The Neuron: AI Explained The Humans Behind AI: How Invisible Technologies Trains 80% of the World's Top Models
30 snips
Nov 3, 2025 Caspar Eliot, an executive at Invisible Technologies, uncovers the human side of AI training, revealing that 80% of top models rely on a dedicated workforce. He explains how models learn through curated data, not just online scraping, and discusses the critical importance of data quality over quantity. The conversation delves into common enterprise mistakes in AI deployment and speculates on the future of jobs—emphasizing that while automation changes roles, human skills will remain indispensable. Expect insights on unconventional job backgrounds and the rising demand for human data expertise.
AI Snips
Chapters
Transcript
Episode notes
Three Human Steps Teach Models
- Large language models learn via three human-driven phases: supervised fine-tuning, reinforcement learning with human feedback, and evaluation.
- Caspar Eliot compares these to taking a model to the library, giving tests, and grading its answers to shape behavior.
Models Predict, Humans Define Truth
- Models are powerful next-token predictors without inherent truth or intent.
- Post-training (reward modeling and evaluation) and data choices shape what they produce and how users perceive 'better'.
Turning Game Video Into Scouts
- Invisible helped the Charlotte Hornets by turning unstructured game video into scout-like insights using computer vision.
- Caspar says video models can create an effective 'scout in every game' to find player movement patterns.