Women's pockets | Next Issue #90
Someone once said, women aren't allowed a lot, including having pockets
Welcome to the October edition of Next — your monthly dose of data stories that make you think, laugh, and maybe even reevaluate your wardrobe choices. This month, we dive into the pocket-sized gender inequality stitched right into your jeans, explore the quietest roads for your next scenic drive, and unpack Taiwan's super-aged future. But that's not all; we've got some must-read updates on language models and handy tips for prompt engineering.
So grab a cup of your favorite brew, and let's jump right in!
Five Stories
Someone clever once said, women were not allowed pockets
I enjoy wearing baggy trousers and jeans. First, the cloth isn’t constantly breathing against my legs. Second, I get bigger pockets! Women have wanted bigger pockets for a long time.
In this post, Jan Diehm & Amber Thomas for The Pudding compare the pocket sizes between men and women across brands. On average, the pockets in women’s jeans are 48% shorter and 6.5% narrower than men’s pockets. It also features a fantastic tool to decide which jeans to buy by choosing what things you expect to fit in. I chose Pixel and only 5% women’s pockets can fit it in.
More pets than children? A Look at Taiwan’s Super-Aged Future
World human population is collapsing, according to many — including Elon Musk. He had once stated that the sale of adult diapers has overtaken baby diapers in some countries, a clear indication of aging population, which I had tested. In this post, the authors take a closer look at Taiwan, of the fastest aging country in the world. They also compare Taiwan with several other countries: United States, China, India, S. Korea, and more. Pretty cool — check it out!
What are the quietest roads in America?
This interactive chart identifies what roads are the most lonely in the United States. Alaska’s State Route 11 is the most scenic and the most quiet. You would get to see Arctic Ocean en route. Utah’s US Route 50 (which I’m grateful to have driven on during my trip to Moab) is the second on the list. It passes through two National Parks — Arches and Canyonlands — on either side of the highway.)
California’s State Route 139 is the only from California in the Top-5, and the only one in the top-10. But fret not, you can use the interactive chart to find the scenic routes in your state! In Tennessee, it is TN-104. #addedtobucketlist
Evaluating LLMs is a minefield
When the news was reported that ChatGPT is getting worse over time, OpenAI’s VP had responded by saying:
However, when other researchers tried to reproduce the results, they failed. Model response to certain prompts changed; modifying prompts minutely lead to the same capabilities. What do we learn? Responses are stochastic, and depend on model updates — not a news at all!
In this slide deck, the authors argue that we do not have established ways of evaluating responses from LLMs. Three key problems:
LLM responses are sensitive to prompts (and stochastic)
Responses are constructs of prompts. Saying ChatGPT is biased is not independent of bias in prompts.
Contamination of evaluation files in training data, like USMLE accuracy might be so high because textbooks are part of training materials.
Prompt Engineering Guide
Prompting has become a lofty job, especially in the near term. Andrew Ng and OpenAI have a course on this: ChatGPT Prompt Engineering for Developers. For those who aren’t interested in a full fledged course, can skim some chapters from this online free guide. There are some neat tricks:
Four Packages
Waywiser package offers a suite of methods for evaluating spatial models, focusing on understanding spatially structured errors and safe prediction zones. It incorporates various statistical metrics like Moran's I and Geary's C, and is particularly compatible with the 'tidymodels' framework. Vignette. Github.
Pointblank package allows for robust data validation and metadata organization in various data structures like data frames, tibbles, and Spark DataFrames. It offers customizable validation pipelines, reporting options, and user-defined failure thresholds, also facilitating workflows for information management. Vignette. Github.
Collapse package is a C/C++-based tool in R designed for fast, flexible data transformation and statistical computing. It offers a broad range of S3 generic functions, OpenMP multithreading, and integrates well with various R packages like 'dplyr' and 'data.table', aiming to make data manipulation and analysis more efficient and programmer-friendly. Vignette. Github.
Easylabel package offers an interactive interface for labeling scatter plots, volcano plots, and Manhattan plots, leveraging 'shiny' and 'plotly' technologies. Users can hover, click, and drag labels, with an option to export the finalized plots directly to PDF for publication. Github.
Three Jargons
Stochasticity: Pertains to the inherent randomness in a system or process, making outcomes statistically analyzable but not precisely predictable.
Aleatoric Uncertainty: A subset of stochasticity, it specifically addresses the irreducible uncertainty arising from inherent randomness in data or processes. Quantifiable through statistics, it remains constant regardless of data or model improvements.
Epistemic Uncertainty: Contrasts with aleatoric uncertainty by focusing on the uncertainty due to incomplete knowledge or flawed modeling. Unlike aleatoric uncertainty, it can be reduced by refining models or gathering more data.
Two Tweets
https://twitter.com/LinuxHandbook/status/1712720794240082125
https://twitter.com/fchollet/status/1719383230389121538
One Meme
That’s a wrap!
Hope you enjoyed today’s letter. See you next month!
To a happy world,
Harsh