entropy of English

This video on model scaling highlights an interesting point, which is to use the best models (nowadays LLM’s) to estimate the entropy rate of the source, in this case, the English language. This isn’t a new idea at all and empirical entropy measurements have been done in the past.

What’s interesting is that past estimates of bits per word of English have been way, way higher. Shannon’s original estimates are 11.82 bits per word, for example, or 2.62 bits per letter.
(Read the article)

hypermiling

There are many who have asked the question, whether coasting in neutral or engaging engine braking is more fuel efficient in coming to a stop. One such discussion led to the following top comments:

Engine braking. It shuts off fuel and lets the cars forwards motion turn the engine.
Coasting means you use fuel to keep the engine idling.
Next.

and

That assumes your car has deceleration fuel cut off.
But normally yeah, engine braking would technically use less fuel than idling/coasting.

Now I don’t know anything about cars but I do have a mental model of a mechanical engine. When no fuel is injected, engine braking should work by having the engine do mechanical work to compress and heat up air (then ejecting it). So RPM matters, as the higher the RPM the higher the energy absorbed per unit time.
(Read the article)

once in ten

Windows 10 is still the mess that Windows 8 was

It doesn’t appear that Windows 10 cleaned up much from Windows 8.

After one day of using the former, it is easy to come to the conclusion that Windows 7 is still the last version that anyone should run on its merits. Nobody in their right mind appreciates the mess that is Windows 8/10. (This is not Microsoft specific either, as Yosemite is also the last version of macOS that seems worth one’s while. The mobile OS angle has addled the brains of program managers at these companies.)

More egregiously, Windows 10 (Home) is also crippleware. It has disabled a number of features out of the box. No Guest account. No Group Policy Editor*. And hilariously, in what must be an attempt to emulate macOS, it automatically restarts after a system update, but saves no program states, with obvious defective results.

* How do I know it’s crippleware? Because vestiges still remain. You can exhume Group Policy Editor from buried .mum files. While you’re at it turning Windows 10 back into some semblance of Windows 7, wouldn’t you like to also disable Cortana?

usps insurance rates

What do USPS insurance rates tell us about its operations? Here is a 2007 document regarding insurance rates for domestic mail:

Prices for insurance coverage changed as follows:
Value up to $50 is $1.65.
$50.01 to $100 is $2.05.
$100.01 to $200 is $2.45.
$200.01 to $300 is $4.60.
The price per additional $100 of insurance, valued over $300 up to $5,000, is $4.60 plus $0.90 per each $100 or fraction thereof.

Crudely taking the mid-point of each bracket up to $300, we get implied loss rates of 6.6%, 2.73%, 1.63%, 1.84%, respectively. The rate converges asymptotically to $0.90/$100, or 0.9% implied loss. The numbers have such a wide range that it’s worth taking a closer look.
(Read the article)

die throwing problem

Here’s a link to a subtle probability problem.

You throw a die until you get 6. What is the expected number of throws (including the throw giving 6) conditioned on the event that all throws gave even numbers?

The “obvious” answer is incorrect.
(Read the article)

Android Pay and mobile payment

Where is the NFC antenna?
Who the heck knows.
http://pay.band/wp-content/uploads/2015/12/apple-verifone-640.jpg

I tried to use Android Pay twice today and despite both stores’ VeriFone terminals showing (and processing through the flow as if) they supported it, the success rate was only 50%. Android Pay is NFC-based. The time it succeeded, the phone immediately made contact. The time it failed, there was no semblance of any NFC connection. I know the phone’s NFC works in other contexts, so it is something in the technology (hardware, software) or workflow itself that makes it so unreliable. That is not even to mention the annoyance of Android Pay’s particularly patronizing attitude in requiring the phone to be unrooted and requiring a screenlock (even a perfectly insecure one) to be set. If any of those conditions fail for even five minutes, all the added cards are removed and have to be re-added (some banks require calling in for verification).

My point is nobody is going to use this idiotic piece of technology that makes things harder.

But mobile payment is not new. More than five years ago, a company called LevelUp started putting in QR-code based readers around town and having used it, it just works much better.
(Read the article)

Passing the Google “SafetyNet” test

If phone is rooted, follow the below steps:

  1. Uninstall Xposed Framework
  2. Uninstall Busybox
  3. Fully unroot (not only uninstall)

Root is required to uninstall the first two.

valuation of miles and points

As with any currency, the true value of various miles and points is determined by their trade value on a unified, liquid, and open market. Unfortunately such a market does not exist publicly — there is an underground secondary market with sometimes stale prices, minimum trade requirements, and counterparty risks, but generally speaking there is no notion of transparent prices. So you often see “Mileage Blogger” sites like this bandying about valuations pulled out of thin air. There should be a much more principled way to determine valuation that helps to make transaction decisions easier.
(Read the article)

Built-in audio variations

In light of Apple’s removal of the 3.5mm jack (for which there is an excellent analysis), I must say that there is something to be said for the inconsistency of analog audio output from built-in audio devices in laptops. Apparently there is quite a bit of variation that I hadn’t realized.

I ran this test using two sets of fairly wideband headphones and got results that were consistent across headphones but different between an HP laptop and a MacBook. The headphones were rated, respectively: (1) 15 – 20,000 Hz, 47 Ω input impedance; and (2) 15 – 24,000 Hz, 35 Ω input impedance. On the HP laptop with “IDT High Definition Audio” (92HD93 chip), I could hear a range from 30 Hz to 18 kHz. On the MacBook Pro with mid-2014 hardware, I could hear a range from 20 Hz to 16 kHz. I was quite surprised at the magnitude of this difference. A headphone amplifier (e.g. one built into the headphones) driven by digital input would eliminate this difference.

Next Page »