New tag: Division of labor

I wouldn't normally devote a post to announcing a new tag; they come up all the time in the natural course of things. However, I've recently found another common thread running through several posts: The division of labor between humans and computers.

One of the lessons of early AI work was that there's not a lot of overlap between what humans naturally do well and what computers naturally do well. I say "naturally" because much of the work in the ensuing decades has been in enabling machines to do things that don't map naturally to their capabilities.

For example, it's not hard to program a computer to calculate a million decimal digits of pi. It takes some cleverness to produce, say, a billion digits reasonably quickly, but the basic problem is not that hard. On the other hand, it's quite hard to get a machine to recognize faces or walk without tripping over, things which are easy for us.

One crucial aspect of engineering is making the best use of the resources you have. If your resource is a computer, try to put it on problems that involve crunching large amounts of data, not, say, perception, judgment, natural language processing or recognizing objects in the natural world. Machines can deal with those problems, too, to various degrees, but not nearly as easily as we can.

I've called this theme dumb is smarter. Division of labor is complementary. It has to do with putting humans in the loop so that the machines only have to do what they're good at.

Pandora's division of labor

A while ago Roku added Pandora to its selection of channels and a shorter while ago I got around to trying it out. I like it, though I don't listen to it all day long (I generally don't listen to anything all day long).

Pandora's main feature is its ability to find music "like" a particular song or artist you select. This is nice not only because it will turn up the familiar music you had in mind, but it will most likely also turn up unfamiliar music that you'll like. As I understand it, that's a major part of its business model. Record labels use Pandora to expose music that people otherwise wouldn't have heard, and Pandora takes a cut.

To that end, it will only allow you to skip so many songs in a given time (though there is at least one way to sneak around this). They pick out likely songs for you and they would like you to listen. You can, however, tell Pandora that you like or dislike a particular selection. Pandora will adapt its choices accordingly.

So how does it work? Pandora is based on the Music Genome Project, which is a nicely balanced blend of

• Human beings listening to music and characterizing each piece on a few hundred scales of 1 to 10 (more precisely, 1 to 5 in increments of 0.5).
• Computers blithely crunching through these numbers to find pieces close to what you like but not close to things you don't like.

This approach is very much in the spirit of "dumb is smarter". Rather than try to write a computer program that will analyze music and use some finely-tuned algorithm to decide what sounds like what, have the software use one of the simplest approaches that could possibly work and leave it to humans to figure out what things sound like.

Even the human angle has been set up to favor perception over judgement. The human judge is not asked to decide whether a given song is electroclash or minimalist techno, but rather to rate to what degree it features attributes like "acoustic guitar pickin'", "aggressive drumming", a "driving shuffle beat", "dub influences", "use of dissonant harmonies", "use of sitar" and so forth. There are refinements, of course, such as using different lists of attributes within broad categories such as rock and pop, jazz or classical, but the attributes themselves are designed to be as objective as possible.

This combination of human input and a very un-human data crunching algorithm is a powerful pattern. Search engines are one example, Music Genome is another, and if there are two there are surely more. In fact, here's another: the "People who bought this also bought ... " feature on retail sites.

Baker's dozen: True Knowledge

In his post on Wolfram Alpha, Mark Johnson mentions True Knowledge as a point of comparison, so naturally that seemed like a good place to try next. TK is supposed to be able to reason inferentially. For example, if you ask "Who are the Queen's grandchildren?" it will be able to find H.M. kids, and from them their kids, and thence the answer.

Game on.

TK wanted me to create a beta account, complete with re-CAPTCHA squiggly words and an email verification step, but it went ahead and logged me right in without verifying. A good thing, as I'd meant to give a different address.

• How much energy does the US consume?

TK answers "Sorry, I don't understand that question." It then wonders if I might be interested in any of a number of recent, utterly unrelated queries, but it also offers a list of standard search engine hits. These don't appear to be the top few hits for the question itself, but rather (I'm guessing) the top hits for several similar questions. It certainly seemed heavier on "how much energy" than other lists I've seen. It's probably not googling for the question verbatim, quoted or not. Hmm ... maybe it's googling for the parts of the question it deems important, something like "how much" energy consume?

• How many cell phones are there in Africa?

Again it's sorry, but the screen looks a little different. It tells me "It sounds like this is something True Knowledge doesn't know about yet. Most of what True Knowledge knows has been taught to it by people like you." and then goes on to paraphrase the question: "How many mobile phones (a telephone device) are geographically located (completely) within Africa (the continent)?" Interesting. Then follow the standard search engine results, probably based on the rephrased form.

But there's more. Right below the "Most of what True Knowledge has been taught ..." message is a button labeled "Teach True Knowledge about this". Sounds good, so I click the button and try to put in the answer from Wolfram Alpha. The tabs are intriguing, including a time period asking when the fact started being true and a couple that appear to provide a glimpse into the technical workings of the engine. Unfortunately, the "Add this fact" option appears to be grayed out, probably because I'm not a confirmed user.

• When is the next Cal-Stanford game?

Overall TK seems a bit sluggish. This is the cost of actually thinking about what you're saying. After pondering a while, TK decides it doesn't understand. The answer is similar to the one for "How much energy ..."

• When is the next Cal game?

Likewise.

• Who starred in 2001?

Well, it gets partway. In particular, it is able to extract just the kind of information I had hoped it would. Here's what it said:

Sorry. I couldn't find a meaningful interpretation of that. The following objects matched your query, but none of them are recordable media (such as TV series, movies, or radio broadcasts)

• the year 2001
• the integer 2,001
• the length of time 2,001 years
• the age 2,001 years

You may be thinking of a particular recordable medium that isn't in the Knowledge Base yet, in which case you can help out by adding it

The "adding it" link was not marked in any way (Say, by underlining it and putting it in blue, mabye?), but now that I see it pasted in here, I see there's a button for adding it. A couple of button-pressing guesses later, I get

2001 can also be used as a way of referring to 2001: A Space Odyssey, the 1968 science fiction film directed by Stanley Kubrick, written by Kubrick and Arthur C. Clarke. If this is actually the recordable medium you are adding, please click the button below.

Looking good ... and next I get

Here are the facts gathered from your information:

Fact 1:	[2001 a space odyssey] commonly translates as "2001"		source: You
Fact 2:	[2001 a space odyssey] is an instance of recordable medium (such as a TV series, movie, or radio broadcast)		source: You
Fact 3:	"2001" can denote [2001 a space odyssey]	source: You

I click the "Add these facts" button. It thanks me.

I retry the question. Same result as before. Most likely the new facts are still rattling through the various caches, or perhaps someone's moderating the input. But if the search succeeds for you later, you'll know whom to thank.

OK, if it's still learning the 2001 -> 2001: a space odyssey link, it presumably knows about 2001 under its full name:

• Who starred in 2001: a Space Odyssey?

And sure enough, there it is, with sources (Wikipedia) cited and even a chance to disagree with its findings.

• Who has covered "Ruby Tuesday"?

TK doesn't understand, but it does provide the Wikipedia entry in its list of regular search results. It also appears someone has asked it "How tall is Barack Obama in nautical miles?"

• What kinds of trees give red fruit?

Likewise (but with a different selection of random questions from other users). As always, the regular search hits are there, so I could always mine that for answers.

• Who invented the hammock?

This time I am asked to confirm its translation of my question. Toward the bottom of a long list of amusing attempts, including "Who is believed by a significant number of people to be the inventor of Hammock Music, based in Nashville, Tennessee, the label imprint under which Hammock released its initial two recordings, Kenotic and the Stranded Under Endless Sky EP?" I see the more relevant "Who is a key person or group involved in the invention of hammock, a fabric sling used for sleeping or resting?"

This sort of thing is the bane of natural language processing. The more you know about it, the more you appreciate the Google approach's^* brilliance in deliberately sidestepping it.

Chasing the link, I find that TK doesn't know, but could I tell it? I'm not going to try to educate it on this one.

• Who played with Miles Davis on Kind of Blue?

No comprende.

• How far is it from Bangor to Leeds?

After it asking me which of a long list of Bangors I meant, and my telling it I meant "Bangor, the city in Caernarfonshire, Wales, the United Kingdom," it tells me 182 km. If I add "in miles" to the query it tells me the answer to twelve decimal places. Perhaps it's impatient with me for asking so many questions and knows that spurious precision is a pet peeve of mine.

• How far is it from Bangor to New York?

This time, instead of giving me a list of Bangors to choose from, it gives a long list of eye-watering rephrases (for example: "How far is it from Bangor, the large town in County Down, Northern Ireland, with a population of 76,403 people in the 2001 Census, making it the most populous town in Northern Ireland and the third most populous settlement in Northern Ireland to The State of New York, the state in the Mid-Atlantic and Northeastern regions of the United States and is the nation's third most populous?"). Fortunately, the one I want is at the top: "How far is it from Bangor International Airport, the public airport located 3 miles (5 km) west in the city of Bangor, in Penobscot County, Maine, United States to the US state of New York?" The answer given is 547 km, or (by my rough-n-ready calculation) about 340 miles.

• How far is it from Paris to Dallas?
  

This time, fascinatingly, there is only one choice available: "How far is it from the French city of Paris to Dallas, the place in Dallas County, Texas, USA?" The answer given is 7928km, consistent with everyone else's answer to that particular form. Even more fascinatingly, it knows about Paris, TX. Asking "How far is it from Paris, Texas to Dallas" gives the rephrase "How far is it from Paris, the place in Lamar County, Texas, USA to Dallas, the place in Dallas County, Texas, USA?" and the answer 152km.

Wow. That was ... certainly interesting.

Clearly, it's a work in progress. Clearly, it's doing a lot of interesting stuff. Clearly, a lot of thought and effort has gone into it. I certainly commend the team for putting the thing together and letting the public have at it. Considered as a journey, it was easily the most engaging of the sites I've visited so far. Considered as a destination, not so much.

If nothing else, if you're wondering why it's so darned hard to make a computer answer questions "the right way", and why dumb can be smarter, a little browsing around True Knowledge might provide some insights.


^* Again, I acknowledge that "Google approach" glosses over a lot of other early work. Google is just the most prominent current exponent of the pure text-based approach as opposed to "semantic" approaches.

Baker's dozen: What did I expect?

With three engines tested (or three and a half, if you include Bing), it's starting to look like another case of dumb is smarter. The pure text-based approach grinds away happily and, even if you don't try to cater to its whims and just give plain English questions, it almost always finds something relevant. Often you'll have to chase links, but all in all you can do quite well.

So far, the semantic approach seems to do no better and may do worse. This might be an artifact of the smaller text base (more or less Wikipedia), but the smaller, more select text base is deliberately part of the approach. If you're lucky, the answer you seek might arrive wrapped up in a bow, but usually not, just as with the pure text approach.

It's also interesting that Ask, which started life as a plain English, sophisticated alternative to Google now seems to look and act pretty much like Google.

And yet ...

There is certainly information in the baker's dozen that could be exploited to give smarter results. The question is whether exploiting this requires magic, or just engineering. I'm not going to make a call on that, but here's what I see:

• How much energy does the US consume?

"How much" says "look for a number" and "energy" says look for a unit of energy (BTU, Joules, kWh etc.) Anything that doesn't seem to give such a number in the context of "US" and "consumes" is probably not useful.

• How many cell phones are there in Africa?

"How many" suggests a number. For bonus points, the number is probably going to be within an order of magnitude of the population.

• When is the next Cal-Stanford game?

"When is the next ..." is a formula suggesting a schedule, timetable or such. A search for "Cal-Stanford game" will probably correlate highly with football (as opposed to, say, "Carolina-Duke game"). If so, that would suggest a football schedule.

• When is the next Cal game?

I found this one interesting. Most Americans will know that "Cal" is one of the short forms of "California". The search engines know that, too, and it throws them off. "Cal" in the context of "game" almost certainly refers to the University of California (Berkeley) Golden Bears. And again, "When is the next ..." rules out the hits for "California Gaming". In this case, even dumb is too smart for its own good.

• Who starred in 2001?
• Who starred in 2001: a Space Odyssey?

"Who starred in X" indicates that X is a film, TV show or play. Powerset appears able to grok this.

• Who has covered "Ruby Tuesday"?

"Who has covered" indicates a song, though not as strongly as "Who starred in ..." suggests an acting role.

• What kinds of trees give red fruit?

"What kind(s) of X" indicates that the answers should be Xs.

• Who invented the hammock?

"Who" indicates that the answer is a person or group of people, but in the case at hand the group is pretty abstract.

• Who played with Miles Davis on Kind of Blue?

"Who played with X on Y" indicates that the answer is a person, probably a musician or member of a sports team ("Who played with Satchel Paige on the Monarchs?" -- another fairly impressive list)

• How far is it from Bangor to Leeds?
• How far is it from Bangor to New York?
• How far is it from Paris to Dallas?

I wrote that "In these, of course, the implicit question is 'Which Bangor?' or 'Which Paris?'" Google Maps handles this well by simple heuristics based (I assume) mainly on size. The "how far is it from X to Y" formula indicates that X and Y are places and the answer is a distance.


Now let me be clear here that when I say that some formula indicates something about the answer, I'm not saying that a search engine ought to be able to exploit that information. I'm well aware it's not as simple as it might look. Rather, I'm saying that if search engines are really going to get smarter, this is the kind of information they'll need to be able to find and exploit.

As to the original question of what did I expect, I'll just say that nothing so far has been surprising, except perhaps that Google handles plain English questions better than I thought it might.

A baker's dozen for Search 2.0

[Re-reading several years later, it's clear that several things have changed.  I'm not going to comment on them specifically since, obviously, this relates to Google's core business.  I would, however, encourage anyone who finds this series interesting to retry the experiment with whatever's currently available.  You should also try the kinds of searches Joe Andrieu mentions in his comment on this post. --D.H. Jan 2016]

As mentioned before, there seem to be a number of next-generation search engines coming out which claim to go beyond merely looking for keywords and delivering up documents. I have no idea how these perform, and the only way to know is to find out. So here are 13 questions, off the top of my head with very little method behind them, which I intend to pose to various engines. I suspect many of them are not really fair questions for search engines, but I could be wrong. Again, that's why we run the experiment.

1. How much energy does the US consume?
2. How many cell phones are there in Africa?
3. When is the next Cal-Stanford game?
4. When is the next Cal game?
5. Who starred in 2001?
6. Who starred in 2001: a Space Odyssey?
7. Who has covered "Ruby Tuesday"?
8. What kinds of trees give red fruit?
9. Who invented the hammock?
10. Who played with Miles Davis on Kind of Blue?

In these next three, of course, the implicit question is "Which Bangor?" or "Which Paris?"

11. How far is it from Bangor to Leeds?
12. How far is it from Bangor to New York?
13. How far is it from Paris to Dallas?

My "dumb is smarter" spidey-sense is tingling again

But first, a digression.

Way back in high school it was time to pick semester projects for computer science class. To give you some idea of how long ago this was, the project was to be written in FORTRAN -- that's FORTRAN in all caps -- on the local university's "time-sharing system." At the time I thought AI was about the coolest thing you could do with a computer^*. I started casting about for a worthy problem.

It seemed all the good ones were taken. Conversing in English in the manner of a non-directive therapist? Done. Moving blocks around in an imaginary world in response to natural-language commands? Solved problem. Chess? Well, folks were working on it, but everyone knew chess was hard. Could a computer ever surpass a human master's deep positional understanding? Doubtful. Certainly this was too much for a semester project.

After some pondering, I hit on the idea of "how come?" puzzles. Example: A man is found murdered in a room with 53 bicycles. How come? Answer: The man is a gambler, the bicycles are playing cards, and two of the 53 cards are the ace of spades. One is expected to solve the puzzle by asking yes or no questions which the computer would answer ("Were the bicycles stolen?" "No.") The yes/no format is nicely circumscribed. The questions less so, but therein lies the challenge, no?

I'm sure the teacher must have entertained, if only for a fleeting second, the idea of letting me actually try this. I'm grateful he didn't go that route. By the time I'd finished bashing my head against the finer points of parsing natural language and representing facts about the world, and realized that the resulting hash could not be made into something coherent, it would have been too late to do an actual project.

Instead, I reluctantly agreed to write an adventure program along the lines of ADVENT. Such "text adventures" were popular on the PCs of the day. You moved and acted in a fantasy world by typing one or two word commands (LOOK ... TAKE ROCK ... THROW ROCK ...RUN) and the game would give you a more or less sensible description of what happened (You see a dragon ... You pick up the rock ... The rock hits the dragon. The dragon appears angry.)

ADVENT was available on the time-sharing system. I forget just how we got away with playing it -- perhaps after classes or as a reward for finishing assignments early? -- but I had managed to find my way through the mazes of twisty little passages and figure out everything but how to solve the final puzzle. More importantly, through experimentation, study and the help of Byte and Dr. Dobb's I had formed a pretty good theory of how such games worked (patience, I'm getting back to the web.stuff) and how to write them.

The project was a success. I got a good grade, my friends thought it was cool and I learned quite a bit about data structures (and this was in FORTRAN, Sonny, none of this newfangled OO stuff). But it was also a bit disillusioning. The program wasn't doing anything deep or smart. It certainly didn't understand English. It understood two lists of words, namely a verb list and a noun list, comprising a few dozen words in all. It could parse a "sentence" consisting of a verb optionally followed by a noun. The model of the world was a few arrays. The prose descriptions were canned, of course, essentially the same from run to run.

And yet it was by far the best thing I'd written up to that point, and much better than anything I could have produced in a full-blown tilt at the AI windmills.

Now, where was I?

Wolfram Research, which has more than earned its keep by producing Mathematica, Mathworld and a smörgåsbord of other useful resources on mathematics in general, is announcing a new portal called Wolfram Alpha, set to open for business sometime this month. It aims to “compute whatever can be computed about the world.” Nothing small, then.

Not being one of the "few select individuals" to see the thing in beta, I can't say exactly what it does, but evidently it's meant to answer general questions on the order of "How many broadband connections are there in Sweden?" and point you at the web resources it used to come up with its answer. Nice.

Also according to some of the select individuals in question, it has mixed success in doing this. When it hits, it often hits a home run. But it seems to strike out a fair bit as well. Hmm ... haven't I seen this movie before?

But for the caliber of the folks behind it, I wouldn't really pay much attention to Alpha. Given that it's Wolfram, I'll certainly give it a try. Not too long ago I spent quite a bit of time trying to track down a seemingly simple figure, something like "the number of hosts on the internet" only to find myself lost in a twisty maze of little tabulations, all different. If Wolfram can pull together such things, accounting reasonably for differences in format, not to mention methodology, I'll certainly be interested.

If they're expecting something to replace Google and Wikipedia, I doubt that will happen, even if the thing works perfectly. On the other hand, if they're shooting for steady traffic from people who occasionally need to know facts that aren't in Wikipedia yet but maybe should be, that might work. I don't know how or whether you make a viable business of that, but Wolfram seems content in at least some cases just to put good stuff out there for the good of both their brand and the ecosystem at large. And more power to them.

[As noted elsewhere, I still use Wolfram Alpha for a certain class of questions --D.H. May 2015]

---

* Taking into account that "AI" and "computer" mean considerably different things now than they did then, I'd still put AI fairly high on the list.

No, I won't help you crack Worldscape

We have a new record holder for most random Google search to bring up a Field Notes post. As of this writing, “how to guess someones password on worldscape”pulls up 80 hits, of which the second is the Field Notes archive for November 2007.

Though the preamble at the top of Google's cached version claims that someones and password "only appear in links pointing to this page" (one day I shall track down what they're driving at with that), both password and the uninflected someone appear in the page, along with how, to, guess, on and worldscape.

In other words, Google is perfectly correct in bringing up this page. Normally, it would have put millions of other articles ahead of such a hit in the list, but in this case there just don't seem to be that many candidates to choose from. Sometimes dumb can only be so smart. Fortunately for our searcher, some of the other hits have to do with the art of guessing passwords, a good thing to read up on next time you have to make one up.

For the curious and/or obsessive, here's the rundown:

• guess shows up mostly in the articles on anonymity, with regard to guessing someone's identity, and otherwise where I say I'm guessing (four times in that apparently uncertain month)
• password shows up in the post on trusted computing
• someone shows up because it's a reasonably common word. It's probably a bit more common than chance in the anonymity articles
• how and to are most likely dropped on the floor, but of course they appear in numerous places
• And the punchline: worldscape refers to the Worldscape Laptop Orchestra, probably not what the searcher was trying to gain access to. But I do see a note there that I need to find a better link. Don't they have a home page yet?

Radio bookmarks

It's fund-raising season at NPR again, and some stations are giving out "radio bookmarks" as premiums. What's a radio bookmark? It's basically a USB device that you can push a button on while you're listening to the radio. It will record enough information for you to go to the station's website when you next go online, find out what was on just then, and even play it back.

In other words, it records the time.

Sort of disappointing when you put it that way, but on the other hand there's no rule that says an idea has to be complex for it to be at least moderately useful. Back in the dotcom days, the same idea was supposed to make Sony a lot of money. Dan Bricklin has a review of it (then called "emarker") from Comdex 2000. The original emarker could store (drumroll, please) up to 10 events, which seems ridiculously small even for the times, but what do I know?

Bricklin makes the point in real time that I was going to make eight years after the fact, that the setup "puts the appropriate intelligence at the right places". In other words, the device itself is dumb. The intelligence comes from the underlying database via the web. I agree with Bricklin that this is the right way to do it, if only because you can easily correlate with other databases as well (video springs to mind). In a way it's a good example of "dumb is smarter", though in this case it's a particular piece, not the system as a whole, that's deliberately dumb.

It's a separate (and interesting) question whether the emarker/radio bookmark "uses the Internet in ways that show the future". Eight years on, I'm thinking not so much, but then the question becomes "why not?". The general idea of connecting a fairly dumb device up to the web to get a smart system overall seems good. Maybe there are other examples hiding in plain sight?

The Economist on crowds

The Economist has a couple of interesting recent articles about crowds. This should hardly be a surprise, given that economics is all about the aggregate behavior of large numbers of people.

The first article deals with crowdsourcing, its benefits and limitations. It mentions several more interesting crowdsourcing examples I forgot (even though I read the article before I wrote that post), including the 1714 longitude prize and Google's initiative to have volunteers in India carry around GPS units to help map India's roads.

It also wonders whether crowdsourcing and business are a good match. There are plenty of examples of volunteers joining in a project for fun and the chance to be part of something important, but this seems inherently hard to monetize. People don't seem so keen on volunteering their efforts so someone else can make money. The counterexample would be the various prize competitions, including the longitude prize and the various X prizes, but in those cases the participants are in it to be part of something important and to make themselves money.

As an aside, if you're interested in the general topic of crowdsourcing before the web, I'd recommend Longitude, by Dava Sobel, and Simon Winchester's The Meaning of Everything, about the Oxford English Dictionary.

The second article doesn't explicitly mention the wisdom of crowds, but cites a study finding that "when individual drivers each try to choose the quickest route it can cause delays for others and even increase hold-ups in the entire road network." Interestingly enough, closing down sections of road can actually make things run faster, not too surprising if you consider that taking a little-used side street generally involves merging back onto the main road at some point. That will tend to gum up the main road, but if the side street saves you enough time you won't particularly care.

This is yet another example of a "local optimum," a fundamental problem with "dumb is smarter" approaches. Deliberately ignoring the big, complex picture and concentrating on small, local features can lead to solutions that look good on the small scale but bad on the large scale. This is another of those lessons we get to learn as many times as we like. Genetic algorithms provide another fairly recent example. The "no free lunch" theorem is also worth keeping in mind.

Weather modeling at the NHC

This is another "not really about the web but I did find it there" posts.

In my post on the National Hurricane Center I said it was fascinating to get to know the personalities of the various computer models. Well, fascinating for a geek, at least. If, like me, you like that sort of thing, the NHC's explanation of what models they use, and how, and why, is a veritable feast. Some of the high points:

• There are a dozen basic models, plus several more ensemble models combining them.
• The forecasters don't rely on any one model in making a forecast. There's no "model of the day". Instead, they consider the results of all of them and make a judgment call.
• The aggregate results of the models are generally more accurate than any particular model. Another "wisdom of crowds" effect, if you will.
  
• The NHC continually reviews its forecasts to see whether the models and the official forecasts are "skillful" A forecast is skillful if it's more accurate than the statistical models, which just look at what past hurricanes have done and don't even try to take current weather conditions like wind and sea surface temperature into account.
• By that measure there are skillful models, but it's a difficult bar to clear. Not quite "dumb is smarter", but dumb is smarter than you might think.
  
• In particular, it's only now becoming possible to predict the intensity of a storm skillfully. Predicting a storm's track skillfully is less of a problem.
• Finally, implicit in the whole report is the understanding that in good science and engineering, it's vital to know what you don't know.

Weather forecasting is a notoriously hard problem. It was one of the drivers behind the discovery of chaos theory. The NHC's technical model summary provides an excellent window into the process of using computer modeling on hard problems in the real world. It also gives some insight about how and whether to use "dumb" statistical models and how and when to try to be "smart". Good stuff!

Another classic example of "dumb is smarter"

When I originally tried to come up with a list of "dumb is smarter" examples, I knew I was missing and at least one, and now I remember what it was.

The Iterated Prisoner's Dilemma is an abstract game, with many real examples or near-examples, in which

• Two players repeatedly face a choice of acting generously, called "cooperating", or acting selfishly, called "defecting".
• At any particular turn, they'll both do better if they both cooperate.
• However, if only one cooperates and the other defects at that turn, the cooperator gets shafted.
• Therefore, looking at a particular turn in isolation, the only rational choice is to defect, with the result that both players do less well than if they'd both cooperated.
• However, since the choice is presented repeatedly, players have a chance to base their present actions on the results of previous turns.

It turns out that under these rules, rational players can agree to cooperate over the long term, even though it makes no sense in the short term ("it turns out that" is math-geek for "I don't want to go into the details"). It's an interesting result.

To probe this further, Robert Axelrod organized a tournament in which computer programs could compete with each other at the game. The tournament attracted considerable interest, and there were many competitors, some quite sophisticated in design.

And now the payoff: The winner, Anatol Rapaport's "tit-for-tat", consisted of four lines. Of BASIC. Its logic was:

• Cooperate on the first turn.
• After that, do whatever the other player did on the last turn.

This isn't an unbeatable strategy. It will lose, by just a bit, to "always defect", but the winner was determined by who had the best total result against all opponents. By always passing up the bigger gain of cooperating, "always defect" gets a low total score. Tit for tat does better than that when playing anyone who cooperates. It gets the best score possible when playing itself (or when playing "always cooperate" or anything else that always happens to cooperate when playing it).

In a later edition of the tournament, a team from Southampton University was able to beat tit-for-tat by having multiple copies of its program collude, but that's a different story (and even then who won depends on how you measure).

What do I mean, "dumb is smarter"?

I previously mentioned Google's page rank as an example of "dumb is smarter" -- doing better by not trying to understand anything deep or do anything clever. Some other examples:

• On balance, chess programs that bash out large numbers of positions do better than ones that try to emulate human judgment.
• SPAM filtering is an interesting arms-race-in-progress, but the most effective technique I've run across is a simple whitelist. Bayesian filtering worked well for a while and all it does is crunch numbers. Any filter based on rules is vulnerable to gaming once people figure out the rules.
• One of the many long-raging unresovled debates in the financial world concerns whether you can do better by carefully picking which stocks you buy and sell and when, or whether you should just make regular purchases of an "index fund" that tracks, say, the Russell 5000 and forget about it. I'm not going to take sides on that one. Rather, the point is that it's a serious debate at all. (Ironically, one of the best-known proponents of the "dumb money is smarter" approach, Warren Buffett, is also considered one of the best stock pickers ever.)
• Every so often, it seems, someone puts out a program that appears able write free-form prose or even take part in a natural-language conversation with a person. Its understanding and grasp of humanity seem almost uncanny. Sooner or later, though, it comes out that the putative AI is really just doing some simple manipulation and the reader is assigning meaning to what closer inspection reveals to be gibberish. Everyone goes back to what they were doing, and real progress in natural language processing continues slowly and painstakingly. The classic examples are Eliza and Parry. Racter and Mark V. Shaney come to mind as well. These days, people write "chatterbots", some of which are one-trick ponies and some of which are considerably more sophisticated.  [See this post on the other blog for a more recent example that grabbed headlines -- D.H. Dec 2018]

I'm not particularly knowledgeable about AI research, but my understanding is that there is a basic division between "strong" and "weak" AI. Strong AI has the rather broad goal of being able to do anything a human mind can do, ideally better. The notion of strong AI can be traced back to the Turing test, though there is debate concerning to what extent the Turing test is a valid measure of intelligence.

Weak AI aims to solve a particular engineering problem, say, beating a human master at chess or finding relevant web pages containing a particular set of terms. A weak AI does not need to claim to be conscious, or to have any particular understanding of the problem at hand. It just has to produce useful results. At which point, we decide that it can't possibly have any real intelligence, since we understand what it's doing and how.

In terms of the challenges of the 20th century, "producing a strong AI" compares to "curing cancer", while producing a weak AI is more like sending a rocket to the moon.

When I say "dumb is smarter", I'm not saying that strong AI is a useless goal, only that it's a very difficult, long-range goal which to this day is not particularly well-defined (though it is better defined than it was forty years ago). As such, it's more likely that progress will come in small steps.

Like anything else, "dumb is smarter" can be taken too far. The best chess programs tend to incorporate the advice of human grandmasters. When Google answers "What is the time zone for Afghanistan?" with a separate line at the top with the time zone, clearly it's doing some sort of special-casing. The absolute dumbest approach is not always the absolute best, but it does seem that the best trade-off is often much closer to it than one might think, and as a corollary, the dumbest approach is often a good place to start.

Searching for a smarter search engine

One look at Google's quarterly reports should be enough to understand why people are still trying to build a better search engine. Google search does a great job. It will come as no shock that I've consulted it repeatedly in practically every post here. A friend once described it as adding (say) 25 points to his IQ, though not everyone agrees with that assessment.

I've cited Google as a classic case of "dumb is smarter". Google doesn't try to do anything one might consider "understanding" the material it's indexing. For the most part it just looks at words and the links between pages. There is some secret sauce involved, for example in handling inflections or making it harder to game the rankings. Mainly, though, Google wins because its PageRank algorithm turns out to do a good job of finding relevant pages and because it throws massive amounts of computing power at indexing everything in sight [There's a lot of secret sauce involved in getting that to work at the scale Google operates on].

Google is the dominant search engine, but that doesn't man there's no room for other engines, particularly engines that take a noticeably different approach or that try to solve a noticeably different problem. Powerset is one such engine. Rather than trying to index the entire web by keyword, Powerset answers English queries about material in Wikipedia. Without delving into a proper product review or comparison, which would have to include at least Google and, say, Ask (formerly Ask Jeeves), I'll just note a few impressions and head on to my real goal of blue-sky speculation [geek note: The "power set" of a set is the set of all that set's subsets; less formally, all the combinations of given set of elements.].

Suppose you want to know when John von Neumann was born. You ask "When was John von Neumann born?" Hmm ... oddly enough, it didn't answer that one directly. It did give the Wikipedia page for von Neuman, which gives the answer (December 28, 1903). "When was Mel Brooks born?" works more as intended, with a nice big "1926" at the top of the results. It also shows a link to a page that says 1928, but seems to know better than to believe it.

Other examples

• "Where is the world's tallest building?" turns up the list of tallest buildings.
• "What is the time zone for Afghanistan?" turns up a list of pages, the first of which mentions the right answer.
• "How much money has been spent on cancer research?" turns up a link giving a figure for the UK, but nothing suggesting an overall figure
• "Why is there air?" brings up the Bill Cosby album of the same name.

Beyond accepting questions posed in plain English, Powerset also aims to give you a richer view of the results it finds. This includes an outline of the page contents and a list of "Factz" gleaned from the text. These take the form of short subject-verb-object near-sentences like (in the case of the "tallest building" article) "dozens measure meter" and "television broadcasts towers". Click on one of these and it highlights a relevant passage in the text, for example "In terms of absolute height, the tallest structures are currently the dozens of radio and television broadcasting towers which measure over 600 meters (about 2,000 feet) in height." or "In terms of absolute height, the tallest structures are currently the dozens of radio and television broadcasting towers which measure over 600 meters (about 2,000 feet) in height."

It's not immediately clear what this is supposed to give me. Powerset says "For most people, places and things, Powerset shows a summary of Factz from across Wikipedia," and to illustrate this, it shows a section of a table of Factz about Henry VII -- whom he married (wife, Anne Boleyn ...) what he dissolved (monestaries, Abbey ...) and so forth. Evidently Henry provides a better example than tall buildings do.

The Factz summary appears to be the sort of thing that Powerset is really driving at. It's certainly the sort of thing that initially drew me to take a look. Rather than just index words, Powerset attempts to extract meaning from the text and present it in a structured way. In other words, it tries to be smart and, in some limited sense, understand the material it's indexing. For example, along with the listing of three-part Factz, it will also display "things" and "actions", with items it deems more significant shown larger.

If we view this smarter approach as an attempt at understanding, however limited, then I'm not sure that the Powerset engine understands all that much. It seems pretty good at distinguishing nouns from verbs, but beyond that, I'm not sure what "dozens measure meter" really signifies. Even in a seemingly simple factual statement like the one quoted, there is more going on than "dozens" "measuring".

It matters that it's dozens of towers, not dozens of meters (or dozens of eggs). It matters that the towers measure more than 600m tall and not less. It matters that the towers are being judged tallest in the limited context of "absolute height". It matters that this is "current", since the Burj Dubai, when completed, will be the tallest completed structure, period. This matters particularly because much of the article is spent wrangling over the meaning of "tallest", a debate which will soon be moot, at least for a while. The Factz approach appears to miss all this, none of which is particularly subtle from a human point of view.

Google, in the meantime, doesn't try to do any of this, but seems to do just fine on the queries above, given verbatim (not in googlese, and without quotes). For "What is the time zone for Afghanistan?" for instance, it said "Time Zone: (UTC+4:30) according to Wikipedia" right at the top. And, of course, Google indexes the entire web ("entire web" defined as "anything you can google", of course), in part because it doesn't spend a lot of time trying to extract meaning. As for the structured view, Wikipedia pages are already outlined, and I'm not sure what Factz give me that ordinary text search doesn't.

Ah well. Understanding natural language isn't just a hard problem, it's a collection of several hard problems, and not a particularly well-defined collection at that.

I don't want to leave the impression that Powerset is useless, and I particularly don't want to denigrate the effort behind it. In fact, I'd encourage people to at least try it. Tastes vary, and some may well find Powerset a nicer way to navigate Wikipedia. Nonetheless, Powerset only serves to confirm my impression that dumb is indeed smarter, and that Google's "we don't even pretend to understand what we're indexing" approach sets the bar remarkably high.

What a concept. Or rather, what's a concept?

One theme I've had kicking around in my head for a while, and may yet write up in earnest, is the concept of "dumb is smarter". The idea is that you can often do better by giving up on the idea of "understanding" the problem your solving and using a blatant hack. For example, Google relies on page rank -- the way a page is connected to other pages -- rather than any abstract understanding of a document, to decide what hits are likely to be "relevant".

The technique of Latent Semantic Analysis is an interesting case. LSA attempts to solve some of the well-known problems with searching based on words alone, particularly synonymy and polysemy.

Synonymy -- different words meaning the same thing -- means you can ask for "house" and miss pages that only say "home". Worse, you don't know what you're missing since, well, you missed it.

Polysemy -- the same word meaning different things -- means you can ask for "house" and get pages on the U.S. House of Representatives when you wanted real estate. This is probably more of an annoyance, particularly since you probably want the more popular sense of a word and not the one that that sense is drowning out.

LSA tries to mitigate these problems by starting with information on what words appear in what documents, then applying a little linear algebra to reduce the number of dimensions involved.

This means that, for example, instead keeping a separate scores for "house", "home" and "senate", there might be one combined score for "house" and "home" and another one for "house" and "senate". A document that contains "house" and "home" but not "senate" would be rated differently from one that contains "house" and "senate" but not "home", which is just the kind of thing we're looking for.

This combined system is called "concept space". Does it deserve the name?

On the one hand, yes, because intuitively it reflects the idea that "house" and "home" can represent the same, or at least related concepts, and because it seems to do fairly well empirically in mimicking how people actually rate documents as "similar" or "different".

On the other hand, clearly no, since all we're doing is counting words and doing a little math, and also because the "concept space" can include combinations that don't have much to do with each other, but happen to fall out of the particular texts used -- maybe "house" and "eggnog" happen to appear together for whatever reason.

The last would be a case of "correlation doesn't necessarily mean cause", and the interesting thing here is that LSA seems to do a decent job of emulating faulty human reasoning. People make that particular mistake all the time, too. As always, one must distinguish "human-like" from "intelligent".

Now that I've tagged it, how do I find it?

I try to assign a few relevant tags to every post, as is common practice. Blogger's tagging facility (along with lots of others, I expect) will suggest tags you've previously used. For example, if I type "bl" into the space, up comes a list containing blacksmithing, BLOBs, blogger, blogs and Eubie Blake. Hmm ... I'd forgotten I'd used "blogger" as a tag before. Curiously, J and only J, pulls up only personal names. I've been known on occasion to run through the entire alphabet to make sure I haven't missed anything.

Conversely, if nothing pops up for a tag, you know you haven't used it. That happened for DMCA on the previous post. That's funny. I'm sure I've mentioned the DMCA before. Aha, there are a couple. All I have to do is search. A couple of quick edits and they're tagged now, too.

Now, if anyone could find those by searching, why add the tag? Indeed, I almost didn't add the tag, because those posts don't happen to be about DMCA. They mostly just mention it in passing. I opted to tag those because the passing mention was apropos of what DMCA is. On the other hand, this post contains the letters "DMCA" but has nothing to do with it.

Likewise, there are a couple of posts, like this one or maybe even this one, that don't mention DMCA specifically, but might be relevant. On review, I'm not tagging those DMCA either. They're both tagged "copyrights" now.

While reviewing that I ran across another post that should have been tagged "copyrights" but wasn't. That's a recurring problem with tagging. Which tags should I use here? Which tags did I put that article under? Did I think to put it under "copyrights"?

All of this leads me to a few general points on tagging:

• It's not particularly new. It's much the same as traditional indexing. In web form, it goes back at least to categories on wikis.
  
• It's time consuming and subjective. Care and feeding of categories is an important part of wiki gardening, for example.
• It's most useful exactly where more automated tools don't work. If you want to find posts here that mention DMCA, just search. "DMCA" is probably not a particularly useful tag. "Annoyances" and "Neat hacks" make better use of the tool.
  
• Likewise, tools like sub-categories or other schemes for tags to aggregate other tags, though useful, aren't foolproof.
• On the other hand, tags are still nice for browsing, particularly the more abstract ones.

Tagging is time-consuming because it's subjective, but this is also what makes it useful and even intriguing ("why did they tag that that way?"). Automated searching remains the workhorse, and will probably continue to.

Crowds and wisdom

Suppose you're measuring, say, the four walls of a room. Have twenty people each do the measurement with off-the-shelf tools. Then do the measurement very carefully with a laser interferometer or whatever. While we're in gedanken mode, assume that the room itself is very precisely joined, so that measuring to a few decimal places actually means something.

The central limit theorem tells us that the measurements will tend to fit a normal (i.e., "bell curve") distribution peaking very near the precise length. If you take the average of the imprecise measurements for each wall, the result will generally be quite close to the precise measurement. If you consider all four walls, the combined result -- the four averages -- will generally be closer to the precise measurement than the best individual set of four measurements, assuming the errors in the imprecise measurements are random.

Now take a typical "wisdom of crowds" example: the Oscar ™ party where each guest guesses who will win. A ballot consisting of the most popular choices almost always does better than the best individual ballot. Clearly this is not quite the same as the measurement problem above. At the very least you'd need a different metric. On the other hand, is this "wisdom", or just statistics at work? Two possible answers, not necessarily incompatible:

Wisdom of crowds is about more than Oscar ™ parties. Surowiecki's original book talks about situations like pedestrians optimizing traffic flows or markets setting prices. The key ingredients for crowd wisdom, he argues, are diversity of opinion, independence, decentralization and aggregation.

The party example is fairly low-powered. It might be explicable in terms of statistics, but something like cars not hitting each other or customers distributing themselves among popular restaurants may not.

Wisdom is more about statistics than we might like to think. The words "wise" and "wizened" come from the same root, having to do with age [In a comment to a post seemingly chosen at random, Earl points out that this etymology is hogwash. Dictionary.com confirms this. Nonetheless, the point still seems valid]. Wisdom is the judgment we (ideally) gain with life experience. It's a matter of gut feel or intuition, not of sequential reasoning. In other words, it seems more likely based on statistics than logic. Do we grow wiser mainly by accumulating more data points?

A stock example is chess mastership. Chess masters typically look at many fewer moves than beginners or even experts, but the moves they look at are better. A master will also typically be better at remembering positions from actual games, as opposed to random placements of pieces, while the rest of us will do about equally well at each. A master is drawing on a large store of game experience and using this to structure the analysis of the position at hand. Clearly there is more involved than a simple "this position looks like that one", but just as clearly that's part of the picture.

Whatever statistical analysis a master is subconsciously doing isn't simple enough to have been captured algorithmically. Computers can beat masters at chess, but they do it by bashing through vast numbers of hypothetical moves. Programs that try to "understand" positions beyond simple rules like "I have more material and my pieces are more centrally located" tend to fail.

If that doesn't muddy the waters enough, you might consider this viewpoint.

"Ten Future Web Trends"

This article on Read/Write Web, which lays out ten likely future trends for the web, has been getting bookmarked a bit lately. It's a perfectly good article as it stands, but here are some comments on it anyway, by way of possibly staking out some sort of overall position, philosophy, weltanschauung or whatever. I'll try to keep the commentary shorter than the article, but I make no promises.

The Semantic Web: The basic idea, from Tim Berners-Lee's Weaving the Web, is that "[m]achines become capable of analyzing all the data on the Web - the content, links, and transactions between people and computers." There are any number of refinements, restatements and variations, and there is probably more than the usual danger of the term being applied to anything and everything, but that's the concept in a nutshell, straight from the horse's mouth (now there's an image).

This is really material for several posts (or books), but my quick take is that the web will indeed gradually become more machine-understandable. Presumably we'll know more precisely what that means when we see it.

I'm not sure whether that will happen more because data becomes more structured or because computers get better at extracting latent structure from not-deliberately-structured data. Either way, I don't believe we need anywhere near all data on the web to be machine-understood in order to benefit, and conversely, I'm not sure to what extent all of it ever will be machine-understandable. Is everything on the web human-understandable?

Artificial Intelligence: Well. What would that be? AI is whatever we don't understand how to do yet. Not so long ago a black box that you type a few words into and get back relevant documents would have been AI. Now it's a search engine. In the context of the web, AI will be things like non-trivial image processing (find me pictures of mountains regardless of whether someone tagged them "mountain") or automatic translation.

(Translation seems to be slowly getting better. The sentence above, round-tripped by way of Spanish with a popular translation engine, came back as "In the context of the fabric, the AI will be things like the process of image non-trivial (encuéntreme the mountain pictures without mattering if somebody marked with label "mountain to them") and the automatic translation". Believe it or not, this looks to be an improvement over, say, a year ago)

The article mentions cellular automata and neural networks, two incarnations of massively parallel computing. I tend to think the technology matters much less than understanding the problem.

It took quite a while to figure out that playing chess is (relatively) easy and walking is fiendishly difficult (particularly if you're supposed to see where you're walking). It also took a while to figure out that matching up raw words and looking at the human-imposed structure of document links works better than trying to "understand" documents in any deep sense. I call this theme "dumb is smarter" and one of these days I'll round up a good list of examples.

As the article points out AI and the semantic web are related. One way to look at it: A machine that could "understand" the web as well as a human would be a de facto AI.

Virtual worlds: In the hardcore version, we all end up completely virtual beings, our every sensory input supplied electronically. Or perhaps we no longer physically exist at all. I'm not willing to rule this sort of thing, or at least the first version, out for the still-fairly-distant future, but in the near term there are some obstacles.

I've argued that our senses are (probably) dominated by sight and sound and that available bandwidth is more or less enough to saturate those by now. But it's pretty easy to fake out the eyes and ears. Faking out the vestibular sense or the kinesthetic senses may well require surgery. Even smell has proved difficult. So the really fully immersive virtual world is a ways away and bandwidth is not the problem.

In the meantime, as the article points out, lots of interesting stuff is going on, both in creating artificial online worlds and in making the physical world more accessible online. Speaking for myself, other than dipping my toes in MUD several years back I'm not virtualized to any significant degree, but Google Earth is one of my personal favorite timesinks.

Interestingly, William Gibson himself has done a reading in Second Life. Due to bandwidth limitations, it was a fairly private affair. Gibson's take:

"I think what struck me most about it was how normal it felt. I was expecting it to be memorably weird, and it wasn't," he says. "It was just another way of doing a reading."

I think this is an example of the limitations imposed by the human element of the web. We can imagine a lot of weird stuff, but we can only deal with so much weirdness day to day.

Gibson also argues that good old fashioned black-marks-on-a-white-background is a pretty good form of virtual reality, using the reader's imagination as a rendering engine. I tend to agree.

Mobile: I've already raved a bit about a more-mobile web experience. To me mobile computing is more about seamlessness than the iPhone or any particular device. Indeed, it's a lot about not caring which particular device(s) you happen to be using at a given time or where you're using them.

Attention Economy: "Paying attention" is not necessarily just a metaphor. The article references a good overview you may want to check out if the term is not familiar.

OK, we have to pay for all this somehow, and it's pretty clear the "you own information by owning a physical medium" model that worked so well for centuries is breaking down. But if no one pays people to create content, a lot less will be created (hmm ... I'm not getting paid to write this).

Because we humans can only process so much information, and there's so much information out there, our attention is relatively scarce and therefore likely to be worth something. Ultimately it's worth something at least in part because what we pay attention to will influence how we spend money on tangible goods or less-tangible services. So we should develop tools to make that explicit and to reduce the friction in the already-existing market for attention.

My take is that this will happen, and is happening, more due to market forces than to specific efforts. That doesn't mean that such efforts are useless, just that markets largely do what they're going to do. They make the waves, we ride them and build breakwaters here and there to mitigate their worst effects.

Web Sites as Web Services: The idea here is that information on web sites will become more easily accessible programatically and generally more structured. This is one path to the Semantic Web. It's already happening and I have no doubt it will happen more. A good thing, too.

On the other hand, I wonder how far this will go how fast. Clearly there is a lot of information out there that would quite a bit more useful with just a bit more structure. It would also be nice if everyone purveying the same kind of information used the same structure. Microformats are a good step in this direction.

My guess is that tooling will gradually have more and more useful stuff baked in, so that when you put up, say, a list of favorite books it will be likely to have whatever "book" microformatting is appropriate without your doing too much on your part. For example if you copy a book title from Amazon or wherever, it should automagically carry stuff like the ISBN and the appropriate tagging.

In other words, it will, and will have to, become easier and easier for non-specialists to add in metadata without realizing they're doing it. I see this happening by fits and starts, a piece at a time, and incompletely, but even this will add considerable value and drive the tooling to get better and better.

Online Video/Internet TV: I don't really have much to add to what the article says. It'll be interesting and fun to (literally) watch this play out. It'll be particularly interesting to see if subscription models can be made to work. If so, I doubt it will be because of some unbreakable protection scheme.

Rich Internet Apps: I occasionally wonder how much longer browsers will be recongizable as such. The features a browser provides -- tabs, searching, bookmarks and such, are clearly applicable to any resource and sure enough, editors, filesystem explorers and such are looking like more like browsers. OS's are getting into the act, too, allowing you to mount web resources as though they were local objects, perhaps doing some conversion or normalization along the way.

Browsers are also growing more and more toolbars, making them look more like desktops, and desktops are growing widgets that show information you used to get through a browser. Behind the scenes, toolkits will continue to go through the usual refactoring, making it easier to present resource X in context Y.

The upshot is that the range of UI options gets bigger and the UI presented for various resources gets better tuned to both the resource and your preferences. Good stuff, and it will continue to happen because it's cool and useful and people can get paid to make it happen.

International Web: Well, yeah!

Personalizaiton: This is a thread through a couple of the trends above, including Attention Economy and Rich Internet Apps. It will also aid internationalization. The big question, of course, is privacy. But that's a thread in itself.

What happened to my bookmarks?

[If you came here trying to recover lost bookmarks for Firefox, mozillazine.org has a knowledge base article on the topic. For Chrome, try this Google search. For IE, try this Google search. For Safari, try this one. For Opera, try this one. The Opera search also turned up this PC Today article for Firefox, IE and Opera. For Ma.gnolia (FaceBook and maybe others?), try this. In any case, please feel free to have a look around since you're here.]

Oh they're still there, but the thing is I care less and less. Just how did that happen?

Back when Firefox was Netscape, I spent a fair bit of time grooming my bookmarks list -- checking for broken links, sorting them, organizing them into folders, making sure they followed me from machine to machine. Now not so much. I'm only starting to use del.icio.us, so I'll have to report later on what effect that might have, but it seems more slanted towards finding cool new things as opposed to old lost things.

What changed? A couple of things.

First, browsers got smarter. Firefox (and others) will remember where you've been recently. To get to my banking site, I have to type "b-a-" into the bar at the top and then hit down arrow a couple of times. This is at least as easy as browsing through my bookmarks, even if I put my banking site at the top (at the expense of whatever else), probably because it works regardless of whether I made a particular effort to remember the site or not.

For news sources, I have RSS/Atom/whatever it is these days. Other interesting sites install themselves in the tool bar and look more like applications than web sites.

That pretty much takes care of the "remember frequently-visited sites" function. If it's frequently-visited, it's in the browser's memory pretty much by definition. If it's particularly well-used, it's probably hooked into the browser one way or the other.

Which leads me to the other bookmark-killer: Google. Early on, I sort of remember thinking that the useful web was mainly a smallish set of known sites and it was up to me to remember what to find where. In such a world it makes sense to use the otherwise memory-impaired early browser's bookmark feature to collect the main portals to the known world. Early search engines also had a higher chaff/wheat ratio than modern ones, discouraging their use somewhat.

These days I accept that I have only a dim idea of what's out there. If I want to find out about something I do what everyone does: put together a couple of search terms likely to nail it down and set Google at it (or Wikipedia, depending).

Google was probably what finally really convinced me that "dumb is smarter" could work in a big way. There's still value in hand-selected indexes and summaries, which is really what a bookmark list is, and the whole Web 2.0-style collaborative tagging thing definitely has value, but a comprehensive, frequently updated search engine will win on coverage and agility every time. That sets a reasonably high bar for anything else.

Like browser history, searching works without any explicit help. I could try to remember whether the Murky News is sanjosemercurynews.com (nope) or mercurynews.com (the actual domain) or sanjosemercury.com (redirects). I could bookmark it and find the bookmark. Or I could just Google "san jose mercury" and get it.

Searches also don't go stale. Bookmarks tend to rot over time as things get reshuffled and relocated, even though the document itself is still out there somewhere. I'm not yet sure how adaptive tags can be. Having one's site prominently tagged will tend to discourage one from moving it.

If bookmarks are not that useful in remembering frequently-visited sites or as a starting point for research, what are they good for? For my money there is one core function they still perform well, namely remembering particularly memorable pages, things you're glad you found but probably won't be revisiting on a daily basis.

I do use bookmarks (and now del.iciou.us) for that, though I don't find myself referring to them much. That's probably because I just don't find myself wanting to replay the greatest hits very much. There's too much interesting new stuff on the web. Old bookmarks are more for rainy days, and it hasn't been raining much lately.