Why is Claude 3.5 Sonnet such a good model for its size?
➕
Plus
8
Ṁ547
2026
88%
Pretraining data composition
63%
Doesn't use any scale.ai training data
49%
Offline policy learning RLHf
37%
Task vectors like golden gate Claude

Claude sonnet (3.5) is a relatively small model (estimated to be 5e24 FLOPs). Yet it beats larger models on GPQA, LMSYS, and many other industry standard benchmarks. While it can’t be known that this market can resolve, it’s possible that academics and OSS will learn in the coming years what was done to achieve this high quality model.

Get Ṁ1,000 play money
Sort by:

Why would not using scale.ai data be beneficial?

@JaundicedBaboon There have been questions about quality