Ravi Mehta’s Post

Retrieval-Augmented Generation (#RAG) offers numerous benefits, particularly when it comes to empowering Large Language Models (#LLMs) with up-to-date information. By integrating a search function that retrieves relevant data from external sources, RAG ensures that LLMs are equipped with the most current knowledge available, thereby enhancing the accuracy and relevance of their responses. In my latest blog, I describe important practical tips for designing a robust RAG pipeline using NVIDIA #AI Endpoints https://lnkd.in/dqYGiihD

A day after I posted about the LLM performance wars heating up, Mistral AI CEO confirms the miqu-1-70b model is in fact a leaked version of a model currently under development with limited release. This story is filled with intrigue: 📈 Performance on certain benchmarks like EQ-bench is the closest open source has come to GPT-4 level 🔢 The model appears to be quantized - a technique to enable top-tier models to run on less sophisticated GPUs 🥷 Source of the leak confirmed as an overexcited customer with an early preview 🤷 Mistral seems to be embracing the leak and subsequent viral media rather than punishing it Source; https://lnkd.in/eWayntrj #llm #opensource #generatieveai #ai

Sunday Morning Reading A collection of Tech and AI related articles that I have saved up to read on a Sunday morning to try and catch the speeding bullet that is Tech and AI right now. https://lnkd.in/er3qz6Nu #tech #AI

I'm thinking about AI again. I know, "All I hear all day long is how great AI is at this and how wonderful AI did that. AI, AI, AI!" (IYKYK) but after spending the last week and a half experimenting with Custom GPTs, I have a lot to say on the computational costs of LLMs and what the future of them—and their adoption—might look like. Read this week's Two Steps Forward to peek into that future for yourself. #llm #openai #ai #nvidia #gpu

AI for Web Devs: What Are Neural Networks, LLMs, & GPTs? https://bit.ly/48Cfz88 Welcome back to this series where we are learning how to integrate AI tooling into web applications. In the previous posts, we got our project set up and did some basic integration. Intro & Setup Your First AI Prompt Streaming Responses How Does AI Work Prompt Engineering AI-Generated Images Security & Reliability Deploying So far we’ve built a very basic UI with a text area that takes whatever the user writes and sends it over HTTP to OpenAI’s API. When the streaming response returns, it updates the page with each bit of text as it arrives. via DZone AI/ML Zone https://bit.ly/41qvfZp January 19, 2024 at 02:27PM

Synthetic data is the answer. Here's why. AI models are better at generating information than you are. Both in terms of quantity, but also in terms of quality. If that makes you feel bad, remember that you are a universe experiencing itself which is pretty cool. Cosmopedia is a good example. It consists of 25 billion tokens of synthetic text data. It would take approximately 5,208 years for a human to generate 25 billion tokens, assuming 250 working days per year writing 2400 words/hour. It took 10k GPU hours to generate the dataset on H100 GPUs. So I guess one GPU hour is half a year of human work. 😎 https://lnkd.in/gwp5TbgF

Activant AI Hardware research is out! Yi Tay has a great article articulating the challenges of handling hardware in training AI models from scratch in startup settings. He discussed several factors, including hardware lottery (Variability in hardware quality), GPU vs TPU training, codebase challenges, and lack of systematic scaling due to resource limitations. Understanding the challenges and details of AI infrastructure is fascinating, and that is the main reason why we at Activant are diving deep into AI hardware. Going forward, we see a future where JARVIS can run on local devices finishing tasks for you with zero data or network concerns. Please reach out if you want to trade notes on AI hardware. We are continuously updating our AI thesis through Activant research and Decoding AI newsletter (Link here: https://lnkd.in/eie43kC3). Let's GO! Sources: - https://lnkd.in/ehSQ87A7 - https://lnkd.in/ebRF8kyg

This is a superb #llm #qlora read, but finishing it will take a good amount of concentration #NLP #AI. Inside you will see the research behind the quantization of larger models enabling less powerful GPUs and even CPUs to run complex fine-tuning. https://lnkd.in/gv73TFf9

pub.towardsai.net: The article discusses the top open LLMs and encourages readers to explore further on Towards AI.

NVIDIA AI Unveils SteerLM: A New Artificial Intelligence Method that Allows Users to Customize the Responses of Large Language Models (LLMs) During Inference  MarkTechPost https://lnkd.in/grKpK7re