A huge thank you to Intel Software and the Intel Liftoff team (Desmond Grealy 🙏) as we get to work training our state-of-the-art dialect model system on Intel GPU Max Series 1100. We are truly rocket strapping with these incredible servers (link to the spec below). More good news being announced shortly, hint: we 💕 Miami Bezoku, better not bigger #innovation #SOTA #dialects Delfina Aleman Cleopatra Bauduy Mana Tech Mohamed El Amine Seddik Omar Choukrani Prachi Jain Akshay Ramakrishnan Priyanshu Sharma Eugenie Wirz 🇺🇦 Ralph de Wargny Ralph Quintero Patrick Castenie Dolores Crazover (She/Her) Cesar Castro AiSalon Hasan Ali Emon Edison Sabala, MBA, MPH Jim Ryan South Florida Tech Hub 🌴Judith Nkwopara Dhruv Diddi Venkatesh Chandrasekar Stefana Raileanu Jyothis V James Ayal Stern Kelli Harris Rajashekar Kasturi Saiful Islam Raul Moas Norge Pena Anita Poletti Lisa Shephard Ines Lucena Jeffrey D. Abbott Wilson Pais Aline Fróes Jeffry Ullman-MBA Tony Jimenez https://lnkd.in/esRMPuPh
Bezoku’s Post
More Relevant Posts
-
Yup, we are stoked 🤙 More to follow...stay tuned to find out how we are going to unlock incredible community led AI innovation with our friends at Intel Software #aiforgood #aiforeveryone Cayman Enterprise City Alan B. Levan | NSU Broward Center of Innovation Charlie Kirkconnell John Wensveen, Ph.D. Cicely Strickland-Ruiz Peter Bas Eugenie Wirz 🇺🇦 Nuri Cankaya Sasha Whitcombe Michael HoShue, MBA Aisha T. McDonald, LMHC Manish Hirapara Jim Ryan Adriana Caballero Ángel J. Gallego Sijo Manikandan Ife Adebara Stacey Correcha-Price Dr. Allen Badeau Keren Flavell Dr Howard Haughton CTP CMath FIMA Arpit Narain, CFA, FRM, CQF Jigyasa Grover Johanna Lupardus Dasha Shunina Jesus Martinez Jamona Hayling Lisa Shephard John Moreno-Escobar Maria Carolina Reina John Hynes Catherine Garrido Nikki Cabus 🌴 Shanine Gilpin, MBA Rob Petrosino Jonathan Wolfe South Florida Tech Hub 🌴 Urban League of Broward County United Way of Broward County
To view or add a comment, sign in
-
-
We are at a cross roads. We need your help, the good people of LinkedIn land. So do we focus on monolingual dialect models or do we lean into the preference in the tech industry for multilingual models ? We are prioritizing dialects (our raison d’etre) but mapping heterogeneous features to previously mapped homogenous vector spaces are proven techniques for building state-of-the-art (SOTA) models. Regardless, powerful accelerators such as lemmatizing (grouping words into a simplified base form) are not affected. So do we go mono-lingual or bi-lingual / multi-lingual to reach our goals of expanding representation and delivering new SOTA dialect models 🤷🏻♀️ We need your help, we welcome any comments and input as we enter a crucial phase of development on, what else, Intel GPU Max 1100 (thank you so much Desmond Grealy) Bezoku, better not bigger Intel AI Intel Software Ralph de Wargny Eugenie Wirz 🇺🇦 Rajashekar Kasturi Robert Chesebrough Eduardo Alvarez John Wensveen, Ph.D. Charlie Kirkconnell Kaitlyn Elphinstone Cayman Enterprise City Alan B. Levan | NSU Broward Center of Innovation Mana Tech AiSalon Mohamed El Amine Seddik Saiful Islam Ali Elfilali Cleopatra Bauduy Dhruv Diddi Prachi Jain Priyanshu Sharma Akshay Ramakrishnan Bill Zhang Francisco Kemeny Ralph Quintero Ife Adebara, PhD Khanh Linh Nguyen Khanh Linh Nguyen Delfina Aleman Patrick Castenie Venkatesh Chandrasekar Felipe Pinzon Uvika Sharma Ángel J. Gallego Jigyasa Grover
To view or add a comment, sign in
-
-
In this wonderful article in the The Guardian from the paper’s science journalist Hannah Devlin (kudos on such a well researched piece joining the dots between TikTok, Machine Learning and Psychology), this is our key takeaway ““They [babies] have just the contours of the language that they will populate it with words,” said Prof Caroline Floccia, a developmental psychologist at the University of Plymouth. “Others start with isolated words and then they will construct sentences from that. Then you can have plenty of kids who are in the middle.”” Read the full article as a great preparation for our new approach for dialect based language models, built using Intel Software services on Intel GPU Max 1100 servers. Bezoku, better not bigger Intel AI Ángel J. Gallego Cleopatra Bauduy Akshay Ramakrishnan Prachi Jain Delfina Aleman Mana Tech Cayman Enterprise City Alan B. Levan | NSU Broward Center of Innovation AiSalon South Florida Tech Hub 🌴 Nikki Cabus 🌴 John Wensveen, Ph.D. Tim Chatfield Tomoko Mitsuoka (三岡 智子) Terri Horton, EdD, MBA, MA, SHRM-CP, PHR Ali Elfilali Mohamed El Amine Seddik Omar Choukrani Saiful Islam Rajashekar Kasturi Kelli Harris Kaitlyn Elphinstone Althea WEST MYERS Reinaldo Fletcher Cayman Islands Government Jigyasa Grover https://lnkd.in/eAssfuGx
To view or add a comment, sign in
-
Bezoku is undergoing a game changing effort to model the heterogenous and hierarchical nature of low resource languages, at scale. During July 2024 we will be testing multiple models in parallel on one or more low resource languages, running on Intel Tiber Developer Cloud. The challenges are numerous, particularly the reduced number of available tokens and annotated corpus, but we have identified new research into techniques that refactor existing Transformer architectures to increase performance without overfitting. Papers and demos will be shared during our quest to create a scalable dialects pipeline to create hundreds, if not thousands, of new models thereby increasing representation outside the ~ 30 languages currently supported by incumbents. As we gear up for this marathon session, basically a 4X of the Intel Liftoff hackathon a few weeks back, a huge thank you to Intel Liftoff management, especially Ralph de Wargny the Liftoff program founder and our Intel Liftoff mentor / developers Desmond Grealy (get well soon) Rajashekar Kasturi and Rahul Unnikrishnan Nair. As ever huge thanks to Eugenie Wirz 🇺🇦, supporting us every step of the way. Without the incredible support from Intel Liftoff, NONE of this would be possible. Stay tuned, and let the innovation commence as we seek to unlock the secrets of human dialects. Bezoku, better not bigger Intel Software Intel AI Intel Business Eugenie Wirz 🇺🇦 Felipe Pinzon Patrick Castenie Mark Martin ⭐️ Mohamed El Amine Seddik Omar Choukrani Saiful Islam Dhriti Jotwani Jigyasa Grover Prachi Jain Priyanshu Sharma Sinchana Bhat Ayal Stern Tim Chatfield Nicolas Rannou Kevin Crain Cleopatra Bauduy Carol Hylton Hasan Ali Emon Dr. Susie A. Ebru G. Kittur Ganesh Abdullah Al Asif Akshay Ramakrishnan Uvika Sharma Edison Sabala, MBA, MPH John Wensveen, Ph.D. Alan B. Levan | NSU Broward Center of Innovation South Florida Tech Hub 🌴 Cesar Castro
To view or add a comment, sign in
-
-
Great, and timely, reporting from Bloomberg Opinion journalist Parmy Olson on the mistakes AI firms are making, setting themselves up for failure by over reaching and hyping solutions that are falling short. "AI isn’t yet a jack-of-all-trades but a master of a few. The sooner business leaders realize they can apply it to an array of niches and not for everything, everywhere, all at once, the sooner they can make the technology useful for them. But they’ll need more level-headed guidance from tech firms, which must resist pitching AI as a general-purpose quick fix and “magic.” Such rhetoric is fuel for a bubble if they don’t." The old saw, "under promise and over deliver", has never been more relevant at this point in the innovation cycle. Bezoku is focused on two capabilities, small and scalable dialect models and privacy preserving machine learning, built with our friends at Intel AI. After the next sprint, we will demo our brand new technology to South Florida customers and partners to listen and learn. Labs are great, but the customer is the ultimate arbiter of success. Bezoku, better not bigger Intel Business Intel Software Eugenie Wirz 🇺🇦 Ralph de Wargny John Wensveen, Ph.D. Charlie Kirkconnell Alan B. Levan | NSU Broward Center of Innovation Cayman Enterprise City Deputy Premier, Cayman Islands South Florida Tech Hub 🌴 Owwll Sunshine Startups Live Jim Ryan Nikki Cabus 🌴 AiSalon Dolores Crazover (She/Her) Cesar Castro Ayal Stern Rajashekar Kasturi Desmond Grealy Tomoko Mitsuoka (三岡 智子) Jigyasa Grover Dhruv Diddi Venkatesh Chandrasekar Patrick Castenie Romaine Foster Althea WEST MYERS Dhriti Jotwani Tamika Lecheé Morales Isabel Tejada Sánchez, Ph.D Nicolas Rannou Cleopatra Bauduy Zhuo Wu Katja Rausch Omar Choukrani Dr. Susie A. Kittur Ganesh https://lnkd.in/ek_biAy2?
Nvidia’s Explosive Growth Masks AI Disillusionment
bloomberg.com
To view or add a comment, sign in
-
Language is the killer technology of our species. Ergo, language technology is a big deal, leveraging thousands of years of human development. Language is context dependent, famously summarized by Charles V, "I speak Spanish to God, Italian to women, French to men, and German to my horse". Bezoku's mission is an affordable, trustworthy, secure language system that works at a human level, as a dialect. Dialects connect communities with a shared cultural vocabulary, often due to a personal attachment to a region. Human identity is deeply entrenched with the local customs, story telling and common ground we share with our friends and family. Dialects create an emotional connection, an intuitive integration that does not require any translation, deep understanding, personal and currently non-existent in the current crop of language models. Starting next week, Bezoku will be breaking ground on our first state-of-the-art (SOTA) dialect model. And we are building it natively (no pun intended) on Intel Tiber Developer Cloud with our friends at Intel Software as a member of the awesome Intel Liftoff program. Do you think your dialect should be on the shortlist ? Share in the comments below. Bezoku, better not bigger Ángel J. Gallego Tomoko Mitsuoka (三岡 智子) Dr. Susie A. Kristy Amaro Patrick Castenie Muhammad Bilal Mark Martin ⭐️ Alexi Schwartzkopff Mohamed El Amine Seddik Omar Choukrani Krysten Brenlla Tamara Polajnar Tawana Petty Sinchana Bhat Tamika Lecheé Morales Lizette Ibarra Tim Chatfield Isabel Tejada Sánchez, Ph.D Cleopatra Bauduy Latrice Thomas Dr. Lisa Dethridge Carol Hylton Minzhi Huang/Hwang Katja Rausch Roxana Marachi, Ph.D. Sara Copic, Ph. D. Sasha Havlicek Ioannis Mollas Stamatis Karlos Aissam Outchakoucht Leona Isabelle Verdadero Ann Kristin Glenster, Ph.D., FRSA Hasan Ali Emon Tarja Stephens Oumayma Essarhi Manel ALOUI Francesca Quaratino Fernanda Jorgensen, MPH Dawn Watts
To view or add a comment, sign in
-
-
Join us and our friends at Intel Software to grow your startup. Contact us for more information particularly if you are in South Florida or the Caribbean, female led or working to increase diversity in technology. Bezoku can help with machine learning engineering and advice on how to leverage the insane power of Intel Tiber Developer Cloud. Bezoku, better not bigger Cayman Enterprise City Alan B. Levan | NSU Broward Center of Innovation AiSalon South Florida Tech Hub 🌴 Mana Tech Fernando Cariello Dasha Shunina WOMEN IN TECH ® Global Raechel Canipe John Wensveen, Ph.D. Charlie Kirkconnell Mark Martin ⭐️ Urban League of Broward County Kaitlyn Elphinstone Althea WEST MYERS Cayman Islands Centre for Business Development Chris Bailey Tomoko Mitsuoka (三岡 智子) Romaine Foster Richard_ C. Imran Adan James Gee Dhriti Jotwani Ayal Stern Tim Chatfield Alejandro D. González Josh Longenecker Carol Hylton CareerSource Broward Tarja Stephens Ebru G. Reinaldo Fletcher Dawn Watts Raul Moas Ralph Quintero Terri Horton, EdD, MBA, MA, SHRM-CP, PHR Data Science Salon Edison Sabala, MBA, MPH Dolores Crazover (She/Her)
Calling all AI startups, join Intel Liftoff now to Leverage Intel’s superpowers for startups and meet a community of over 300 AI founders! Free, virtual, global. With access to the new Intel Tiber AI Cloud, featuring CPUs, GPUs, the Gaudi 2 AI accelerator platform, and soon Gaudi 3! Apply here: https://lnkd.in/eH5jpmc9 #intelliftoff Intel Software Intel AI #genAI #generativeAI #ai #ml NVIDIA AMD Amazon Web Services (AWS) Google Cloud Microsoft Azure Microsoft NVIDIA AI #nvidiainception Berkeley SkyDeck Stanford University Cerebral Valley Y Combinator UCLA New York University Cornell Tech Google for Startups Microsoft for Startups AWS Startups
To view or add a comment, sign in
-
-
RAG innovation, super interesting and worth experimenting with on Intel Tiber Developer Cloud 🌟 Intel AI Intel Software Prachi Jain Priyanshu Sharma Akshay Ramakrishnan Yitae Jeong Rajashekar Kasturi
Corrective Retrieval Augmented Generation (C-RAG) introduces a plug-and-play method that enhances the accuracy and dependability of language models by refining how they incorporate and leverage retrieval-based information during the generation process. This is an important step in the right direction. 🔵 RAG: Giving Context to Language Models In the business use of GenAI, RAG has been an important development. It gives the AI access to internal private data, enhancing the precision and relevance of Language Models by supplementing them with contextually pertinent information. Despite RAG's considerable contributions to model accuracy, there is an acknowledgement within the GenAI community that there is still a considerable gap in delivering the accuracy needed in regulated environments. If you get the basic Information Retrieval part of RAG wrong, the generated answer will likely be incorrect too. 🔵 C-RAG: A Step Towards Precision CRAG elevates the RAG framework by integrating a retrieval evaluator — a mechanism that assesses the relevance and quality of the information retrieved in response to queries. A low relevance score can trigger a web search to validate and supplement the information passed to the generation process. 🔵 G-RAG: Adding Graph Reasoning to Correction The fusion of Knowledge Graphs with CRAG offers a multitude of technical advantages: 🔹 Contextual Relevance: By providing a structured and semantically rich organisation of information, Knowledge Graphs can enhance the contextual relevance of RAG's retrieval process. 🔹 Curated Data Source: While CRAG expands the information through web searches, Knowledge Graphs offer a stable, curated data reservoir, thereby reducing dependency on the sometimes erratic quality of web-based content. 🔹 Efficiency in Retrieval: The structured nature of Knowledge Graphs facilitates quicker and more efficient information querying. 🔹 Leveraging Private Data: Integrating RAG with dynamically updated Knowledge Graphs enables the consistent utilisation of private data, securely stored behind firewalls, in the generation process. Text and structured data can be merged into one unified layer that feeds the GenAI. 🔹 Good Old-Fashioned Reasoning: Ontologies can be combined with the creative power of generative methods, supplementing them with the reliability of symbolic reasoning. 🔵 Working Memory Graph I propose a conceptual framework termed the 'Working Memory Graph' (WMG), which synergises LLMs and Knowledge Graphs. The WMG integrates LLM embedding vectors with ontologies and factual data from the Knowledge Graph. The WMG translates natural language queries into graph structures, employs Graph Retrieval-Augmented Generation (GRAG) to collate facts, and then applies traditional reasoning techniques to ensure the generation of high-quality, factually accurate content. ⭕CRAG Paper: https://lnkd.in/eEBxFJBq ⭕ Working Memory Graph: https://lnkd.in/eQF4PE27
To view or add a comment, sign in
-
-
In this excellent book, first published in 2011, by the author JP Davidson and made into a BBC series with Stephen Fry, a prescient section can be found on page 74, which was aimed at the United Nations, but would be applicable to today for large language models, “That Tower of Babel is increasingly under threat as the homogenization of languages seems to be pushing us back to the days of a few proto-languages and that… spells the death of many of our micro-languages [dialects], each one a repository of knowledge, a definer of culture and identity, exquisitely structured and vastly complex.” As a small language model company, focused on collaborating or developing excellence in syntactic and semantic morphological data and models, our goal is to capture not just the grammar and vocabulary, but also infuse each model with cultural and identity markers that are excluded from hegemonic platforms, where incumbent technology operates with an algorithmic monoculture. As we continue to build our Homomorphic Encryption as a standard capability, our next demo will showcase our linguistic skills and end to end engineering to bring SOTA dialects to market, running on, what else, Intel Tiber Developer Cloud and deployed everywhere. Bezoku, better not bigger. Intel Business Intel Software Intel AI Ralph de Wargny Abbie Fawcett Kevin Crain Rajashekar Kasturi John Wensveen, Ph.D. Alan B. Levan | NSU Broward Center of Innovation Ángel J. Gallego Mohamed El Amine Seddik Omar Choukrani Prachi Jain Priyanshu Sharma Akshay Ramakrishnan Dr Howard Haughton CTP CMath FIMA Saiful Islam Isabel Tejada Sánchez, Ph.D Khanh Linh Nguyen Oumayma Essarhi Pallav Bhansali Ralph Quintero Tomoko Mitsuoka (三岡 智子) Uvika Sharma Yitae Jeong Zhuo Wu Susan Gibson Dr. Susie A. South Florida Tech Hub 🌴
To view or add a comment, sign in
-
Global Community Manager @ Intel Corporation | Intel® Liftoff
1wGreat work