Best of Twitter - Week of September 20, 2021
My roommate is taking intro stats and learning about standard deviations. In retrospect saying "beware summary statistics!" and shoving The Datasaurus Dozen in her face was probably not a helpful instinct, but it was fun to remember these silly graphs
autodesk.com/research/publi…
The ability for Apple to create M1, for Tesla to create Model S, for Google to innovate on Google Workspace are all predicated on the fact that non-competes are not enforceable in California as each team is led by a previous leader at a top competitor (Intel, Mazda, Microsoft)
Humans, perfected by evolution to intuit purpose and intention, as a side effect of tool use: "Haha cats evolved to be visually accurate predators but they get hyperfixed on laser pointers"
Same humans: "Why did God make us?? What is the meaning of life?? What's my purpose??"
Every time I see someone get tripped up by trolley problems, nature of free will arguments or "if a tree falls and no one hears it..." questions I think "haha they love the brain laser pointer, silly beasts"
"Drug users use a lot of drugs! So why don't psychiatric patients get bladder injuries? It's because you get bladder injuries when you're taking more like 90,000 mg of ketamine a month, and not when you're taking 280 mg."
astralcodexten.substack.com/p/drug-users-u…
Fascinating experiment in top scientific conference shows the large part of randomness in peer review:
Submitted papers went to 2 review committees.
👉50% of acceptance decisions were different!
👇 Correlation between the 2 review scores was only 0.55.
arxiv.org/abs/2109.09774
My goodness. China's cyberspace watchdog, the CAC, just published a long (and unprecedented) set of draft regulations for recommendation algorithms. The short version: they will be tightly controlled. Key points below. 1/
Why do large models do worse? In the image, small sizes of GPT3 give true but less informative answers. Larger sizes know enough to mimic human superstitions and conspiracy theories.
New CDC Study: "the portion of 5-11-year-olds who are classified as overweight or obese is now 45.7 percent, up from 36.2 percent before the pandemic."
If we prevented all cardiovascular disease, average lifespan would increase only 1.5 years.
Prevent all cancer? We get an extra 2.1 years.
Prevent aging? We get many extra decades.
Yet gov't biomedical research spends almost nothing on preventing aging....
H/T @davidasinclair
The 24 largest tax-exempt hospital systems made a collective net profit of $11.9 billion in the second quarter alone. They used to be called non-profits; now the correct label is tax-exempt for obvious reasons.
Great analysis by @bobjherman
Take, for example, the Antikythera mechanism. It is described as the oldest example of an analog computer. Equivalent complexity didn't appear again until the 18th century. arstechnica.com/science/2021/0…
Quarantined section
EcoHealth: "We will analyze all SARSr-CoV..sequences for..proteolytic cleavage sites in S2 and for..potential Furin cleavage sites..[W]e will introduce appropriate human-specific cleavage sites and evaluate growth potential in Vero cells and HAE cultures."
drasticresearch.files.wordpress.com/2021/09/defuse…
^ quotes from unfunded DARPA proposal from Wuhan lmao
The bigger GPT models are worse paper is pretty meh: https://www.youtube.com/watch?v=aX8phGhG8VQ