#38 Stanislas Polu (Co-fondateur @ Dust) : Assistant personnel basé sur l’Intelligence Artificielle

6 Dec 2023 · 44 min
43:40

In this French‑language Data Driven 101 podcast, host Marc Sancelme interviews Stanislas Polu, co‑founder of Dust and former Stripe and OpenAI veteran. Polu explains Dust’s product: a platform that lets enterprises build custom, task‑specific AI assistants using Retrieval‑Augmented Generation (RAG) over company data sources (Notion, GitHub, Slack, Google Drive, plus custom APIs). He discusses the business need – scaling AI productivity across an adoption curve of early‑adopter power users and reluctant employees – and the product challenges of onboarding, data‑privacy silos, and UX that makes assistants feel like tools rather than generic chatbots. Polu argues for RAG over fine‑tuning, warning that fine‑tuning on company data can induce hallucinations because the base model’s knowledge graph lacks that data. He details model choices (GPT‑4 32k, Claude 2, OpenAI Ada embeddings), cost considerations, and the importance of chunking (256‑512 token windows) and hybrid semantic/keyword search. The conversation moves to market dynamics: low entry barriers for product‑focused startups, lock‑in through custom assistants, competitive pressures from OpenAI’s scale, and the need for a strong go‑to‑market strategy to achieve product‑market fit. Polu describes the current stage with design partners (French firms of 100‑500 employees) showing sticky usage, and outlines the next steps: reducing onboarding friction, quantifying acquisition cost and time, expanding into the US market, leveraging reputation‑based SaaS growth, and preparing for a Series Z round once metrics are solid. He also shares anecdotes from his time at Stripe and the quirky acquisition path that led him there, illustrating the high talent density at OpenAI. The interview ends with reflections on LLM strengths (code, translation, summarisation, data‑driven support) versus weaknesses (deep reasoning, creativity), and future plans for Dust’s expansion, especially into the US market.

Chapters

More from Stanislas

youtube.com

Dust x Clay - GTM in the Age of AI - full conversation

9mo ago
youtube.com

The Outer-Loop Era - Stanislas Polu - Dust

11mo ago
youtube.com

AI Agents Are Here: What They Can Already Do—and What’s Next

1y ago
youtube.com

AMA with Stanislas Polu, Co-Founder of Dust: Revolutionising Tech with AI Agents

1y ago
youtube.com

Getting the Most From AI With Multiple Custom Agents ft Dust’s Gabriel Hubert and Stanislas Polu

1y ago
youtube.com

Getting the Most From AI With Multiple Custom Agents ft Dust’s Gabriel Hubert and Stanislas Polu

1y ago
youtube.com

Agents @ Work: Dust.tt — with Stanislas Polu

1y ago
youtube.com

dotAI 2024 - Stanislas Polu - LLMs reasoning and agentic capabilities over time

2y ago
youtube.com

No GPU Before PMF // Stanislas Polu // AI in Production Conference Lightning Talk

2y ago
youtube.com

[EXTRAIT] Décryptage du fonctionnement d’un LLM en 5min - par Stanislas Polu, CEO de Dust

2y ago
youtube.com

Qdrant x Dust: How Vector Search helps make work work better - Stan Polu | Vector Space Talk #010

2y ago
See all content