AI Engineer Circle

Set of companies that are now hiring for remote positions in AI space. #ai #generativeai #jobs #remotework

Interested in starting your career in AI? Don't miss the deadline to apply for the AWS AI Scholarship ($4,000 value) Students will get hands-on AI experience with real-world projects, learn from industry mentors, access career resources. 𝗦𝘁𝗲𝗽 𝟭: Sign up for the AWS DeepRacer Student competition and opt into the AWS AI & ML Scholarship. Applications are open now. 𝗦𝘁𝗲𝗽 𝟮: Successfully complete the requirements of the DeepRacer Student League to receive a unique application code for a Udacity nanodegree. Deadline: September 30, 2024. 𝗦𝘁𝗲𝗽 𝟯: Get a scholarship for the 4-month "AI Programming with Python" Nanodegree program ($4,000 USD value). Program starts on October 14, 2024. ➡️ Apply here: go.aws/3YnIieK Open to students around the world. No previous AI knowledge is required. Share this scholarship opportunity with your friends! ——— Follow Viktoria Semaan for more AWS & AI insights and opportunities #ai #scholarship AWS Training & Certification

Exciting news from OpenAI! Open AI announced their new model GPT-4o Mini! This new model is significantly smarter and cheaper than GPT-3.5 Turbo. It supports text and vision, and will soon include video and audio inputs and outputs. Featuring a 128k context window, GPT-4o Mini is a treat for developers. Performance Highlights: - Scores 82% on MMLU for textual intelligence and reasoning - Excels in math, coding, and multimodal reasoning - Surpasses smaller models like Gemini 1.5 Flash (79%) and Claude 3 Haiku (75%) Affordable Pricing: - Just $0.15/1M input tokens and $0.6/1M output tokens - Ideal for long context use cases, such as large document retrieval and generation (RAG) OpenAI has delivered a high-quality model with impressive capabilities relative to its size, making GPT-4o Mini a worthy successor to GPT-3.5 Turbo and a strong candidate for ChatGPT's free version. GPT-4o Mini also promises improved multi-lingual support and excels in reasoning and instruction-following tasks. More can be read from here https://lnkd.in/d2u5K6JY

Start your learning journey with these FREE AI courses from Microsoft! Whether you're a beginner or looking to advance your AI skills, these courses cover everything you need to get started and excel in Azure Open AI Check them out: 1. Fundamental AI Concepts: https://lnkd.in/dSbhE7C3 2. Fundamentals of Generative AI: https://lnkd.in/dVAjnymE 3. Fundamentals of Azure OpenAI: https://lnkd.in/dA6BYJTR 4. Prompt Engineering with Azure OpenAI: https://lnkd.in/dhExyuRY 5. Implementation of RAGs: https://lnkd.in/d8SD9nzm 6. Build RAG based copilot: https://lnkd.in/dpZkP7HQ 7. Create a knowledge store: https://lnkd.in/dZq5dR6X 8. Fine-tune foundational models: https://lnkd.in/dHPW6ivK 9. MLOps: https://lnkd.in/dcTntKni 10. Deploy model to NVIDIA Triton: https://lnkd.in/d6DwXV74 11. Monitoring ML models: https://lnkd.in/dBdm9z-j 12. Securing ML Models: https://lnkd.in/dq82tXQ4 🆓 These courses are FREE to take and are provided by Microsoft, so we strongly advise you to enroll and take advantage of these valuable resources. ♻ If you find this information useful, please share it with your friends and network. #ai #azureOpenAI #aiengineercircle #aiengineer

This is a huge step in generative AI. So far, there has been single-token prediction, but Meta has released a research paper on multi-token prediction [here](https://lnkd.in/dZJFSCzF). Most modern LLMs have a simple training objective: predicting the next word. While this approach is simple and scalable, it’s also inefficient, requiring several orders of magnitude more text than what children need to achieve the same degree of language fluency. In April, Meta proposed a new approach to build better and faster LLMs by using multi-token prediction. With this method, language models are trained to predict multiple future words at once instead of the traditional one-at-a-time approach. This enhances model capabilities and training efficiency while allowing for faster speeds. Today, Meta released pre-trained models for code completion under a non-commercial/research-only license, available on [Hugging Face](https://go.fb.me/dm1giu). I think this will make a huge difference in the Gen AI applications. #metaai #ai #multitokenprediction #aiengineercircle

Today's demo of Kyutai Labs' fully end-to-end audio model is a significant milestone that many overlooked. Even thiugh the demo was less polished, the product’s highlights are far more impactful: 👉Scalability and Simplicity: The model training pipeline and architecture are both straightforward and highly scalable. Kyutai's small team of just over 8 people built it in four months, leveraging synthetic data as a major enabler. 👉Local Device Focus: With a strong emphasis on local device integration, Moshi is poised to be ubiquitous soon. Unlike larger companies, nonprofits like Kyutai prioritize making smaller models run locally. The Moshi demo is already accessible online, whereas OpenAI's 4o remains in limbo. 👉Low Latency and High Quality: Achieving sub-300 ms latency while maintaining Llama 8B or higher quality responses is a game-changer for interactivity. The model can answer questions before they're fully asked and even react to interruptions—a testament to its predictive coding capabilities. Kyutai Labs has nailed the fundamentals. This interactive voice technology will soon be everywhere, becoming an essential commodity. The future of voice tech is here and it's incredibly exciting to witness. Checkout the demo here #ai #moshi #voiceai #aiemgineercircle

RAG is a significant buzzword and an essential domain to master in the field of Generative AI for any AI Engineer. Therefore, we are creating this mini-series for our community. 🆕 RAG Series - Mini Bites 01 ➡ What is RAG? The process of bringing the External information and inserting it into the model prompt is known as Retrieval Augmented Generation (RAG). ➡ RAG has 2 phases: 1. Document Loading / Indexing   - Documents are processed and loaded into a database, usually offline:     1. Load     2. Split     3. Store 2. Retrieval & Generation   - Appropriate data is retrieved and added to the context of the prompt together with the original query. ➡ Document Loading phase 1. Extract:   - Documents could be of different:     - File formats: .csv, .pdf, .doc, etc.     - Data formats: text, tables, images, movies, etc.   - These must be extracted and put into a format that can be processed for the next stages. 2. Chunking:   - The extracted text data is broken into smaller pieces of data for the next processing step.   - This increases the accuracy of the relevance of the retrieved document chunk and also will be suitable to adjust the context window of the LLM. 3. Embedding:   - Converting the chunk into a dense vector as a numerical representation of the meaning of the text in the chunk. 4. Loading/Ingestion:   - Adding the embeddings to the original data of a database. 5. Database:   - Provides storage for the embedding and data.   - Often it is a vector database due to embeddings.   - Graph databases and traditional databases are also used. ➡ Retrieval & Generation Phase 1. Query:   - Embedding: The question or request asked by the user is also converted into a dense vector, a numerical representation of the meaning of the query using the same embedding model used for embedding the documents in the first phase. 2. Retrieval:   - This is the process of finding the K entries in the database that are the closest to the query vector. The vector represents the meaning of the text, so in other words, this is getting the related document chunks from the database which are closely related to the query.   - Many ways to find the closest query, with the most famous one being cosine similarity. 3. LLM Response Generation:   - The retrieved K chunks are added to the context together with the original query of the user and this is called “Augmented”. Now the LLM will have the new retrieved information on top of the information it has been trained on.   - The LLM can now answer based on your private new data. Following image is the architecture of the simple RAG application. #RAG #GenAI #AI #AIEngineer #AIEngineerCircle

Privacy and confidentiality of the data used in RAG is a concern for many companies & individuals. Using the LLMs locally is one of the prominent solution for this issue. This repository provides an easy way to run Gemma-2 locally directly from your CLI (or via a Python library) and fast. It is built on top of the Transformers and bitsandbytes libraries. It can be configured to give fully equivalent results to the original implementation, or reduce memory requirements down to just the largest layer in the model! Try out this local Llm here. https://lnkd.in/dZwe-eQE

This is an interesting video to understand CPU & GPU

Exciting News from LlamaIndex! LlamaIndex has launched the alpha version of `llama-agents`, an open-source framework for building multi-agent AI systems. This new tool is perfect for creating complex AI applications, such as collaborative assistants and question-answering systems. Key features include: - Distributed architecture - Standardized APIs - Flexible orchestration - Easy deployment and scalability - Built-in observability tools Ready to check in? Install via pip (`pip install llama-agents`) and start building today! Learn more: https://lnkd.in/dkw2Kynd