# Older blog entries for raph (starting at number 290)

bytesplit

bytesplit: I mean you no ill will. However, by now it has become painfully clear that you and Advogato are not a good match for each other. I ask of you, please find another place for your writings.

Heather's got the 600 now. Yet another battery was tanking - the useful life was down to about 10 minutes. So I tried a technique suggested by Javier Valero, and so far it seems to work.

I've put up a page summarizing the ThinkPad 600 battery problem and possible solutions.

Proofs

I have lots more feedback than I can digest right now. But I will point out Robert Solovay's interesting reference to two flavors of second-order arithmetic. One is the "pure" Z_2 axiomatization with naturals, predicates over naturals (or functions from naturals to bool, if you prefer), and quantifiers over both. The other, with equivalent strength, has functions from naturals to naturals. The latter is probably a lot more directly useful for what I'm trying to do, because new functions can be defined as plain old lambda terms, and application is primitive, i.e. requires no hoops to be jumped through.

I'm more than a bit intrigued, though. Since these systems are indeed equivalent in logical strength, but differ widely in the convenience of their definitional systems, it's clear that choice of definitional framework is very important. How carefully has this space been mapped out, though? Is it possible that there are other definitional tricks that could make the systems even more usable, or is macro expansion plus function definition and application somehow complete in that all other definitional frameworks can be reduced to it tightly?

I'm not sure what such things might look like, but I see glimmerings. For one, I'd imagine that statements of the form "abstract tree S is a parse of string S by grammar G" might be easier to express and prove than in the usual logics. In any case, these statements are exactly the kind of thing I'd expect to be primitive in a human-friendly "source language", and then compiled automatically to a proof "object code", so maybe in the end it doesn't much matter.

Paul Snively has been blogging about proofs as well. I've skimmed his references, but so far am not sure how relevant they are.

I haven't felt much like interacting with anybody the past few days. It's probably the days getting shorter, and it will certainly pass.

Ghostscript

The 7.31 beta candidate tarballs are up. Please give them a shot.

I was hoping to get subpixel text antialiasing in this release, but didn't quite manage. Hopefully next time.

Proofs

I got some very nice email from Robert Solovay, answering many of my questions about second-order arithmetic.

I'm starting to get a handle on definitions. All proof systems have a mechanism for them, but mathematically they tend to be not very interesting, so mathematicians don't talk about them much. Even so, they're clearly important.

In general, definitions create a mapping from a source language to a target language, mostly by expanding out rewrites of the form <symbol> = <expansion>. The target language has just the axioms of the underlying system, like ZFC set theory or Z_2 arithmetic. So the concept of truth or provability is basically the same between the two languages: a statement in the source language is provable iff its expansion is provable in the target language.

So now the important thing is to make sure that there are usable inference rules in the source language. Ideally, all the inference rules that work in the target language are also sound in the source. But this doesn't necessarily have to be the case. For example, if you define the expansion of, say, y = 1/x to be x * y = 1, then you have 1/0 not equal to anything. This is cool in a way, but you have to use somewhat different axioms for equality.

Partiality is of course a big issue here. Joe Hurd recommended this paper on how to get partial effects in a total logic such as HOL. A different approach is Jan Kuper's quite readable thesis, which gave me the courage to consider tweaking the equality axioms in the source language so that you can write terms that don't exist. Also see Freek Wiedijk's First order logic with domain conditions.

There is one way in which definitions may be interesting: as a means of making proofs portable to different axiomatizations. The idea is that a module of definitions has an interface that exports proved theorems, but might keep certain details private. A good example is pairing. Your theorems basically say that you get out what you put in, and are the same in Z_2, HOL, and ZFC, but the actual construction (hidden) is quite different. When you plug in a different definition module with the same interface, you still expect your proofs to go through.

I'm not sure about the mathematical consequences of this approach. I get a feeling that it won't work in all cases, but if it works in all non-contrived ones, that's good enough. It's also entirely possible I'm reinventing category theory, in which case I better break down and get a book on it. I got a nice email from Sarah Mount recommending a few, as well as this paper updating Reynolds' result that polymorphism is not set-theoretic.

The feedback and intellectual excitement I'm getting from my exploration into proof systems is turning out to be quite addictive!

Second-order arithmetic

Just a quick note: many, if not most, of the questions I posed yesterday about second-order arithmetic are answered in the books referenced. I've skimmed them and am still grokking the details.

These books also give nicer pairing functions than the one I came up with on my own, but it's nice to note that they're recognizably similar.

Voda's book in particular bulds up a great deal of definitional machinery, much of it geared towards computer science applications.

12 Oct 2002 (updated 12 Oct 2002 at 09:00 UTC) »
Aaron Swartz on the Eldred case

AaronSw has written a fine first-person account of the Eldred vs. Ashcroft case seeking to overturn the the latest 20 year extension to copyrights. Also see Aaron's excellent visualization of the evolution of copyright lifetime since 1790.

I have an idea for a poster or t-shirt: Mickey Mouse is crushing diamonds underfoot. The caption reads, "Diamonds last a pretty long time, but Copyright is Forever(R)".

More words on proofs

There's a lot of material, through email and my own thinking. I'll try to get through it. This entry might be long.

Second-order arithmetic

chalst recommends second-order arithmetic as a nice "small" system. I poked around a bit and am definitely intrigued. It's been well-studied for a long time, which of course means that the main results are in dusty books rather than Web pages. Here's what I did find, though: a book (with a first chapter and content-rich reviews online), a concise definition (with axioms). There are also some drafts of a book by Paul Voda on computability theory. These are interesting, but it will take some slogging to figure out how to make it practical.

None of the "fifteen provers" in Freek's list use second-order arithmetic. Why not? Voda's CL seems to be, but I don't see other people talking about it. Why is it not on Freek's list?

Can you use HOL to prove the consistency of second-order arithmetic (as ZFC can do with HOL, but see below)? If so, HOL would seem to be a strictly "larger" system. What other important things are in HOL but not second-order arithmetic? My intuition tells me that integration is one of them, but of course it could be wrong (I'm not an intuitionist).

Conversely, can second-order arithmetic prove the consistency of primitive recursive functions? If so, I have a pretty good idea of examples you can do in second-order arithmetic but not primitive recursion: Ackermann's function is the most famous.

Definition schemes

If you look at the axioms for second order arithmetic, it's clear that they'd be painful to work with directly. You can't even write a tuple of integers (but see below). Instead, you have to construct all these basics from the primitives.

It's clear that any usable formal system must have a system for introducing definitions for things like tupling. At minimum, such a scheme must be safe, in that you can't introduce inconsistency through definitions. Metamath's mechanism for definitions lacks enforcement of this property, but almost all the definitions introduced in set.mm are of a restricted form for which safety is obvious (there's lots more discussion in the Metamath book). Fortunately, ZFC is powerful enough that these restrictions are not too painful. In particular, you can express concepts of extraordinary richness and complexity as functions, then apply and compose them quite straightforwardly.

HOL's definition scheme is a lot more complex. Basically, you provide a proof that implies that the new definition won't let in any inconsistency, then the system lets you define your new constant or type. For types, you do this by giving a construction in terms of already existing types. However, once you sneak the definition past the prover, it doesn't tie you to that specific construction. In fact, you could replace the construction with another, and all proofs would still go through. I consider this an important feature, well worth the extra complexity it entails.

Metamath's set.mm doesn't have this feature; rather, it nails down specific set-theoretical constructions of all the basics (fortunately, these are for the most part the standard ones used by mathematicians). In practice, all the proofs in set.mm are carefully designed so that the construction can be replaced, but again there's nothing to check this short of plugging in different constructions and trying it. It would be possible to rework set.mm so that reals are a tuple of 0, 1, +, -, *, /, etc., and all theorems about reals are explicitly quantified over this tuple. It sounds pretty painful and tedious, though. (Incidentally, Norm Megill sent me a nice email sketching how to do it).

I get the feeling that for second-order arithmetic, the situation is even more urgent. My intuition is that you can't just add another layer of quantification as you can in ZFC (and probably HOL-like type systems), because that would "use up" one order.

Bertrand Russell famously said that definitions had many advantages, namely those of "theft over honest toil". It looks to me like a search for a system of "honest theft" will be rewarding.

A good definition scheme can help with some other concerns, as well. Primarily, it should be possible to construct the basics compatibly in a number of logical systems. Sequences of integers, for example, are straightforward in second-order arithmetic, HOL, and ZFC, although in each of these cases the most natural construction is quite different.

I'm also intrigued by the possibility of instantiating an axiom set with a concrete type, perhaps multiple times within the same universe. The example that comes most readily to mind is integer arithmetic as a subset of the reals. If you look at set.mm, naturals appear (at least) twice; once in the usual set-theoretic construction, and a second time as a subset of reals (which, in turn, are a subset of complex numbers). The Peano axioms are provable in both. Thus, a proof expressed in terms of Peano axioms only should prove things in both constructions. It could be a powerful technique to stitch together otherwise incompatible modules, although of course it's equally probably I'm missing something here.

A little puzzle

One of the things I did not find was a library of useful constructions in second-order arithmetic, so I thought about them myself. One of the most basic constructions, pairs of naturals, makes a nice puzzle. This should be doable for most readers.

I came across one such construction in the datatype package for HOL. It is: pair(x, y) = (2 * x + 1) * 2^y (where ^ is exponentiation). This is most easily understandable as a bit sequence representation: the bit sequence of x, followed by a 1 bit, followed by y 0 bits.

But second-order arithmetic doesn't have exponentiation as a primitive (although you can do it). Can you write a closed form expression as above using only + * > and an if-then-else operator ("pred ? a : b" in C)? Obviously, such an expression has to meet a uniqueness property: no two pairs of naturals can map to the same result.

I've got a pretty nice, simple construction, but I have the feeling an even simpler one (perhaps without the if-then-else) is possible. In any case, I'll keep posting hints if nobody comes up with an answer.

The Web as a home for proofs

In any case, the choice of logical systems and the conversions between them, while interesting and important, is but one aspect of the evil plan. The other is to make the proof system live comfortably on the Web.

I think "hypertext" is a good analogy here. We've had plain vanilla text for a very long time. Closed-world hypertexts resident on a single computer system (such as HyperCard), also predated the Web by many years. A lot of people dreamed about a global hypertext for quite some time, but it took about thirty years for the Web in its present form to happen.

Likewise, proof systems have come a long way. In the early days, even proofs of trivial theorems were quite painful. Over the years, technology has improved so that truly impressive proofs are now practical. Even so, I can't just download my math from the web, do some proofs, and upload the results. I'm convinced that this is exactly what's needed in order to build a library of correct programs and their associated proofs.

How did the Web succeed where so many bright people failed before? It's certainly not through doing anything interesting with the text part of the hypertext problem. In fact, HTML as a document format is, at best, a reasonable compromise between conflicting approaches, and, at worst, a steaming pile of turds. What won the day was a combination of simplicity and attention to the real barriers to successful use of hypertext.

In particular, the Web made it easy for people to read documents published to the system, and it made it easy for people to write documents and get them published. In the early days of the web, touches like choice between ftp:// and http:// protocols (it was sometimes easier to get write access to an FTP repository than to set up a new server), port :8080 (for the relatively common case of being able to run user processes on a machine but not having root access), and of course the use of DNS rather than some more centrally managed namespace significantly lowered the barriers to publish.

The analogous situation with proofs is clear to me. It should be easy to use a proof once it's already been published, and it should be easy to publish a proof once you've made one. A great deal of the story of theorem provers is about the cost of proving a theorem. This cost is effectively reduced dramatically when it is amortized over a great many users.

I'll talk a bit about ways to achieve these goals, but first let's look at how present systems fail to meet them. Keep in mind, they weren't designed as Web-based systems, so it's not fair criticism. It would be like complaining that ASCII doesn't have all the features you need in a hypertext format.

Metamath

In many ways, Metamath is a nice starting point for a Web-based proof system. In particular, there's a clean conceptual separation between prover and verifier, enforced by the simple .mm file format. The fact that set.mm has been nicely exported to the Web for browsing is also a good sign.

A big problem with Metamath is the management of the namespaces for definitions and for theorems. There are two issues. First, the names of existing objects might change. Second, there's nothing to prevent two people from trying to use the same name. The namespace issue is of course analogous to programming, and there's lots of experience with good solutions there (including Modula's module system, which serves as the basis for Python's).

Another big problem is the lack of a safe mechanism for definitions, as discussed above. Of course, on the Web, you have to assume malicious input, while in the local case you only have to worry about incompetence.

HOL

HOL solves the definition problem. In fact, the definition scheme is one of the best things about HOL.

However, the distinction between prover and verifier is pretty muddy. HOL proofs are expressed as ML code. This isn't a good exchange format over the Web, for a variety of reasons.

Fortunately, people have started exploring solutions to this problem, specifically file formats for proofs. See Wong and Necula and Lee. I haven't seen much in the way of published proof databases in these formats, but I also wouldn't be surprised to see it start to happen. Interestingly, Ewen Denney uses these results (motivated by proof-carrying code) as the basis of his HOL to Coq translator prototype.

I'm unaware of any attempt to manage namespaces in HOL (or any other proof framework), but again this could just be my igorance.

How to fix namespaces

Whew, this entry is long enough without me having to go into lots of detail, but the namespace problem seems relatively straightforward to solve. In short, you use cryptographic hashes to make the top-level namespace immutable and collision-resistant. Then, you have namespaces under that so that almost all names other than a few "import" statements are nicely human-readable.

There's more detail to the idea, of course, but I think it follows fairly straightforwardly. Ask, though, if you want to see it.

Category theory

Yes, I've come across category theory in my travels. I have an idea what it's about, but so far I don't think I need it. My suspicion is that it might be helpful for proving metatheorems about the system once it's designed, but right now I'm mostly exploring. Ideally, the system will be simple enough that it's obviously sound, but of course it's always nice to have a rigorous proof.

Even so, if I get a strong enough recommendation for an accessible intro to the subject, I'll probably break down and read it.

Translating HOL to ZFC

chalst also warns that the translation from HOL to ZFC is not trivial, and cites a classic paper by Reynolds. Of course, being Web-bound I can't quickly put my hands on a copy, but I think I understand the issue. Here's a quick summary:

In untyped lambda calculus, representing the identity function (id = \x.x) in ZFC set theory is problematic. In particular, applying id to itself is perfectly reasonable in the lambda calculus, but a function in ZFC theory can't take itself as an argument, because that would create a "membership loop", the moral equivalent of a set containing itself.

For any given type, the (monomorphic) identity function over that type is straightforward. In Metamath notation, given type T it's I restricted to T.

So the trick is to represent the polymorphic identity function (\forall A. \x:A. x) as a function from a type to the identity function over that type. Before you apply a function, you have to instantiate its type. You never apply a polymorphic function directly, and you never have polymorphic functions as arguments or results of other functions (or as elements of types). So no membership loops, and no problems in ZFC.

Is this what you meant, or am I missing something?

Well, that's probably enough for today. And I didn't even get into Joe Hurd's email :)

Proof feedback

I've been getting some really good feedback from my series of diary entries on formal proof systems. The latest batch are from Norm Megill (of Metamath fame) and Joe Hurd, who's done quite a bit of work with HOL.

I've used up almost all of my diary time responding to these threads (privately), but the juicy bits will find their way back here.

BitTorrrent

As Aaron Swartz suggests, I downloaded Red Hat 8.0 using BitTorrent. From casper, the performance was amazingly good: around 800-1000 kB/s. I've long felt that free software ISO's were a killer app for this code. Unfortunately, it seems not to be in anybody's business interest to push this angle, because the relative convenience factor of getting a boxed CD is pretty much gone if you can get full-bandwidth downloads.

Incidentally, the GUI version crashed randomly on me, but the command line version behaved fine.

Red Hat 8.0

I like it. I like it a lot.

I installed it on a spare partition on spectre, just to give it a whirl. This time, I kept a log of usability issues and bugs. I ran into quite a few, but I think overall the system hangs togther pretty well. Obviously, people put a lot of work into making the desktop a good experience as a whole. This is rare and highly welcome in the oft-fragmented world of free software.

I was pleased that it detected (almost) all my hardware correctly, including the Audigy soundcard and the dual CPU's. I'll have to add the (serial) Wacom pad by hand, but that's not bad at all.

There's a good chance I'll put this on my laptop too.

GnomeMeeting

One of the perks of RH 8.0 is that installing GnomeMeeting 0.93.1 was near-trivial. As I understand, this is pretty difficult software to get running from scratch. I wasn't looking forward to herding a Gnome 2 development environment onto my machine. To have it Just Work is a nice treat.

I'm not sure the rest of our team will be able to use GnomeMeeting effectively just yet. Among other things, the H.323 nonsense prevents it from working behind a vanilla firewall. That's a problem TeamSpeak seems to have solved by using a nonstandard protocol.

I went to #gnomemeeting on irc.gnome.org and got personal help from Damian Sandras, principal author of the program. He was very nice, and of course this level of community involvement is exactly the reason why free software is so good. I want GnomeMeeting (or at least some free VoIP solution) to win.

ZFC vs HOL

I got a very nice email from Robert Solovay in response to my last entry. He patiently explained to me that you can prove the consistency of HOL using ZFC theory, which means that ZFC is strictly more powerful. GĂ¶del's second completeness theorem means that no system can prove its own consistency. Thanks.

A lot of this stuff still seems fairly murky to me, but I feel the waters clearing. I had some minor epiphanies over the past few days. For one, I now see a clean separation between two parts of making a Web-friendly mathematical system. One part is the translation between various proof systems, so as not to limit people too much. The other part is to arrange the inputs, outputs, and immediate files to make it maximally easy to share over the Web. These are quite separate: it's not hard to imagine proof format translators still stuck in the teletype interaction mode characteristic to HOL, or a Web-friendly system based on any one of the major logical systems. The latter seems easier to me, actually. I'm still a novice at the heavy mathematics behind logic and proofs, but I feel I know a lot about the Web and about how people can collaborate over the Internet.

The other insight has to do with the issue of making "conservative" proofs designed for portability. I think I can now answer the question: what are the natural numbers?

Well, what are the natural numbers?

This may be a bit sketchy, because the ideas still feel a bit raw.

At heart, the natural numbers are the Peano axioms. But just stating the axioms isn't quite enough to make them useful in a proof system. Among other things, it's dangerous to just add axioms lest they introduce inconsistency.

So, you build a construction of the naturals using native primitives of your proof system. For each axiom, you prove a theorem. In the Metamath approach, you then do everything in terms of this construction.

But what HOL does for type definitions is better. Essentially, by making a construction, you're proving that the axioms are consistent with the primitive logic. It doesn't really matter which construction you use, as long as you end up with the same axioms.

In fact, what I just said is a key meta-result. The usual interpretation of proofs in these formal systems is that they prove a single statement. But here, an abstract proof using only the Peano axioms can easily be transformed into a concrete proof, just by substituting in the actual construction.

Meta-logic in general makes me uncomfortable, but in this case I think it's what's needed to address deep problems of portability. In particular, I think it solves the problem of multiple different axiomatizations of the same basic concept. Your "construction" can go between two equivalent axiom sets just as easily as to a construction in your primitives. You keep your proofs abstract, but at the end you check them in some specific formal framework: ZFC if all you care about is the truth of the proposition, and "smaller" systems such as HOL if you want to check portability.

I don't think I've communicated the idea well. It's also a bit half-baked. I've only talked about naturals, but there are other things, especially functions, that are harder to treat this way. But I do think I'm on to something here.

VoIP

The quest for a good VoIP solution continues. One step which may be considered either encouraging or discouraging is TeamSpeak. Most importantly, it's not free software, but it does seem to work well enough to use. It comes from the gamer community rather than any existing VoIP constituency, which leads to some funky terminology. When someone joins a channel, a sci-fi female voice announces, "new player." On the plus side, there's an emphasis on performance and usability (especially with NAT's, which basically Just Work).

I do want to try GnomeMeeting again, but so far it hasn't worked for me. One of the most encouraging things about GM is that the next version will support the new Speex codec, which basically does for voice what Vorbis does for recorded audio. I took the standalone coder for a spin and was very pleased.

Proof-carrying code and code-carrying proofs

Michael Norrish sent me a nice email on my recent entry on proofs and code. The question of whether the proof checker should include a virtual machine, so that proofs can include programs is in some ways the dual of proof carrying code (and bib from 2000), in which programs come in "fat binaries" with proofs attached. A runtime can then check the proof, and execute the code with some assurances, primarily safety. PCC obviously needs a good, clean, efficient proof format, and equally obviously motivates a sharp separation between proof generation and proof checking. It's worth following.

Code-carrying proofs are by no means a new idea. In fact, Coq already implements them. It's not clear to me how deeply they're actually used. Among other things, you'd think they'd be useful for avoiding size blowup when importing theorems from other systems, but at least Freek didn't seem to use this in his HOL to Coq translator (Encoding the HOL Light Logic in Coq).

Translations between proof frameworks

When you look for it, there's been quite a bit of work done on translating from one proof framework to another. HOL has been translated into both Coq (and see another prototype) and NuPrl. These are not unique, although as it happens such translations can be quite tricky. For one, logical systems really do differ in expressive power. This is a bit unintuitive for those of us who grew up on the Church-Turing thesis, which effectively states that all programming languages have the same expressive power, because you can always write a translator from one to another. But logical systems differ in, for example, the number of ordinals. A proof that so-and-so-many ordinals exist can be ported to a "larger" system, but not to a "smaller" one. I can easily imagine a network of such formal systems, but I do not know how to draw the arrows. For example, I'm pretty sure that you can represent HOL in ZFC set theory, but my hunch is that the reverse is not possible.

However, I think a lower-tech solution may well be practical. Not all proofs make use of the full power of the logical framework. If a specific proof only uses the subset common to both frameworks, it should probably be portable.

I think this issue is related to the way Metamath nails down constructions. Since other systems naturally choose different constructions, you clearly can't just port proofs over. In Metamath, you can add sets and integers and get meaningful results. Obviously, porting any such proof to a different system will fail. However, if a specific proof uses only the integer axioms, and doesn't refer to the construction directly, it should be portable.

The way forward seems to be a significant portable library of the basic constructions: integers, reals, pairs, sequences, trees, etc. These basic constructions should be done in the most natural way in the framework. Then, when importing proofs, you hook up these constructions. There are some subtle details, mostly around partiality. What exactly is 1/0?

Proof efficiency

Another concern Michael raised is the relative inefficiency of a translated proof. Especially in the HOL framework, small proofs in the "source language" (which is really ML with a nice library) expand out to a large number of primitive inference steps. In the usual HOL scenario, the inference steps go very fast. If you're writing them out to a file, then sending it over the net to a different system, you might start caring about how many steps there are.

I'm not very concerned about this, for a number of reasons. For one, I believe it should be possible to greatly optimize the resulting proofs. For two, storage and bandwidth is relatively cheap and is getting cheaper. For three, a special-case data compressor is likely to be very effective, if one is needed.

But it's not clear one is needed. Both Automath and Metamath have significant libraries of theorems, and both can be checked on modern hardware in about a second. To me, this suggests a huge amount of safety margin between the optimal and the acceptible.

The Ghostscript codefest is over now. It went well, I think. We fixed a bunch of bugs, came to understand issues in such areas as embedded controllers for color laser printers a lot better, had a lively discussion about Fitz, and spent a little time hanging out and getting to know each other. I'm convinced that this is quite important, especially when the developers are geographically dispersed.

Billy

I got a number of responses on the Billy bookcases, generally expressing disbelief that they really were sold with such a basic defect as the holes drilled in the wrong place. No, they really were. rillian and jack both looked at them, and found pretty direct evidence - for the bolts that hold the thing together, there are two holes; one thinner and drilled all the way through, and one shallow but with greater diameter for insetting the head of the bolt. On correct pieces, the centers align exactly. On the two defective ones, they're misaligned by about a half inch.

I seem to have a knack for getting defective products. Perhaps it's just because it bothers me more than most other people.

Microkernels

I didn't mean to imply that microkernels are bad. In fact, there are definitely some good kernels based on the microkernel idea out there (QNX comes to mind), and a number of even more exciting research projects either based directly on microkernels or inspired by them (User-mode Linux, exokernel and eros come to mind).

What I was criticizing is the belief that, other things being more or less equal, a microkernel will turn out better than a comparable monolithic kernel. Experience now indicates otherwise. In particular, it seems you have to work quite a bit harder to get top-flight performance from a microkernel design. And, the really good ideas from microkernels (modules) have been ported to the monolithic style.

Beliefs of the form "X is a good way to build a system of type Y" are not easy to prove or disprove, because there are so many other variables that affect the quality of the result. If the question is important, the way to settle it is to build many systems using both X and non-X, and see if you can pick out a quality trend. If possible, though, it's better to focus on more crisply testable beliefs.

I'll give an example of one of those: a dynamic specializer is an effective way to get good performance out of highly dynamic interpreted languages. In particular, I was skeptical about Psyco, but at the rate it appears to be going, it looks like it will prove this belief. I hope this demonstrates that "show me the code" is useful for overcoming unwarranted skepticism as well as filtering out junk ideas.

28 Sep 2002 (updated 28 Sep 2002 at 08:30 UTC) »

A little light on technical content today, largely because I've been preparing to host a weekend codefest for Ghostscript developers. I want to study Bram's recent thoughts on trust metrics, and respond to an email from Michael Norrish on "code-carrying proofs" vs "proof-carrying code". Ah well, later.

Show me the code

rlevin wonders whether "show me the code" is perhaps too arrogant a stance. I don't find arrogance particularly appealing, but I think there is a good reason why this tenet is central to the free software ethic.

The problem is that the world of computing is filled with many, many opinions about things. Hard sciences are driven more by data and falsifiable experiment, but in the sciences with the word "science" in the name, it's still mostly opinion. However, there are good and bad opinions. Many years from now, when we look back, some will look sensible, but most won't. But now, I believe, critical thinkers among us can start to separate the wheat from the chaff.

The hard sciences, of course, evolved from proto-sciences dominated by opinion. Over many years, they developed a culture organized around propagating the good ideas more efficiently than the bad ones. Central to this culture is the idea of experiment. In particular, if a particular opinion can be tested by doing an experiment, it's a lot more interesting. Philosophers used to spend a fair amount of time debating how many angels can fit on the head of a pin. It's clear to us now that this is a matter of opinion. You cannot do the experiment.

As an example of an opinion of an entirely different sort, many chemists in the early 18th century were of the opinion that the relationship between a calx (such as rust) and its corresponding metal (iron) was the addition of a substance, some kind of "inflammable earth". Others believed that it was rather the subtraction of a substance that formed a metal from its calx. What decided the matter was a series of quantitative experiments carefully designed to settle the matter. Today, of course, we know the answer.

In computing, many opinions are of the form "method X is a better way to build systems of type Y than its competitors". I can choose from hundreds of examples, but I'll choose one fairly well known, not too controversial, but with some bite left in it. Microkernels, so believe some people, are a vastly superior way to build operating systems than monolithic kernels. Ten years ago in academic circles, such an opinion would have been considered entirely mainstream. Indeed, the arguments were quite persuasive: elegance, modularity, robustness. Who but a reactionary would disagree that they are the wave of the future?

Well, you can see where I'm going with this. The way to settle the question is to "show me the code". Today, the best kernels for general-purpose use are BSD (still) and Linux. Comparing, say, Darwin and Linux head-to-head is virtually no contest. Not only does Linux eat Darwin for lunch performance-wise, but it's also quite a bit more stable in actual use. Ordinary users see OSX kernel panics commonly, whereas in Linux-land, unless you're a kernel developer they're a pretty reliable sign your hardware has died.

There are still a few crazy holdouts who hold the opinion that microkernels are better. Perhaps they're right, and the superior track record of monolithic kernels is due to accidental factors, such as Linus's superhuman code chops. If so, I'll be convinced when they show me the code. After all, the early experiments to try to settle the phlogiston vs oxygen debate had many flaws, and trying to draw a conclusion from them may have been misleading.

But opinions are damn cheap. Any idiot can have them, and many do. Even smart people are tempted by pretty ideas. We're surrounded by a flood of wrongheaded opinions. The best way to convince me that your idea is one of the few good ones is to show me the code.

Yes, this attitude makes life harder for those with good opinions, but lack time or resources to write code themselves. Lord knows, I've felt the sting of that lash myself with my ideas on trust metrics. If I can't spare the time to code up at least a prototype myself, how can I expect other people to invest their time into my ideas? Since all my ideas are brilliant, it would of course be a lot more efficient if others would just take them up without me having to waste my time on coding, but in the off chance that one of my ideas isn't quite as good as I believed, for other people to take the attitude of "show me the code" might, I admit, save quite a bit of trouble.

Aeron

I bought a couple of Aeron chairs for the codefest, and so far I love mine. It's still not easy to come to grips with spending the money, but dammit, I earned the money from my own work, and if I should choose to spend it on something I'll enjoy, I shouldn't have to justify it in terms of higher productivity because of better ergonomics or whatnot. But, of course, it's easier if I do.

Ikea

I got a bunch of Billy bookcases from Ikea a few months ago, and tried to put the last four together yesterday, in preparation for the codefest. But two of them had the holes drilled in the wrong place, and I can't put them together properly.

I suppose I should try to fix the problem, either by getting a replacement from Ikea (not particularly easy), or by hacking (drilling new holes, filling with epoxy). I don't really want to invest my time and energy either way, though, particularly because the material is cheap. Particle board is in many ways the worst construction material very heavy, and no strength. What I'll probably end up doing is slapping some nails into them and ending up with improperly built bookcases, really more suitable for storage of old papers than showcasing my collection of books, almost all of which are meaningful to me and some of which are valuable. I won't be happy with them in my studio, but they'll do.

281 older entries...