Opening Keynotes - AIE Paris 2025 (Day 1)

Key Concepts:

• AI Agents: LLMs with tools in a loop, directed by a goal.
• Agentic Models: Models that "think" with tools, exhibiting increasing autonomy.
• MCP (Model Communication Protocol): A protocol for agent communication.
• LLM OS: The infrastructure and tools needed to build and deploy AI agents.
• Context Engineering: Managing and utilizing context effectively for AI models.
• Fast Agents: AI agents optimized for speed and low latency.
• AI Governance: Policies and practices for managing AI systems and data.
• Data Silos: Isolated data repositories within an organization.
• AI Flywheel: A continuous cycle of building, deploying, observing, and improving AI systems.

1. Introduction and Welcome (Raou Jabri, MC):

• Raou Jabri welcomes attendees to AI Engineer Paris, emphasizing the event's focus on learning from European and international AI experts.
• He highlights Paris as a hub for AI innovation, citing Mistral AI, Black Forest Labs, and Qout.
• The event builds on the success of previous AI Engineer events in San Francisco and New York.
• He thanks platinum sponsors Neo4j and Docker, gold sponsors Sentry, Arise AI, Deep Mind, and Alolia, and other sponsors and partners.
• Attendees are encouraged to visit the expo to network and explore opportunities.
• He mentions the welcome party and encourages attendees to get their tickets for a free drink.

2. Community and Global Movement (Benjamin Dumpy, Co-founder of AI Engineer):

• Benjamin Dumpy expresses his desire to extend the AI Engineer brand beyond self-produced events.
• He emphasizes the importance of partnering with organizations that share the same vision, drive, and professionalism.
• He welcomes attendees as fellow members of a global community.
• He thanks KOYB for partnering to organize the event.

3. AI in Paris and COYB's Role (Yan Leger, Co-founder and CEO at COYB):

• Yan Leger welcomes attendees to Paris, the home city of COYB.
• He highlights the unique nature of the event as the first AI Engineer conference organized outside of the US.
• He mentions the diverse audience, with 70% coming from outside of France.
• He introduces the lineup of over 30 speakers discussing foundational models, MCP deployment, and coding experiences.
• He mentions Leelio from Mistral AI as a speaker.
• COYB provides high-performance serverless infrastructure for AI applications, focusing on deploying agents and inference services across various hardware.
• He thanks the COYB team and the AI Engineer team.

4. Station F and the Parisian AI Ecosystem (Marwan, Head of Startups at Station F):

• Marwan welcomes attendees to Station F, a massive startup campus in Paris.
• He highlights Station F's role in gathering talented people and creating connections.
• He mentions that Station F hosts 1,000 startups participating in 30 programs, including AI.
• He congratulates COYB as one of the top startups at Station F.
• He notes that 70% of Station F startups have AI components.
• He mentions Hugging Face as a successful Station F alumni.

5. State of AI Agents (Swix, Co-founder of AI Engineer):

• Swix provides an update on the state of AI agents, focusing on the "year of agents."
• He references predictions of 2025 as the year of agents, citing Satya Nadella, Roman, Greg Brockman, and Sam Altman.
• He discusses the evolving definition of agents, referencing Simon's definition: LLM with tools in a loop, directed by a goal.
• He emphasizes the importance of intent, memory, planning, authority, control flow, and tools in agent engineering.
• He highlights the "epic infrastructure buildout" with hundreds of billions of dollars being invested.
• He notes that ChatGPT is projected to reach 1 billion users quickly.
• He references Andrej Karpathy's view that the next 10 years will be focused on building agents.
• He presents a slide of major model launches, highlighting Mistral, Quinn 3, Coder GM 4.5, and Frontier Labs.
• He describes agentic models as "thinking with tools," referencing a piece on GT5 developer preview.
• He notes the increasing autonomy of agents, citing Replit's agent launch with 200 minutes of autonomy.
• He mentions the increasing use of RL (Reinforcement Learning) in LLMs, with similar compute allocation for post-training as pre-training.
• He discusses agent products, agent protocols, and agent labs.
• He lists notable agents and highlights the rapid evolution of the agent field.
• He emphasizes the importance of agent protocols, particularly MCP.
• He mentions Google's A2A and Zed's ACP (Agent Client Protocol) as interop layers for terminal agents.
• He discusses the emerging consensus around agent APIs with standard libraries of tools like code execution sandboxes, web search, and document libraries.
• He updates Karpathy's LLM OS thesis, identifying search, code execution, document library, multimodal input/output, and MCP as key components, with memory and orchestration still missing.
• He notes the explosion of code agent labs and references his blog post on cognition and code AGI.
• He announces an AI Engineer summit focused on coding agents in New York in November.
• He raises open debates:
  • Do we need evals? He references Boris from cloud code saying that evals are not necessary.
  • How to do context engineering very well? He references the multi-agent debate and the lack of a standard way of viewing context engineering.
  • Fast agents: He highlights Cerebra's code as an example of fast agents with 2,000 tokens per second.
• He predicts more development in email clients, browser agents, voice calling, vibe coding, low code, and education agents.

6. Bringing AI to the Enterprise (Leelio Lavo, Head of Engineering at Mistral AI):

• Leelio Lavo introduces Mistral AI, founded by scientists with the goal of bringing AI to the enterprise world and advancing open source models.
• He mentions the release of Mistral 7B and Mistral 8*7B, which triggered experimentation and new usages in the community.
• He highlights the company's focus on foundational models, Mistral Studio (API wrapper), and custom solutions.
• He discusses the challenges enterprises face in adopting AI:
  • Data: Massive, unstructured data in silos with poor AI governance.
  • Observability: The "black box" phenomenon and the need for trust and safety.
  • Skills Gap: Lack of internal and external expertise.
  • Model Performance: Issues with model alignment and reliability.
• He emphasizes the need for a strong context engine to understand the link between entities in different data silos.
• He discusses the importance of observing how people use AI to understand workflows and automate tasks.
• He highlights the need to improve models by collecting data and optimizing them for specific tasks.
• He discusses the importance of model performance, expertise, and change management in achieving AI maturity.
• He emphasizes that AI is more than just a tool; it's a capacity that can transform organizations.
• He discusses how AI can transform organizations through growth (new revenue) and efficiency (cost reduction).
• He summarizes the key takeaways:
  • Build AI your way: Customize AI to your use cases.
  • Leverage the community: Utilize open source and community contributions.
  • Break barriers: Get data where it is and avoid vendor lock-in.

7. Q&A with Leelio Lavo:

• Raou asks about how Mistral helps enterprises structure data. Leelio responds that AI can be used to parse and classify data silos, creating more AI projects.
• Raou asks about balancing building for enterprise customers with an open source focus. Leelio responds that pre-trained models will become a commodity and open source, but post-training and continuous pre-training require expertise and compute power that many customers lack.

8. Closing Remarks (Raou Jabri, MC):

• Raou thanks Leelio Lavo for his presentation.
• He announces the welcome party and encourages attendees to attend.

Synthesis/Conclusion:

The AI Engineer Paris event aims to foster collaboration and knowledge sharing within the AI engineering community. Key themes include the evolution of AI agents, the importance of infrastructure and tooling, the challenges and opportunities in bringing AI to the enterprise, and the need for a holistic approach that considers data governance, model performance, and human expertise. The event highlights the rapid pace of innovation in the AI field and the potential for AI to transform various industries.