Nicolas Rasamimanana

LATEST IN AI : Quick Read Nvidia launches Nemotron, a 70B model that outperforms GPT-4o and Claude 3.5 Sonnet. Technical Highlights : The model features 70 billion parameters, offering efficient handling of text and coding queries. It builds on Llama 3.1 architecture, based on transformer technology, ensuring coherent and human-like responses. Performance Benchmarks : Nemotron-70B achieved high scores on alignment benchmarks such as Arena Hard (85.0), AlpacaEval 2 LC (57.6), and GPT-4-Turbo MT-Bench (8.98), surpassing its larger counterparts. Efficiency Focus : Despite having fewer parameters compared to GPT-4o, the model's performance demonstrates the efficiency of smaller, well-optimized models. Open-Source Availability : Nvidia has made the model, reward models, and training datasets open-source on Hugging Face, encouraging further testing and innovation. This launch reinforces Nvidia's growing influence in AI beyond hardware, showcasing the potential of efficient, smaller-scale LLMs. NVIDIA #futureofai #aiinmedicine

8

1 commentaire

The FIRST NEURAL NETWORK It is well know that neural networks are the modular unit of #AI, and are powering the ongoing technological revolution. However, what many people do not realize is that the very first neural network dates as far back a 1805. In 1805, Adrien Marie Legendre published a method which given a set of data points, determines the line of best fit in the least squares sense. The so called "method of least squares" minimized the sum of squared distances between the data points ("training data") and the candidate lines. Carl Friedrich Gauss published the same idea in 1809 but noted that he had the result since as early as 1795. The method of least squares, commonly known today as "LINEAR REGRESSION" is indeed the first known fully end-to-end differentiable method that learns a model from training data. And since that is exactly the defining property of neural networks, it follows that linear regression is the first known neural network. #artificialintelligence #mathematics

223

17 commentaires

“𝘞𝘩𝘰 𝘴𝘦𝘦𝘴 𝘵𝘩𝘦 𝘩𝘶𝘮𝘢𝘯 𝘧𝘢𝘤𝘦 𝘤𝘰𝘳𝘳𝘦𝘤𝘵𝘭𝘺: 𝘵𝘩𝘦 𝘤𝘢𝘮𝘦𝘳𝘢, 𝘵𝘩𝘦 𝘴𝘺𝘯𝘵𝘩𝘦𝘵𝘪𝘤 𝘱𝘳𝘪𝘰𝘳, 𝘰𝘳 𝘣𝘰𝘵𝘩?”  Paraphrasing Picasso's quote in Cafca's 1️⃣ intro to show the importance of synthetic data. Significant process since: • the decade old 2️⃣ "𝘋𝘺𝘯𝘢𝘮𝘪𝘤 3𝘋 𝘈𝘷𝘢𝘵𝘢𝘳 𝘊𝘳𝘦𝘢𝘵𝘪𝘰𝘯 𝘧𝘳𝘰𝘮 𝘏𝘢𝘯𝘥-𝘩𝘦𝘭𝘥 𝘝𝘪𝘥𝘦𝘰 𝘐𝘯𝘱𝘶𝘵 (2015)", and, • even the recent 3️⃣ "𝘈𝘶𝘵𝘩𝘦𝘯𝘵𝘪𝘤 𝘝𝘰𝘭𝘶𝘮𝘦𝘵𝘳𝘪𝘤 𝘈𝘷𝘢𝘵𝘢𝘳𝘴 𝘧𝘳𝘰𝘮 𝘢 𝘗𝘩𝘰𝘯𝘦 𝘚𝘤𝘢𝘯 (2022)", Cafca's pertinent use of a 𝘀𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝗽𝗿𝗶𝗼𝗿 allowed its team to: • 📉reduce capture requirements to only 3 unrestricted mobile images,  • 🧔‍♂️generalize to real-world faces, capturing intricate facial details, and • 👓(personal favorite) even show progress in glasses capture ! Still not in a plug-n-play format, but I guess that's coming 🔜 . #nerf #SIGGRAPHAsia #3d #ai #machinelearning #ml #metaverse 1️⃣ https://lnkd.in/dFX7mqFK 2️⃣ https://lnkd.in/djMWz6t2 3️⃣ https://lnkd.in/dWGEZC4e

8

One more intresting presentation of Paul Tremblot de la Croix "How Generative AI Can Help Startups Succeed" at the #VDS2024. And my note on how AI can drive success: 💬 speakers emphasized that data is the foundation. Generative AI applications are only as good as the data they're built on. The ability to combine real-time data with generative capabilities is a game-changer for startups. 🕵‍♂️ shared the importance of customization. Don’t adopt generative AI just because it's trendy. Tailor it to your business needs, ensuring it’s the right tool for your use case. 👮‍♀‍ discussed data privacy and security concerns, emphasizing GDPR and other regulations. Startups must handle customer data carefully and integrate strong security protocols when deploying generative AI tools. Key points: 1. Generative AI + RAG = Personalization. By using AI with RAG, startups can retrieve relevant real-time information or past interactions, ensuring personalization not just based on predictions but also on actual preferences, history or external knowledge sources. 2. Customization & Use Case Fit. It's crucial to pick the most efficient AI model, considering cost, latency and accuracy. Generative AI isn't one-size-fits-all! 3. From MVP to Production. Building an AI-powered product is just the first step. Measuring, tracking and fine-tuning AI in production is essential for maintaining customer trust👌 I appreciate your meaningful insights shared about scaling AI strategically. Also on keep security tight and focus on solving real customer problems and think about there needs 😇 #AI #Startups #VDS2024 #DataPrivacy #Cybersecurity

4

Use the right tool for the job. It's again shown that ML (and LLM by extension) are great technologies, just not at the core of planning optimization and Operations AI algorithms (*). If you use a hammer on a screw - or a screwdriver on a nail - it will show in your results. (*) they do have uses in the planning optimization product experience.

51

1 commentaire

Presenting Orpomis Prime 7B model series ! These models are a result of trying different merge strategies of the following models: - mistral-orpo-beta from KAIST AI - Hermes-2-Pro-Mistral-7B from NousResearch In my experiments, dare ties merge leads to a significantly better-performing model on the openllm benchmarks In general, the model performance is quite solid in text generation as well as math reasoning. Full result comparison: https://lnkd.in/gb92DXwG Dare ties Model: https://lnkd.in/gMS-H4Sq Model stock Model: https://lnkd.in/giwEJ7Gs

9

Finitism in #AI datasets can be mitigated if a more significant portion of the worldwide internet populace can create data, and if there were a method to incentivize these interested contributors to create such data. A big drum roll for Pertin-ant 🐜here... #AI #AGI #LLMs #trainingdata

6

Ever wonder if the so-called long-context LLMs is really working ? Or just getting lost-in-the-middle ? This new benchmark, LongMemEval, tests the long-term memory of chat assistants. What's in it ? LongMemEval has: - 164 topics, & - 500 manually created questions and it tests 5 core memory abilities: - information extraction - cross-session reasoning - temporal reasoning - knowledge updates - abstention "Prior benchmarks often fail to capture the nature of user-AI interactions, only have limited memory ability coverage, or often feature too short chat history..." To address this, "LongMemEval requires chat systems to parse the dynamic interactions online for memorization, and answer the question after all the interaction sessions." They also discover several interesting things about how to deal with the memory in long-context LLMs, which may be of interest to those involved in such research But most importantly, they seem to verify what I suspect a lot of already know--truly interactive and useful long context LLMs are still a long-ways away. paper: https://lnkd.in/gTjTxAum code: https://lnkd.in/gC5CnQTq source: https://lnkd.in/g_C_iJDM

34

1 commentaire

Sober Sunday: Meta’s Secret to AGI: Highlights from Yan LeCun’s Latest Talk 💡 Yann LeCun, Chief AI Scientist at Meta Facebook, recently shared his insights into Artificial General Intelligence (AGI)...spoiler alert, we're not close and here’s what you need to know (though I highly recommend you watch the summary,it's only 12 min!): 👓 Future of Smart Assistance: LeCun envisions a world where virtual assistants with memory and advanced reasoning support productivity, making everyone a "boss" with their own digital workforce. 🚧 Current AI Limitations: Today’s AI lacks essential human-like traits: - Persistent memory - Common sense and intuition - Complex, adaptive planning 🔍 Marr's Paradox – A Core Challenge: Tasks that are simple for humans, like tidying up or even dishwashing, are still too complex for AI, while high-level math and calculations are simple for machines. A rethink in AI development is essential. 📊 The Data Challenge for AGI: A large language model (LLM) trains on trillions of words, requiring data equivalent to 350,000 human years of reading. Even so, LeCun argues that this isn’t enough. A child processes 16,000 hours of sensory experience in their first few years—far richer than text data. 🎯 Objective-Driven AI: Instead of merely reacting, LeCun's proposed model of Objective-Driven AI could: - Predict outcomes and adjust actions to achieve specific goals - Navigate and adapt to unfamiliar situations - Mark a shift from pattern-based responses to proactive, goal-oriented thinking 🗣️ Sound Bite: "We're never going to reach human level intelligence by just training on text" – Yan LeCun 📅 Long-Term Pathway to AGI: LeCun emphasizes that achieving AGI is far from imminent, requiring years of technological advancement, with developments in computational power, adaptable models, and nuanced world understanding. In this exciting yet challenging journey, Meta’s vision offers a promising approach to unlocking true AGI. Note for Data Enthusiasts: The Road to AGI Needs You! 🚀 To all the data professionals out there—this is huge news! 🌟 As Yan LeCun pointed out, for AGI to move beyond today’s limitations, we’ll need far more rich, diverse, and real-world data than we currently have. The journey to AGI isn’t just about smarter algorithms; it’s about feeding these systems the comprehensive, nuanced information they need to truly understand and interact with our world. Get ready—the path to AGI is as much a data journey as it is an AI one! #ArtificialIntelligence #AGI #DataScience #AIInnovation #FutureOfAI #DataDriven #MachineLearning #DeepLearning #DataProfessionals #AIResearch #MetaAI #YanLeCun #FutureOfTech #SmartAssistance #TechLeaders #DataStrategy AI Engineers, Inc. Data Science Community The Machine Learning Enthusiast - Luis Valencia Tech Innovation Hub

2

1 commentaire

🦙 Meta AI Releases Llama 3: 8B & 70B Models! 🚀 Meta has released two new Llama 3 models : 8B parameter model outperforms 70B Llama 2 🤯 70B parameter model is competitive with Gemini Pro 1.5 and Claude 3 Sonnet 💪 (more models are coming soon!) 405B model in the works, nearing GPT-4 level 👀 🔑 Key Takeaways: Trained on 15T tokens, 7x more than Llama 2 8K context window, shorter than some recent models Performs well on reasoning benchmarks like GSM8K Supports commercial use with some license restrictions 🧪 Trying it Out: Available via Ollama, HuggingFace, API services Can be run locally with quantization Responds well to chain-of-thought & ReACT prompts Some multilingual support but dedicated model likely coming 📜 License Notes: - Can't use Llama 3 to train other large models - Must include "Llama 3" in name of fine-tuned versions - Review prohibited use cases before commercial use The Llama 3 models look promising and signal continued rapid progress in open LLMs. Expect to see powerful fine-tuned versions and new capabilities soon! 🦙🔥 #Llama3 #OpenLLMs #LLMProgress

74

2 commentaires

Join us for our second edition of HelicALL! In this webinar, we will be discussion the latest single-cell foundation models and their most popular applications! Helical Maxime Allard Mathieu Klop

17

Thank you Max Welling and Yann LeCun, I could not have said it better myself!  It's time for AI4Science to take center stage. Combining cutting-edge AI/ML techniques with scientific principles can unlock new frontiers. With the ability to work securely with limited data and without reuse, the potential is limitless. #AI4Science. #AI4sustainability. #ScienceBasedAI.

16

4o ? More like 4oh-no amirite? The latest model from OpenAI, named 4o (o is for "omni") can do so much that that it might make you think "oh no, I'm about to be replaced by an AI." The word "awesome" really fits for describing the new capabilities and what this advancement could mean for human life as we know it. It's one thing to get text or a picture generated, but the 4o model can joke with you, help you get ready for an interview, or tutor your children! It can also read facial expressions, make inferences based on context, and offer words of encouragement. So since this new virtual personal assistant can do everything... what are we supposed to do now? Are teachers replaced? Writers? Stylists? No job is safe, because the computer can do it all. Not quite. The computer cannot care. It's wild that I have to say this - I thought it was only stuff for sci-fi, but here were are. Computers do not care. They cannot express compassion. They cannot empathize. They can simulate these things but cannot perform them, because they are not human. All of the advancements we make in technology are nothing but cold and heartless science if we don't take into account that the needs being addressed are human needs. Humans can care. We're wired for it from the ground up - if we see someone in pain, sometimes we can feel it even though we're just seeing or hearing about it. We are fallible, imperfect, even a bit messy - but we can offer true compassion. So... do you really want to get coaching and tech help from a machine? How many youtube videos can you watch about a topic before you start to wonder "if I could just ask this person a simple question I'd understand this so much more"? Even a talking computer assistant can only regurgitate what it has been fed, even if may be able to rephrase and summarize complex content all on its own. It cannot understand your motivation, it can't empathize with your lived experience as a person. These things matter! If you'd like help getting over the limiting beliefs that are holding you back from being a confident and competent user of your day-to-day business tech, and you're ready to work with a human who cares, book a call with me today! https://askmaria.tech

4

Crafting AI Apps in 45 days is that possible ? Week 2 is over and exciting things are happening at Syntax Erreur as we dive into buildspace. Here’s a sneak peek into our journey and how you can benefit: ## What We’re Building: Innovative AI-Powered Toys: We're developing a series of cutting-edge toys that showcase the potential of AI. Rapid Prototyping: Perfecting our skills to create MVPs in less than 45 days.🚀 ## Why This Matters to You: Support for Startups: Our Buildspace experience enhances our ability to assist early-stage startups, offering fast and efficient tech solutions. Community Inspiration: This project is more than toys; it's about pushing tech boundaries and inspiring innovation. ## How You Can Get Involved: Follow Our Journey: Stay updated with insights and progress from our buildspace adventure. Collaborate with Us: Are you an early-stage startup? Let’s turn your ideas into reality. DM us! Learn More: Explore our work and philosophy at Syntax Erreur Join us on this exciting journey and be part of the innovation wave! #Buildspace #AI #AppDevelopment #Startups

113

11 commentaires

Modeling the Real World Part 3 – Realistic Goals and Limitations (Biopharma example applies to all industries) From Yann LeCun’s announcement of META’s commitment of his 500-person team to a ten-year effort to build an initial real-world model, to Sweden’s announcement of their goal to build a biomedical digital twin of every citizen, interest in modeling the real world is growing. However, reality requires assessing our current limitations before setting goals that are exciting but out of reach, especially in modeling life. What is achievable? The NAS published a Physics of Life Report which concluded, “An important lesson from the long and complex history of neural networks and artificial intelligence is that revolutionary technology can be based on ideas and principles drawn from an understanding of life, rather than on direct harnessing of life’s mechanisms or hardware.” https://lnkd.in/eG5VmUiW An example is mitochondrial ATP production. The image is a 3D model of ATP Synthase - molecular motor (enzyme) that powers all living cells (humans, animals, plants, or microbes), generating chemical energy in the form of ATP molecules. It would be nice if we could harness this mechanism to power energy-glutenous computer chips, but…😊 There are, however, principles we can learn from mitochondria when implementing the Biomimetic Digital Twin Ecosystems methodology discussed in the white paper Road to Intelligent Computing - https://lnkd.in/e4ePqPWx To begin: 1.     Scope the model to focus on critical outcomes – mitochondria has it’s own DNA, in humans its only the 37 genes needed to fulfill its purpose 2.     Multipurpose design – the primary purpose is producing energy, but it also helps manage the cell’s life cycle Applications: ·        Focus the biomimetic digital twin model design on critical outcomes ·        Identify method library components for reuse in the digital twin ecosystem Example Genomic dark data (unseen relationships) discovery biomimetic digital twin ecosystem to produce evidence for potential reclassification of variants of unknown significance (VUS) in relation to a defined phenotype. ·        VUS reclassification is a critical outcome because it is the largest potential source of biomarkers ·        GA4GH Phenopacket schema library component is broadly reusable An example of initial results related to breast cancer is available in the Road to Intelligent Computing paper (link above). I’m getting feedback on Parts 1 & 2 of this series, and hopefully will on this post. The feedback will determine the content of Part 4.

1

Modeling the Real World Part 3 – Realistic Goals and Limitations (Biopharma example applies to all industries) From Yann LeCun’s announcement of META’s commitment of his 500-person team to a ten-year effort to build an initial real-world model, to Sweden’s announcement of their goal to build a biomedical digital twin of every citizen, interest in modeling the real world is growing. However, reality requires assessing our current limitations before setting goals that are exciting but out of reach, especially in modeling life. What is achievable? The NAS published a Physics of Life Report which concluded, “An important lesson from the long and complex history of neural networks and artificial intelligence is that revolutionary technology can be based on ideas and principles drawn from an understanding of life, rather than on direct harnessing of life’s mechanisms or hardware.” https://lnkd.in/eG5VmUiW An example is mitochondrial ATP production. The image is a 3D model of ATP Synthase - molecular motor (enzyme) that powers all living cells (humans, animals, plants, or microbes), generating chemical energy in the form of ATP molecules. It would be nice if we could harness this mechanism to power energy-glutenous computer chips, but…😊 There are, however, principles we can learn from mitochondria when implementing the Biomimetic Digital Twin Ecosystems methodology discussed in the white paper Road to Intelligent Computing - https://lnkd.in/e4ePqPWx To begin: 1.     Scope the model to focus on critical outcomes – mitochondria has it’s own DNA, in humans its only the 37 genes needed to fulfill its purpose 2.     Multipurpose design – the primary purpose is producing energy, but it also helps manage the cell’s life cycle Applications: ·        Focus the biomimetic digital twin model design on critical outcomes ·        Identify method library components for reuse in the digital twin ecosystem Example Genomic dark data (unseen relationships) discovery biomimetic digital twin ecosystem to produce evidence for potential reclassification of variants of unknown significance (VUS) in relation to a defined phenotype. ·        VUS reclassification is a critical outcome because it is the largest potential source of biomarkers ·        GA4GH Phenopacket schema library component is broadly reusable An example of initial results related to breast cancer is available in the Road to Intelligent Computing paper (link above). I’m getting feedback on Parts 1 & 2 of this series, and hopefully will on this post. The feedback will determine the content of Part 4.

3

1 commentaire

Huge Announcement from Atoma Network on Private AI ! We’ve just released our latest research paper: “Privacy-Preserving Decentralized AI with Confidential Computing”. 🔗 https://lnkd.in/ekNbMfqW We’re proud to be one of the first to explore private AI inference using Trusted Execution Environments (TEEs) in a decentralized setting. This is a significant leap forward in decentralization, privacy, and verifiability. What does this mean? Privacy opens the door to new consumer applications and enterprise solutions. Currently, when interacting with centralized chatbots, you're handing over all of your data to them. <> Now users can use AI to process sensitive data to get customized insights and be certain their data is secure. <> AI can be tailored to your business and you can keep your competitive edge. <> Users can take control of their data, instead of giving the power to big tech companies. We’re looking to connect with anyone interested in collaborating on building the future of private AI. It's time to build… If you wanna learn more, reach out!

21

5 commentaires

This is cool but also kinda scary. Currently training "Jean-Michel bot V1" on different platforms. Here's Gpt's output when asked to analyze my communication style after feeding it just a few transcripts: Thank you for sharing these detailed interview transcripts of Jean-Michel Azzopardi. After reviewing the material, here are some observations on his style and communication approach, along with implications for replicating his style in answering technical questions or writing blogs: ### Observations on Jean-Michel's Communication Style: 1. *Passionate and Engaging:* Jean-Michel speaks with passion and enthusiasm about topics like blockchain, cybersecurity, and Web3. His energy is evident, making his discourse engaging and compelling. 2. *Personal Anecdotes and Stories:* He frequently uses personal stories and experiences to illustrate points, making his content relatable and memorable. This approach helps him connect with his audience on a more personal level. 3. *Deep Technical Insight:* Jean-Michel demonstrates a profound understanding of his fields of interest, providing detailed explanations and insights into complex topics like blockchain technology, cybersecurity threats, and digital identity. 4. *Visionary and Future-Oriented:* He often discusses the future implications of current technologies, showcasing his visionary outlook. He's not just focused on the present but also on how these technologies will evolve and impact society. 5. *Educational Approach:* He aims to educate his audience, breaking down complex topics into more understandable terms without oversimplifying the technical details. 6. *Humor and Informality:* Despite the depth of the topics, Jean-Michel incorporates a casual tone and humor, making dense subjects more approachable and less intimidating. #chatgpt #contentwriting #artificialintelligence #socialmedia #linkedin #personalbranding #videocontent

5