Ankur Singh’s Post

It's amazing to see how the rise of Large Language Models (LLMs) is changing the way we code and the tools we choose. For example, I have started writing doc strings and comments so that I can get better copilot recommendations, Unit test generation, and also help LLMs correctly choose the tool during function calling. Furthermore, I've noticed a trend where more developers are picking tools that play well with LLMs. For instance, like: - Pandas over Polars or DuckDB, - Elastic Search over other full-text search engines. - SQL over custom query languages like PQL. Why? These established tools have: - Longer market presence - Abundant StackOverflow threads - Stable APIs As a result, LLMs excel at understanding and generating code for these tools, making them more attractive in our AI-assisted development workflow. What are your thoughts on this emerging pattern? Have you noticed similar trends in your development environment? #AI #SoftwareDevelopment #LLM #CodingTrends

IMO, NexusRaven-V2-13B is a game-changer in the world of function-calling LLMs! 🌟 Here's why it's the best: - Unmatched Function Calling Capabilities: Supports Single Call, Parallel Calls, No Call, Multiple Functions, Nested Functions, and JSON schemas effortlessly. - Surreal Developer Experience: Say goodbye to formatting hassles and fiddling with system prompts. - Open Source and Lightweight (13B) All the good stuff with none of the bad! NexusRaven-V2-13B is truly a breakthrough, making development smoother and more efficient. Model Card: https://lnkd.in/gTergUKt Check the comment section for helpful resources to get started. Also, feel free to connect and share your thoughts! 👇💬 #AI #MachineLearning #OpenSource #DeveloperExperience #NexusRavenV213B #Innovation #TechNews

📢"Part 2: Function Calling" is out! Function calling is cool, but why should only OpenAI have all the fun? Learn how to recreate function calling APIs using open-source models like Hermes-Pro-2 (fine-tuned for function calling) and Llama-3-8B-instruct (no function calling). You will be surprised to see how easy it is to implement function calling. All you need is a model that excels at following instructions. Check out the blog for details – link in the comment below. #FunctionCalling #OpenSource #AI #MachineLearning #HermesPro2 #Llama3 #TechInnovation #OpenAI #agents

🚀📢 New Blog Post : Structured Output from LLMs Excited to share my latest blog post, where I dive deep into the world of structured output from large language models (LLMs). Discover all the tips and tricks you need to build robust and reliable systems. 🛠️💡 What you'll find inside: - Steps to get structured output from LLM - Hidden gems and resources for deeper understanding - Practical examples to enhance your projects Stay tuned for my upcoming blog on function calling in LLMs, coming out in a few days! 📅🔜 Don't miss out! Check out the blog in the comment section below. Let me know if I missed anything or if you have any feedback. Happy reading! 📖✨ #AI #MachineLearning #LLM #StructuredOutput #TechBlog #FunctionCalling #StayTuned #Innovation #AIResearch #NewPost

🚀 Excited about the Agentic AI? You should be! Check out this lightning talk by Andrew Ng showcasing Vision Agent. 🌟 It's a glimpse into the future of application development that's truly inspiring. 🔗 Link to the video in the comments. #AI #MachineLearning #Innovation #TechTalk #FutureOfAI #AndrewNg #VisionAgent #ApplicationDevelopment #AgenticAI

🚀 The AI landscape is rapidly evolving, and here are some key trends to watch: 👉Inference Hardware Revolution: While GPUs excel in training and batch jobs, latency is crucial for customer-facing applications. Expect to see more dedicated, mainstream hardware, like NPUs, optimized for inference from all chip manufacturers. 👉Unified Multi-modal LLMs: Current multi-modal LLMs often rely on separate encoders for each modality. However, just as transformers have replaced CNNs and RNNs, an unstructured approach is likely to prevail. Models like Gemini, GPT-4o, and Meta's Chameleon are leading the way by being natively multimodal and trained end-to-end across various modalities. 👉The Complexity of RAG: RAG's promise of natural language querying is challenging to achieve with just a simple retriever and vector database. Our software's nuances require a sophisticated approach. Integrating agents into RAG systems will become essential, and soon, you won't see RAG systems without them. Stay tuned as these developments reshape the AI field! 🌟 #AI #MachineLearning #DeepLearning #Inference #MultiModalAI #RAG #Innovation

Fine-tuning vs RAG is an interesting debate. There are strong arguments for both. But it's important to note that isn't an either-or problem. Ideally, in most cases, you should be using both. Here's are my two cents: - Unless privacy is a concern, avoid fine-tuning. - Queries about your data, use RAG + Agents. - With larger context windows, prefix caching, and low latency, most motivations for LLM fine-tuning will be addressed. Read the blog the details (link in comment) on why I think so and the mental model I use to when thinking about fine-tuning and RAG. PS: Emmanuel Ameisen's talk “Why Fine-Tuning is Dead?” inspired me to write this piece. #llm #finetuning #rag

🚀 Gradio is evolving into what Cog (by Replicate) and Jina (by Jina AI) aimed to be: a solution for packaging ML models as production-ready services. It's not just a UI library! Gradio offers features like: - Authentication - Batch/Stream prediction - Concurrent request handling - Job queue - API Endpoints (with Docs) Having used Cog and Jina, I can vouch for their capabilities. However, Gradio fills a crucial gap: Interactive UI. ML is inherently experimental, and sharing an app with stakeholders for vibe checks, debugging, and Q&A is invaluable. Gradio's intuitive UI, minimal code requirements, and automatic generation of production-grade APIs and docs—with streaming, queuing, and concurrent request support—are game-changers. This setup suffices for most industry use-cases. The latest release amplifies this further with enhanced Python and JS client libraries, making API endpoint calls seamless. Monitoring is now integrated, and the new `render` decorator enables dynamic layouts. #MachineLearning #AI #Gradio #TechInnovation #MLops #GenAI #HuggingFace #Cog #Jina #PrototypeToProduction

In case you are finetuning LLMs on custom data, you check this out. GGUF format is 🔥, exporting your model to GGUF ensures easy compatibility with various tools in the ecosystem. So, instead of wasting time hunting for the steps each time, I decided to jot them down for future me. If it helps others too, bonus! Link: https://lnkd.in/gPfgKxse #LLM #GenAI #Axolotl #llamacpp #ollama #finetuning