Big non-tech companies are quietly fine-tuning open-weight models in-house
ivan_bezdomny · x · 2026-09-14
A pattern is emerging: the biggest non-tech companies fine-tune open-weight models on their sensitive data and proprietary evaluations instead of relying on closed APIs.
The playbook: a small internal post-training team designs the datasets and evaluations, hires contractors under NDA, then contracts the actual post-training to post-training startups — similar to what Harvey did. The author plans an article and video noting that finetuning isn't complicated but is fragile: default parameters often don't fit your problem, and it's non-obvious which metrics to watch or how much data you need.
More from Venture
- Anthropic's gross margins top 80% as frontier labs turn profitable early — bookwormengr · 2026-09-14
- Anthropic tells investors it will be profitable for second straight quarter — ResultBackground2450 · 2026-09-14
- Snarky take: AI models can do anything except generate positive cash flow — SatelliteNetSec · 2026-09-14
- Nvidia Partners With Palantir to Chase the $500 Billion Sovereign AI Market — Beth_Kindig · 2026-09-14
- NVIDIA Inception offers AI startups free training, cloud credits and VC intros, no equity — EAccelerate_42 · 2026-09-14
- Indie puzzle game Prism Rules: 75 optics levels, $10 one-time buy — measure_plan · 2026-09-14