Matt Granmoe’s Post

Automated agents are going to be a big part of the future of AI. With larger context windows of LLMs it is more and more possible. I have written a full fledged application on how we can use OpenAI’s APIs and automation tools like Playwright and Selenium to create an agent. https://lnkd.in/gmkRb2MG

Companies MUST learn from OpenAI before they are all swallowed up in the AI wave of innovation. It took me 30 minutes to build and launch a GPT in the GTP Store: See video here ---> (https://lnkd.in/gPQdBm8f). When I first applied for the OpenAI API key, I just got it. 👉 Other app stores its a nightmare - most approvals are automated biased algorithms that discriminate against people like us. Even today I have been rejected on the LinkedIn API - the reason: "unknown" 👉 Meta - you must supply ID, business documents and DNA sample to be rejected 5 times and approved the 6th time you submit the same documents. 👉 Twitter: Abandoned since they started charging $1000 for the API 😳 Other app stores just get me pulling my hair, they want screenshots, videos, documents, declarations 🤧 Honestly as a developer we spend more time dealing with approvals for apps that actually building apps, which make no sense at all 🤷♀️ and as a black developer, half your apps will be rejected for no reason. It makes more sense to build in safety control measures within the API instead of making people sign useless declarations.

I've successfully integrated OpenAI with my Spring Boot project, unlocking powerful AI capabilities within our application. By seamlessly combining OpenAI's cutting-edge technology with Spring Boot's robust framework. This integration empowers to deliver personalized user experiences, improve efficiency, and stay future-ready in the ever-evolving landscape of technology. Dive into my GitHub repository to explore the code and join the conversation on how AI is shaping the future of software development! #OpenAI #SpringBoot #AIIntegration #GitHub #Java https://lnkd.in/gyk9PduJ

This weekend, I got hands-on with Gemma.cpp, Google's lightweight C++ inference engine that lets me run the the impressive open source Gemma model family from Google DeepMind locally on my machine! I'm thrilled by the potential to run advanced AI models locally, prioritizing privacy and opening up new possibilities for development. 🔐 It's like having a powerful AI playground right on my machine! Here's how I got started: Setting Up: I used my local workstation running Ubuntu 22.04. 🐧 Cloning and Building: I cloned the Gemma.cpp repository with this command: git clone https://lnkd.in/g3eebHwW Selecting the Right Model: Kaggle provides access to various Gemma models. I opted for the 2b-IT-SFP model, leveraging 8-bit switched floating point to optimize performance without excessive resource demands. ⚖️ Preparing the Weights: After downloading the pre-trained model weights, I untarred them within the Gemma.cpp build directory. 📦 Then, I compiled and built the gemma application using CMake and Clang - a surprisingly straightforward process. 🔨. I navigated to the build directory and ran the following commands: cmake .. make -j8 gemma Running Inference: The magic happened when I executed this command: ./gemma --tokenizer tokenizer.spm --compressed_weights 2b-it-sfp.sbs --model 2b-it This command brought me straight to the prompt with gemma.cpp in ascii art, allowing me to experiment with Gemma models right on my own machine. 🪄 Why does this matter? Consider these possibilities: Offline Functionality: Imagine AI-powered applications that function seamlessly even without an internet connection (for example when you are flying). ✈️ Fast AI Experimentation: Gemma.cpp opens doors for experimenting with powerful language models locally, fostering creativity and innovation. 🧪 I'm excited to see how others leverage Gemma.cpp and Google's Gemma models 🤝 #Google #Gemma #GemmaCpp #GoogleAI

Some nerdy Friday evening fun: How many of you know (really) what a token is? Maybe you think, easy: Tokens are pieces of words. Well, yes, but what kind of pieces and why? Despite the widespread use of the term 'tokens' in the AI community, many seem to lack understanding of its true nature. Let's shed some light on this. The initial trigger for working with tokens in LLM (around 2016) was the Byte Pair Encoding (BPE) algorithm initially designed for data compression (1994): Input: arbitrary text with words, numbers, whitespace, and punctuation. Output: list of tokens representing the same text, where each token is an integer identifier that can be looked up in a vocabulary to reproduce the input text. And a 100K token model just means it's using a dictionary with 100K integers entries, each representing a piece of a word/text. Here are 4 lines from OpenAI's dictionary: IEdOVU5FVA== 100199 aWRVc2Vy 100200 IHN1cGVydmlzZQ== 100201 KENvbnRhY3Q= 100202 Any LLM input gets tokenized because, with such numbers, it is simpler to calculate. Let's make an example - The tagline of my company is: Saphirion - We make supply chain prices predictable. Which are 52 characters and translate to 12 tokens. With OpenAI's tokenizer you can see how a text is broken down into tokens: https://lnkd.in/dayiSTik

🚀 Introducing FLUX.1 by Black Forest Labs 🚀 Explore the cutting-edge FLUX.1 models designed for diverse applications: FLUX.1 [pro]: Top-tier performance for state-of-the-art image generation. Available via API and partners like Replicate and fal.ai. Perfect for enterprise solutions. FLUX.1 [dev]: Open-weight model for non-commercial use. Efficient and high-quality, available on HuggingFace, Replicate, and fal.ai. FLUX.1 [schnell]: Fastest model for local development and personal use. Openly available under Apache 2.0 license on Hugging Face and GitHub. Check out the video to see FLUX.1 in action! 🎥 💬 We’d love to hear your thoughts! Leave a comment below and let us know which FLUX.1 model you’re most excited about. For more information, visit beyondaisolutionsllc.com or contact me at Rome@beyondaisolutionsllc.com

Excited to share our new, incredibly easy to use GitHub action, which will run your evals on every PR, and cache repetitive LLM calls for you automatically using our proxy. (link in comments) I highly encourage every AI team to run evals on EVERY commit. Unless you do so, you leave yourself vulnerable to regressions that slip in between the times you run evals ad-hoc. Of course, there are engineering and cost challenges associated with this. For example, you don't want to blow up your OpenAI bill running evals. That's why the action, by default, uses our proxy which end-to-end encrypts LLM calls and caches them for you. Expect more iteration on this as we figure out what works with our users. I expect a lot of folks to switch over since this is just so much easier than running evals manually.

Exciting Update on StudyBuddy 📚 🔍 Last week, I dived into techniques to boost #RetrievalAugmentedGeneration systems in my article, "Improving RAG AI Systems." 🤖🌐 Check it out here: https://lnkd.in/dxWCTC_t Now, let's talk about the latest StudyBuddy update! 🌟 I've implemented the Query Expansion with Multiple Queries technique to enhance StudyBuddy's accuracy in answering your questions. 💬✨ Leveraging LangChain's "MultiQueryRetriever" class and some custom prompting, I've fine-tuned the question outputs that are used for the relevant document retrieval process. 🛠️💡 Now, when you input a query, the RAG system generates three additional similar queries which when used to retrieve documents from the Pinecone vector database, create a comprehensive knowledge base for a more robust answer! 🤯🔗 Explore StudyBuddy on the public repo: https://lnkd.in/dvy9r6Yc 🚀📖 Your comments and feedback are always valued! 🙌🏽 💬 #AIInnovation #ArtificialIntelligence #StudyBuddy #OpenSourceAI

https://lnkd.in/ge82G_Da An absolute game changer for AI data structuring. You can now provide a Pydantic or Zod schema and OpenAI’s JSON response will adhere to the provided typings with a purported 100% accuracy… amazing! #openai #nodejs

🚀 Exciting news for data analysts and programmers! 📈 Google Colab has introduced a new feature that significantly enhances the performance of Pandas code. This feature can even accelerate the pandas code/queries generated by Colab AI, ChatGPT or any LLM-based chatbot, without needing to learn new paradigm. Here’s a quick summary of what this feature entails: ➡ Performance Boost: Pandas operations are now 50 times faster on Google Colab without any code modifications. ➡ RAPIDS cuDF Integration: Google Colab has integrated RAPIDS cuDF, which enables GPU acceleration for Pandas, making data processing tasks much quicker. ➡ Easy Activation: To activate this feature, users with NVIDIA GPU-enabled Colab notebooks can simply add the following command at the beginning of their notebook: ➡ COMMAND: %load_ext cudf.pandas This command loads the cudf.pandas extension, allowing users to take advantage of GPU acceleration for their Pandas code. It’s a straightforward way to enhance the performance of data manipulation tasks in Colab notebooks. Check out the link below to learn more about this exciting feature and try it out for yourself! 🔗NOTEBOOK: cudf_pandas_colab_demo.ipynb ⬇ https://lnkd.in/gCQ3Vz7U #Google #AI #Programming #Code #Python #LinkLayer