Moore’s Law improvement?

Brad Templeton has a post suggesting a new law on semiconductor growth. Basically, Moore’s “Law” isn’t really a law at all, since it’s been slowing down. Brad suggests a new law, in which the period of doubling doubles every 40 years. I commented:

I’ve always thought that the weight placed on Moore’s Law (and the ensuing hystrionics when it’s discovered to have “slowed down”) was pretty silly. It’s clearly not reasonable to expect exponential growth to continue indefinitely. Resource limitations will put pressure on growth. In this case, we’re talking about physical limits, as well as limits on the resources needed to construct chip fabs capable of making the denser chips.

A more reasonable “law” would model the growth as a logistic function, the way population growth is modeled in biology/ecology. A logistic is (essentially) an exponential that “slows down”. (sound familiar?) In fact, your characterization of an exponential whose rate is cut in half every 40 years may in fact be a logistic. Anyway, if I had the data, I’d love to regress a logistic curve onto it and see what comes out. Since the logistic eventually levels off, the model would give a prediction of the limit of semiconductor growth.

You can read about the logistic curve at MathWorld. Curves like the logistic show up in other bounded-growth analyses, like the Hubbert curve for oil production, in which the amount of production at any particular time is essentially the derivative of a logistic-like sigmoid function, and can probably be nicely modeled using the logistic.

If anyone has data on the maximum chip transistor density over time since the 60’s, I’d take a crack at regressing a logistic onto it in MatLab, or whatever. Another interesting set of data to use would be Hans Moravec’s data on computing power per unit cost. The data goes back to 1892. The Moravec paper I linked has the raw data in the appendix. I’ll take a crack at it, when/if I get time (probably not for a couple weeks.

The Professoriate and the Truth

Tech Central Station is carrying the text of an interesting lecture on the state of higher education. What’s interesting about it is that it was given at a conference of educators, and it rails against the intolerance of dissent against the liberal political ideal that currently prevails in college campuses.

I, myself, am more or less liberal, but unlike many others at UT and elsewhere I feel that college campuses should be places where controversial ideas on race, sexuality, religion, etc, should be discussed openly; those who espouse unpopular and politically incorrect ideas should be welcomed to debate, not shouted down; and the purpose of tenure is to allow professors to research in controversial areas and to hold controversial opinions on any topic without fear of political reprisal against their job. Although it may be old news to some, it’s still refreshing to see a professor stand up at a meeting of professors and attack ideas that are currently held as doctrine at universities.

Forum for AI: Pat Beeson and Joseph Modayil

This Friday’s Forum for AI features Pat Beeson and Joseph Modayil, fellow grad students from my lab, talking about robot map building using hybrid metrical/topological maps. These are maps that describe large-scale space as a coarse-grained topological graph of places connected by path segments. The small-scale space around each place is represented by a fine-grained metrical map that describes the local area in detail. Here’s description and an excerpt from the abstract:

Hybrid Mapping Models:
Bridging the Gap between Robot Sensors and Symbolic Spatial Representations
Patrick Beeson and Joseph Modayil
Friday, April 23, 2004
11:00 a.m.
TAY 3.128
ABSTRACT:
… Our lab, along with other researchers, have looked into how to build, check, and order topological map models given reliable actions and perceptions in the world. The problem with topological maps has always been in the way they ground the robot’s sensory experience. This is often done in an ad-hoc way for each individual robot sensory setup. There is often little overlap in the approaches used throughout the community.

Today, our lab is joining a handful of other researchers in developing hybrid mapping techniques. We are promoting the use of metrical models to describe “small-scale space”: akin to a human’s ability to have a detailed model of their local surround. Topological maps are useful for representing “large-scale space”: graph structures imply compact, hierarchical, symbolic descriptions useful for planning, communication, and storing multiple hypotheses.

Our talk will focus on how these two ontologies interact. In particular, we discuss what new concepts are necessary to move from detailed metrical descriptions of bounded local regions to graph-like structures of large, complex environments. We will present published work that shows several of the advances we have made in recent months as well as present the open questions that still need to be examined in more detail.

Music sharing is viral marketing

After recent news that music sales were down again in 2003, the music industry has renewed their complaint that digitally enabled sharing of music is hurting their sales. The thing they refuse to recognize, at least outwardly, is that the sharing of music has always been one of the most important marketing forces for the music industry. I don’t mean online peer-to-peer sharing, I mean sharing in the more general sense, like the time-honored tradition of making mix-tapes and giving them as gifts, or a friend handing another friend a CD and saying “listen to this, I think you’ll like it”.

The big issue here is how serious music fans decide what music to buy. I’m talking about the people who maintain large CD collections and spend a lot of money on music — the customers that the music industry should be holding close to their hearts. These people not only spend a lot of money themselves, but they influence their less musically-inclined friends. These people tend to have idiosyncratic tastes, and are picky to the point of snobbishness. They don’t buy music based on music industry mass-marketing. They buy it based on hearing it and liking it, and the way they hear new music is by sharing it with friends. Radio used to play a part, too, but consolidation has turned music radio into a steaming pile of crap, so what’s left? Networks of like-minded friends sharing music are what’s left.

Recommender systems based on collaborative filtering, like the one used by Amazon.com have tried to assume the role played by social networks by keeping track of individuals’ tastes and then recommending to each user those things ranked highly by other like-minded users. The problem with these systems so far is that they don’t have a rich enough representation of the music they’re recommending to account for the idiosyncrasies of listeners’ tastes. For example, I love all the AC/DC albums up to and including Back In Black, but I think their stuff after that is mostly crap. On Amazon I rated all the early stuff highly, so of course it recommended all the later stuff that I don’t like. If I rate all the later stuff very low it’s unlikely that it will be able to form any kind of high level concept that will be useful in recommending things to me. If Amazon can come up with a system that takes those ratings and recommends old Krokus albums to me, because they capture the flavor of Bon-Scott-era AC/DC, then they’ll have something! But I’m not going to hold my breath for that. This kind of thing happens a lot. I love Cake’s Comfort Eagle, but I’m not a such fan of Fashion Nugget. I could go on and on. The essential point is that recommendations from like-minded friends aren’t enough. I need to actually listen to the music before I decide if I like it. I don’t just mean listening to 30-second snippets on Amazon or the iTunes Music Store. I mean really listen to the music. Listen to whole songs, in high fidelity, from start to finish, to see how they develop. Of course, I don’t want to pay just to sample, because a large part of what I sample is crap, and I don’t want to pay for crap.

Recently I’ve happily discovered a bunch of new stuff to listen to, thanks to a handful of instances of this kind of music sharing:

• After lamenting to a friend that I haven’t been discovering new music that I like, he sent me a mp3s of a few songs by bands that he’s been listening to: The Strokes, The White Stripes, and And You Will Know Us By The Trail Of Dead. I liked some of the stuff, and I’m not so sure about others, but I’ll probably end up buying at least one album from this.
• A couple of days ago, I was running iTunes in my office, and it discovered someone else’s shared iTunes library and made it available to me. I have no idea whose library it was, but their tastes overlapped with mine in several areas, so I started playing some of the stuff that they had and I didn’t. I didn’t like everything I tried, but I loved several songs from Garbage’s Beautiful Garbage, so I bought the album from the iTunes Music Store.
• An entry on the Boing Boing blog recommended Skeewiff’s awesome electronic remix of the Soggy Bottom Boys’ “Man of Constant Sorrow” from O Brother, Where Art Thou?. I’d never heard of Skeewiff before today, but I’ll be checking them out soon. I would never have discovered them if I’d had to pay just to hear one song.

The music industry needs to recognize that mainstream radio sucks ass, and MTV and VH1 don’t play music videos anymore, so there’s not much left but word-of-mouth and music sharing as ways to really market music.

Google’s GMail + Orkut = Automatic Spam Whitelist?

A couple months ago, Nature had an article on using social networks to filter spam. The basic idea is that people tend to send mail to people whom they know, and the system could induce users’ social networks from message headers, and detect spam because the messages aren’t sent within a social network. If you’ve already got social network software and email, why bother inducing the social network? Why not just integrate the two, and make an automatic spam whitelist? Google is in a position to do just that, with GMail, its impending free email service, and Orkut, its social network software. That might actually give me a reason to join Orkut.

Google’s Free Email: A giant free text database

Google is starting free email service to compete with Yahoo Mail and Hotmail/MSN. There are lots of interesting ramifications of this.

Google is the king of targetted text ads. Obviously they’re planning on putting ads on every message.

The free service will be supported by ads that Google automatically places in users’ e-mail messages. The ads will be targeted to the readers of the messages and will correspond to the text in the messages. For instance, an e-mail message about your trouble with your DSL connection might have an ad from SBC Communications in it.

Kinda creepy, when you think about it. Your lover sends you an explicit love note, and the ad sidebar offers a set of “appropriate” links.

The really interesting part is that they’re offering 1GB of storage space per user.

Rosing said the company arrived at the 1 GB figure by estimating how much storage an average user would need to store up to a decade’s worth of e-mail. “It’s a new paradigm where you don’t delete your e-mail,” said Rosing. Instead, users would leave all their messages on Google’s server and search through them as needed.

This is an interesting feature, and it seems like a strange business decision, given the resources required. Of course, typical users will take a long time to fill up 1GB. On the other hand, anyone who needs to hide data can just create an anonymous GMail account and mail more or less whatever they want to it. Need to back up some data? Just email it to your GMail account! I think the space will get used pretty fast. Of course, Google already has a collossal server farm, it’s possible that they just have terabytes of storage lying around, and offering all this space basically costs them nothing, but there’s still got to be some kind of investment to make all that space addressable from basically anywhere. What’s in it for them?

• The idea that all your mail is indexed, and you never throw any of it away is a step toward your own personal google. Think about it, anything you want to keep track of, or index, you can just mail it to your GMail address. If they add the other portal services, like calendar, address book, bookmarks, RSS feed aggregation, etc, and index all of it, you’ve got a search able index of your online life.
• Google gets an enormous database of plaintext documents (email) to play with and run their indexing and machine learning algorithms on. Sure, they have the web, but the web has to be crawled, and it’s very different in character from email: less conversational, less about one-to-one communication, marked up with all manner of content-free style info. If Google has, for example, some unsupervised algorithm that attempts to learn semantic categories or topics from text, what an awesome bunch ot training data they’d get from years worth of email for millions of people.

Of course, one thing to remember when saving 10 years worth of email on a public server, is that it’s not really private. Yes, Google may have a strong privacy policy (or not — I don’t know), but no privacy policy will protect your email from a subpoena. I’m not just talking about criminal cases here. Civil courts can subpoena diaries and other writings if they might be appropriate. Imagine, say, the MPAA is suing you for alleged copyright infringement, if they thought you had discussed sharing pirated movies with your friends in email, they could subpoena and pore over your GB of mail, and unlike if you were saving the mail on your PC at home, there’s not really any way you could prevent them from getting it.

PS: Wired has also picked up on the privacy implications of GMail.
PPS: Slashdot is linking a story from The Register (a noted anti-Google source), on the greater implications, including the fact that you won’t necessarily be able to delete your email, even if you delete your account.