AI summary by TwinMind
How Data Happened | Chris Wiggins & Matthew L. Jones | Talks at Google
Talks at Google
43m 19s
Introduction and Book Overview
- Sonia Suban (Google software engineer) introduces Dr. Chris Wiggins (Columbia Applied Mathematics) and Dr. Matthew L Jones (Columbia Faculty of Arts & Sciences)
- Discussion of book "How Data Happened: A History from the Age of Reason to the Age of Algorithms"
- 45-minute session format with audience Q&A via chat
Book Genesis and Motivation
- Collaboration began decade ago, formal class creation started 2015
- Originally focused on history of data science, evolved based on student feedback
- Designed for interdisciplinary audience: engineers and humanistic students
- Class became foundation for book after identifying novel material for both technical and non-technical backgrounds
Book Structure - Three Parts
Part One: Pencil and Paper Era
- Data analysis for understanding truth using manual methods
- Covers ~200-year span to illuminate current political/ethical challenges
Part Two: Digital Computation Era
- Begins at Bletchley Park with digital computation creation
- Integration of hardware, diverse communities, technical skills
- Industrial scale data processing emergence
Part Three: Present Day
- Current milieu analysis (not chronological)
- Corporate data use rise
- Understanding powers constraining/guiding data-empowered algorithms in daily life
Data Collection and Decision-Making Blind Spots
- Only some facets of data collected, creating worldview-shaped understanding
- Stephen J. Gould's "scientific racism" concept discussed
- Example: Insurance industry vs. sociologist Du Bois debate on insurability data analysis (late 19th/early 20th century)
- Hoffman (insurance): data showed essential differences between groups
- Du Bois: data revealed historical causes of stratification
- Danger: numbers create false objectivity, masking subjective design choices in data collection
Bayes' Theorem in Cryptography and Philosophy
- Chapter 6 focuses on Bayes' theorem importance in WWII codebreaking
- Distinction: "being philosophical about Bayes vs. using Bayes' Rule formula"
- COVID diagnosis example: sensitivity/specificity vs. probability of illness given test results
- Bletchley Park reintroduced Bayes' Rule for practical code-breaking applications
- NSA declassified papers showed practical Bayesian attitudes contradicting contemporary philosophical/statistical positions
- Direct line from cryptographic mass data analysis to modern computational statistics
Social Physics Concept
- Belgian statistician Quetelet (creator of BMI) working post-French Revolution
- Applied Newtonian physics tools to human behavior data
- Argued social phenomena followed patterns (normal curve)
- Characterized human activity (divorce rates, criminality) as physics with mechanics/dynamics
- Focus on group behavior rather than individual choice
Correlation vs. Causation Historical Example
- Yule's 1899 paper "On the Causes of Pauperism in England"
- Performed multivariate regression (supervised learning equivalent)
- Inferred positive coefficient between welfare recipients and poor population
- Concluded welfare causes poverty (political statement from correlation)
- Footnote 25 acknowledged "causation should be read as associated with"
- Excellent predictive models don't guarantee causal understanding
Hypothesis Formation Evolution
- Pre-large data era: significance testing prestige in science reorganization
- Current crisis in significance testing methodology
- Genesis from different statistical philosophies:
- William Gosset (Guinness Corporation): economical beer-making experiments
- Fisher (British statistician): scientific truth replacement theory, randomization, p<0.05 suggestion
- Neyman (Polish-American): decision theory for practical action
- Fisher-Neyman "warfare" for decades over fundamentally different worldviews
- Textbook flattening created standard scientific procedures
- Radically different from modern machine learning/data science approaches
Field Making Discussion
- Mathematical statistics field creation (adjective "mathematical" distinguished from other statistics)
- Academic departmental statistics vs. algorithmic hypothesis testing
- Algorithmic approach: collect data → confidence interval → p-value → <0.05 determination → next hypothesis
- Both Fisher and Neyman would have hated this algorithmic approach
- Biggest insult for both: calling someone "Bayesian"
Early AI vs. Modern NLP
- Early AI vision: human intelligence not from data accumulation/sensory input
- Deeply anti-empirical approach, focused on mathematical ability, chess, logic
- Turing had more capacious vision including sensory world
- Mid-20th century: language as small number of rules (Chomsky influence)
- Google researchers' "Unreasonable Effectiveness of Data" paper
- Complete inversion: language as empirical science rather than physics-like rules
- Modern AI transformation from rule-based to data-driven approaches
Data Privacy Evolution
- 1970s: bipartisan US conversation about data privacy
- Privacy laws constrained federal government but not corporations
- Individualistic privacy understanding emerged
- Privacy as personal concern vs. ecological concept affecting communities
- Individual action solutions (cookie opt-outs) vs. systemic protections
Current Privacy Concerns and Solutions
- Three-player power distribution: individuals, corporations, state
- Multinational companies subject to multiple jurisdictions (EU regulations affecting US users)
- Principles → standards → rules → code → product decisions chain
- Individual software engineers/product owners have outsized impact through private ordering
- Corporate power includes deplatforming capabilities
- "Unstable three-player game" between corporate, state, and individual power
Future Vision for Data Science/ML
Chris Wiggins' Excitement:
- Causality as explicit field with modeling assumptions
- Deep learning/large language models continued progress
- Extremely performant but uninterpretable methods
- Theory development for understanding why methods work
- Healthcare applications: privacy-preserving data access, population stratification challenges
Matthew Jones' Perspective:
- Large language models as intelligence exploration tools, not human substitute
- Integration with affected communities (scientific collaborations, community development)
- Recognition of political/ethical implications at all levels (federal to municipal to schools/jails)
Audience Q&A Highlights
Large Language Models Integration
- Weekly lab updates incorporating LLM material
- Environmental costs of training considerations
- Infrastructure metaphor: neither good nor bad, but collective building requiring community input
- Eliza chatbot example: artificial intelligence without machine learning
- Reinforcement learning optimizing "appearance of spontaneity"
Health Data Privacy Evolution
- Belmont principles framework: respect for persons, benefits vs. harms, justice
- Latanya Sweeney's re-identification research (Harvard professor, former MIT student)
- Anonymous database re-identification using join keys (age, zip code, gender)
- Genomics/DNA data sharpening privacy conversations post-late 1990s
- Genotype-phenotype mapping opportunities vs. potential harms
Collaboration History
- 10-year professional relationship beginning with machine learning history talk
- Journalism school teaching program: data analysis tools for journalists
- Columbia student initiative suggesting joint class
- Evolution from small seminar to larger lecture course
- Student feedback driving connection between history and ethics/politics
- Organic book development from educational context
