Giggso

https://lnkd.in/gJ_knz9N One more AI base.model, but this is interesting.. it's built off mamba Basically to quote AI21 "successfully incorporating Mamba alongside the Transformer architecture and advancing the hybrid SSM-Transformer model to production-grade scale and quality. " What does that mean? Well Mamba was introduced to address transformer pitfalls, mainly surrounding long text sequencing due to quadratic self attention complexity. What does that mean? Attention is all you need, is the fundamental of what we do today with gpts, technically it's complex and loses its historic thoughts fast. Google or YouTube for Attention is all you need by (Vasvani et al 2017) explains in detail. Now, getting to #Mamba and #SSM SSMs use a technique called Exponential Moving Average (EMA) to summarize past data. EMA assigns more weight to recent observations compared to older ones. It's like taking an average where recent data points count more. However, SSMs have a limitation: they forget past data at a fixed rate with each new observation. This is not always be ideal. Mamba improves upon SSMs by introducing a new concept: an additional output that determines how much to forget the past. This output acts like a "forgetting factor" that tells the EMA how much weight to give to previous observations. Essentially, it allows the EMA to be more flexible in how it forgets past data. In simpler terms: SSMs are like forgetting past information at a constant speed, regardless of how important it is. Mamba is like having a smart assistant that decides how much to forget past information (key difference) based on its relevance to the current situation and Jamba is Mamba for production grade work, to quote AI21 "Jamba can handle up to 140,000 tokens while running on a single GPU with at least 80GB of memory (like a high-end Nvidia A100). That translates to around 105,000 words, or 210 pages — a decent-sized novel." for now that's what I know. Will update more #modelops #genai #llm #llmops #transformers

https://lnkd.in/gXGrMAPZ #GenAI go with custom GPT.

https://lnkd.in/gQ3yC4HM There it is.. TOFU for unlearning and this is a giant leap. Excited for more.

2023 was an exciting year in AI and ML! The rapid advancements in LLMs and General AI were astounding. Despite daily updates, remember that these are mainly experimental. The real-world application requires safety and governance. From the many reads this year, I've curated five articles and two books that stood out and were insightful to me. Hope it helps you too 🙂 AI & Data Visualization: Psychology of Data Visualization (Toptal): The most fascinating read on the human side of data, understanding how visuals influence perception and decision-making. I strongly recommend this for any UX design planners for Data teams. https://lnkd.in/g6ZhmjG4 Regulating AI: Beyond the "Black Box" (Substack): Good ideas on responsible AI development and ensuring ethical use of powerful tools.  https://lnkd.in/gN-v4gDq Learning & Growth: JLA: A Framework for Evaluating Learning Analytics Tools (Learning Analytics):  https://lnkd.in/gGNZ8EZv Unlocking insights & opportunities with the power of data-driven analytics. https://lnkd.in/g9efVZVK Exploring AI Bias: Latent Factors in NLP Models (arXiv): The technicalities of bias detection and mitigation in natural language processing. https://lnkd.in/gxPqRF_X If you are a learner and following my posts 75+ NYT Graphs for Student Analysis (NYT): Put theory into practice with this fantastic resource for hands-on exploration of real-world data. https://lnkd.in/gaxXTb8V 2023 Book Recommendations: If you are interested in Operations Management principles for efficient and effective business operations. The McGraw-Hill 36-Hour Course: Operations Management by Linda Brennan: https://a.co/d/c3xt2B6 Interesting to read about the fast-paced world of financial technology and its transformative impact on the industry. Fintech: The New DNA of Financial Services by Pranay Gupta , T. Mandy Tham  https://a.co/d/giFopCI Happy reading and a very Happy New Year 2024 to all! #dataanalytics #AI #fintech #OperationsManagement #AIML #modelops #aisafety #trustworthyAI

https://lnkd.in/gi9CAWn9 I do not typically forward news clips, but found this particularly interesting. There is a seismic shift on the cards, 2024 will be peppered with 3L's... Losses, Lay offs and Legal suits. 1. Mitigation of losses and increasing productivity via AI 2. Leading to Layoffs and transition 3. When enterprises do the above two without proper audit controls, metrics and human empathy... Law suits will follow. AI is here to stay, using guard rails and auditing it in every step is the only pathway to success with it. Compliance, Controls and Cognizance is the 3C remedy to the above. #modelops #ethicalai #trustworthyai #guardrails #mlops #aisafety

There you go, one more Open Source LLM... #apple has released a multi modal LLM , open source #ferret https://lnkd.in/g3Ec9JFe Here is the context Apple has developed a technique where they use the flash storage in mobile devices (like iPhones and iPads) to store parts of large language models (LLMs). These parts are then loaded into the device’s main memory (DRAM) as needed. This method is very efficient, reducing the amount of data moved from the storage and doing it in small, continuous pieces. As a result, they can run models that are twice the size of the device’s main memory. This also speeds up the processing time - it’s 4.5 times faster on the device’s central processing unit (CPU) and 20-25 times faster on the graphics processing unit (GPU). This approach, called ‘LLM in a Flash’, could pave the way for Apple to introduce advanced AI capabilities, possibly even their own version of GPT, in future iOS updates.

I’m proud to share Giggso has been included on the @Renaissance Venture Capital Fall 2023 Startup Hotlist! The Hotlist features 50 of the most promising early-stage companies in Michigan, curated from hundreds of recommendations from VC funds, universities, accelerators, and angel networks. Check us out at www.giggso.com #modelops #llmops #nocodeai #mlops #trustworthyai #aisafety #modelmanagement #inclusiveai