Ron Yorita wrote a great thesis. I need to make it easier to find the various parts of his great work.
So, I just bought a copy of Screenflow, an apparently reputable piece of screencasting software for the mac. They sent me a license code. Now it’s time to enter it. Here’s the window:
Hmm… says here all I need to do is… enter my admin password, and then the license code.
Why in heaven’s name would Screenflow need to know my admin password here?
My guess is that it’s because it wants to put something in the keychain for me. That’s not a very comforting thought; it also means that it could delete things from my keychain, copy the whole thing, etc. etc.
This is a totally unacceptable piece of UI. I think it’s probably both Apple’s and Telestream’s fault. Really, though, I just paid $100 and now I have to decide whether to try to get my money back, or just take the chance that Telestream isn’t evil.
Good news, students. I can’t accurately predict your final grades based solely on your first two assignments, quizzes, and labs.
I tried, though…
First, I took the data from Winter 2015. I separated the students into a training set (70%) and a validation set (30%). I used ordinary least-squares approximation to learn a linear weighting of the scores on the first two labs, the first two assignments, and the first two quizzes. I then applied this weighting to the validation set, to see how accurate it was.
Short story: not accurate enough.
On the training set, the RMS error is 7.9% of the final grade, which is not great but might at least tell you whether you’re going to get an A or a C. Here’s a picture of the distribution of the errors on the training set:
The x axis is labeled in tenths of a percentage point. This is the density of the errors, so the y axis is somewhat unimportant.
Unfortunately, on the validation set, things fell apart. Specifically, the RMS error was 19.1%, which is pretty terrible. Here’s the picture of the distribution of the errors:
Ah well… I guess I’m going to have to grade the midterm.
TL;DR: Molis Hai
Randomly generated passwords:
More randomly generated passwords:
- wargestood hury on,
- wealenerity," stp
- twould, aftilled himenu
- Whaideve awasaga
- andir her hing ples. F
- spe it humphadeas a
- to and ling, ace upooke,
- Mr. Syd, why.’ tred. "D
Yet More randomly generated passwords
- brothe aponder and," reasun
- ther atternal telle is be
- his me, he foundred, id
- allant our faces of rai
- time! What it of vail
- sourned," reate." Manybody.
- they would reck," read-doom
- raise thack ther meant,
Which of these look easiest to remember?
All three of these sets of passwords are randomly generated from a set of 2^56; they’re all equivalently secure. The second ones and the third ones are generated using markov models built from the text of Charles Dickens’ A Tale Of Two Cities, where transitions are made using Huffman Trees.
The secret sauce here is that since traversing a Huffman tree to a common leaf requires fewer bits than traversing that same tree to reach a deep leaf, we can drive the generating model using a pool of bits, and use varying numbers of bits depending on the likelihood of the taken transition.
This means that there’s a 1-to–1 mapping between the sequences of bits and the corresponding English-like textual fragments, thus guaranteeing the security of the passwords (or, more precisely, reducing it to the problem of generating a cryptographically secure sequence of bits, which many smart people have thought hard about already).
Another reasonable way to describe this process is that we’re just “decompressing” randomly generated crypto bits using a model trained on Dickens.
The only difference between the second and third pools is that the second one uses a 2nd-order markov model—meaning that the choice of a letter is driven by the prior 2—and that the third one uses a 3rd-order model, resulting in more Dickensian text—but also in longer text.
Naturally, you can push this further. When you get to a 5th order model, you get passwords like this:
- not bitter their eyes, armed; I am natural
- me. Is that. At fire, and, and—in separable;
- reason off. The nailed abound tumbril o
- and many more." “See, that,” return-
- falls, any papers over these listen
- do you, yes." "I beg to takes merc
- paper movement off before," said, Charles," rejoin
- that. She—had the season flung found." He o
Much more Dickensian, much longer. Same security.
You can try it out yourself; Molis Hai contains a small JS implementation of this, and a canned set of 2nd-order trees.
Please note that there’s nothing secret about the model; we’re assuming that an attacker already knows exactly how you’re generating your passwords. The only thing he or she is missing is the 56 bits you used to generate your password.
For a more carefully written paper that explains this a bit more slowly, see the preprint at ArXiv.
Naturally, you can use any corpus you like. I tried generating text using a big slab of my own e-mails, and aside from a serious tendency to follow the letter “J” with the letters “o”, “h”, and “n”, I didn’t notice a huge difference, at least not in the 2nd-order models. Well, actually, here’s an example:
- 0.91, Also: Zahid We rigor
- argustorigoring tent r
- Myrics args foling") (
- can’s fortalk at html-unds
- having avaScript" 0.88489232B
- John? I doe.cal fluore let a
- botheird, creally, there thic
- to ind [(solutell wil
It’s probably true that Charles Dickens wasn’t quite so likely to type “avascript” as I am. Or “html”.
To read the Racket code I used to generate the models, see github.
And for Heaven’s sake, let me know about related work that I missed!
Why are these things stuck in my head? They pop out all the time, and I can’t for the life of me figure out why.
- The Bilestoad - a game for the Apple ][e. Playing this game looks like piloting shrimp competitively.
- Captain Midnight - I started singing the theme from this game about two days ago. But why?
- that other western shootout game whose name I can’t even remember but whose theme song is stuck in my head forever.
One of my children is in third grade. As part of a “back-to-school” night this year, I sat in a very small chair while a teacher explained to me the “Math Practices” identified as part of the new Common Core standards for math teaching.
Perhaps the small chair simply made me more receptive, taking me back to third grade myself, but as she ran down the list, I found myself thinking: “gosh, these are exactly the same skills that I want to impart to beginning programmers!”
Here’s the list of Math Practices, a.k.a. “Standards for Mathematical Practice”:
- Make sense of problems and persevere in solving them.
- Reason abstractly and quantitatively.
- Construct viable arguments and critique the reasoning of others.
- Model with Mathematics.
- Use appropriate tools strategically.
- Attend to precision.
- Look for and make use of structure.
- Look for and express regularity in repeated reasoning.
Holy Moley! Those are incredibly relevant in teaching programming. Furthermore, they sound like they were written by someone intimately familiar with the How To Design Programs or Bootstrap curricula. Indeed, in the remainder of my analysis, I’ll be referring specifically to the steps 1–4 of the design recipe proposed by HtDP (as, e.g., “step 2 of DR”).
Let’s take those apart, one by one:
I can’t think of any relevant tags for this, and I’d like to think it’s because this is such a broadly applicable idea that it spans most categories.
The idea is this: many systems take inputs and produce outputs. Often, users of these systems would like to know WHY these systems produced these outputs.
I have two examples:
First, home automation. There’s a bunch of work on home automation and how it’s perceived, and one thing that comes across clearly is the frustration and unhappiness that users experience when the system acts in a way that isn’t what they expect.1
In response to this, it seems that the most obvious first step is to have a WHY button. In this case: why did you just turn all the lights off? Why is the thermostat cranked up to a hundred?
Second Example: programming, and more specifically debugging. This is a fairly obvious domain, and there’s already lots of work on time-travel debugging. The basic idea is the same: you have an outcome, it’s not the outcome you want, and you want to try to understand why. In this case, it’s not that we don’t know that we want a WHY button, it’s just really hard to implement.
In between these two extremes, there are lots of other examples. One of the one that gives me the most trouble is in system configuration. Why isn’t SpamAssassin running? (Should be not impossible.) Why is distnoted taking so much memory? (Harder.) Why does my mouse freeze whenever I hit a key in an xterm window behind VNC? (okay, that one is just a debugging question).
Actually, these last issues are interesting ones, because they dip into the space of “search”. That is, I would try to solve all of these last three by just using a search engine. At the moment, though, search doesn’t help me figure out why my program isn’t working (much) or why my lights aren’t on.
SO: what is it that makes the WHY button possible? In general, it appears to me that the answer is simply: declarative programming. When your program is written in a declarative language, it’s much more likely that you can get a good non-brain-bending “why” out of it.
Go, declarative languages!
Okay, now you can share all of the existing Human Interface research that already covers this topic.
1 : Brush, A. J., et al. “Home automation in the wild: challenges and opportunities.” Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, 2011.
This summer I started raising wild yeast. It was more or less a biology experiment, and it was extremely hit and miss, mostly miss, for the first few months.
For one thing, it took me quite a while to figure out that you actually do have to dump 4/5 of the starter down the drain every day; I was trying to feed an ever-growing bowl; not nearly enough food-to-yeast ratio, and the resulting bacteria were really… um… not the ones I was looking for.
Fortunately, I enjoy strange smells, but the rest of my family started moving out of the room when I would start feeding my yeast.
Anyhow, I’ve now figured out more or less how to keep a culture alive.
The next big step was reading Chad Robertson’s Tartine Bread. It turns out that if you have a reasonably lively yeast culture, and a little patience, it’s totally possible to make bread that the rest of your family actually likes a lot. Here’s a loaf from two weeks ago:
I also got an interesting lesson last week when I’d been away for two days and was trying to jump-start my starter again; the starter was actually just fine, but my “help” was actually doing terrible things to it; I started feeding it every eight hours, and I realize now that the fast feeding—or, more precisely, the fast 4/5-killing—was annihilating the population of my microbiological zoo.
Anyhow, it’s hard to knock the stuff out completely, and this week it’s bubbling away as well as ever.
Next week, maybe I start experimenting with higher hydration.
Ooh, just came across this today. I really liked this logo… I did this back in 2012, if the date stamp is to be believed. I think this struck a nice balance between the letter “r” and the lambda (reversed, yes).
So, I’d like to do some statistical analysis. I hear that R is really good at this. Let’s download it and take a look.
(Ten minutes later)
AAAHHH! MY EYES! THEY’RE BLEEDING!
What about Matlab? It’s the same story.1 As a programming languages person, these languages make me … well, angry.
Well, after thinking about this for a while, it seems to me that what I hate most about these languages is their complete lack of a small semantic core.
is this all just library design? Most of the things I really hate can easily be constructed in any dynamic library through a suitable application of
Terrible Library Design (tm)
… except that when it applies to things like vector dereference, it feels like fairly ‘core’ syntax.
Example time! First, R does this crazy thing in distinguishing logical from numeric vectors.
> a  "a" "b" "c" "d" > a[c(2,4,3)]  "b" "d" "c" > a[c(FALSE,TRUE)]  "b" "d"
In the first of these two array derefs, we’re using the indices from the vector to decide what elements of
a to take. In the second case, though, the index expression is a ‘logical vector’ and is therefore tiled to the length of the original one, and used to decide whether to take the corresponding element.
If you imagine this as part of a language semantics, you’d see this horrible side-condition attached to these rules, where array deref’ing works in totally different ways depending on the kind of argument it gets.
To say nothing of the silent tiling, which seems like an open invitation to horrible bugs.
But wait, we can compound this problem with some nasty coercion:
> a[c(4,c(FALSE,TRUE,TRUE))]  "d" "a" "a"
What on earth is going on here? First of all, vectors get silently flattened, so that
c(3,c(4,5)) is the same as
c(3,4,5) — ugh — but then, the logical values are coerced into numeric ones, so the index vector that’s produced is actually
c(4,0,1,1), which is then used to index the vector
a. But why are there only three values? Oh, well, there’s no index 0, so let’s just skip that one, shall we?
Honestly, I guess the real problem is in thinking of something like R as a programming language; it’s not. It’s a statistical analysis tool, with a rich text-based interface. After all, would I get upset if Photoshop used ‘b’ for blur and ‘s’ for sharpen and I couldn’t nest them the way that I wanted, using parentheses? Probably not.
And finally: apologies for everything I’ve written. I’ve used R for about fifteen minutes, and this piece is really just me blowing off a small amount of steam. Not well written, not well thought-out. Meh.
Actually, maybe not; I spoke with a friend yesterday, and I get the impression that Matlab may not be as horrible as R, here. ↩