DataSmith auto agent outperforms model architecture tweaks using 200x fewer tokens via data interventions
josh_wills · x · 2026-08-19
DatologyAI introduced DataSmith, an auto agent for data research, demonstrating that data quality is the ultimate compute multiplier.
Key Results:
- Performance: With only 734M post-training tokens (200x fewer than the default SmolLM3 recipe), DataSmith outperformed SmolLM3-3B's post-training performance by 3 percentage points.
- Superiority: It outperformed harnesses that also intervened on architecture and optimizers solely through data interventions.
Design Features:
DataSmith is designed to experiment substantially more than default harnesses, reason more, generate more research ideas, explore novel datasets, and spawn more subagents to find better data configurations.
More from coding & agent
- Single developer migrates 850K lines with agents in 6 months — tristanbob · 2026-08-19
- Monitoring 2k+ MCP Servers: 7,190 Safety-Relevant Changes, Read-to-Write Flips Undetected by Allow-lists — mcpindex · 2026-08-19
- Generating UIs by having LLMs emit raw HTML/JS? That's a risk pile, says Googler — rseroter · 2026-08-19
- YC Startup Launches Graphify, a Knowledge Graph Engine for Enterprises — ycombinator · 2026-08-19
- Open-source SRT whiteboard animation skill: hook it to Codex and ship 100 videos a day — huangyun_122 · 2026-08-19
- Founder's $20/month stack: Claude coding plus free SaaS runs a whole startup — clarashih · 2026-08-19