Francois
Showing 1–4 of 4 transcripts.
- Y Combinator1h 51m
Why Inference Matters 12x More Than Training | YC Paper Club
Dan Fu, Jeff Morgan, Joe Fioti, Lianmin Zheng, Francois
DeepSeek, Together AI, and other industry leaders are transforming AI economics by shifting focus from training to inference optimization, where compute costs now dominate due to high token volumes and long model lifespans. Through innovations in kernel compilers, disaggregated system architectures, and model-specific techniques like Mixture of Experts and quantization, providers are achieving gross margins near 80% while delivering open models at a fraction of the cost of proprietary alternatives. Strategic advancements in runtime scheduling, KV cache retention, and hardware-at-compiler levels address critical bottlenecks in decode throughput, enabling robust, low-latency inference across hybrid local-cloud environments.
- Y Combinator1h 37m
AI Safety Fact vs. Fiction | YC Paper Club
Clark Barrett, Mykel Kochenderfer, Francois
The YC AI Safety Club convened to reframe artificial intelligence discourse from sensationalist extinction scenarios to rigorous engineering standards advocated by figures like François, Yann LeCun, and Jensen Huang. Speakers presented concrete mitigation strategies, including Clark Barrett's automated formal verification of cryptographic code and Michael Kochenderfer's adaptive stress testing methods modeled after aerospace safety protocols. The event concluded by proposing a market-driven regulatory framework where independent reporting agencies and zero-trust security principles replace ineffective hardware bans while prioritizing cyber defense against nation-state threats.
- Y Combinator30 min
New LLMs Are Unlocking Robot-Use Agents
Francois, Hanmei, Vincent, Jay
Frontier researchers and startups like Waddle Labs are advancing a paradigm shift toward general-purpose "robot use agents" that leverage coding-based policies and cross-modal data to perform complex physical tasks without specialized fine-tuning. By applying the "Bitter Lesson" principle to consolidate diverse datasets, these systems utilize in-context learning and memory consolidation architectures to rapidly adapt to new environments, with current demonstrations successfully executing multi-step manipulation challenges. Industry projections indicate that as latency decreases and internal world representations converge, economically viable, general-purpose robots capable of performing competent human-like tasks will emerge within two years.
- Y Combinator1h 14m
World Models, JEPA And The Path To Sample-Efficient RL
This event analyzes the critical bottleneck of sample efficiency in artificial intelligence, contrasting current deep learning models' massive data requirements with the human brain's ability to learn from minimal experience through superior world modeling. The discussion details how advancements in non-differentiable control theories, video diffusion architectures, and Joint Embedding Predictive Architectures are shifting strategies from model-free behavior cloning to synthetic, simulation-based planning for complex robotic and autonomous driving tasks. By addressing scaling challenges in high-dimensional action spaces and architectural limitations like the Transformer's inefficiency in time-domain compression, the presentation outlines a roadmap toward general-purpose robotics and AGI by 2026 through the integration of "awake sleep" mechanisms and physics-informed predictive systems.