Matt Bamberger's journal

Substrates part 1

Fri, 26 Oct 2007 02:57:54 GMT

The term "substrate" comes up often in discussions about AI, especially in the context of comparing AI to biological intelligence. The meaning of the term is pretty straightforward: the substrate is the hardware that embodies a particular intelligence. Human intelligence, then, runs on a meat substrate: all the neurons that make up the human nervous system, along with their support infrastructure. In traditional computing, we normally use the term "platform" to refer to the same concept. For example, I'm writing this article in Word (the software), which is running on an Intel Core2 / Windows Vista computer (the platform, or substrate).

One of the earliest theoretical results of computer science was Alan Turing's development of an imaginary computing platform called the Turing Machine. A Turning Machine is about as primitive as a computer can get: it consists of a very long paper tape with numbers written on it and a mechanical head that moves up and down the tape reading and writing numbers according to some simple rules. Despite the rudimentary nature of the Turing Machine, Turing was able to prove a strange and profoundly important theorem:

Aside from speed and memory constraints, all possible computers are functionally identical to a Turing Machine, and therefore to each other.

One real-world implication of this is that it's possible for one computer to exactly emulate another. You've seen this in action if you've ever run an emulator for an old computer game. MAME (Multiple Arcade Machine Emulator) is an example of this: it's a program that allows modern computers to run games that were originally written for old video arcade machines. Even though Space Invaders was written to run on a specific computer system (an old stand-up arcade machine), it can be run on any modern computer for which a version of MAME exists (which is pretty much all of them). This is true even though modern computers are vastly different from 1980s arcade machines, and even though the authors of MAME may have no idea how Space Invaders actually works.

What this means for AI, of course, is that if the human brain is a Turing Machine, it's possible to run its software (ie, human intelligence) on any sufficiently powerful computer, regardless of how greatly its underlying architecture may differ from that of the human brain. In part 2, I'll discuss why the brain is probably a Turing Machine, and explore some of the non-obvious implications of that fact.

Deconstructing AI

Mon, 10 Sep 2007 23:32:15 GMT

Imagine that you want to colonize Alpha Centauri. Your task is daunting, but it's conceptually fairly straightforward: you're gonna need propulsion systems, life support, radiation shielding, power sources, etc. Some of those individual problems may be very hard, but it's clear what they are, and in every case, there's a pretty obvious starting point. If you want to spent $10B on the project, you can do a reasonable job of hiring people, building teams, and getting started. You may or may not succeed, but you know how to get started.

AI is different, and therein lies one of the fundamental problems with developing AGI.

Although lots of people have opinions, the truth is that none of us have anything more than guesses as to what technologies or components are needed to build an AGI, let alone how to put them all together. About the only certainty is that faster computers and better development tools are helpful. Beyond that, who knows? Here's a very incomplete list of some possible technologies that might or might not be useful: HTM, pattern recognition, NLP, neural networks, Bayes nets, large atom tables, predicate logic parsers, large databases of common knowledge, neuron and brain simulation, etc....

There are a number of plausible AGI designs floating around, each with a list of component technologies and a high-level plan for gluing them together into a working whole. The problem is that both the high-level plans and the component technologies vary wildly from one design to another.

On a related note, if you want to colonize Alpha Centauri, there are a number of obvious incremental steps. Launching a successful moon shot will almost certainly be helpful. Establishing a Mars colony will get you tangibly closer to your goal. With AI, the incremental steps are not obvious. Does the DARPA Grand Challenge get us any closer to an AGI, or is it purely a distraction? Does working on Machine Translation teach us anything at all?

Smart, well-informed people have different opinions, but I think that I probably speak for the majority when I say that as far as I can tell, pretty much all current narrow AI work contributes little or nothing to an AGI effort in the foreseeable future. Many of them are useful in their own right, and many of them contribute to our general scientific knowledge, but none of them are obvious stepping stones toward the ultimate goal of AGI.

Does all of this mean that AGI is a currently intractable problem? Maybe. Peter Norvig make a good case at this weekend's Singularity Summit II that working on AGI doesn't currently make sense because the underlying science simply isn't sufficiently advanced yet. Other smart people feel that AGI has simply been neglected, and that the problem is tractable by the right team.

I think we can all agree, though, that AGI is at the very least an incredibly daunting challenge.

Flight patterns

Tue, 07 Nov 2006 21:07:03 GMT

A nifty video of US air traffic patterns.

Infovore in a hurry

Tue, 26 Sep 2006 06:53:14 GMT

If you're already getting your full RDA of data from RSS, you probably already know all of this. If you're not, it's time to wake up and smell the 21st century. RSS is a dramatically better way of getting all kinds of information, and has profoundly changed how I find out about the world. Not only does it give me access to all kinds of information that I couldn't realistically obtain otherwise, but it makes me a much more efficient information consumer. At this point, I find the idea of life without RSS pretty unbearable.

Here's why RSS is cool:

It's ubiquitous. All of the periodic media I consume is available as RSS. I use the same tool to read the New York Times and to read Monk's latest misadventures</a>. For those of you who use LiveJournal, RSS is like LiveJournals's friends page, only it gives you access to the other 95% of the universe as well.
It's push media. I don't have to check the feeds I'm interested in to see whether they've changed. Some of the feeds I read have a dozen new articles a day, and some don't update for months at a time. With RSS, all of the new articles magically show up in my reader as soon as they're posted.

Here's how it works. Pretty much any online source worth reading these days has an "RSS feed". Typically, the RSS feed is marked with a little orange button, like so: . The RSS feed is simply a web page with a bunch of scary-looking XML code that contains information about recent updates. To read an RSS feed, you'll need a reader. There are many good ones; I happen to like Bloglines because it works well and is web-based. You tell your feed reader which RSS feeds to monitor, and it'll notify you whenever one of them posts something new.

Now that you have your shiny new toy, you'll need something to read. To get you started, here's a partial list of what I read:

Mainstream periodicals

The Economist. The only newspaper that matters.
The New York Times. Okay, the NY Times matters a little. The NY Times has over a dozen RSS feeds, broken down by topic.
The New Scientist. Probably the best general science periodical.
The Seattle Times. The Seattle Times really doesn't matter at all. They're local, though, so I read their real estate and food feeds.

Random tech stuff

Accelerating Future. Michael Anissimov's Singularity / Transhumanist blog.
Adam Bosworth. Very infrequent, always worth reading.
Bram Cohen. Equally brilliant and insane. Infrequent and fascinating.
Greg Linden. Lots of good stuff about web tech and personalization.
Google Talkabout. Local Google team, fairly interesting blog.
Joel On Software. Excellent developer blog.
Journal of complexicism. Infrequent but interesting.
Kurzweil AI. Excellent aggregator of Singulariy-related news.
Longevity Meme. Somewhat cranky but interesting aggregator of life-extension news.
Official Google Blog. Google's cool.
Schneier on security. Outstanding security blog.
Treonauts. For those of us who take our Treo phones a little too seriously.

Random fun links

43 Folders. Getting Things Done fansite.
Cool Tools. Exactly what it sounds like.
Damn Interesting. Ditto.
Digital Photography Review. Probably the best source of digicam info on the web.
Penny Arcade. Gaming comic strip with absolutely no redeeming social value.
Proceedings of the Athanasius Kircher Society. All kinds of cool stuff, often related to science and/or science history.

Brain simulation

Mon, 01 May 2006 06:25:34 GMT

There are two basic approaches to building an AI. The traditional approach is to write an artificial intelligence from scratch. A less sophisticated but perhaps more tractable approach is to simply simulate the workings of an actual human brain. For a variety of reasons that I'll detail below, I don't think this is the best approach, but I think it's an excellent fallback. If traditional approaches fail to deliver a working AI in a timely fashion (which they very well may), there's an excellent chance that brain simulation will be able to deliver a good enough AI within the next few decades.

Why brain simulation is likely to succeed

Unlike traditional AI, brain simulation requires solving relatively few big problems. In order to simulate a brain, it's probably sufficient to understand in detail how each of the 50 or so different types of neurons work, and how the different parts of the brain are connected. That's a lot of knowledge, but it's a well-constrained, well-understood problem space, and one in which we're making very rapid advances. The key point is that you can completely simulate a human being without necessarily having any deep understanding of how human intelligence actually works.
The requisite computational power is almost here. The amount of computational power required to completely simulate a human brain at the neural level is probably about 20 petaflops. Computers with that capacity will be available within a few years, and will be ubiquitous within the next two decades.

Nonetheless, brain simulation is harder than it sounds

Although running a simulation of a brain is hard, I don't think it's the hardest part of making brain simulation actually work. There are three non-obvious but significant obstacles to producing something useful:

Assuming that we have the knowledge required to completely simulate all the components of a human brain and all their interactions, we still have to know the exact structure and connectivity map of the brain. That is to say, it's not sufficient to be able to run a simulation: you also have to know the exact starting state. This is difficult because it requires either a complete high-fidelity scan of an existing fully-grown brain, or else the ability to simulate with high fidelity the growth of a human brain (which is a separate and equally complicated field), plus the ability to mature the simulated brain (which presumably involves "raising" it from infancy).
By definition, a fully-functional simulation of a human brain is a fully-fledged person, in ways that a traditional AI would not necessarily be. There are a variety of hard ethical issues associated with experimenting on human beings.
It's not clear to what extent a human brain is capable of functioning without a body and rich sensory input. It's possible that running a virtual brain will also require running a virtual body and/or virtual world, each of which is a substantial problem in its own right.

Brain simulation is more powerful than many people realize

On the face of it, the ability to create a fully functional human being isn't that profound an accomplishment. My partner and I have accomplished that very thing on two separate occasions, using only common tools we found around the house. However, simulated brains have some very interesting properties that potentially make them far more capable than real humans, even in the absence of any ability to understand or tinker with their core capabilities.

Virtual brains can run very quickly. A virtual brain running on a computer doesn't have to run in realtime. If you can run a brain in realtime on a given computer, you can run the same brain a thousand times faster on a computer that's a thousand times faster. A fast AI isn't any smarter than a slow one, but it can accomplish the same amount of work a lot quicker. A computer programmer running at 1000X realtime can accomplish in one year what a real programmer would take 1000 years to accomplish.
You can clone AIs. One obvious implication of this is that you can populate your AI population exclusively with geniuses. A less obvious implication has to do with knowledge-sharing and task balancing. When a new version of the flu appears, you can instantly suspend all of your architect and engineer AIs, and reboot them as copies of your leading flu expert. Another use for AI cloning is getting around boredom. A lot of jobs require a high level of expertise, but involve a great deal of boring and repetitive work. By rebooting your doctor AI before each new patient, you get a doctor who's always fresh and focused on the task at hand (and who can instantly morph into whatever specialist the patient turns out to need).

Limitations of brain simulation

The above notwithstanding, there are some serious problems with brain simulation that make it problematic.

Simulated people are by definition just like people. That means they're cranky and sneaky and prone to behaving badly. Those are bad properties in an AI.
It's hard to extend the core capabilities of a simulated brain. One of the advantages of the simulation approach is that it requires very little deep understanding of how human intelligence actually works. The downside of that is that it makes it very hard to improve the AI. Even fairly simple tasks like increasing the AI's memory or directly transferring knowledge or skills from one AI to another become very hard problems. Virtual brains will be super-fast, super-capable people, but they're unlikely to be genuinely superhuman.

Overall, I see brain simulation as a good backup choice. I'd much rather see a traditional approach succeed, because its end result is likely to be both safer and more powerful. In the eventuality that traditional approaches don't succeed, however, I think that virtual brains will very probably become feasible within the next few decades.

Artificial Development's CCortex project is, as far as I know, the largest and most interesting virtual brain project currently underway.

Moving on

Sat, 03 Sep 2005 06:17:00 GMT

For various reasons (mostly too geeky to admit to in public), I'm moving my posts from here to my website: www.mattbamberger.com. In the incredibly unlikely event that anyone's actually been reading this, you'll need to look there for my future updates.

More on vocabulary size

Sun, 14 Aug 2005 05:48:52 GMT

In The Language Instinct, Stephen Pinker notes that linguists use the term "listeme" to refer to a linguistic unit whose meaning must be memorized. For purposes of estimating human capacity, total number of listemes known is probably a more interesting metric than number of words known. Pinker guesses that the average high school graduate knows about 60,000 listemes, and above-average graduates know perhaps twice that many.

An even bigger computer...

Sun, 14 Aug 2005 05:27:11 GMT

From the SIAI newsletter:

Japan's Ministry of Education, Culture, Sports, Science and Technology has announced plans to begin building the world's fastest supercomputer in 2006, to be completed by 2011. The supercomputer would be capable of operating at 10 petaflops, or 10 quadrillion calculations per second, roughly 73 times faster than IBM's Blue Gene, the world's present record holder. This value would exceed the estimates by a number of experts for human brain-equivalent computing capacity. Kyodo news reports that the total cost of the project will fall somewhere between 80 billion and 100 billion yen, or $714 million and $893 million.

As noted, that would put this machine in the human brain range. By my earlier math, it would take two of these to match a human brain, for a total cost of about $1.8 billion.

Basic AI reading

Mon, 01 Aug 2005 07:05:02 GMT

My two favorite AI books:

How the Mind Works, by Steven Pinker

The best basic theory I've found about how the brain works. In a nutshell, Pinker argues for a relatively large number of specialized modules, each optimized for solving particular kinds of problems. There's nothing revolutionary in this book, but it's an excellent synthesis of much current thought on the topic.

On Intelligence, by Jeff Hawkins

Unlike Pinker, I think Hawkins is wrong about a lot of things, including the value of trying to exactly replicate human cognition, and the functional uniformity of the cortex. Nonetheless, he makes an excellent and interesting case for pattern matching playing a central role in human intelligence.

Basic singularity reading

Mon, 01 Aug 2005 06:56:34 GMT

Here are some good basic sources on the singularity and related issues.

The Technological Singularity

Although not the first use of the term technological singularity, this was the one that really got things rolling. It's a great 1993 essay by Vernor Vinge that lays our the basic premise of the singularity.

Wikipedia on the singularity

A good overview of what's been said about the singularity.

The Age of Spiritual Machines

Ray Kurzweil's most recent book does an excellent job of laying out his vision of the singularity. He believes in a soft takeoff (ie, a relatively gradual transition occuring over decades rather than hours). Among other things, he presents a detailed and compelling case for Moore's law applying much further into the past and the future than is commonly assumed.

Why the Future Doesn't Need Us

A classic 2000 article by Bill Joy. Unlike most other futurists, Joy is profoundly alarmed by the catastrophically destructive potential of bio-engineering, nanotechnology, and artificial intelligence. His proposed solution is one of relinquishment: we should immediately stop all work not only on the specifically dangerous technologies, but on those entire branches of knowledge.

How much do we know?

Sat, 09 Jul 2005 07:10:11 GMT

We've already estimated that the total amount of interesting state in a human brain is less than 14,000 TB. How does that map to the amount of high-level information we can retain? Let's take a look at some of the things we know:

We know language. An incredibly well-educated person might know a few hundred thousand words. Let's say 400,000 words, with 1K of information each (including spelling, pronunciation, emotional connotations, usage, etc.) That's 400 MB of language information (assuming that the bulk of our language data consists of vocabulary). By comparison, the complete text of the OED runs to about 350 MB.

We know facts about the world. A commonly cited estimate is that a human expert in a given field knows between 50,000 and 100,000 "chunks" of information about that field (a chunk might be something like "the incremental linker is unreliable if you've changed your pre-compiled headers", or "antelope always let down their guard when they're drinking.") A generous estimate is that a person might be an expert in 1,000 fields (oncology, driving, tasting food, family members, etc.) That's a knowledge base of up to 100 million chunks. If we allow 1KB per chunk, we end up with 100 GB of facts.

We know what we've read. Let's say you like to read. A lot. Maybe you can read 10 words per second. If you read long words, that's 100 bytes per second. If you read for every 2.5 billion seconds of a long human life, and you remember every word you ever read, you'll need to remember 250 GB.

We know about the things and people we've encountered. Let's say you encounter some new entity (person, object, etc.) every minute of your life, and you remember 1 MB of information about it (context, image, associations, etc.). That's a total of 42 TB over the course of your life.

Let's go nuts. Let's say you remember everything that ever happens to you, at DVD quality. If you keep the camera running while you're asleep, you'll end up with 1,752 TB of information by the time you're done.

In reality, of course, we don't remember very much at all, and much of what we do remember we quickly forget. A fair amount of work has been done on this, and a common estimate is that we're capable of learning about 2 bits of information per second across a wide variety of domains. Over a human lifespan, that works out to a total of about 600 MB.

The point of all this is that the amount of high-level information that we actually remember is many orders of magnitude less than the brain's theoretical capacity. That might be because the brain is incredibly inefficient, or it might be because most of the brain is unused or is dedicated to low-level non-cognitive tasks, or it might be because much of the brain's capacity is dedicated to storing indexed and processed forms of the raw information it contains.

Let's build a brain, part three: software size

Sat, 09 Jul 2005 06:39:26 GMT

On the face of it, people seem pretty complicated. You'd think that it takes a lot of data to make a person, but you'd be wrong.

The whole human genome consists of about 3.2 billion base pairs. Each base pair contains two bits of information (A, C, G, or T), meaning that the entire genome only contains 800 MB of information. You can fit the complete blueprints for a person on a decent-sized thumb drive.

It actually gets even better. There's some disagreement about the exact number, but as far as we can tell, between 95% and 98% of the genome is completely unused "junk DNA". That means that the complete DNA blueprints for building a person contain no more than 40MB of real information.

By comparison, the code portion (ie, the .data fork) of Microsoft Word is about 750KB, or about 1/50th the size of the entire functioning human genome. The whole of Microsoft Office is probably about 5 times that size, or 1/10th the size of the entire functioning human genome.

It's not clear what the true information density of the genome is (it certainly contains a non-trivial amount of redundant information), nor what percentage of it is remotely related to digestion. It's interesting to note that the human genome is 98.4% identical to that of chimpanzees (ie, only 640KB of information separates us from chimps), and that well over 90% of all mouse genes have very similar counterparts in humans.

Let's build a brain, part two: state size

Mon, 04 Jul 2005 06:37:51 GMT

In addition to needing a lot of raw computational power, an artificial brain will need to store a lot of state. The three main elements we'll need to store state for are the neurons, the synapses, and the glial cells. Let's take a look at how much state each of these requires:

The brain contains about 100 billion neurons, each of which has a fair amount of state. Let's give each one a very generous 10K of storage. That comes to a total of 1,000 TB.

There are about three glial cells for each neuron. It's not clear how much meaningful state is associated with each one, but let's be generous: we'll give each one 10K also. That comes to a total of 3,000 TB.

Most of the action is in the synapses. We're figuring about 1,000 synapses per neuron, each of which will need some state. Let's allow 100 bytes per synapse. That comes to a total of 10,000 TB.

Our very generous grand total for storing all the state in the human brain is a mere 14,000 TB. That's a very manageable amount: disk storage is currently running about $1,000 per terabyte. That means that today's prices, we can store a complete human brain for only $14 million.

Let's build a brain, part one: computational capacity

Sun, 03 Jul 2005 06:39:24 GMT

An obvious starting point in building an AI is the human brain. Let's take a look at the capacity of the human brain.

It turns out that an average human brain has about 100 billion neurons. The number of synapses per neuron is less clear, but it's probably around 1000. To a first approximation, each neuron can cycle about 200 times per second. So, doing a little math, we see that:

100 billion neurons x 1000 synapses x 200 cycles per second equals 20,000 teraflops.

So, to simulate a human brain running in real-time, we need about 20,000 teraflops of computational capacity. Let's see how current hardware stacks up:

Computer: Blue Gene/L
Capacity: 360 teraflops (56 needed to model a brain)
Cost: $100 million
Cost to model a brain: $5.6 billion

Computer: 2005 personal computer
Capacity: 500 megaflops (40,000,000 needed to model a brain)
Cost: $1000
Cost to model a brain: $40 billion

Obviously, there's a lot of slop in these numbers, depending on your assumptions about what constitutes a flop, how complicated synapses are, how quickly neurons really fire, etc. Nonetheless, it's clear that simulating a human brain in real-time is currently within computational reach, although it's a major undertaking.

Word lists

Mon, 27 Jun 2005 06:59:45 GMT

Here's a fun little page: Kevin's Word List Page. I found this when I was looking for a list of English words for feeding to a pattern matcher. It's a set of databases of English words from various sources.

In addition to being a great vocabulary list, the SCOWL database comes in several sizes with some interesting documentation. For each list, Kevin includes a set of simple words included in that list. This turns out to be an interesting way to estimate the size of your vocabulary. For example if you know words like both, first, little, and six, then your vocabulary probably contains at least 5,000 words. Here are some samples:

5,000 words: both, first, little, six
55,000 words: blender, chump, gazebo, township
103,000 words: centralist, gobstopper, lyrebirds, teletext
211,000 words: atomisms, dogberry, laterality, pentane
628,000 words: anatropal, ectoretina, overacceleration, trophoplasmic