> It means he can't physically get a better time

... using the same techniques the TAS does.

It's possible (although unlikely) that a new glitch/exploit could be discovered that allows for faster times.

This is, as always, a local maxima— new strats may come along which change everything again. Summoning Salt's WRP videos do a great job of showing how the strategies for a particular game evolve over time as new things are uncovered:

https://www.youtube.com/channel/UCtUbO6rBht0daVIOGML3c8w/vid...

I came here to paste the link to Summoning Salt, and an upvote alone isn't enough to emphasize how much I recommend his channel. Here's the link to his retrospective on Super Mario Bros specifically (now a bit out of date, of course): https://www.youtube.com/watch?v=RdAkY7RfajY

Summoning Salt makes great videos. The SMB1 video has already been linked, but I also love his video on the Choco Mountain track in Mario Kart 64. It has some great livestream captures which show records being broken, and contextualizes them well enough to give them some pathos:

https://www.youtube.com/watch?v=Y99Wj-NStok

Summoning Salt makes absolutely incredible content. I'm interested to see how (if at all) he covers this event.

There aren't any techniques he can use that aren't available to a TAS runner. TAS runners aren't just using reinforcement learning to let a computer learn how to play; they're thinking about possible techniques and glitches, and specifically programming their tools to try and exploit them.

What's common, especially in the SMB speed running community, is that the human and tool-assisted runners loosely collaborate. e.g., if one of the humans has an idea, before they spend days or months trying to figure out for themselves if it's even workable, they'll get someone to inexpensively proof-of-concept it with a tool-assisted speed run.

A new glitch has been discovered the same day as the new record run: https://www.youtube.com/watch?v=OcVG8IDstuA

For sure. This is what happened with Super Mario World. Someone discovered a glitch that dropped it down to like 20 minutes. Then another one was found that dropped it to 10. Then 5. Then 4. Now, you can beat SMW in under 42 seconds if you do everything correctly. Or you can spend two hours building and playing Flappy Bird in it instead. Either one works

The Flappy Birds stuff is utter madness.

Given 60fps games, with 5 minutes of play, and 8 NES buttons, it seems reasonable to search through just under 4,608,000 possible plays to find the shortest run!

5 mins * 60 fps * 2^8 button possibilities = 4,608,000 button possibility frames = 4,608,000 possible plays

Close! You swapped an operation there: it'd be (2^8 button possibilities) ^ (5 minutes * 60 FPS), or about 10^43348.

A bit of a bigger space to search ;)

Only 2^7 button possibilities on a real controller, because you can't press up and down, or left and right, at the same time.

Worn controllers can allow it to happen, however, so I don't think it's worth discounting.

but then wouldn't TAS also be able to use it?

Yes, but this is referring to a specific TAS (the current fastest* known one), not a possible TAS.

http://tasvideos.org/Game/nes-super-mario-bros.html

[*] It's also worth noting that there are 2 categories for TAS timing: Humanly possible vs TAS-only. The former meaning that it's possible in theory for a human to achieve under "normal" conditions, the latter implying the opposite e.g. using left + right inputs at the same time which is not possible on a normal NES pad. The TAS in question was the current theoretical limit for humans.

I thought glitches were not allowed?

Speed runner come up with different "categories" for runs.

The most common is "Any %," also the most basic, it means finish the game, usually anything goes, glitch away.

You can also have no glitches category, 100% complete, no warps, no powerups, all <game item>s, etc. Sky's the limit. The community creates the categories.

My favorite are randomized speed runs, where everyone gets the same randomized version and have to speed run it against each other. Normally these are done as real time races and the results of the randomization aren't known to the player until they start. Not all games support it (think games like Zelda and Metroid where you have to collect powerups to push get past obstacles to collect more powerups to get past more obstacles). The randomizer has to make sure the game is still beatable (so no locking a required item behind an obstacle that requires the item). It is also an area that TAS can't help and it has more to do with strategy than physical perfect button presses. Also, since every time you uncover a new randomized item it changes what is the most potential optimized path, it require a deeper knowledge of the game compared to speed running where there is normally a given optimized path that people compete against.

> Not all games support it (think games like Zelda and Metroid where you have to collect powerups to push get past obstacles to collect more powerups to get past more obstacles).

There is, however, OoT Randomizer[1] which is really fun to play and watch. You can get end stage items at the beginning depending on your seed.

[1] https://github.com/AmazingAmpharos/OoT-Randomizer

I just realized I worded that poorly. I meant that games like Zelda and Metroid do support it, other games don't because there isn't much to randomize (such as platformers, given these tools generally can't create random levels, just swap item locations).

Any% refers to completion percentage, not the number of glitches allowed. But yes, some categories allow for glitches and some don't.

Over the years the Super Mario Bros community in particular has become more accepting of glitches, as otherwise both TAS and human runs would have hit the limit of how fast is possible. By allowing glitches there is now always a possibility that the runs can be further improved.

>Any% refers to completion percentage, not the number of glitches allowed.

Indeed, which is why I said "usually."

I usually see "Any %" and "Any % no glitches," but I just dabble in watching speedruns. If I see "Any %" with no other qualifiers, I'm going to assume glitching is allowed.

I think it can be a bit arbitrary at times when a technique is considered "glitching" or not. The community decides. Which is fine.

I saw a speedrun of Super Ghouls and Ghosts that had a clever limitation - "pick up all weapons that show up." If you're familiar with the game, if you pick up a weapon you lose your previous weapon and some weapons are way worse than others.

If a glitch is severe enough it will result in a new category of competition being created. A glitch that saves a second here or there would likely be accepted in the main category. Frequently speedrun games also have categories for how much has been completed in the run, like Any% meaning you just got the end credits up, or 100% meaning you collected all the items on the way or whatever

A notable example of this is Mario 64, where the lowest stars possible category has been renamed multiple times as ways of defeating bowser with fewer and fewer stars are discovered. I believe we are at 0.

And now that the star count is at zero (which humans have pulled off in runs), there is a distinction between "two-key zero-star" and "one-key zero-star", of which the latter is likely only possible in TAS. Video here: https://www.youtube.com/watch?v=j5j6l9ULxmI

"likely only possible in TAS" has this habit of becoming possible.

SethBling turning Super Mario World into Super Flappy Birds on a console comes to mind. https://www.youtube.com/watch?v=hB6eY73sLV0

Exactly why I made the conscious choice to use the word "likely". :)

Conscious choice consensus

It depends on the game and category. Category rules are established by the community and the top priority is runs being fun to watch and play. Usually, any glitches are fair game. However, if any game-breaking glitches are found, the category will usually be split into one that allows and one that disallows the glitch. For example, Portal speedruns were split into "out of bounds" and "inbounds" categories once glitches were discovered that allowed you to beat the game mostly without interacting with the actual game at all.

Same situation for Super Metroid, which seems to be one of the most popular games for speedrunning. The "anything goes" category seems to be "Any% Glitched" [0], which allows for techniques that modify in-game memory values -- the current record is 12m19s. The most popular category, and fun to watch (IMO) is "Any%" [1], which does not allow for such memory tricks and has a current record of 41m18s.

But the latter is still not quite a "glitchless" run, based on the average person's definition of "glitch". For example, I doubt it was intended behavior for Kraid to die after 4 missiles [2] based on the player shooting him during a specific animation. I never played SM much beyond beating it as a kid, but I enjoy its speedruns because besides the variety of techniques and clutch moments required, there's at least one boss who behaves randomly, and requires the player to improvise. Also, in the course of new records being set, there have been a variety of strategies (e.g. different decisions on which path to take) employed, and there still seems to be room to discover more optimal strategies.

[0] https://wiki.supermetroid.run/Any%25_Glitched

[1] https://wiki.supermetroid.run/Any%25

[2] https://youtu.be/ezFcybTLOOE?t=27m29s

> there's at least one boss who behaves randomly

Phantoon, bane of speedrunners.

Not at all. Speed runs rely heavily on glitches in the game to take advantage of.

For many games the community create additional glitchless categories. Other common ones are warpless (don't use intended level skips) and 100% (collect/do everything).

Watch the video — all sorts of glitches

For anyone else wondering:

TAS = Tool-assisted game movies http://tasvideos.org/Game/nes-super-mario-bros.html

Tool Assisted Speedrun. It means you perform input to the game programmatically meaning you can set the input perfectly on a per-frame basis.

what is the point of that

If normal speedruns are like playing speed chess, tool-assisted speedruns are like finding the theoretically-optimal solutions to chess problems (https://en.wikipedia.org/wiki/Chess_problem). Both of them require intense devotion to the game, though the first emphasizes physical ability and execution while the second emphasizes exhaustive thinking.

You could consider bicycling tool assisted running. It's just a different take on competing.

Some people also use TAS's to find, reproduce and understand glitches for use in normal speedrun.

People use them as another tool for finding glitches and exploring the possibility space of the game's systems. They're basically using debugging tools and in some cases even looking at memory allocations etc.

As for full length tool assisted runs, I just find them to be a fascinating new lens to examine these games under. It's cool (depending on who you ask...) to see what the game looks like when another computer program is piloting it perfectly towards the goal of 'finish as fast as possible'

Just for fun, or to have a "theoretical optimum" to compare human speedruns to.

Because NASCAR is fun, even if you are going at super-human speeds. And there is skill in maneuvering a car at super-human speeds.

TAS players play the game frame-by-frame and seek theoretical perfection.

It started out for pure entertainment. Now, they sometimes work with realtime speedrunners to optimize their strategies.

If you're just watching it's arguably more fun to watch, since it theoretically represents absolutely perfect play.

What proves that he's not switching to a TAS from times to times?

That's a technique called splicing, specifically it refers to "splicing" two videos in the middle, in this case one with real input and one faked. It is surprisingly possible (and often easy) to detect splices based on minute details.

Splicing isn't always cheating, it can be used for other reasons.

I can't prove that he isn't, strictly speaking, but Kosmic is a highly respected runner and has the skills to make this record.

If it needs to be verified, a camera on his hands would be enough for experts to be able to see that. Also if the inputs are visualised on screen.

What if he uses prerecorded footage of his hands, created by using some kind of TAS programmed to perform a 4:55 run in human style and then outputs those button presses to a screen. So instead of playing the game, he just presses what he sees on the screen. It would take a lot of work but would arguably be a lot easier than actually accomplishing the real run. Then, the footage is just TAS footage and the camera on his hands is also prerecorded. Then he just has to read out the comments on stream as this charade is going on and make sly responses to it

I guess my point is that, although I trust the veracity of this record, one should not underestimate the potential for cheating.

> So instead of playing the game, he just presses what he sees on the screen. It would take a lot of work but would arguably be a lot easier than actually accomplishing the real run.

First, matching prerecorded input with the frame-precision required would be nearly as difficult as doing the run, if not more so. At this point, he's already trying to produce very specific inputs that require frame-precision; he's not reacting to what's on the screen and adapting on the fly.

Second, some TAS techniques are incredibly difficult for a human to replicate, and there'd be an expectation that he can replicate those tricks at least sometimes.

Third, some of the techniques for detecting this kind of cheating include carefully watching the inputs to see if they're ever just the slightest bit late or early.

Mostly a lack of evidence

Lack of evidence can't really ever prove anything.

I think part of is the nature of it. Why would you get sufficiently into this scene, care enough about this one game to post video after video of you speedrunning it, to then cheat? The challenge for someone doing it manually is to do it manually; the challenge for someone doing it TAS is to do it TAS. Without the challenge there is literally no reason to do it. So while lack of evidence can't really prove anything, there really isn't that much motivation to cheat. If it was a complete nobody claiming they had manually set a record, yeah, that's suspicious, but someone well involved in the scene either is playing the long con for something that they themselves don't care about (and which they can't really revel in since any sort of in person interaction would show they're frauds), which is unlikely, or they're legit.

>Without the challenge there is literally no reason to do it.

What about money? Fame? Aren't these reasons?

And there are countless examples from history of 'the best' in whatever area getting caught at cheating just to stay 'the best'. I'm not really sure what you're getting at with this.

Actually plenty of people who have been deep in the community have cheated. It happens. But given that this run was performed live during a stream where he is reacting to chat it is pretty unlikely that he was doing playbacks of the entire stream.