Juantomás García Molina

Earlier this week Meta dropped their new Llama iteration, Llama 3.1, with the 405b model achieving comparable levels to closed model competitors (Claude 3.5 Sonnet and GPT-4). This was then followed by a release from our favorite frenchies at Mistral, Mistral Large 2, on par with Llama 3.1, making it two open models matching closed models' performance. As strong advocates for open source models, they tend to pose an issue to closed alternatives. Does OpenAI still have a moat if open, less costly versions achieve the same specs? The company is reportedly running out of cash and has seen the shipment of their worthwhile features delayed. However, they still have cards to play. Earlier today they announced the introduction of their own search engine (search gpt) to take on Google and Perplexity. From the Elonverse side of things, Tesla's computing abilities are still ramping up and the company is in talks to invest $5 billion in xAI. If not in a ring, Elon and Zuck seem to have found their battlefield...

3

Take your first steps with #ApacheKafka and Flink on Confluent! In this video of the #DataAICommunity, Òscar del Álamo introduces how to build #DataEngineering stream pipelines and queues with Apache Kafka and how to query and process this data in real-time using #ApacheFlink. Plus, you'll learn how to easily deploy an instance on the cloud with the Confluent platform. Discover how to do it and start experimenting with Confluent's free entry-level plans! Watch the video now 🎥👇🏻 #data #betterjoinERNI #betteraskERNI

10

What exactly is AI alignment? You can find the answer to that question and many more on Codex Presents AGI! Full Episode on YouTube: https://lnkd.in/e4f_8xB6 #agi #ai #genai #podcast #techtalk #codex

9

🚀 A new read for my Menorca vacation! 🏖️ I just received an advance copy of "Building AI Applications with Microsoft Semantic Kernel" by Lucas A. Meyer, courtesy of Packt! 📚✨ As I soak up the sun in this beautiful paradise, I can’t wait to dive into this promising book and explore the depths of Microsoft’s Semantic Kernel. 🌞📖 Having worked extensively with Python’s LangChain, I’m eager to see how Semantic Kernel stacks up, especially since it’s available in both .NET and Python. For someone like me who thrives in the .NET ecosystem, this could be a game changer! 💻🔧 Stay tuned, I’ll be sharing my thoughts and insights as I delve into the intricacies of this powerful tool. #SemanticKernel #DotNet #Python #AI #MachineLearning

34

8 comentarios

Many discuss AI and innovation, but the conversation often overlooks the practical implementation of AI-based systems. To address this critical aspect, my colleagues Roberto Hernández Ruiz and Ferran Aran Domingo have launched the first post in a series dedicated to exploring the productionalization of AI-based solutions. GFT Technologies #artificialIntelligence #solutions https://lnkd.in/dr9qDDwH

43

2 comentarios

"To ingest the incremental data from Kafka to S3, we considered three excellent data lake and ingestion solutions: Apache Hudi, Apache Iceberg, and Delta Lake. In the end we chose Hudi for its excellent performance with our update-heavy workload and its open-source nature & native integration with Debezium CDC messages." Excellent blog by Notion's data platform team on how they built their data lake (house) to meet the surge in 10x data growth due to user & content growth to cater to AI, Search, and other product use cases. 🔗 Blog: https://lnkd.in/dEYxAZDQ #dataengineering #softwareengineering

62

1 comentario

Available in Langflow: Dynamic Model Selection 🧠

12

2 comentarios

Interesting post by João Salgado about partitioning vs. sharding in #Postgres which lists trade-offs for each. It's great looking at the list of cons for partitioning and seeing that #TimescaleDB actually helps with most of them - Complexity ⬇️ because we transparently manage partition creation and lifecycle for you as data is ingested - Optimization Challenges ⬇️ because we have planner hooks that help with handling data exclusion (you still do need to pay attention to your partitioning key though). - Storage Overhead ⬇️ (although Postgres doesn't really have much overhead for partitions) because you can just choose to compress your data (we often see a 95x reduction in size). Timescale 🚀

18

1 comentario

📣 New notebook contribution Fine-Tune Google's Gemma on Vertex AI with Hugging Face's DeepLearning container (DLC)! ⛳ Want to customize Google's Gemma to your specific needs using Vertex AI on Google Cloud? 🔥 This new notebook from Álvaro Bartolomé del Canto, Simon Pagezy and the Hugging Face team walks you through fine-tuning Gemma with SFT & LoRA on Vertex AI, all within a single GPU environment using Hugging Face's DeepLearning container (DLC). 🔥 🔗 Dive into the code: https://lnkd.in/dugfrP5G 🤗 We are working together to make Vertex AI a great place to customize and deploy your open models! --- If you want to know more on how to train & deploy Hugging Face models on Vertex AI, check out the documentation  👇 Also stay tuned 🔔 ! More content about open models on Vertex AI is coming 😉 And if you find this post helpful 🔥, like, share and let’s connect if you have questions 🤗 #GoogleCloud #VertexAI #HuggingFace #Gemma #FineTuning #MachineLearning #AI #NLP #PyTorch

83

3 comentarios

Structure you "unstructured data" with Nuclia • AI Search & RAG as a Service and JSON Output https://lnkd.in/dTHZbgtb

8

The next generation of LLMs will arrive later this year. As a reference, Llama 3 by Meta was trained on 25.000 H100’s. Also, of course training started at 4:20 https://lnkd.in/dyYZqz5n)

10

Currently part of the Virtual Protocol's 15 - Day Webscraping Challenge with Shikaniah Castillo! Learning how to finetune LLMs is just merely scratching the surface when it comes to creating this Generation's Top AI Girlfriend/Boyfriend XD. One of the most underrated realizations that we had in training LLMs is that Webscraped Data with a few tweaks from chatgpt is often better than datasets available in kaggle and Huggingface in terms of training LLMs to embody fictional characters. As the top Challengers from the competition, here is our guide for other competitors and people towards how we used Beautifulsoup 4 to webscrape a D&D Wiki to gather data about official D&D Spells to possibly create a AI Dungeon Master. https://lnkd.in/gykHgWM8 #medium #webscraping #beautifulsoup #python #pandas #numpy #datastructures #ai #machinelearning #llms #chatgpt #openai #huggingface #algorithms #article #tech #aigirlfriend

5

💥💥💥 Large Concept Models: Language Modeling in a Sentence Representation Space The LCM team, Loïc Barrault, Paul-Ambroise Duquenne, Maha Elbayad, Artyom Kozhevnikov, Belen Alastruey, Pierre Andrews, Mariano Coria, Guillaume Couairon, Marta R. Costa-jussà, David Dale, Hady Elsahar, Kevin Heffernan, João Maria Janeiro, Tuan Tran, Christophe Ropers, Eduardo Sánchez, Robin San Roman, Alexandre Mourachko, Safiyyah Saleem, Holger Schwenk Abstract LLMs have revolutionized the field of artificial intelligence and have emerged as the de-facto tool for many tasks. The current established technology of LLMs is to process input and generate output at the token level. This is in sharp contrast to humans who operate at multiple levels of abstraction, well beyond single words, to analyze information and to generate creative content. In this paper, we present an attempt at an architecture which operates on an explicit higher-level semantic representation, which we name a concept. Concepts are language- and modality-agnostic and represent a higher level idea or action in a flow. Hence, we build a "Large Concept Model". In this study, as proof of feasibility, we assume that a concept corresponds to a sentence, and use an existing sentence embedding space, SONAR, which supports up to 200 languages in both text and speech modalities. The Large Concept Model is trained to perform autoregressive sentence prediction in an embedding space. We explore multiple approaches, namely MSE regression, variants of diffusion-based generation, and models operating in a quantized SONAR space. These explorations are performed using 1.6B parameter models and training data in the order of 1.3T tokens. We then scale one architecture to a model size of 7B parameters and training data of about 2.7T tokens. We perform an experimental evaluation on several generative tasks, namely summarization and a new task of summary expansion. Finally, we show that our model exhibits impressive zero-shot generalization performance to many languages, outperforming existing LLMs of the same size. The training code of our models is freely available. 👉 https://lnkd.in/deMs-WWr #machinelearning

3

1 comentario