Issam El Alaoui

🚀 Mistral AI: Pixtral Large and Le Chat Platform Introducing Pixtral Large: Mistral AI, the French AI startup, has launched Pixtral Large, a groundbreaking 124-billion parameter multimodal model designed to process both text and visual data. - 🧠 Core Architecture: Combines a 123-billion parameter text decoder with a 1-billion parameter vision encoder. - 📖 Expanded Context Window: Supports up to 128K tokens, enabling it to handle 30 high-resolution images or a 300-page book. - 🏆 Benchmark Leader: Outperforms GPT-4o, Gemini 1.5 Pro, and Claude-3.5 Sonnet on key benchmarks like MathVista (69.4% accuracy), ChartQA, and DocVQA. Le Chat Platform: Unlocking New Possibilities Mistral’s Le Chat platform integrates Pixtral Large, introducing innovative features for diverse applications: -🔍 Web Search with Citations: Real-time data retrieval with source transparency. -🖋️ Canvas: Tools for live document creation, collaborative editing, and version control. -📄 Advanced OCR: Efficiently processes PDFs, tables, and equations to extract insights. - 🎨 Image Generation: Powered by Flux Pro from Black Forest Labs. - 🤖 Task Agents: Automates tasks like summarization and invoice processing. - 🌟 Performance Highlights Pixtral Large sets new standards on benchmarks like ChartQA, DocVQA, and VQAv2, showcasing cutting-edge capabilities in document analysis and visual data interpretation. The model is available under the Mistral Research License for non-commercial use, with commercial licenses for enterprise adoption. 📢 Additional Updates: Mistral AI also unveiled Mistral Large 2.1, a 123-billion parameter model optimized for general-purpose tasks, along with enhanced APIs to streamline developer integration.

6

This is an exciting comparison! 🚀 Both Claude 3.5 Sonnet and o1 pro represent cutting-edge advancements in AI, showcasing unique strengths tailored for different use cases. Let’s dive into the details: 🤖 Versatility in Task Handling Claude 3.5 Sonnet is a powerhouse for creative tasks, everyday coding, and long-form conversations. Its larger context window makes it ideal for handling extensive documents and multi-turn dialogues, offering speed and flexibility for general-purpose AI needs. 🔍 Advanced Reasoning with o1 pro o1 pro, on the other hand, is designed for deep reasoning and multi-step problem-solving. It excels in complex areas like advanced mathematics, scientific research, and intricate programming challenges, delivering precise and well-reasoned outputs. 💡 Specialized Strengths 1️⃣ Coding and Debugging: Claude is efficient for straightforward coding, but o1 pro’s advanced reasoning skills make it a better fit for challenging algorithmic tasks. 2️⃣ Mathematical Problem-Solving: o1 pro leads here, thanks to its ability to handle complex calculations and logical sequences. 3️⃣ Image-Based Reasoning: While Claude is capable, o1 pro provides enhanced performance, particularly for tasks requiring visual and contextual understanding. 🌟 Scenarios for Each Model Claude 3.5 Sonnet is perfect for content creation, marketing, and lightweight coding. Meanwhile, o1 pro shines in research-heavy environments, academic explorations, and tasks demanding analytical depth. Choosing between them depends on your specific requirements. 🔧 Innovation in AI Models The integration of advanced capabilities like image-based reasoning and multi-level problem-solving in o1 pro signals a step forward in AI evolution. These models are shaping the future of AI applications across industries. Thank you for sharing this comparison! 🙌 It’s exciting to see how models like Claude and o1 pro cater to diverse needs, pushing the boundaries of what’s possible with AI. #ClaudeAI #o1Pro #AIComparison #AdvancedAI #OpenAI #AIInnovation #FutureOfAI

3

You must know these 𝟯 𝗺𝗮𝗶𝗻 𝘀𝘁𝗮𝗴𝗲𝘀 of 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗮𝗻 𝗟𝗟𝗠 to train your own 𝗟𝗟𝗠 on your 𝗽𝗿𝗼𝗽𝗿𝗶𝗲𝘁𝗮𝗿𝘆 𝗱𝗮𝘁𝗮. # 𝗦𝘁𝗮𝗴𝗲 𝟭: 𝗣𝗿𝗲𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗶𝗼𝗻 You start with a bear foot randomly initialized LLM. This stage aims to teach the model to spit out tokens. More concretely, based on previous tokens, the model learns to predict the next token with the highest probability. For example, your input to the model is "The best programming language is ___", and it will answer, "The best programming language is Rust." Intuitively, at this stage, the LLM learns to speak. 𝘋𝘢𝘵𝘢: >1 trillion token (~= 15 million books). The data quality doesn't have to be great. # 𝗦𝘁𝗮𝗴𝗲 𝟮: 𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 (𝗦𝗙𝗧) 𝗳𝗼𝗿 𝗱𝗶𝗮𝗹𝗼𝗴𝘂𝗲 You start with the pretrained model from stage 1. This stage teaches the model to respond to the user's questions. For example, without this step, when prompted, "What is the best programming language?", it has a high probability of creating a series of questions such as: "What is MLOps? What is MLE? etc." As the model mimics the training data, you must fine-tune it on Q&A data to align the model to respond to questions instead of predicting the following tokens. After the fine-tuning step, it will respond to "What is the best programming language?" with "Rust". 𝘋𝘢𝘵𝘢: 10K - 100K Q&A example 𝘕𝘰𝘵𝘦: After aligning the model to respond to questions, you can further single-task fine-tune the model to specialize the LLM on a specific use case. # 𝗦𝘁𝗮𝗴𝗲 𝟯: 𝗥𝗟𝗛𝗙 Demonstration data tells the model what kind of responses to give but doesn't tell the model how good or bad a response is. The goal is to align your model with user feedback (what users liked or didn't like) to increase the probability of generating answers that users find helpful. 𝘙𝘓𝘏𝘍 𝘪𝘴 𝘴𝘱𝘭𝘪𝘵 𝘪𝘯 2: 1. Using the LLM from stage 2, train a reward model to act as a scoring function using (prompt, winning_response, losing_response) samples (= comparison data). The model will learn to maximize the difference between these 2. After training, this model outputs rewards for (prompt, response) tuples. 𝘋𝘢𝘵𝘢: 100K - 1M comparisons 2. Use an RL algorithm (e.g., PPO) to fine-tune the LLM from stage 2. Here, you will use the reward model trained above to give a score for every: (prompt, response). The RL algorithm will align the LLM to generate prompts with higher rewards, increasing the probability of generating responses that users liked. 𝘋𝘢𝘵𝘢: 10K - 100K prompts #machinelearning #mlops #datascience . 💡 Follow me for daily content on production ML and MLOps engineering.

128

4 Comments

Want to test Meta, Google, and Mistral models all on one platform ? **Groq.com** Why Groq? Because speed matters. GroqCloud supports a diverse array of models, enabling you to select the best for your needs: ➡ llama3-8b-8192 by Meta (Context Window: 8,192 tokens) ➡ llama3-70b-8192 by Meta (Context Window: 8,192 tokens) ➡ mixtral-8x7b-32768 by Mistral (Context Window: 32,768 tokens) ➡ gemma-7b-it by Google (Context Window: 8,192 tokens) Groq's commitment to GenAI inference speed is transforming real-time AI applications with its LPU. What is an LPU? ➡ Language Processing Unit, is a revolutionary processing system designed specifically for LANGUAGE applications. Unlike traditional CPUs (Central Processing Units) and GPUs (Graphics Processing Units), which handle a broad range of tasks. LPUs are optimized for the rapid processing demands of LLMs This specialized architecture allows for more efficient handling of sequential data, significantly reducing the time per word processed. ➡ That means you get faster outputs You can run Llama-2 70B at over 300 tokens per second per user!

24

🚀 𝐑𝐞𝐚𝐥-𝐓𝐢𝐦𝐞 𝐀𝐧𝐚𝐥𝐲𝐭𝐢𝐜𝐬 𝐃𝐚𝐬𝐡𝐛𝐨𝐚𝐫𝐝: 𝐁𝐢𝐠 𝐃𝐚𝐭𝐚 𝐢𝐧 𝐀𝐜𝐭𝐢𝐨𝐧! 🌐 As part of our Big Data module, my teammate Amine Saihi and I successfully developed a 𝐑𝐞𝐚𝐥-𝐓𝐢𝐦𝐞 𝐀𝐧𝐚𝐥𝐲𝐭𝐢𝐜𝐬 𝐃𝐚𝐬𝐡𝐛𝐨𝐚𝐫𝐝 using cutting-edge tools and technologies: 🔹 𝐁𝐢𝐧𝐚𝐧𝐜𝐞 𝐀𝐏𝐈 : to integrate live market data via WebSocket protocol.  🔹 𝐀𝐩𝐚𝐜𝐡𝐞 𝐒𝐩𝐚𝐫𝐤 & 𝐊𝐚𝐟𝐤𝐚 : for real-time streaming and data transformation.   🔹 𝐒𝐜𝐚𝐥𝐚 : for Kafka producers and Spark streaming transformations (with checkpointing).    🔹 𝐄𝐥𝐚𝐬𝐭𝐢𝐜𝐬𝐞𝐚𝐫𝐜𝐡 & 𝐊𝐢𝐛𝐚𝐧𝐚 : for efficient storage and powerful visualizations.   🔹 𝐃𝐨𝐜𝐤𝐞𝐫𝐢𝐳𝐞𝐝 𝐊𝐚𝐩𝐩𝐚 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 : for seamless containerized deployment.   🔹 𝐆𝐢𝐭 : for collaboration to streamline development and code versioning.   🔹 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐞𝐝 𝐰𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬 : Push code → Trigger CI/CD pipeline → Push to DockerHub for container image management and delivery. This project was a tremendous learning experience, combining the power of distributed systems with real-time analytics to build scalable, efficient, and user-friendly solutions. We invite you to explore the project and contribute on 𝑮𝒊𝒕𝑯𝒖𝒃 :   👉https://lnkd.in/eMEPvzQn A huge thank you to our professor Abdelmounaim Kerkri for his guidance and mentorship throughout this journey! 🙌 We are grateful for the opportunity to explore the world of big data and handling massive data, and for teaching us big data fundamentals such as Hadoop MapReduce, which have been instrumental in shaping our understanding of this field. #BigData #RealTimeAnalytics #ApacheSpark #Kafka #Elasticsearch #Kibana #Scala #Docker #Teamwork #BinanceAPI

140

21 Comments

I think Yann LeCun has been right, there is no world model inside transformers... -------------------------------------------- ***Comment with your best prompts below*** Original Prompt: "A perfectly executed open stance forehand on clay with a recovery step. The orientation is from the side perspective of the court with the athlete."

3

2 Comments

Kyutai's Moshi, ability to engage in natural, emotion-sensitive conversations represents a significant leap forward, enhancing user experiences across various fields such as healthcare and education. These advancements are not just technical improvements but steps towards creating more accessible and inclusive tools that empower users and drive further innovation. As AI continues to evolve, it promises to reshape a lot in our world significantly, fostering smarter applications and more intuitive human-machine interactions. I'm excited😃 for what the future holds. #dk #davidkariuki #ai #artificialintelligence

6

Aya Expanse 8B and 32B models are here. They set a new standard for the quality of AI interaction in 23 languages. It took several multilingual breakthroughs to get to this stage: * Our research lab Cohere For AI used data arbitrage - a new method for synthetic data generation, in which the model samples from multiple “teacher” models to generate high-quality data for specific languages. * In the final stages of training, they applied preference training with DPO and RLOO models toward multilingual performance. It was tested against leading open-weights models. * Finally, they combined the weights of top-performing model candidates at each stage, improving both versatility and accuracy across languages. Excited for you to check them out!

77

5 Comments

This is how you can parse unstructured text *locally* and *cheaply* (no GPU required) in less than 5 min using the latest models from NuMind (YC S22) Many thanks to your team Etienne Bernard 👏 How to (5 min read) : https://lnkd.in/e2sbGnzn

3

Kyutai, a french AI lab with $300M in funding, just unveiled Moshi, an open-source GPT-4o competitor. Moshi is a real-time multimodal model that can listen, hear, and speak. - Code, model, and paper will be release soon. - Demo here (for US): https://lnkd.in/gVX-DJSk ↓ Check out https://AlphaSignal.ai to get a daily summary of breakthrough models, repos, and papers in AI. Read by 180,000+ engineers and researchers.

1,195

61 Comments

Very solid mathematical foundation of non-linear RL using Calabi–Yau space and  Ricci-flat metric!

9

😎 OG ML engineers! Remember the Theano era? Here is my ML origin story 𝗝𝗼𝗶𝗻𝗶𝗻𝗴 𝘁𝗵𝗲 𝗠𝗟 𝗵𝘆𝗽𝗲 𝘁𝗿𝗮𝗶𝗻 Picture this: it's around 2016, and I'm a Bachelor student in Communication Systems at EPFL. As graduation looms, I realize I need to complete a Bachelor project to seal the deal. At the time, machine learning and data science were gaining serious traction, and I couldn't help but take notice. That's when I stumbled upon a project that seemed like a match made in heaven: 𝗳𝗮𝗰𝗲 𝗿𝗲𝗰𝗼𝗴𝗻𝗶𝘁𝗶𝗼𝗻 𝗱𝘂𝗿𝗶𝗻𝗴 𝗯𝗿𝗼𝗮𝗱𝗰𝗮𝘀𝘁 𝘀𝗽𝗼𝗿𝘁 𝗲𝘃𝗲𝗻𝘁𝘀 (basically detecting and identifying players faces). 😅 𝘛𝘩𝘦 𝘤𝘢𝘵𝘤𝘩? 𝘛𝘩𝘦 𝘱𝘳𝘰𝘫𝘦𝘤𝘵 𝘥𝘦𝘴𝘤𝘳𝘪𝘱𝘵𝘪𝘰𝘯 𝘮𝘦𝘯𝘵𝘪𝘰𝘯𝘦𝘥 𝘵𝘩𝘢𝘵 𝘢𝘱𝘱𝘭𝘪𝘤𝘢𝘯𝘵𝘴 𝘴𝘩𝘰𝘶𝘭𝘥 𝘩𝘢𝘷𝘦 𝘴𝘰𝘮𝘦 𝘦𝘹𝘱𝘦𝘳𝘪𝘦𝘯𝘤𝘦 𝘪𝘯 𝘮𝘢𝘤𝘩𝘪𝘯𝘦 𝘭𝘦𝘢𝘳𝘯𝘪𝘯𝘨. 𝗧𝗶𝗺𝗲 𝘁𝗼 𝗮𝗯𝗮𝗻𝗱𝗼𝗻 𝘁𝗵𝗲 𝘀𝗵𝗶𝗽? 𝗻𝗼𝘁 𝘆𝗲𝘁! I applied for the project and started devouring everything I could find on the subject, determined to demonstrate at least a bit of experience. Neural networks, recurrent neural nets, autoencoder, you name it, I read about it. By some miracle (and maybe a bit of charm 😉), I managed to secure an appointment with the lab. And guess what? They gave me the Bachelor project! (To be fair, they were pretty cool and didn't dig too deep into my so-called "experience" in the domain.) With the project set to start after the winter break, I knew I had to level up fast. So, what did I do? I spent my entire holiday break following the famous Andrew Ng's course on Coursera. 𝗚𝗲𝘁𝘁𝗶𝗻𝗴 𝗺𝘆 𝗵𝗮𝗻𝗱𝘀 𝗱𝗶𝗿𝘁𝘆 When the semester (and the project) began, I found myself using the 𝗧𝗵𝗲𝗮𝗻𝗼 framework to implement an autoencoder architecture with triplet loss. The goal was to take two different images of the same athlete, use the autoencoder to build embeddings, and then train a distance model that would give a low number for the same person's face and a high number for different faces. I won't lie, it was a 𝘁𝗼𝘂𝗴𝗵 𝗳𝗶𝗿𝘀𝘁 𝗠𝗟 𝗽𝗿𝗼𝗷𝗲𝗰𝘁, but I managed to get some interesting results in a restricted context: the model worked well on a subset of players and high-quality image samples with clear faces and minimal motion blur. 𝗖𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝗺𝘆 𝗠𝗮𝘀𝘁𝗲𝗿'𝘀 𝘁𝗼𝗽𝗶𝗰 Inspired by my Bachelor project, I decided to pursue my studies in machine learning. As luck would have it, EPFL had just introduced the Data Science master's program. I jumped at the opportunity, essentially becoming a beta tester for the program. 😎 𝗗𝗮𝗿𝗲 The moral of the story? Grab opportunities even if you don't feel 100% ready. You'll figure it out along the way! This is especially true when applying for job, don't be afraid to overshoot. It's a fantastic way to 𝗽𝗿𝗼𝗴𝗿𝗲𝘀𝘀, as it will put you in situations where you have no choice but to rise to the challenge. 💡 That's my ML origin story! What's yours? #career #machinelearning

23

3 Comments

What is the secret sauce behind Meta's LLaMa-3 ? And what are the technologies used by Meta in it's new Meta AI Assistant ? 🤔 🔍 𝗗𝗮𝘁𝗮 𝘂𝘀𝗲𝗱 𝗳𝗼𝗿 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗟𝗟𝗮𝗠𝗮 𝟯 Meta's LLaMa-3 marks a significant leap forward in AI capabilities, driven by meticulous data curation and massive training datasets. With over 𝟭𝟱 𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻 𝘁𝗼𝗸𝗲𝗻𝘀 of pretraining data, LLaMa-3 dwarfs its predecessor, LLaMa-2, by a factor of 𝟳 !! 🤯 Meta's technical report highlights a relentless focus on data quality, including a heightened emphasis on code examples for enhanced reasoning capabilities. The team invested significant time and resources in filtering and studying data sources 🔍 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝘄𝗶𝗻𝗱𝗼𝘄 𝘀𝗶𝘇𝗲 𝗳𝗼𝗿 𝗟𝗟𝗮𝗠𝗮 𝟯 One of the remarkable features of LLaMa-3 is its expanded context window size, facilitating richer interactions and deeper understanding. Across its variants, including the 8B and 70B parameter models, LLaMa-3 boasts an impressive context length of 8K tokens (double the context length of LLaMa-2) 🔍 𝗠𝗲𝘁𝗮 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 The introduction of the Meta AI Assistant brings forth a new dimension of AI-powered capabilities to messaging platforms like Messenger and WhatsApp.  * Users can now seamlessly generate high-resolution images from text prompts using Meta's 𝗘𝗺𝘂 image generation model, akin to the capabilities of OpenAI's 𝗗𝗔𝗟𝗟-𝗘. * Additionally, the integration of 𝗟𝗟𝗮𝗠𝗮-𝟯-𝗯𝗮𝘀𝗲𝗱 𝗰𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗔𝗜 allows for engaging and dynamic interactions within messaging environments. Whether it's generating visual content or engaging in natural language conversations You can try out image generation / chat with LLaMa-3 here 👇 : 🔗 https://www.meta.ai/ #meta #ai #llama3 #imagine #image_generation #metaai #assistant #dalle #emu #chatgpt #whatsapp #messenger

2

📢 Hugging face is directly integrated with GenAI developers favorite tool "langchain"🦜 Get set go!🚀 💡 pip install langchain-huggingface https://lnkd.in/ddHreVrs #huggingface #GenerativeAI #langchain

22

🚀 Excited to share my latest project: DocTalk 🤖📚! This really isn't anything revolutionary. It's a simple, but elegant (in my opinion) implementation of what is probably currently the most common business application of Generative AI and Large Language Models. DocTalk is a Streamlit-based web application that allows you to upload and interact with your documents (PDF, DOCX, TXT) using natural language queries. Leveraging OpenAI's GPT-4o model, it provides context-aware responses based on the content of your documents. By using a Retrieval-Augmented Generation (RAG)-style prompting architecture, model responses are grounded in the uploaded data as direct context, rather than relying solely on model pre-training data to get inference. This aims to reduce model hallucinations. In particular, one feature I'm proud of is what I'm calling a 'query validation' step. This is a functionally secondary - but actually runs prior to the main chatbot - API call to classify whether the user's query should require document context to generate a response. One downside of using a RAG architecture is that there is often an excess of tokens in each prompt that are not needed for the model to respond. In these cases, this 'query validation' step aims to minimize input token waste (cost), ultimately reduce chat response time (fewer input tokens to chatbot in total), and has the additional benefit here of the app seeming more 'intelligent', only providing response context citations when appropriate to do so. Check out the project on GitHub (app screenshots included in the README): https://lnkd.in/gffTMiCa Explore the live app: https://lnkd.in/g3cWTs5a What should I build next? Let's discuss. #AI #NLP #Streamlit #OpenAI #MachineLearning #DataScience #DocTalk

37

4 Comments

Mistral 7 seems "smarter" than the latest Llama 3 at understanding instructions 🤔 I run 3 tests designed to represent real-life implementations in end-user applications : 1/ Question answering on a topic without any context. This is designed to assess the ingrained knowledge compressed into the model and its ability to hallucinate or self-censor its answers. 2/ Parsing unstructured text into a constraint structure This test is very useful to process written content coming from basically anywhere (CRM, emails, documents, code, audio transcripts, ...) 3/ Generating structured content based on misc sources This test is designed to assess the ability to both produce relevant answers from a complex context *and* stay within format contraints The preliminary results from llama-3 are somehow disturbing as the model does not seem to understand as well as mistral 7B. This is surprising given the benchmark scores... I'll share what I find...

3

3 Comments