Roboflow: vision AI faces a long-tail problem — BNSF runs bespoke models on 32,500 miles of track
PTrubey · x · 2026-11-18
At the TomorrowX conference, Roboflow highlighted a key difference between language and vision AI: while language training generalizes well, every vision application needs unique training on its own dataset, and no base model will likely ever cover all of vision. Combined with latency requirements pushing models on-device or on-prem, this explains the industry shift toward purpose-built, application-specific vision models.
The slide shows Roboflow's deployment at BNSF Railway: 32,500 route miles across 28 states, with cameras at every checkpoint for wheel inspection, grade-crossing safety, container ID, and yard inventory. The author adds that humans handle the long tail via continuous learning—another reason the lack of continuous learning remains an efficiency bottleneck for AI.
More from Infra
- Open-source Personal AI Computer: DIY 2–8 GPU builds with full BOM, CAD and guides — dee_hw · 2026-11-18
- Blink CPU speed test: model2vec 3.8x faster solo, Sentence Transformers 2.2x faster at batch 1024 — tomaarsen · 2026-11-18
- Runway cuts real-time 57B video model serving from 8 H100s to 4 with inference optimizations — runwayml · 2026-11-18
- Over half the power for America's largest data centers comes from gas and coal — The Decoder · 2026-11-18
- Nvidia halts US-approved H200 shipments to China after $2.4B quarter — kimmonismus · 2026-11-18
- Analyst: Hyperscaler Nvidia spend may be flat once segment reclassification is accounted for — RihardJarc · 2026-11-18