The Future of AI: Building the Most Impactful Technology Together
summary
Argues that fully open models are achievable and necessary, showing concrete work: rebuilding DeepSeek’s R1 training pipeline as Open R1, publishing FineWeb dataset, FineMath and FineVideo, and tooling (Data12, Nanotron, LightEval). Points to public clusters (Lumi, Alps, Leonardo) with ~10,000 GPUs and estimates GPT‑4 scale training in ~one month on that hardware. Cites leaderboard convergence (Llama3 caught up in ~1 year), a competitive 7B “Olympic coder,” DeepSeek’s 8,000‑token reasoning traces, and DAB‑STEP‘s low 14–15% multi‑step success rate.
AI-written summary · about
chapters (5)
- 0:00Open AI collaboration
- 1:34Build open language models
- 14:18Fully open models today
- 25:15AI trends and applications
- 41:55Q&A
related talks
17mOpen Innovation: Unlocking AI's Potential
27mKeynote: Open Source in the Age of AI
34mMapping the Open Model Landscape in 2025
20mTrue Exponential: Building a Generative AI Company
22mThe Imperative for Open Source AI - The Path Forward to Take Down the AI Oligarchs
10mThe Future for AI-First Enterprises is “Open” - How to Succeed and Not...
43mBeyond the AI Hype: What's Real, What's Next
14mReliable Agentic AI, the Open Way