AI Engineer Paris 2026 Main Stage: Google DeepMind, ElevenLabs, Hugging Face & Stripe | Day 2

Published 2026-09-24 · Duration 8:44:15

Summary

This summary covers advanced topics in production AI systems, focusing on scaling generative media, optimizing LLM inference, and building real-time AI features. Key technical deep dives include customizing generative models (e.g., Flux for action prediction), achieving low-latency speaker diarization, and optimizing model deployment by integrating Hugging Face's `transformers` library into `vLLM`. Furthermore, the talks emphasize that modern AI development requires a global, agent-centric playbook, necessitating localized payments and robust fraud monitoring.

Download summary

Key takeaways

AI Model Customization is Multi-Layered 23:20
Customizing models like Flux involves optimizing multiple surfaces: Prompt Upsampling (using VLM/LLM to steer user intent), Moderation (via configurable APIs or embedding into weights using LoRA), and adding new modalities (e.g., Flux Action for robot control).
Optimizing LLM Inference with vLLM 41:40
The `transformers` modeling backend allows running Hugging Face models directly in `vLLM`, enabling access to advanced features like KV cache, continuous batching, and expert parallelization while maintaining the original model's code base.
Real-Time AI Requires Causal Segmentation 46:40
Achieving low-latency speaker diarization for voice agents requires moving from batch processing to causal segmentation, which processes audio in streaming chunks to maintain accuracy while minimizing latency (e.g., achieving 400ms latency).
AI Product Playbook Must Be Global and Agent-Ready 50:00
Successful AI companies must treat their day-one market as the entire world, localizing pricing, payment methods, and adapting their product design to be discoverable and purchasable by AI agents.

Technical details

Generative Media & Action Prediction 1400s
Black Forest Labs demonstrated customizing the Flux model for action prediction (Flux Action). This involves mid-training and post-training fine-tuning to predict robot movements (e.g., a 6-dimensional vector for a robot arm) by treating the model as a World Action Model, which normalizes action vectors directly into the latent space.
LLM Inference Optimization (vLLM) 2500s
The `transformers` modeling backend integrates Hugging Face models into `vLLM`'s runtime. This is achieved by patching the model's configuration, registering attention shims (e.g., `VLM attention`), and surgically replacing layers (e.g., QKV fuser, RMS norm) to gain access to `vLLM`'s performance features like KV cache and continuous batching.
Voice AI & Diarization 2800s
Speaker diarization involves segmenting conversations and assigning turns to speakers. For real-time applications, the process must shift from batch processing (sliding windows) to causal segmentation, which processes audio in streaming chunks to maintain accuracy and achieve low latency.
AI Payments and Developer Experience 2600s
Stripe demonstrated how AI agents can streamline the setup of complex financial integrations. Developers can use prompts to configure payments (e.g., invoicing, subscriptions) and receive recommendations for specific integrations (like Stripe Checkout) and tax knowledge (Stripe Tax) without leaving their development environment.

Mentioned resources

Flux (Generative Model)
Flux Action (Model/Research)
vLLM (LLM Inference Engine)
Hugging Face Transformers (ML Library)
Stripe Tax (Financial Tool)
pyannoteAI / PianoAI (Voice AI Toolkit)

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.