brchr

“I never had a doctorate, so I had to make do with Quicksort.” —Sir Tony Hoare (unpublished interview for Algorithms to Live By)

FYI, There is absolutely an Internet Hall of Fame and anyone would be welcome to nominate Jay! https://www.internethalloffame.org

This is only half correct.

It is true that some/many Rolex AD’s will allocate the most desirable watches to customers with an existing purchase history, and that some customers therefore buy less desirable models in order to earn goodwill with the AD.

However, it is not the case that the most desirable watches are necessarily (or even on average) the most expensive models. For instance, it is generally the steel models that are the most desirable and command the highest markup from MSRP on the secondary market. The Submariner, the Daytona, the GMT-Master II: almost all of Rolex’s most iconic, most in-demand, most "flippable" watches are the full steel versions, which are the cheapest versions of those model families.

To give a concrete example, it is generally considered easier to get a full-gold GMT (~$43k) or a two-tone (half steel, half gold) GMT (~$18k) at an Authorized Dealer than it is to get the full steel version ($11k).

It is possible to reproduce one of the key claims in this post -- the "Russian tail" in the early voting tallies -- straight from the raw data hosted on the Clark County, NV website. This code can be run in a Colab notebook:

  # Download and extract zip file
  import requests
  import zipfile
  import io

  # Get raw data from Clark County website
  zip_url = "https://elections.clarkcountynv.gov/electionresultsTV/cvr/24G/24G_CVRExport_NOV_Final_Confidential.zip"

  # Download the zip file
  response = requests.get(zip_url)
  zip_file = zipfile.ZipFile(io.BytesIO(response.content))

  # Extract to the current working directory
  zip_file.extractall()

  # Close the zip file
  zip_file.close()

  import pandas as pd
  import matplotlib.pyplot as plt
  import numpy as np

  # Read the actual data, skipping the first three header rows and excluding downballot races
  df = pd.read_csv('/content/24G_CVRExport_NOV_Final_Confidential.csv', skiprows=3, usecols=range(21), low_memory=False)

  # Find the Trump and Harris columns
  trump_col = "REP"
  harris_col = "DEM"

  # Convert to numeric
  df[trump_col] = pd.to_numeric(df[trump_col], errors='coerce')
  df[harris_col] = pd.to_numeric(df[harris_col], errors='coerce')

  # Filter for early voting
  early_voting = df[df['CountingGroup'] == 'Early Voting']

  # Group by tabulator and calculate percentages
  tabulator_stats = early_voting.groupby('TabulatorNum').agg({
      harris_col: 'sum',
      trump_col: 'sum'
  }).reset_index()

  # Calculate total votes and percentages
  tabulator_stats['total_votes'] = tabulator_stats[harris_col] + tabulator_stats[trump_col]
  tabulator_stats['harris_pct'] = tabulator_stats[harris_col] / tabulator_stats['total_votes'] \* 100
  tabulator_stats['trump_pct'] = tabulator_stats[trump_col] / tabulator_stats['total_votes'] \* 100

  # Create subplots
  fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(10, 8))

  # Plot Harris histogram
  ax1.hist(tabulator_stats['harris_pct'], bins=50, edgecolor='black', color='blue', alpha=0.7)
  ax1.set_title('Distribution of Harris Votes by Tabulator (Early Voting Only)')
  ax1.set_xlabel('Percentage of Votes for Harris')
  ax1.set_ylabel('Number of Tabulators')

  # Plot Trump histogram
  ax2.hist(tabulator_stats['trump_pct'], bins=50, edgecolor='black', color='red', alpha=0.7)
  ax2.set_title('Distribution of Trump Votes by Tabulator (Early Voting Only)')
  ax2.set_xlabel('Percentage of Votes for Trump')
  ax2.set_ylabel('Number of Tabulators')

  plt.tight_layout()
  plt.show()

Vint Cerf & Bob Kahn (TCP/IP), Paul Baran (packet switching), Tim Berners-Lee (WWW), Marc Andreesen (Netscape), Brewster Kahle (Internet Archive), Douglas Engelbart (hypertext), Aaron Swartz (RSS, Creative Commons), Richard Stallman (GNU, free software movement), Van Jacobson (TCP/IP congestion control), Jimmy Wales (Wikipedia), Mitchell Baker (Mozilla), Linus Torvalds (Linux)...

...but you’re missing the point of my comment, which is simply to acknowledge and honor (my late dear friend) Peter.

Peter Eckersley (1978-2022) was posthumously inducted into the Internet Hall of Fame for his founding work on Let’s Encrypt. The Internet is a better place because of Peter (and his many collaborators and colleagues).

My understanding of the evidence is that it suggests the mechanism was made in Rhodes.

Leslie Lamport gives a short oral history of the Bakery Algorithm in Episode 3 of "Algorithms at Work": https://www.audible.com/pd/Episode-3-Concurrency-How-to-Coor...

Author of "The Alignment Problem" here, to say: Of course this question depends on your semantics of "harm," "AI," and "alignment," but by most definitions (and certainly by mine) the answer is overwhelmingly yes, many.

These harms can be diffuse at massive scale, and acute at small scale.

One example of each: (1) https://www.science.org/doi/abs/10.1126/science.aax2342 One of USA’s largest health insurers builds ML system for patient triage. It optimizes for a proxy metric of health need (namely, cost) rather than health need itself; consequently it deprioritizes and systematically excludes millions of people from access to health care.

(2) https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg Autonomous Uber car builds their braking system on top of a vision model that optimizes for object classification accuracy using categories of {"pedestrian", "cyclist", "vehicle", "debris"}; consequently it fails to determine how to classify a woman walking a bicycle across the street, as a result killing her.

In both cases, optimizing for a naively sensible proxy metric of the thing that was truly desired turned out to be catastrophic.

Watch movements are generally sensitive to magnetic fields, and can become magnetized and lose accuracy. Some watch models explicitly advertise their level of resistance to magnetism, for instance the Rolex Milgauss, which is designed to withstand 1,000 ("mille") gauss.

For the donation of a single dollar, you are given access to all of his videos. I would not describe that as "exclusive," although I understand what you are saying.

I can see why you might think that on a first impression, but I genuinely think you’d be surprised! (Book has been discussed on HN quite a few times, e.g., https://news.ycombinator.com/item?id=25717949.)

Current title ("Peter Norvig Leaves Google to Join Stanford HAI") seems to overstate slightly. According to Peter: "I still have my affiliation with Google, but will spend most of my time at Stanford."

Oh no, are you getting hit with an adwall? I’m able to open the following links in incognito windows with no trouble:

https://www.nytimes.com/2021/06/04/opinion/ezra-klein-podcas...?

https://www.nytimes.com/2021/06/04/podcasts/transcript-ezra-...

Brian Christian here. Honored to be the guest on the Ezra Klein show this week and by Ezra’s kind words: "Brian Christian’s recent book The Alignment Problem is the best book on the key technical and moral questions of A.I. that I’ve read."

Happy to take any questions, etc., if folks are interested! AMA.

There are some semantic issues, e.g., the distinction between "mandatory" and "discretionary" expenditures, but the data is pretty easy to explore.

Essentially you are "both right". > 50% of total spending is Social Security, Medicare and Medicaid. Whether that is "social services" is another semantic question.

And > 50% of the discretionary budget is defense.

But have a look for yourself: e.g., https://en.wikipedia.org/wiki/United_States_federal_budget

Five days ago, the National Association of Black Journalists revised their style guide:

'For the last year, the National Association of Black Journalists (NABJ) has been integrating the capitalization of the word "Black" into its communications.

However, it is equally important that the word is capitalized in news coverage and reporting about Black people, Black communities, Black culture, Black institutions, etc.

NABJ's Board of Directors has adopted this approach, as well as many of our members, and recommends that it be used across the industry.

We are updating the organization's style guidance to reflect this determination. The organization believes it is important to capitalize "Black" when referring to (and out of respect for) the Black diaspora.

NABJ also recommends that whenever a color is used to appropriately describe race then it should be capitalized, including White and Brown.'

https://www.nabj.org/news/512370/NABJ-Statement-on-Capitaliz...

This appears to have been part of what prompted a large number of newspapers to change their style guides this past week, including USA Today, NBC News, MSNBC, the LA Times, the Seattle Times, the Boston Globe, the San Diego Union-Tribune, and the Washington Post.

This comment is both unhelpful (dismisses the original claim using zero evidence) and completely incorrect.

Please look at the trendlines for "heroin deaths" and "synthetic opiod deaths" in the following graph of US CDC data, and note that from the turn of the millennium to 2017, opioid-related deaths in the US have increased more than 10x:

https://www.drugabuse.gov/related-topics/trends-statistics/o...

I have always found selfish routing to have interesting implications for a world in which traffic is more coordinated and centrally managed -- be it through present-day apps like Google Maps and Waze, or some near-future technology with autonomous cars communicating with one another. The happy Roughgarden/Tardos result that selfish routing is only 4/3 as bad as coordination is great, but it also means there’s not much slack to be picked up: fully coordinated routes would only be 3/4 as long as the status quo.

(Of course that's not the entire story on autonomous cars, for instance, because they will also get into fewer accidents, react to changing traffic conditions faster, can potentially drive more tightly at high speed, etc.)

Anyone interested in selfish routing should check out Braess's paradox, which is wonderfully unintuitive and strange:

https://en.wikipedia.org/wiki/Braess's_paradox

One of the most crucial parts of my reading workflow is opening things in new tabs. (It allows one to explore a topic by BFS rather than DFS.) Asking me "Don't you want to use our app?" is like asking, "Don't you want your browser to only support a single tab?" The answer is no...for all the reasons browser makers discovered in the 1990s.

It’s exactly the cutoff at zero that puts the "kink" into the function which makes ReLU nonlinear. :)

WWDC 2009: 13" MacBook Pro and iPhone 3GS announced

WWDC 2010: iPhone 4 announced

WWDC 2012: Retina MacBook Pro announced

WWDC 2013: new Mac Pro, Time Capsule, AirPort Extreme, and MacBook Air announced

WWDC 2017: iMac, MacBook and MacBook Pro, iMac Pro, 10.5" iPad Pro and HomePod announced

This sentence looks fine to me vis-a-vis subject-verb agreement: "[personnel + produce + precision] go."

How do you think it ought to read instead?

According to the new plans page [1], "shared link controls" for things like passwords and expiration dates will now no longer be available on the Plus plans like they have been for the last 5+ years? Am I reading this correctly? If so, that is a deeply unsatisfying regression buried in this announcement.

[1] https://www.dropbox.com/plans

One of the key results on the role of randomness in polynomial identity testing is what's called the Schwartz–Zippel lemma.[1]

As the wiki article states, "Currently, there is no known sub-exponential time algorithm that can solve this problem deterministically. However, there are randomized polynomial algorithms for testing polynomial identities."

[1]https://en.wikipedia.org/wiki/Schwartz–Zippel_lemma

(Note that this is in the context specifically of polynomial function equivalence.)

In the 1960s, the biostatistician Marvin Zelen proposed using something very much like the Pólya urn for clinical trials, calling it the "play the winner" rule [1]. This has had a major effect in causing a rethinking of the traditional randomized controlled trial, and these ideas are still making their way through the medical community today [2].

[1] https://www.jstor.org/stable/2283724

[2] https://www.fda.gov/downloads/MedicalDevices/DeviceRegulatio...

Looks like this is a 2005 study from William Jones, et al.

If folks are interested in this line of research, Jones has a 2007 book called Keeping Found Things Found that takes a more comprehensive look at personal information management.

Absolutely.

I find it really interesting that modern CPUs are designed for "interrupt coalescing," where they group interruptions into clusters. It’d be great if modern desktop/mobile OSes were designed for that: "Only notify me once every 5/10/30/60 minutes, but then tell me everything."

Okay, so we can say that:

1. CO2 is currently the highest it has ever been since human civilization has existed (10k years, roughly).

2. CO2 is currently the highest it has ever been since humans have existed (200k years, roughly).

3. CO2 is currently the highest it has ever been since primates have existed (50M years, roughly).

But for the sake of argument let’s say:

4. CO2 may not be the highest it has ever been since mammals have existed (200M years, roughly).

5. CO2 is not the highest it has ever been since multicellular life has existed on land (500M years, roughly).

Fair enough. Maybe 1-3 are less concerning to you than they are to me. When you say "the climate has been warmer in the past and CO2 levels higher" and are referring to a period prior to the existence of primates, it doesn't necessarily make me feel that much better. :) That said I’m nothing close to being expert in this stuff.

It’s interestingly difficult to think about time over so many orders of magnitude.