🚀 Mikkel Jensen’s Post

Sick in bed. Done with work for the day. What to do? Code more on my own projects obviously. I am not only sick with the flu but I am dealing with my long term Codeholic tendencies. I keep promising my wife I will spend less time coding after work but I can't help myself. But for reals I am trying to build an AI Gmail responder that is trained on your own emails and still working on the email extractor, training data generator. Public gitHub repo to come along with a video and blog post. With much more details on the process and how you can replicate it yourself.

A nanoGPT pipeline packed in a spreadsheet https://lnkd.in/ePfa-wpR #learning #LLMs #GPT #spreadsheet

The most effective way to learn to code is by practicing. One of the best approaches I've come across is to study the entire codebase and simplify complex classes and functions for better understanding. The focus here is not on efficiency but on gaining a thorough knowledge of how things work. You should aim to be proficient enough to explain concepts to others or provide advice when needed. How to Start? Let's say for example you are intrigued with Contrastive Language-Image Pre-training (CLIP). 1. Read the Paper: Begin by reading the research paper from start to finish to understand the key concepts. A solid mathematical understanding is essential. 2. Understand the Diagram: Interpret the diagrams in the paper and make sure you understand them well. 3. Find Open-Source Implementations: Check GitHub for any open-source implementations of the paper. If you find one, start by reading the `README.md` file to get a high-level understanding of the repository. 4. Locate the Training Script: The training script is crucial as it contains the core components and the model to be trained. For example:   `from open_clip import get_input_dtype, CLIP, CustomTextCLIP` 5. Locate the Model: Find where the model is defined in the repository. If it’s not clear, check the `__init__.py` file, which often imports key components without specifying the source module. For instance: `from .model import CLIP` 6. Study the Source Script: Once you locate the source script of the CLIP class, study it thoroughly. Note the various functions and classes it uses. 7. Create a Flowchart: Visualize the workflow and start working on it accordingly. Three Key Pieces of Advice: 1. Keep the Paper Handy: Always have the experiment paper open next to you. Observe, understand, and learn how to implement formulas and equations. 2. Work with a Strategy: Don't work aimlessly; have a clear plan. 3. Maintain a Clear Vision and Consistency: Stay focused on your goal and be consistent in your efforts.

🚀 Discover the power of RAG-nificient Styles! 🚀 Looking for a tool that makes style guides easy to navigate and super accessible? Check out RAG-nificient (https://lnkd.in/eBvR7aKg), a repository that harnesses the potential of retrieval-augmented generation (RAG) to deliver instant, relevant answers from your guides. No more time wasted searching! 🔍 🔧 Key Features: - Full Docker Compose setup for quick deployment - Postgres and Supabase logging for interaction tracking - Qdrant vector database for fast, smart querying - Integrates Ollama models for flexible generative AI ✨ Get up and running in no time with a comprehensive ingestion and API setup, plus a sleek Streamlit front end to test your queries. Don’t miss out on leveling up your productivity! 💼 Also, if you're interested in expanding your knowledge on LLMs, join the LLM Zoomcamp by DataTalksClub! This course offers hands-on learning to build your own AI-powered solutions, just like RAG-nificient. #llmzoomcamp #LLM #datascience Ready to explore? Let’s go!

This one has been a regular question - from clients, VCs and friends who have been interested in Gen AI Yes - in most cases where you need to leverage the ‘langauge’ part of LLMs a fine tuned model works great. Llama, mixtral and a bunch of open source so a great job heavy lifting language generation loads. And then there are cases where these admittedly amazing models, trained on generic data lack the finesse to execute special tasks. They pretend to do it - but you’ll see a gap. That is where pre training comes in - your data, your strategy and your AI. Compute costs are high today - and there is a general view that LLMs cost hundreds of millions to build. That will change - smaller models (like tinyllama, phi 1.3 and Pints.ai 1.5) will evolve - and compute costs will decline. And organizations will take charge of their AI infrastructure by leveraging large proprietary models and augment it with private models trained and exposed to strategic and confidential data.

Using useReducer + useContext for state management has a mild learning curve. Sharing my experience of implementing this method here.

Incremental Reading • Incremental reading is a learning technique that involves importing articles, extracting important fragments, converting them into questions and answers, and reviewing them using spaced repetition. • It allows efficient processing of large amounts of information by focusing on the most important parts and skipping less relevant sections. • The five basic skills of incremental reading include importing articles, reading and decomposing them, converting key information into questions and answers, reviewing the material, and handling information overflow. • Importing articles can be done through copy and paste, mass import from supported browsers, dedicated imports from sources like Wikipedia and YouTube, local file imports, and mail imports. • Incremental reading promotes efficient learning, better retention, improved comprehension, and increased creativity, among other advantages.

𝐶𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦 𝑖𝑠 𝐾𝑒𝑦: 𝐿𝑒𝑡'𝑠 𝑀𝑎𝑠𝑡𝑒𝑟 𝑅𝑒𝑎𝑐𝑡 𝑇𝑜𝑔𝑒𝑡ℎ𝑒𝑟 💪 Remembering all the topics we learn can be challenging. Often, we study 30 topics in a month but forget the first one by day 30 🤯. This happens because we focus on memorizing rather than understanding. To truly grasp knowledge 📕, especially for something as vital as React, we need to delve deeper. That’s why we’re taking a Deep Dive into React 🏃♂️, exploring each topic comprehensively. 𝐃𝐚𝐲 𝟕: 𝐮𝐬𝐞𝐑𝐞𝐟 𝐇𝐨𝐨𝐤 ⚓ After exploring the useEffect hook, today we're diving into another powerful but often underutilized hook: 𝑢𝑠𝑒𝑅𝑒𝑓. The 𝑢𝑠𝑒𝑅𝑒𝑓 hook is used to store mutable data without re-rendering the component. So, if you want to store state but there is no need to re-render the UI, you can use this. In simple words, it can replace useState Hook in some cases and can improve the efficiency of your code along with some other use cases as well. Here are some use cases for useRef: • DOM reference • Storing mutable state • Storing previous state In the video below, I demonstrated all three use cases: 𝐒𝐭𝐨𝐫𝐢𝐧𝐠 𝐏𝐫𝐞𝐯𝐢𝐨𝐮𝐬 𝐒𝐭𝐚𝐭𝐞: We used prevState ref to save the previous value of input element. You can see that the previous value is saved and updated correctly. 𝐃𝐎𝐌 𝐑𝐞𝐟𝐞𝐫𝐞𝐧𝐜𝐞: When we click the "Focus Input" button, the input field gains focus. This is an example of using useRef to reference a DOM element. 𝐂𝐨𝐮𝐧𝐭𝐢𝐧𝐠 𝐑𝐞𝐧𝐝𝐞𝐫𝐬: We count the renders of the component when typing in the input. Here, useRef is used to keep track of the render count, while useState (i.e., name) increases the render count. 𝑁𝑜𝑡𝑒, 𝑡ℎ𝑒 𝑟𝑒𝑓 𝑖𝑡𝑠𝑒𝑙𝑓 𝑤𝑜𝑛'𝑡 𝑡𝑟𝑖𝑔𝑔𝑒𝑟 𝑎 𝑟𝑒-𝑟𝑒𝑛𝑑𝑒𝑟. 𝑊𝑒 𝑠𝑒𝑒 𝑡ℎ𝑒 𝑢𝑝𝑑𝑎𝑡𝑒𝑑 𝑣𝑎𝑙𝑢𝑒 𝑖𝑛 𝑡ℎ𝑒 𝑈𝐼 𝑏𝑒𝑐𝑎𝑢𝑠𝑒 𝑤𝑒 𝑎𝑟𝑒 𝑎𝑙𝑠𝑜 𝑢𝑠𝑖𝑛𝑔 𝑢𝑠𝑒𝑆𝑡𝑎𝑡𝑒. By efficiently using these hooks, we can make our code much more efficient and maintainable. This was all about the useRef hook. 𝑆𝑡𝑎𝑦 𝑐𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑡 𝑎𝑛𝑑 𝑘𝑒𝑒𝑝 𝑙𝑒𝑎𝑟𝑛𝑖𝑛𝑔! 𝐿𝑒𝑡'𝑠 𝑚𝑎𝑠𝑡𝑒𝑟 𝑅𝑒𝑎𝑐𝑡 𝑡𝑜𝑔𝑒𝑡ℎ𝑒𝑟. 🚀

𝐓𝐢𝐩𝐬 𝐨𝐧 “𝐇𝐨𝐰 𝐜𝐚𝐧 𝐈 𝐢𝐦𝐩𝐫𝐨𝐯𝐞 𝐦𝐲 𝐋𝐨𝐠𝐢𝐜-𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐬𝐤𝐢𝐥𝐥𝐬?” 💻 . “Ye Logics nhi banti mere se bass, Baqi syntax sara ata hai mujhe” 😅 … I know that you all have this feeling once or may you have now in your life. But don’t worry, this happens when you are at the beginning stage of anything that you know nothing about. 💪 You can follow these Tips to better your Logic-Building skills. 𝗦𝘁𝗲𝗽 𝟭. 𝗗𝗮𝘁𝗮 𝗳𝗹𝗼𝘄: 𝘜𝘯𝘥𝘦𝘳𝘴𝘵𝘢𝘯𝘥 𝘵𝘩𝘦 𝘥𝘢𝘵𝘢 𝘷𝘢𝘭𝘶𝘦𝘴 𝘢𝘵 𝘦𝘢𝘤𝘩 𝘴𝘵𝘦𝘱 Try to dry run your code at each or different steps to check if it is working correctly or not and you can also analyze how your for loop is working and what it is lacking. Today “if-else” conditions always seem easy to handle and implement but it matters a lot in building foundation logic and a lot of people find it difficult to use them in different loops. At this step, just try to understand the flow. Don’t worry about logic, as these are important in building complex functions in the future. 𝗦𝘁𝗲𝗽 𝟮. 𝟵𝟬-𝟭𝟬: 𝘛𝘦𝘢𝘤𝘩𝘪𝘯𝘨 𝘔𝘦𝘵𝘩𝘰𝘥𝘰𝘭𝘰𝘨𝘺 Teaching is different from the YouTube learning method. In teaching, there is a methodology in which a Teacher does 90% of the whole work and assigns 10% of the work to students in the form of assignments. With this methodology, a base for Students is set for learning and understanding the concepts behind logic-building. This later becomes 80:20, then 70:30, then 60:30 until students can solve problems on their own. Unfortunately, YouTube Tutorial Hell makes it impossible for beginners to best their logic-building rather than just copying the YouTuber’s steps and methodologies. 𝗦𝘁𝗲𝗽 𝟯. 𝗨𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱 𝗧𝗢𝗗𝗢 𝗣𝗥𝗢𝗝𝗘𝗖𝗧𝗦 𝘁𝗵𝗼𝗿𝗼𝘂𝗴𝗵𝗹𝘆 Foundation is important and try to take it seriously. Don’t directly try to do and achieve big milestones when you don’t have sufficient knowledge about that specific thing. You can try to work on these functions (Create,React,Update,Delete,Search,Filter) at first and grab a strong hold on building fundamental logic concepts. 𝗦𝘁𝗲𝗽 𝟰. 𝗔𝗱𝗱𝗲𝗱 𝗩𝗮𝗹𝘂𝗲𝘀: 𝘈𝘥𝘥 𝘴𝘰𝘮𝘦 𝘷𝘢𝘭𝘶𝘦 𝘰𝘯 𝘺𝘰𝘶𝘳 𝘰𝘸𝘯 You always need some information or overview to start and you may need to know what to do and how to do a thing. For that purpose, you can watch a project-based tutorial video. Then try to Add some features from your own, no matter if it’s small or big. When you face these challenges all on your own, then this will boost your logic-building. image from: Rakesh Kumar R #SoftwareEngineering #ComputerScience #Programming #TechCareer #LogicBuilding #HigherEducation #CareerAdvice #TechDebate