Going In Deep On Data | YC Paper Club

Video summary

YC speakers argue expert data is the AI bottleneck, while Mercury 2 reaches 1,000 tokens per second and multilingual models need measured language mixes.

At YC Paper Club, Francois Chaubard argued that curated, up-to-date data—not model architecture or GPUs—is increasingly the AI bottleneck, noting that data businesses have generated over $100 billion in market capitalization. Snorkel’s Vincent Chen showed how expert knowledge can be encoded as software and introduced Senior SWE-bench, which evaluates coding agents on realistic tasks and measures “tasteful” engineering; Fable, Opus, and Soul were tied for first. Inception Labs’ Volo Kuleshov said its diffusion-based Mercury 2 models generate more than 1,000 tokens per second, and described Data Forge, which synthesizes training and evaluation environments from real-world interactions. Finally, Shane Longpre presented multilingual scaling research: Thai has just 0.6% as many tokens as English in one corpus, and language transfer can be asymmetric. His team’s measured scaling laws help select data mixtures, model sizes, and training strategies.

Chapters

1. 0:00Francois Chaubard: Why data matters—from $100 billion in market creation to production error analysis
2. 5:16Francois Chaubard: Salesforce UI shifts, weak S&P forecasts, and specialized AI data providers
3. 9:51Vincent Sunn Chen: Scaling expert supervision from 30-second labels to complex agent environments
4. 13:50Vincent Sunn Chen: Data programming turns expert heuristics into scalable, denoised labels
5. 17:31Vincent Sunn Chen: Senior SWE Bench tests realistic coding tasks and senior-engineer judgment
6. 21:08Vincent Sunn Chen: Validation agents convert expert specs into tested code requirements
7. 25:41Vincent Sunn Chen: Tasteful pass ranks Fable, Opus, and Soul; open grants support new benchmarks
8. 28:44Inception: Mercury 2 exceeds 1,000 tokens per second as Dowo Forge builds real-world RL environments
9. 34:54Inception: Dowo Forge improved Mercury 2.5 by over 23% on separate environments
10. 40:26ATLAS: Madlad 400 has only 0.6% as many Thai tokens as English
11. 44:05ATLAS: Spanish benefits from Portuguese, Italian, and French, while Japanese interferes
12. 48:55ATLAS: Script matters more than language family, and scaling laws guide four-to-eight-language expansion

What the full analysis includes

12 chapters · 63 key points · 12 section summaries · mind map

The remaining 60 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →

This video is 53 min. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.