Falsehoods Programmers Believe About Reality
summary
Names, time, geography and categories repeatedly break systems: claim that more data = better model is false, and large language models are described concretely as probabilistic token predictors — not mirrors of human nuance — with examples like a SharePoint+LLM search failure, student-name matching that never reaches 100%, and EV routing that had to account for broken chargers. Practical prescriptions include designing abstractions for a stated purpose, auditing who benefits, avoiding overcapture of data, and treating the map is not the territory when measuring outcomes.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:03Computers and objective reality
- 2:26Falsehoods about reality
- 7:09More data vs better representation
- 10:39LLMs as mirrors limitations
- 13:48Measurable vs immeasurable reality
- 16:58Reality is subjective
- 18:47Artificial categorization
- 20:31Forcing reality into categories
- 22:57Airspace algorithmic homogenization
- 24:46Overcapture dangers
- 25:48Abstractions as tools
- 27:15Designing with purpose
- 27:58Building useful systems
related talks
01h 05mAlgorithms Cheat
41m"Keynote: Brains Are Beautiful Liars"
35mMental Models to Use and Avoid as a Data Scientist
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
34mEverything is a graph, including LLM Applications (and that’s handy)
16mA Word About Programming: Applying a Natural Language Vocabulary Acquisition Model(…)
35mAvoiding Bad Database Surprises: Simulation and Scalability
33mLearning to program is hard, and how to fix that