Vanishing Gradients Episode 6: Bullshit Jobs in Data Science (and what to do about them)
10 snips
Apr 4, 2022 AI Snips
Chapters
Transcript
Episode notes
Fancy Models Won't Replace Finance's Excel
- Jacqueline built an agent-based simulation to forecast a coupon campaign but stakeholders wanted simple Excel proofs for finance.
- She had to recreate simplified formulas despite the complex model because executives needed auditable numbers.
Anomaly Alerts Break Without UX And Thresholds
- Early in her career Jacqueline built an anomaly-detection "Adler Alert" that worked but overwhelmed teams with too many uncertain metrics.
- The product failed because thresholds, UI, and signal overload made it unusable.
Most Business Problems Need Simple Models
- High-profile deep-learning work gets glorified but most production problems are solved by simple models.
- Jacqueline estimates 80–90% of useful data science could be solved with logistic or linear models.