From OpenAI to Opensource AI: Navigating Between Commercial Ownership and Collaborative Openness en

This presentation explores the evolution of generative AI, highlighting the trajectories of various models such as GPT-4, and examining the dynamics between commercial interests and the ethics of open collaboration. We offer an in-depth analysis of the levels of openness of different language models, assessing various components and aspects, and exploring how the (de)centralization of computing power and technology could shape the future of AI research and development.
Additionally, we explore concrete examples like LLaMA and its descendants, as well as other open and collaborative projects, which illustrate the diversity and creativity in the field, while navigating the complex waters of intellectual property and licensing.

- Talk presented at State of Open 2024 in London, UK: https://www.youtube.com/watch?v=tVTtyLuiYQk
- As a keynote at the "I LOVE Tech" event in Timisoara, Romania: https://www.slideshare.net/slideshows/unlocking-ainavigating-open-source-vs-commercial-frontiers/266821723
- At AI_Dev Europe Summit in Paris, France: https://sched.co/1c1mU

Vers une re-decentralisation d'Internet : panorama des technos et initiatives actuelles en

Je vous invite à un voyage captivant à travers les initiatives et technologies émergentes (telles que celles du Web3, du Web3.0, et du Fediverse) qui pourraient façonner un avenir d'Internet décentralisé. Découvrez comment vous pouvez vous impliquer dans ce mouvement en faveur de la (re)decentralisation du Web et d'Internet. Rejoignez-nous pour explorer les opportunités offertes par ces avancées, et contribuez à façonner un Internet où le contrôle des données est entre les mains des utilisateurs.

The hitchhiker's guide to software architecture and design en

Designing a new platform is always tricky to set up.

How to start? What is the best strategy to adopt while designing a platform? What kind of architecture should we deploy: event streaming, orchestration, or choreography?

For a brand-new platform: "Donut @ Home", we will proceed a live architecture study.

After analysing the customer needs, brainstorming, and exchanging our ideas, we will choose among all the potential solutions the *least worst* option.

You will be asked to validate our design and the different implementation examples.

At the end of this talk, you will have tips and tricks for thinking about it and starting working on architecture studies in complete peace of mind.

Why should we say Internet3 rather than Web3? Let's demystify blockchains and decentralized apps! en

Do words like Blockchain, proof-of-stake, smart contract, dApp, token, NFT, DAO or Web3 leave you perplexed, dizzy or even annoy you? Come and discover the concepts and structuring components that are hidden behind it all!

Let's take a step back from buzzwords and speculative bubbles and focus on what is at stake in the background: the redecentralization of the Internet (and not just the Web!). This presentation is the map that I would have liked to have when starting my lonely exploration of the enormous Web3 ecosystem.

Its objective is not to feed endless debates on the usefulness or business value of these technologies and usages, but rather to better understand the ongoing effort to rebuild the Internet and give back control to users over their data. You'll the be able to make your own opinion on the future of all this!

This presentation is a didactic effort to introduce a complex and bubbling ecosystem. It offers a visual and progressive mapping of Web3 (or decentralized Internet) technical stacks in order to position concepts and technologies. It aims to share enough background to be able to step back and see the global picture, without going too deep in technical details.

Dialoguer avec vos documents : découvrez la magie du RAG avec et sans OpenAI fr en

Découvrez le monde fascinant du Retrieval-Augmented Generation (RAG), une avancée révolutionnaire dans les Large Language Models (LLMs).

Rejoignez-nous pour un voyage pratique à travers l'implémentation du RAG avec OpenAI et des solutions Open Source alternatives.

Apprenez comment maîtriser les embeddings et vectorstores, et déverrouillez les secrets d'une mise en œuvre concrète adaptée à vos besoins.

Ce que les participants vont apprendre
- Compréhension approfondie des concepts clés du RAG : embeddings, vecteurs et vectorstores.
- Compétences pratiques pour implémenter le RAG avec des solutions commerciales ou Open Source.

Chat with your Documents: Discover the magic of RAG with and without OpenAI fr en

Discover the fascinating world of Retrieval-Augmented Generation (RAG), a revolutionary advancement in Large Language Models.
Join us for a practical journey through the implementation of RAG with OpenAI and alternative opensource solutions like Hugging Face models, Ollama, Chroma, Flowise, AnythingLLM, RAGFlow and Dify.
Learn how to master embeddings and vector stores, and unlock the secrets of a concrete implementation tailored to your needs.

What audience will learn:
- In-depth understanding of key RAG concepts: embeddings, vectors, naive versus advanced RAG techniques
- Practical skills to implement RAG using commercial or open-source solutions

Conference given in French at the DevDay 2024 event in Belgium with positive feedbacks. Slide are available here (but without the many demos): https://www.slideshare.net/slideshow/devday-2024-dialoguer-avec-vos-documents-decouvrez-la-magie-du-rag-avec-et-sans-openai/273102671

The Future Beyond LLMs: Exploring Agentic AI en fr

At the recent Vivatech conference, Yan LeCun, Chief AI Scientist at Meta, declared that Generative AI has only five years left to live. He argued that Large Language Models (LLMs) are not the future of artificial intelligence due to their limitations in four key areas: understanding the real world, maintaining persistent memory, reasoning, and planning.

Given this prognosis, let's take a comprehensive look at the budding field of Agentic AI and its future prospects. What exactly are AI Agents, and how do they operate? How do they compare to and interact with LLMs and functionalities such as function calling, chain-of-thought processing, assistants, tools, or actions?

In this talk, we'll delve into the unique features of Agentic AI, including perception, state estimation, goal setting, planning, and action selection & execution. We will define various levels of Agentic AI and form a map to help navigate this emerging landscape. By categorizing current agent-based or agent-related solutions with practical examples, we'll provide an overview of the current state of Agentic AI:
- LLM with Function Calling and Tools (OpenAI's GPT-4 or Gemini), these models will serve as a base reference to illustrate current capabilities and limitations in function calling, tool usage, and action execution.
- Agent Framework like LangGraph, AutoGen or CrewAI.
- LAM Framework like LaVague, designed for developing AI Web Agents and facilitating Web Automation.

Prepare to uncover how today's Large Language Models (LLMs) are being complemented —and perhaps even superseded— by Large Actions Models (LAMs) in the near future.

Agentic AI is a rapidly evolving field with new breakthroughs emerging constantly. My talk aims to bring much-needed clarity to this innovative domain by defining what AI agents and Large Action Models are, and how they compare to existing Large Language Models.

By providing structured insights and real-world examples, I'll help the audience understand the current state and future potential of Agentic AI. I will ensure that my presentation remains up-to-date, incorporating any new breakthroughs that may arise before the event.

Selecting this talk will equip attendees with a clear roadmap to navigate and leverage the next generation of AI technologies.

L'avenir au-delà des LLM : Explorons l'IA Agentique en fr

Lors de la récente conférence Vivatech, Yan LeCun, Scientifique en Chef de l'IA chez Meta, a déclaré que l'IA Générative n'a plus que cinq ans à vivre. Il a soutenu que les Modèles de Langage de Grande Taille (LLMs) ne sont pas l'avenir de l'intelligence artificielle en raison de leurs limitations dans quatre domaines clés : la compréhension du monde réel, la mémoire persistante, le raisonnement et la planification.

Étant donné ce prognostic, examinons de manière approfondie le domaine en plein essor de l'IA Agentique et ses perspectives d'avenir. Que sont exactement les agents IA et comment fonctionnent-ils ? Comment se comparent-ils et interagissent-ils avec les LLM et des fonctionnalités telles que l'appel de fonctions, le traitement en chaîne de la pensée, les assistants, les outils ou les actions ?

Dans cette présentation, nous examinerons les caractéristiques uniques de l'IA Agentique, y compris la perception, l'estimation d'état, la définition d'objectifs, la planification et la sélection et exécution d'actions. Nous définirons les différents niveaux de l'IA Agentique et formerons une carte pour aider à naviguer dans ce paysage émergent. En catégorisant les solutions basées sur des agents ou liées aux agents actuelles avec des exemples pratiques, nous fournirons un aperçu de l'état actuel de l'IA Agentique :
- LLM avec appel de fonctions et outils (comme le GPT-4 d'OpenAI ou Gemini), ces modèles serviront de référence de base pour illustrer les capacités et les limitations actuelles dans l'appel de fonctions, l'utilisation d'outils et l'exécution d'actions.
- Frameworks multi-agents comme LangGraph, Autogen ou CrewAI.
- Framework LAM comme LaVague, conçu pour développer des agents Web IA et faciliter l'automatisation du Web.

Préparez-vous à découvrir comment les Modèles de Large Language Models (LLMs) d'aujourd'hui sont complétés — et peut-être même remplacés — par les Large Action Models (LAMs) dans un avenir proche.

L'IA agentique est un domaine en évolution rapide, avec de nouvelles percées émergentes en permanence. Mon intervention vise à apporter la clarté nécessaire à ce domaine innovant en définissant ce que sont les agents IA et les LAMs, et en les comparant aux LLMs existants.

En fournissant des perspectives structurées et des exemples concrets, j'aiderai le public à comprendre l'état actuel et le potentiel futur de l'IA agentique. Je veillerai à ce que ma présentation reste à jour, en intégrant toutes les nouvelles avancées qui pourraient survenir avant l'événement.

Choisir cette intervention équipera les participants d'une feuille de route claire pour naviguer et tirer parti de la prochaine génération de technologies IA.

Mastering Image Generation with Stable Diffusion en

Unlock the power of Stable Diffusion for controlled image generation with an engaging, hands-on session led by Hajer Mabrouk and Raphaël Semeteys.
This talk will explore how generative AI can produce accurate, diverse, and inclusive images of yoga poses—essential for use cases where stock images or internet-sourced photos fall short.
By leveraging ComfyUI’s modular workflows, embeddings, inpainting, ControlNets, LoRA and FaceID for style and pose control, you will discover how to generate precise images with ease. From text-to-image generation to hybrid image-to-image workflows, we’ll showcase how to overcome key challenges like pose consistency and identity preservation.
The session will include live demos, practical tips, and real-world insights from the development of YogĀrkana, a dedicated yoga website.
You will leave with actionable knowledge on how to customize Stable Diffusion workflows to meet their own creative, commercial, or inclusive design needs.

Discover how to generate precise, diverse, and inclusive images using Stable Diffusion and ComfyUI. This session will show you how to control poses, styles, and character consistency for use cases like yoga pose visualization. Learn to master text-to-image and image-to-image workflows, leverage LoRA for lightweight customization, and create visually rich, inclusive image libraries. With live demos and actionable insights, this session will inspire your next creative AI project.