Sell data or fine-tune? Businesses face a strategic choice on proprietary data
johnschulman2 · x · 2026-09-01
The post discusses a critical strategic choice businesses face: what to do with years of accumulated proprietary data.
Two Paths:
- Sell Data: Provide it to vendors or frontier labs. This improves next-gen models for everyone, including current and future competitors, eroding your edge.
- Self Fine-tuning: Fine-tune existing models internally to maintain the competitive edge built over years. Tinker aims to make this practical for businesses without large ML infra teams.
Case Study:
- Citing Thinky Machines' Inkling model: fine-tuned on 1, 25, and 100 hours of proprietary alphanumeric entity data.
- Results: On VoiceCodeBench, the task success rate increased from 56.33% (untrained) to 79.00% (100-hour fine-tune), with significant gains in entity recovery (CTEM) as well.
More from Companies & People
- OpenAI Codex sees exponential growth amid Anthropic user decline — FinanceYF5 · 2026-09-01
- Bootstrapping with no network: Using a free automation library for cold outreach — Far_Fault_5899 · 2026-09-01
- Developer hasn't thought about type systems in 6 months due to AI — ethanniser · 2026-09-01
- GOSIM Shenzhen 2026 set for October: 150+ speakers on agentic AI and robotics — juntao · 2026-09-01
- Pacing the frontier: AI speed control term gains recognition. — sjgadler · 2026-09-01
- Developer View: Moving Fast Requires Forgetting All Restrictions — KlausCodes · 2026-09-01