--- title: "Know the latest buzzword before it becomes a buzzword šŸ" newsletter: "MLOps Community" date: 2024-06-27 source: https://aaif.live/newsletters/mlopscommunity/2024-06-27-know-the-latest-buzzword-before-it-becomes-a-buzzword --- # Know the latest buzzword before it becomes a buzzword šŸ *Plus, AI doing the real heavy work, more multimodal mastery, the need for speed (benchmarks), some guidance on guidance, and hidden gems.* *MLOps Community ā€” Agentic AI Foundation, 2024-06-27* ## Accelerating Multimodal AI // Ethan Rosenthal // MLOps Podcast #242 AI moves fast, but buzzwords move faster. Introducing a new one coined by Ethan: multimodal feature store. He acknowledges the buzzword bingo but explains how he sees it as a concept to handle diverse data types for AI training and address challenges faced by his team. We also chat about the ā€˜musical differencesā€™ between researchers and engineers, proposing solutions like a unified codebase for training and deployment. Plus, we cover practical strategies for ML teams, including cloud resource management and GPU optimization. Who knows, this could be the buzzword of today that becomes the future of tomorrow! Video [https://home.mlops.community/home/videos/accelerating-multimodal-ai] || Spotify [https://open.spotify.com/episode/2czL0OAmDdn53gr85RsWiM?si=rm5Gb97oTyaQpBN4QLGLlg] || Apple [https://podcasts.apple.com/us/podcast/accelerating-multimodal-ai-ethan-rosenthal-242/id1505372978?i=1000659796293] [https://podcasts.apple.com/us/podcast/accelerating-multimodal-ai-ethan-rosenthal-242/id1505372978?i=1000659796293](https://podcasts.apple.com/us/podcast/accelerating-multimodal-ai-ethan-rosenthal-242/id1505372978?i=1000659796293) [https://home.mlops.community/home/videos/accelerating-multimodal-ai](https://home.mlops.community/home/videos/accelerating-multimodal-ai) [https://open.spotify.com/episode/2czL0OAmDdn53gr85RsWiM?si=rm5Gb97oTyaQpBN4QLGLlg](https://open.spotify.com/episode/2czL0OAmDdn53gr85RsWiM?si=rm5Gb97oTyaQpBN4QLGLlg) ## CLOUD FOR AI PRACTITIONERS FROM NEBIUS https://nebius.ai/?utm_medium=email&utm_source=bay-area-june&utm_campaign=main-campaign Nebius is a cloud platform specifically designed to train AI models, providing you with the latest and greatest GPUs such as the NVIDIA H100 to manage intensive AI workloads. Features: * Built for large-scale ML workloads: Multihost training with thousands of H100 GPUs, 3.2Tb/s InfiniBand network. * 50% cheaper compared to major public cloud providers. * Fully managed Kubernetes: Simplify deployment, scaling, and management of ML frameworks. * ML-focused libraries, applications, frameworks, and tools. * Detailed documentation, resource management in a user-friendly cloud. Try Nebius now [https://nebius.ai/?utm_medium=email&utm_source=bay-area-june&utm_campaign=main-campaign] ## ML and AI as Distinct Control Systems in Heavy Industrial Settings // Richard Howes // Podcast #243 It sometimes feels like GenAI is the spotlight-grabbing Jobs to practical AIā€™s workhorse Wozniak. Thatā€™s why it was great talking to Richard about integrating AI and ML in heavy industries like oil and forestry. We discussed practical applications such as predictive maintenance, regulatory compliance, and IoT devices for equipment monitoring. We also covered control philosophies, system diagrams for stakeholder clarity, leveraging external data for resource exploration, and enhancing data scientist effectiveness through comprehensive context sharing. Next step: create a video of Wozniak eating spaghetti while monitoring pressure and temperature for low-impact drilling. Video [https://home.mlops.community/home/videos/ml-and-ai-as-distinct-control-systems-in-heavy-industrial-settings] || Spotify [https://open.spotify.com/episode/6vmaJiA3445UEJVRzCB7AY?si=p8jBhegxTkmHHT6qc3S5NQ] || Apple [https://podcasts.apple.com/us/podcast/ml-and-ai-as-distinct-control-systems-in-heavy/id1505372978?i=1000660220722] [https://podcasts.apple.com/us/podcast/ml-and-ai-as-distinct-control-systems-in-heavy/id1505372978?i=1000660220722](https://podcasts.apple.com/us/podcast/ml-and-ai-as-distinct-control-systems-in-heavy/id1505372978?i=1000660220722) [https://home.mlops.community/home/videos/ml-and-ai-as-distinct-control-systems-in-heavy-industrial-settings](https://home.mlops.community/home/videos/ml-and-ai-as-distinct-control-systems-in-heavy-industrial-settings) [https://open.spotify.com/episode/6vmaJiA3445UEJVRzCB7AY?si=p8jBhegxTkmHHT6qc3S5NQ](https://open.spotify.com/episode/6vmaJiA3445UEJVRzCB7AY?si=p8jBhegxTkmHHT6qc3S5NQ) ## Job of the Week [https://forms.gle/dmEBWvmJFAnkSyma6](https://forms.gle/dmEBWvmJFAnkSyma6) ## Beyond Text: Multimodal RAG for Video // Anup Gosavi // Community IRL Meetup #83 Silicon Valley ## MLOps Community IRL Meetup How multi-modal is it, if itā€™s only a text output? This talk introduces a multimodal retrieval and generation architecture that enhances efficiency by pre-processing and indexing video content and utilizing vector databases for storage. Tools like Dirosa for audio, Whisper for transcription, and FFmpeg for video compilation are used to streamline the storage, retrieval, and streaming of video data. This approach reduces latency and operational costs while addressing the complexities of generating video output from multimodal queries. Work in progress, but they're hitting fast-forward on it! Watch it here [https://home.mlops.community/home/videos/beyond-text-multimodal-rag-for-video] [Watch it here](https://home.mlops.community/home/videos/beyond-text-multimodal-rag-for-video) ## Exploring LLMs Speed Benchmarks At the risk of sounding like a gym-bro, what do you bench? This blog isnā€™t about a workout routine but instead provides an independent analysis of speed benchmarks for three 7-billion parameter language models: LLama2 7Bn, Mistral 7Bn, and Gemma 7Bn. The tests, conducted on Azure A100 GPUs using different setups and libraries, measured token generation speed with an emphasis on uniform environments and consistent configurations to evaluate the modelsā€™ efficiency and adaptability. Have a read to make some speedy gains. With thanks to Aishwarya Goel and Rajdeep Borgohain for their contribution. Guided Generation for LLM Outputs [https://home.mlops.community/home/blogs/guided-generation-for-llm-outputs] No matter how great, we all need a little guidance sometimesā€”even LLMs. This blog explores techniques to control and direct the outputs of LLMs. We'll cover methods like regular expressions, JSON schemas, and context-free grammars to ensure outputs follow specified formats. We'll also share efficiency improvements offered by these techniques, along with comparisons to existing methods and practical examples using tools like the Outlines library and llama.cpp. Emphasis is placed on the importance of parameter tuning (e.g., temperature, top-k sampling) to balance randomness and predictability in generated text, enhancing output quality and relevance. Have a read for some guidance about your guidance. With thanks to Kopal Garg for their contribution. [Exploring LLMs Speed Benchmarks](https://home.mlops.community/home/blogs/exploring-llms-speed-benchmarks-independent-analysis) ## Hidden Gems ## IRL Meetups https://www.meetup.com/amsterdam-mlops-community/events/301828602/Amsterdam [https://www.meetup.com/amsterdam-mlops-community/events/301828602/] - July 10 (šŸ“£shout out to Prosus) Denver [https://mlops.community/event/pizza-pints-expert-speaker-intro-to-machine-learning-toolkits-with-kubeflow/] - July 30 [Amsterdam](https://www.meetup.com/amsterdam-mlops-community/events/301828602/) Thanks for reading. See you in Slack [https://go.mlops.community/slack], YouTube [https://www.youtube.com/channel/UCG6qpjVnBTTT8wLGBygANOQ?view_as=subscriber], and podcast [https://home.mlops.community/public/content/] land. Oh yeah, and we are also on X [https://twitter.com/mlopscommunity]. The MLOps Community newsletter is edited by Jessica Rudd [https://www.linkedin.com/in/jmrudd/]. --- Source: https://aaif.live/newsletters/mlopscommunity/2024-06-27-know-the-latest-buzzword-before-it-becomes-a-buzzword