Mystery Bail Theater

While working on a project, I was wondering how to most efficiently implement string internalization. String “interning” is when two strings that compare true always have the same reference. This is quite nice as you can test if two strings are equal by just comparing their pointers.

It seems everyone uses hash tables for this, and indeed when I asked someone with more VM experience, he suggested using hash tables as well. I was still curious though, especially since SourceMod has a fancy-pants compressed “double-array” trie implementation. We knew that insertion time was poor, but I secretly hoped that this cost would be amortized over the much faster retrieval case.

I took a dump of ~5MB of plugin text from the forums and extracted all syntactically valid identifier tokens (including ones found in comments and strings). In total I had 553,640 strings, of which 19,852 were unique. This seems pretty reasonable as people tend to use similar names for identifiers. I set up four benchmarks:

• std::map<std::string, int> — Typical STL container (red-black tree I think)
• SourceMod::KTrie<int> — SourceMod’s Trie container
• SourceHook::THash<const char*, int> – SourceHook’s “TinyHash,” chained hash table
• SourceHook::THash<std::string, int> – Same, with more allocation involved

As a note, “TinyHash” is one of the first ADTs I tried to write, and it shows. Its stock find() function iterates over the whole table, instead of computing a hash, which is really slow. I filed a bug on this and hacked up my local copy to benchmark.

Each test gets run twice. The first “batch” run goes over all identifiers and inserts them if they are not already in the container. The second “retrieval” run does the same thing, except now there will be no string insertions since the table is populated.

Here’s the results, in milliseconds. I used the Intel C++ Compiler, version 10, on my Macbook Pro (2.4GHz Core 2 Duo, 4GB RAM).

Whoa! Everyone was right. Well, I knew that insertion on our Trie was pants, but I was surprised to see that the cost was in no way amortized. In fact, it wasn’t even faster on retrieval! Why? I don’t know (yet), but if I had a guess, it’s because our Trie has to peek at memory a lot. David Gregg explained to me that the advantage of tries is sorting. He also theorized it might be faster on long strings, which is in line with my original benchmarks ages ago, where I used very long randomized strings. (Tries are also good for compression.)

Well, I has sad. Looks like we’ll be ditching KTrie for a chained hash table. At least for string interning. Someone is welcome to request a hash table ADT for SourceMod scripting as well.

I also benchmarked non-PGO ICC and non-PGO g++-4.2 if you would like to see other compilers.

Note 1: Go article returns next week.
Note 2: Firefox 3.5 comes out tomorrow (June 30th!)

dvander 2 comments Articles, Coding

Andreas Gal gave the TraceMonkey talk last Thursday at PLDI, in Dublin. This talk was a bit different than normal as it focused on type specialization, rather than control flow. The crux of TraceMonkey, after all, is that it generates statically typed code that operates on native values. There is no need to unbox/box data on trace as the type is known. The “tracing” aspect can be considered as a means to this end.

We thought the talk was well received. Interest in optimizing dynamic languages is sure spreading.

I defer to Dave Mandelin’s blog for a real follow-up of PLDI 2009.

dvander No comments Mozilla, Tracing

Last week I gave an introduction to Computer Go, and left off mentioning that the current hot topic is Monte-Carlo methods. These methods use repeated random gameplay to sample a position’s value.

How do programs use random gameplay to reach a good conclusion? While you can sample specific positions, you only have finite time. How do you balance simulating one interesting position over another?

A few years ago a breakthrough algorithm was published, called UCT (Upper Confidence Bounds applied to Trees). The algorithm attempts to minimize the amount of sampling needed to converge to an acceptable solution. It does this by treating each step in the game as a multi-armed bandit problem.

Bandit Problem

A bandit is a slot machine. Say you have a slot machine with some probability of giving you a reward. Let’s treat this reward as a random variable . You don’t know the true value of , but by playing the slot machine repeatedly, you can observe an empirical reward, which is just an average ().

Now consider having slot machines, each expressed as a random variable . Once again, you don’t know the true value of any . How can you choose the right slot machine such that you minimize your losses? That is, you want to minimize the loss you incur from not always playing the most optimal machine.

Well, if you knew the true values of each machine, you could just play on the one with the greatest reward! Unfortunately you don’t know the true values, so instead you conduct trials on each machine, giving you averages . You can then use this information to play on the observed best machine, but that might not end up being the most optimal strategy. You really need to balance exploration (discovering ) with exploitation (taking advantage of the best ).

A paper on this problem published an algorithm called UCB1, or Upper Confidence Bounds, which attempts to minimize regret in such multi-armed bandit problems. It computes an upper confidence index for each machine, and the optimal strategy is to pick the machine with the highest such index. For more information, see the paper.

UCT

Levente Kocsis and Csaba Szepesvarí’s breakthrough idea was to treat the “exploration versus exploitation” dilemma as a multi-armed bandit problem. In this case, “exploration” is experimenting with different game positions, and “exploitation” is performing Monte-Carlo simulations on a given position to approximate its . UCT forms a game tree of these positions. Each node in the UCT tree stores and a “visit” counter. The confidence index for each node is computed as:

Each run of the algorithm traverses down the tree until it finds a suitable node on which to run a Monte-Carlo simulation. Once a node has received enough simulations, it becomes “mature,” and UCT will start exploring deeper through that position. Once enough overall simulations have been performed, any number of methods can be used to pick the best action from the top of the UCT tree.

How this all worked confused me greatly at first, so I made what is hopefully an easy flow-chart diagram. In my implementations, UCT starts with an empty tree (save for the initial moves the player can make).

UCT converges to an acceptable solution very quickly, can be easily modified to run in parallel, and can be easily interrupted. It has seen massive success; all competitive Go programs now use it in some form.

Levente’s paper: click here

Next week: Monte-Carlo simulations and my project.

dvander No comments Articles

As a degree requirement, my university requires students to complete a “major qualifying project” (MQP) in their field of study. Usually these projects are done in teams of two to three people, and often they take place abroad. MQPs last seven weeks and they culminate with the group writing a thesis-length paper (80+ pages) and giving a twenty minute presentation.

One of the projects available last quarter was about teaching computers to play Go. It took place in Budapest, Hungary, because one of the foremost researchers in the field (Levente Kocsis) works at the renowned Hungarian Academy of Sciences.

I really enjoy Go. It is an amazingly deep game where complex play is derived from very simple rules. Computers have a very difficult time playing it. Amateurs can defeat the best programs even when they run on supercomputers. Playing Go is very much an open problem in artificial intelligence.

The project was researching Go with Levente Kocsis. I was really eager to get on board! Because of my (bad) academic standing I had problems with the bilious idiots in the registrar and administration, but eventually I got in. In the end, no one else joined. I had to tackle the whole project myself. That’s implementation, research, testing, and writing that paper — all in seven weeks. Well, that’s a lot for one student with absolutely no AI experience.

If I had known this ahead of time, and I had also known that I’d be going into real compiler research, I might not have done this Go project. On the other hand, I loved Budapest. It’s a great city, and I’d go back in a heartbeat. I also learned a lot. So in another one or two blog posts, I will try to share a bit about the project, before it completely slips my mind.

Introduction

Go is a two-player board game. The board is a 19×19 grid. Players take turns placing black and white stones on the intersections. The object of the game is to surround as many empty intersections as possible. The game ends when both players agree there are no more moves, at which point the player with more “territory” is the winner.

So why is Go really difficult for computers, while a desktop can beat grandmasters at Chess?

The standard Chess algorithm is α-β pruning over minimax trees. Unfortunately Go’s search space is too large for this to work efficiently (see chart). The average number of moves per position in Chess is around 35, whereas in Go it is 250. Even if you play Go on a small board (such as 9×9, for beginners), there are around 40-60 moves per position.

There are additional complications. In Chess there are opening books and end-game solvers. In Go there are opening styles and patterns, but they evolve rapidly over time and are not rote as in Chess. It is also possible to quickly evaluate the value of a Chess position, whereas in Go the game must be played until the end to get an accurate scoring.

So, that’s what doesn’t work work. What does work?

Monte-Carlo Introduction

If there’s one thing a computer can do quickly, it’s picking moves at random until there is a position that can be scored. With enough random sampling, the true value of a position can be approximated within a certain accuracy. This is the basis of all Monte-Carlo algorithms, and it has become the foremost area of research for Go.

The most common application is Monte-Carlo Tree Search. The idea is to play random games from various positions. The results can then be used to explore the most interesting paths in the game tree. The idea seems ridiculous both in its simplicity and stochasticity, but it works.

Next week: Bandit problems.

dvander No comments Articles

To mirror Dave Mandelin’s blog post, we’ve gotten a paper on trace compilation accepted to PLDI, one of the top programming language research conferences.

Andreas Gal is presenting the paper on June 18th in Dublin. You can read it by clicking here (requires PDF reader).

Trace compilation is the new optimization technology appearing in Firefox 3.5. If you want to see how it feels, give 3.5 beta a go (it’s stable and nearing release).

I’ll have a bigger and better post about this and sundry things soon. The short of it is, I’m back into trace/language research after a six-month reprieve. It is good to be back!

dvander 2 comments Mozilla (non-PMO), Tracing

People reading this blog probably know about PM. He was one of the original SourceMod authors, and forked off his work as the amazing SourceHook, which became the backend to Metamod:Source.

I finally got a chance to meet PM today. He was going from Slovakia to Germany, and stopped in Budapest, Hungary for a few hours. We did a little touring around the Castle District (Várnegyed). There are some great panoramic views of the city, since it is high up on the hills of the Pest side. It’s also a very touristy area, people were speaking English (and according to PM, German).

Picture or it didn’t happen.

It’s rare I get a chance to meet people from the AlliedModders community, and it’s great to put a real person to the handle/avatar on the forums. I can only recall a few other times to do this, I got to meet Freecode, OneEyed, and teame06 at various LAN events.

I first heard from PM in 2003 or so. I was writing a Counter-Strike mod and it was exposing a very bad bug: if you removed a weapon on the ground, Counter-Strike would crash, either immediately or much later. I couldn’t figure out what was going wrong and was at my wits’ end, until I got a random message from PM with a mysterious but working fix.

In early 2004 he and I both ended up being the core developers of AMX Mod X. Back when I was still trying to figure out what pointers were, or what malloc() really did, or what the difference between the stack and heap were (you know, all these beginner things), PM was hacking away at the Small virtual machine and doing big rewrites of some of the most terrible AMX Mod code. (I eventually figured this stuff out, and looking back I wonder why PM let me do anything.)

So, meeting one of the most awesome Half-Life/AlliedModders community developers after six years was really cool. I hope I get more chances like this in the future. And thanks to PM for putting up with our boring traipsing around the city!

dvander 8 comments AlliedModders, Articles, Slice of Life

I spent a good portion of this week upgrading various parts of the AM “infrastructure” (read: server held together with twine and duct tape).

2009-03-08
 * 16:32 GMT-5. PHP FastCGI upgraded from 5.2.8 to 5.2.9.
 * 16:44 GMT-5. MySQL upgraded from 5.1.30 to 5.1.32.
 * 17:38 GMT-5. MediaWiki upgraded from 1.10.1 to 1.14.0.
 * 17:44 GMT-5: ViewVC upgraded from 1.0.5 to 1.0.7.
 * 17:55 GMT-5: phpMyAdmin upgraded from 2.11.3 to 3.1.3.
 * 18:55 GMT-5: PHP FastCGI recompiled against MySQL 5.1.32.
 * 19:27 GMT-5: Apache for users upgraded from 2.2.8 to 2.2.11.
 * 22:08 GMT-5: Apache for vhosts upgraded to 2.2.8 to 2.2.11.
2009-03-09
 * 00:13 GMT-5: Bugzilla upgraded from 3.1.4 to 3.2.2.
 * 01:16 GMT-5: Buildbot upgraded from 0.7.8 to 0.7.10.
 * 01:59 GMT-5: Mercurial upgraded from 1.0.2 to 1.2.
 * 04:39 GMT-5: Debian upgraded from 4.0 (etch) to 5.0 (lenny).
 * 04:42 GMT-5: Mercurial and Buildbot switched from Python 2.4.4 to 2.5.2.
2009-03-10
 * 00:44 GMT-5: vBulletin upgraded from 3.6.7 to 3.8.1-PL1.

DS helped me on the Buildbot/Mercurial side. Hopefully everything still works for the most part. Many of these upgrades were sorely needed, fixing known bugs or security issues.

All of this was delayed so long because, well, no one wanted to do it. We maintain private patches against Apache, Bugzilla, vBulletin, and ViewVC. Keeping track of those changes is difficult.

This time around we started using Mercurial Queues. We keep the official software in a Mercurial repository and our custom changes in a queue (patch series). When it’s time to update, we pop the series, commit the new software, then push the series back.

This workflow is nice and all but Mercurial queues aren’t really a part of the repository. So if someone wants to check out the repository and stage development somewhere else, they have to export the queue manually and re-import manually on the other side. Ideally, these patches would be lightweight development forks off the main repository.

We used to keep separate repositories and do three-way diffs. That was a huge pain. Never again. It still seems like we’re missing something though.

dvander 1 comment AlliedModders, Articles

I spent the better part of the last week or so on a rather annoying TraceMonkey bug. A few weeks ago I mentioned that we extended multiple type-specialized trees to include specialization on global types. Suddenly a rare but dangerous class of bugs appeared.

The Bug

When a trace exits back to the interpreter, it does so through a guard. These guards contain an “exit typemap,” which tells the interpreter how to take types from the native CPU and box them back into generic containers for the interpreter. The bug was that some global variables were being boxed back as the wrong type. For example, the 32-bits in an integer would be stored in a slot interpreted as a 64-bit double, or an object would be tagged as a string. When the interpreter went to unbox these again, it got garbage. That’s bad. The only way this can happen is if the exit typemap contains wrong type information.

Lazily Specialization

Global variables are tracked lazily. The overarching Trace Monitor keeps track of which global variables have been seen on trace. For example, say we have a type-stable tree, Tree 1. It has no global variables, and thus has empty exit and entry typemaps (for globals). Later, the global variable X is encountered. It’s now being tracked, but Tree 1 doesn’t have a type entry for it.

If Tree 1 wants to execute again, it will lazily update its entry typemap. The exit typemaps on the other hand are fixed and cannot be modified. So now Tree 1 looks like this:

Tree 1 Entry: X = Object
Tree 1 Exit:

When exiting the tree, we merge the exit and outermost entry types together, giving a complete typemap. More on this later. When entering Tree 1, X is unboxed and boxed as an Object, even if it never gets used. This is because Tree 1 could call into a nested tree that does use X.

Problem #1

Let’s say later we build Tree 2. It’s the same loop as Tree 1, but it’s type-unstable. The typemaps look like this:

Tree 2 Entry: X = String
Tree 2 Exit:  X = Object

TraceMonkey’s multitrees kicks in, and now Tree 2’s exit will go directly to Tree 1’s entry, since their types link together. When Tree 1 exits, we combine the outermost typemap with the exit typemap. In this case, Tree 2’s entry is the outermost typemap, but it contains the wrong type! The type of X is an Object, but now it’s being boxed as a String. Crash.

Note: This is different from normal type instability, because type unstable traces represent a path where a type is actually changed. In this scenario, the type is being mistaken, which is deadly.

Failed Solutions

Solving this was complicated by nested trees. Trees can be deeply nested, and each tree along the way could be missing global types, so it seemed like we needed to recover missing global types incrementally up the call stack. That is:

1. Start with global types from the innermost exit.
2. Add missing globals from each tree on the call stack, starting at the deepest, and ending at the outermost tree.

Since the outermost tree is the original tree we started with, it is guaranteed to have all missing types, so it was the last resort. Achieving this ended up with three different implementations as the patch progressed, but the idea was: as we enter trees, push them onto a stack. As we exit trees, pop the stack. If we hit a guard (which exits), immediately reconstruct the full global typemap using that stack of trees. By the time we exit back to the interpreter, a global typemap will have been prepared already.

Alas, this wasn’t enough, because…

Problem #2

Say we have two trees on the same loop, Tree 1 and Tree 2. Their typemaps are:

Tree 1 Entry: X = Object
Tree 1 Exit:  X = String
Tree 2 Entry: X = String
Tree 2 Exit:  X = String

In this situation, Tree 1’s exit is linked to Tree 2’s entry. Later, a new global appears, Y. We execute Tree 2 with X = String and Y = Object. Then we execute Tree 1 with X = Object and Y = String. Because of lazy specialization, the typemaps now look like this:

Tree 1 Entry: X = Object, Y = String
Tree 1 Exit:  X = String
Tree 2 Entry: X = String, Y = Object
Tree 2 Exit:  X = String

This is problematic, because Tree 1 is still connected to Tree 2, but their typemaps are incompatible! If we run Tree 1, Y will be unboxed as a String and reboxed as an Object, without the underlying type actually changing. The solution to Problem #1 doesn’t fix this.

Solution

Linked trees should never have incompatible typemaps. What counts as a linked tree? Any two trees that are connected via a nested call (nested loops), or any two trees that are connected via type instability (multitrees), are “friendly.”

In pseudocode:

LazilySpecialize(Tree):
  AddMissingGlobalTypes(Tree)
  FOREACH Tree IN FriendlyTrees
    IF MissingGlobals(Tree)
      LazilySpecialize(Tree)

Now when a tree exits, it suffices to use only the exit typemap and the typemap of the innermost entry (that is, the tree the exit immediately came from). This neatly solves Problem #1. If any one tree gets a new global, the entire graph of connected trees is updated, solving Problem #2.

This is probably one of the more difficult TraceMonkey bugs I’ve had to reason through. Nested trees always seems to complicate how you reason about tracing, and in this case it ended up not mattering. And though the problems are conceptually disjoint, they also feed into each other: fixing only #1 led to #2 breaking.

dvander No comments Articles, Mozilla (non-PMO), Tracing

It’s the question no one asked!

SourceMod Motivation

As I’ve too often said, the original SourceMod failed because we tried to do too much without understanding project management, the scope, or even how to code most of what we wanted to do. When we began talking about restarting the project, we did so from a research point of view. We asked the community: should we spend time writing a new language, or just get something to market on Pawn?

The community voted the latter, so we approached the project with the following questions: What would SourcePawn look like if it were modernized a bit? What would AMX Mod X look like if it had a much cleaner and more consistent API?

We wrote a new VM and JIT for Pawn. We added fast strings, basic function pointers, and dynamic arrays to the language. We boxed it into a fairly simple plugin with a small API and watched it take off, adding more as it got requested. Commands, events, menus, FFI, et cetera. It was fun designing all this.

What’s left to do that’s interesting? There’s a lot of open bugs for features people want. Perhaps the only really big ticket item left for SourceMod is script-level detouring of virtual functions, which has been in the works for a long time (and is about 50-60% done).

Most of what needs to be done though is hacking away at the indescribably infuriating Source engine, and all of the closed source games that run on it. This is boring. Sure, we’ll continue to do it. But it’s not interesting. It’s rote reverse engineering against the same old stuff. It’s monotonous and tedious, not to mention fragile. If we were developing a full product (game), we’d have access to everything. But we don’t, and we’re stuck in what’s ultimately a script kiddy’s role.

SourcePawn is Limited

SourceMod has stretched Pawn to its limits. It’s a crufty language. There is no heap. Memory management is manual. Strings are byte buffers. There are no types. There is no data integrity or safety. There are no structures or classes, just C-style functions and opaque integer values.

Basically, writing SourceMod plugins feels vaguely reminiscent of WinAPI except you don’t have any of the advantages of C.

Compiling plugins is asinine. It causes so many problems – arbitrary compatibility issues, secret backdoors, license violations, site and forum maintenance difficulty, et cetera.

Pawn is very fast because it’s so limited, but it’s also hard to squeeze more optimizations out. Everything has to go through natives, even small trivial functions. The bytecode is very low level, and though Pawn as a language has immense JIT potential, our JIT is very basic because the bytecode is too hard to analyze.

So… where do we go from here?

I want to start looking at more involved questions. The type of questions SourceMod might have asked if we had chosen the former path when restarting the project. I’ve boiled this down to two issues:

• What would a language look like if it were specifically designed with trace compilation, game scripting, and embedding in mind?
• What would Pawn, a language designed to be simple, look like as a truly modern scripting language, and not something from 1980?

Relatedly, there are two interesting questions for SourceMod:

• What would SourceMod (as an “AMX Mod X done right”) look like with a more modern, object-oriented language available?
• Can SourceMod be self-hosted?

I have mentioned this before. Most people say “Just use Python, or JavaScript, or Mono, or X!” where X is some random language that I don’t want to use. That’s not something I’m interested in. If you want to see what your favorite language looks like in Source, go ahead and embed it. This isn’t a case of NIH. It’s something I very much want to pursue to see where it goes.

The Next Project

The next project has been started. It’s a new programming language. I’ve chatted a bit about it in IRC, mostly to get some bikeshedding out of the way. As the project is not official yet, there’s no official name.

I have no idea how far this will go, or if it will go anywhere. But it’s been started. Goals, or rather, properties of the language will be:

• Dynamically typed.
• Strongly typed.
• Object oriented.
• Garbage collected.
• Static lexical scoping.
• Nested block scoping.
• Explicit variable declaration.

I’ve started work on a garbage collected, register-based bytecode interpreter. I won’t start the actual compiler until I’m pleased enough with backend progress, and I won’t start the JIT until I’m pleased enough with overall maturity.

Everything will be incremental. Parity with Pawn is enough to let people start playing with things, and that would be a huge, huge step.

If anything ever comes of this, it will be interested to see how it can fit into SourceMod. While we can never get rid of Pawn because of compatibility, there’s lots of room for potential. IRC goer theY4Kman embedded Python in a SourceMod extension. We could try something similar as a starting point, or we could do something like EventScripts where Python and the “old language” of the older, darker times lives alongside.

Who is working on this?

For the most part, just me (and, if they so desire, other SourceMod developers). This is a pet project and I don’t want to be too distracted. I also don’t want to tie SourceMod to its success, so it has no relation to SourceMod’s roadmap in any way. But like SourceMod, it’s something I will attempt to work on incrementally in my free time.

If progress on this is something people are interested in, I can keep talking about it. If you’re really interested in it, feel free to find me on IRC (irc.gamesurge.net, #sourcemod). More than likely I’ll be occasionally throwing random design questions around out loud.

dvander 4 comments AlliedModders, Articles

Since I’ve returned to the wonderful and freezing east coast in December, I’ve had two encounters with AAA while staying with my parents.

Encounter 1
Around December 20th or so our car couldn’t get up an unplowed hill in New Hampshire. It was a two-wheel drive Mazda (I forget the model), and for whatever reason we couldn’t gun it enough to get to our house.

We called AAA to see if the car could be towed up the hill. No go, they said, if we couldn’t get it up then they couldn’t get it up either. Furthermore that didn’t count as something that would be covered, since the problem was with the road, not the car.

We said, “well eventually we’ll be stuck long enough out here that we’ll run out of gas.” The reply? “Well, in that case we’d come bring you more gas.” Uh.

I can believe them that that towing the car up the hill seemed implausible. But they could have offered something, like, towing the car somewhere safe and then bringing us up the hill.

Instead, we had to walk. In 10F degree (-9.5C) snowing weather. Up a steep hill, in the dark, for what we later clocked to be 0.7 miles (1 km). While carrying all of our luggage. It felt like the desert scene from Spaceballs, but we didn’t have any food or supplies at the house, so there were more necessities than usual.

I have severe chronic asthma, which though has gotten much better over the years, still needs work. It must have taken me around fifteen minutes to breath normally again after deeply inhaling freezing cold air for so long.

Later the next day, a family acquaintance with a plow came and plowed the street, then managed to gun the car straight up the hill with no problem. Gee, I wonder if AAA could have done that.

The incident caused my parents to, a mere ten days later, get a used 2006 Toyota Highlander Hybrid for its four wheel drive.

Encounter 2
Cars are supposedly supposed to turn off their internal lights once they’ve been closed and locked. For one reason or another that didn’t happen in the Toyota Hybrid, and the battery drained. We didn’t have jumper cables for that car yet so we called AAA.

AAA refused to jump the car because it was a hybrid. They hadn’t developed a procedure for hybrids yet. Huh? Hybrids have been out for a decade. What has AAA been doing all this time?

But that’s irrelevant, because it’s a normal battery like any other car. The owner’s manual even tells you how to jump it, and it’s exactly the same as any other vehicle.

So AAA tows the car back to our house, but the guy AAA contracted wants to try jumping it anyway out of curiosity. He connects positive to positive, then negative to negative – wait what?

You’re not supposed to connect the dead negative terminal – it’s supposed to go to ground. So AAA apparently doesn’t know how to jump a normal car. The guy was nice though, was glad to know the correct procedure, and what do you know? The hybrid jumped fine.

Conclusion
I’m not really sure I have one, other than that AAA clearly isn’t going to go out of their way to help you. They have some tome of procedures and will obey it to the letter. Maybe they’re afraid hybrids will detonate into some sort of mushroom cloud explosion. Perhaps in 2050 they’ll know how to jump them (or, any car).

dvander 2 comments Articles, Things that Annoy Me