer(input_text, return_tensors=&quot;pt&quot;).to(&quot;cuda&quot;)<br>outputs = model.generate(**inputs, max_length=150)<br>print(tokenizer.decode(outputs[0], skip_special_tokens=True))</pre><h3>5. Experimenting with Different Prompts</h3><p>To explore the model’s capabilities, try different types of prompts:</p><h3>Summarization Example</h3><pre>input_text = &quot;Summarize the importance of open-source AI models.&quot;<br>inputs = tokenizer(input_text, return_tensors=&quot;pt&quot;).to(&quot;cuda&quot;)<br>outputs = model.generate(**inputs, max_length=100)<br>print(tokenizer.decode(outputs[0], skip_special_tokens=True))</pre><h3>Code Generation Example</h3><pre>input_text = &quot;Write a Python function to calculate Fibonacci numbers.&quot;<br>inputs = tokenizer(input_text, return_tensors=&quot;pt&quot;).to(&quot;cuda&quot;)<br>outputs = model.generate(**inputs, max_length=200)<br>print(tokenizer.decode(outputs[0], skip_special_tokens=True))</pre><h3>Creative Writing Example</h3><pre>input_text = &quot;Compose a short science fiction story about AI and the future.&quot;<br>inputs = tokenizer(input_text, return_tensors=&quot;pt&quot;).to(&quot;cuda&quot;)<br>outputs = model.generate(**inputs, max_length=300)<br>print(tokenizer.decode(outputs[0], skip_special_tokens=True))</pre><p>By experimenting with various inputs, you can better understand how DeepSeek LLM handles different tasks and where it excels.</p><p>Google Colab to play with</p><p><a href="https://colab.research.google.com/drive/1EhYQ4xMB2J-AcdsEfzW-5CdDyyBko6PF?usp=sharing">Google Colab</a></p><h3>Next Steps</h3><p>Now that we have DeepSeek LLM set up and have explored its basic capabilities, the next step is <strong>fine-tuning it for specific tasks</strong>. Stay tuned for the next article, where we’ll explore <strong>Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO)</strong> to align DeepSeek with domain-specific requirements.</p><h3>References &amp; Citations</h3><ul><li>DeepSeek LLM Research Paper: <a href="https://arxiv.org/abs/2401.02954">DeepSeek LLM: Scaling Open-Source Language Models with Longtermism</a></li><li>OpenAI Scaling Laws: Kaplan et al., 2020, <em>Scaling Laws for Neural Language Models</em> (<a href="https://arxiv.org/abs/2001.08361">arXiv</a>)</li><li>LLaMA-2 Performance Benchmarks: Meta AI, 2023, <em>LLaMA-2: Open Foundation and Fine-Tuned Chat Models</em> (<a href="https://arxiv.org/abs/2307.09288">arXiv</a>)</li></ul><p>Stay on the cutting-edge of AI! 🌟 Follow me on Medium, connect on <a href="https://www.linkedin.com/in/abhishek-g-m/">LinkedIn</a>, and explore latest trends in AI technologies and models. Dive into the world of AI with me and discover new horizons! 📚💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ddae3d6d667" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>