Links for 2022-12-07
Deep learning algorithms, shown 85,000 pictures of human retinas, can identify gender with 87% accuracy. Clinicians are not aware of any differences between men’s and women’s eyes. Authors used a professionally-put-together UK dataset (Biobank) and then separately validated the model with an entirely different dataset from a cooperating researcher from Germany. https://www.nature.com/articles/s41598-021-89743-x
New technique could diminish errors that hamper the performance of super-fast analog optical neural networks. https://news.mit.edu/2022/scaling-analog-optical-computing-1129
Intuitive physics learning in a deep-learning model inspired by developmental psychology https://www.nature.com/articles/s41562-022-01394-8
Forecasting elections with non-representative polls: Polling Xbox live players to generate election forecasts in line with ordinary opinion polls https://www.microsoft.com/en-us/research/wp-content/uploads/2016/04/forecasting-with-nonrepresentative-polls.pdf
The Aberfan disaster was the catastrophic collapse of a colliery spoil tip killing 116 children and 28 adults as it engulfed Pantglas Junior School and a row of houses. https://en.m.wikipedia.org/wiki/Aberfan_disaster
“Roman dodecahedra date from the 2nd to 4th centuries AD. No mention of dodecahedrons has been found in contemporary accounts or pictures of the time.” https://en.wikipedia.org/wiki/Roman_dodecahedron
Even people who loathe discrimination tend to have a blind spot for discrimination based on attractiveness. https://psyarxiv.com/5uz8g/
“Students could most easily & usefully learn to "think for themselves" by debating controversial topics, where a standard accepted claim is challenged by contrarians. But schools prevent this, as this tends to move opinion toward those contrarian positions.” https://www.overcomingbias.com/2022/11/heresy-makes-you-think.html
"This longitudinal investigation from birth to adulthood indicates head growth as a proxy of brain development and intelligence. Repeated early head circumference assessment adds valuable information when screening for long-term neurocognitive risk." https://www.cambridge.org/core/journals/journal-of-the-international-neuropsychological-society/article/head-growth-and-intelligence-from-birth-to-adulthood-in-very-preterm-and-term-born-individuals/AC59EB120E5897AD5A78B260F9DA0A4F
ChatGPT can explain things in the form of a Seinfeld script. https://threadreaderapp.com/thread/1599408644382023683.html
Long AI timelines guy updates towards shorter AI timelines. Here is why: https://www.lesswrong.com/posts/sbb9bZgojmEa7Yjrc/updating-my-ai-timelines
The mistakes ChatGPT makes are really fascinating. Here it insists that the word “giraffe” has 8 letters, even after it wrote a Python program that correctly enumerates the letters.
Interestingly, when you formulate the sentence differently, it doesn't make this mistake. “How many letters are there in the word "giraffe"?” yields the answer 7.
If you then ask a follow-up question with the original wording, it keeps the correct answer. But if you start with the original phrase, followed by the restatement, it incorrectly insists that the word has 8 letters.
This suggests two ways to fix ChatGPT:
1. Make use of programming to verify answers.
2. Make it less stubborn. That is, make it actively doubt its own answers when challenged, rather than coming up with increasingly implausible explanations for why the original answer it gave was correct.