p
e
r
s
o
n
a
l |
THPS Rules!
(30 Mar at 18:05) |
Oof, so busy! I finished up my paper(s) for SIGBOVIK and submitted them. Phew. I'll post 'em here after the embargo ends. I think the papers may be the canonical form of this particular project, but I'm starting on a visual version, which will probably become a video some time this month. (First up: I need to prep a live version for Unnecessary Detail, as mentioned in the previous post.) But right now I'm also on the West coast (in a car traveling from Los Angeles to San Diego) for a short family trip which was cut even shorter by the Spirit of Bad Aircraft Management of Spirit Airlines. Immediately upon arriving at the hotel I looked out our window and saw a little park and thought, "it would be fun to skateboard in that park," (I am not a skateboarder) and then, "did I already skateboard in that park 20 years ago in Tony Hawk's Pro Skater?" and I looked it up and yes, the Los Angeles level in THPS3 is based on that very park.
I thought I wasn't going to be able to make the SIGBOVIK live event at all because of this trip, but it looks like I will be back and might try to make it, depending partly on whether I can get any material together for a talk in time.
I spent most of my free time in March on hacking and writing for this silly paper, but sometimes the brain needs a break, and I continued with Teardown. I think this game is great. It is impressive technically and graphically. The sandbox is fun; I had a great time painstakingly disassembling an enormous blast furnace until the framerate became intolerable. (The way the physics works, the entire blast furnace can be held up by the connection of a single voxel. This is obviously totally unrealistic but it is pretty fun to try to blowtorch around an entire building and then try to hunt down why it is still standing.) But I was also impressed with how they managed to make the missions compelling too. There are a couple of ergonomic annoyances (like: There are limitations on what keys can be rebound to what, so I had to play with a controller. And the quick-save is great, but given that you might spend an hour setting up a heist in a level, it would be nice if you could make an in-level save that was a little less quick, just in case you accidentally stayed up to 2am). Getting close to the end of that one. I also played through Gunlocked, which was a good small Roguelike shoot-em-up. The powerups were really well done; it just could've used a bit more variety in the bad guys.
OK, getting a bit carsick here so I'm going to put the laptop away. See you soon. |  |
|
| Can't wait to see what you've done this time for SIGBOVIK! |
You know in the original ReLU paper they were more than hinting that ReLU was a switch. But everyone latched onto the function viewpoint.
That was the wrong thing to do. ReLU either conducts 1 to 1 or outputs zero. Therefore it is a switch, with switching decision (x>=0)?
Then weighted sums are being connected to and disconnected from each other. And those connected weighted sums can be simplified by linear algebra at each stage. You can gain a very clear insight into how ReLU neural networks operate.
You can look up switchnet, or WHTeBooklet or a frozen neural network on Archive.org. Or somewhere in this booklet I talk about it: https://sites.google.com/view/algorithmshortcuts/walsh-hadamard-transform |
| I've been stalking you on SourceForge for a while (but tried my best not to spoil the fun), and I think it'll be quite challenging to turn your work into video this time. Please don't feel discouraged by the amount of editing work this might require! Of course I have no rights to represent your fans, but I'd still like to say that we'll fully support you, at least emotionally. Take your time, don't let this burn you out, we'll wait. At the end of the day, I really want this to become a video, and your videos are by far the best videos I ever watched on YouTube. |
| Video recording of Tom7's SIGBOVIK presentation about his new typography system is at https://www.twitch.tv/videos/2111841043?t=0h8m53s |
| The proceedings volume of SIGBOVIK 2024 is now avaliable at http://www.sigbovik.org/2024/proceedings.pdf . This includes Tom7's articles and much more. Have fun reading, everyone. |
| I want to see the original metroid guide! |
The papers for Badness 0 are so much better than the presentation suggested! Had I known, I'd just have skipped the presentation entirely. That said, here are a few comments.
In page 132 and 146, in “since mathematics does not obey the rules of project management”, the space after “mathematics” seems to be wider then the other spaces in the same line. I wonder why this is.
In the colorful chart on page 134 and 149, given that the x axis is the number of CPU threads used, what does the column labeled “00” mean? I don't care what the colors and numbers in the chart even mean, but this part somehow bothers me.
Still on page 134, in the description of that same chart that I just said seems to contradict the chart, where it says “The x axis of the graph is the number of CPU threads”, the space after “x” is too narrow.
In page 138 and 139, in text “The tt function” and “value "FixederSysLight" to the layout” and “FixederSys family.[25] Functions like b and it apply”, all of which are in lines where the spaces are spread wider than normal, the space after text typeset in FixederSys seems to be too narrow. (The space after the footnote mark also appears to be too narrow.)
In page 141 and 157, in “But it can certainly be more satsifying.”, the last word is either a typo or some pun that I don't understand.
On the same pages, re the achievement system in BoVeX, I'd like to remind you that you still haven't awarded yourself an Xbox style achievement box for completing Pac Tom level II in expert mode. I find this more jarring now that I know you still care about giving achievements.
Pages 142 and 158, re “challenge your paper's reviewers to a game of chess against a strong engine embedded within your document”. A suitable starting point for such a thing would be Óscar Toledo's nanochess from https://nanochess.org/chess.html . One of its versions runs on a 6502 CPU, which is great for this project because you can say that it's a trivial task to just get any of the thousand 6502 emulator implementations from the internet and wrap it as a BoVeX runtime primitive, but then actually spend weeks perfecting the details of a 6502 emulator that you implement on your own from scratch. I should have mentioned this back in 2019 when you wrote Elo World, but I didn't really understand your request for “easy to describe algorithms for playing chess” until you published the paper, and then it was too late.
|
| @Anonymous The original metroid guide is in the papers' bibliography, https://gamefaqs.gamespot.com/snes/588741-super-metroid/faqs/10114. |
| looking forward to the video. Actually checked your "channel" to see if I had somehow missed the notification. |
jonas: Thanks for the bug reports :) I think all the cases of extra spaces are a known problem where certain constructs (often it's a comment) result in adjacent text nodes that don't get fused. It's surprisingly hard to fix well, since the right fix would be to finish the somewhat delicate hand-written bytecode that concatenates layout nodes. I just ran out of time to get that done before the deadline, although since the deadline I've rewritten that part of the compiler so that it's not as bad to hand-write bytecode, so I plan to fix this one soon. The spaces burn my eyes too!
Regarding the ones where space seems too narrow, sometimes this is apparently the actual metrics for Palatino italics when mixed with roman (but I agree it looks bad and should be tweaked). Other times it is a consequence of hacks (in the paper's code) to work around the problem (with the language implementation) described above. Most should be fixable. Of course I absolutely deserve to be taken to task on these things.
Fixed the typo! It's a miracle if that's the only one, since I didn't have any spell checking support!
Colorful graph: Yeah, this is described imprecisely. 00 actually means 1 thread, and 01 means 2 threads, and so on. I just looked and the data are correct, but I didn't do the +1 when generating the image. It's fixed in case I regenerate that.
Others: Thank you for your enthusiasm! There is a video coming. I have essentially all the content now and I am still alive. |
|
|
|
|