On September 24th, 2018, Super Mario Bros.
history was made. - I cannot believe it! The game was beaten in 4 minutes and 55.913 seconds in what might be the final second break ever. Let’s see how the run was achieved, how
it could be beaten and if the legendary 4:54 is possible. - I'm just really good at it. Back in February, darbian had the world record
at 4:56.528, when Kosmic beat it with a 4:56.462. Third place runner somewes attempted to beat
that record and on May 11th, set a new personal best of 4:56.495, two frames behind Kosmic. - Yes! My goodness! A PB! How long has it been? He kept playing, and two weeks later, on May
25th somewes set a new world record with a time of 4:56.245. At that point, somewes had one goal in mind:
to be the first person to get a 4:55. He tried and tried for four months, but there
was one problem: 4:55 is hard. It’s really, really hard. Here’s what happened during those four months. Somewes recorded a total of 6022 attempts. 2141 of them got past 1-1. 795 attempts got past 1-2 but only 442 of
them were fast enough to get into World 4 on 4:55 pace. This means World 1 alone is so hard to do
fast enough that he had a 7.3% success rate. Of those 442 runs into World 4, 271 of them
made it to 4-2 and only 81 of them into World 8. In other words, for every 75 attempts, one
single run would reach World 8 on 4:55 pace. 43 runs made it past 8-1, 12 runs made it
past 8-2 and only two runs made it to the final level. Both of those attempts ended at the same place: on this pixel perfect, frame perfect wall jump. The first one missed the pixel and the second one hit the right pixel,
but somewes did not time his jump correctly. And that’s how, over 6000 attempts later,
4:55 still eluded him. Meanwhile, Kosmic had been playing as well but he was only playing to beat somewes’s record,
not for 4:55. He had a few close calls too, but had been
unsuccessful thus far. This brings us to the night of September 24th. After about an hour of attempts Kosmic managed to tie somewes’s world record
with a 4:56.245 of his own. - I don't know! I don't know how fast it is, I hope that turnaround
doesn't kill me. After celebrating his newly acquired tied
record, he decided to do a few attempts at 4:55, just for fun, with no expectation to
even make it to World 8 once. And that’s when this happened. This time, we’re only going to look at the
most significant points in the run, but the full unedited speedrun is available here. I’m also not going to explain most of the
tricks used in the run in full detail, as they have already been thoroughly explained
in my previous video. If you have never seen a
Super Mario Bros. speedrun before, you should watch that video first, because
even though it covers Kosmic’s 4:56.462 from February, all the information is still relevant for 4:55. Here’s a quick rundown of the main things
you need to know. First, the game runs at 60 frames per second,
so one frame is 1/60th of a second. The game will only load the next level every
21 frames, or 0.35 seconds. That amount of time is called a framerule. For the first seven levels, the goal is to
always stay on the fastest framerule, as losing one framerule means losing 0.35 seconds. On the last level, every frame counts because
timing stops when touching the axe, which isn’t influenced by framerules. It is possible to clip into walls, but it requires
a position even more precise than one pixel. Units of distance smaller than a pixel are
called subpixels. These clips allow for tricks like the flagpole
glitch, bullet bill glitch, and also save time in 1-2 and 4-2. They are all extremely difficult to perform and usually require multiple
frame perfect button presses to work. Sometimes, you can trick the game into sending
you to a different room than intended by scrolling the screen past a certain point. This is used in 4-2 and 8-4. You can also perform a walljump by hitting
the wall exactly between two blocks and jumping on the same frame. Finally, Mario accelerates faster in the direction
opposite of which he’s facing, so facing backwards is faster
when accelerating from a stop. That should cover the most important things. For all the details, my previous video got
you covered. - That's a good start. So, Kosmic managed to get the 1-1 flagpole
glitch. He then pulled off the newest addition to
Super Mario Bros. speedruns: the 1-2 pipe clip. Let's go! I hope it's fast! This trick was the final piece of the 4:55
puzzle. To get this, you first need to beat 1-2 without
slowing down at all, which is already pretty hard. This is crucial because any deviation from
the optimal strategy will change Mario’s subpixel position, which in turn can make
the clip harder or impossible. Then, you need to land right here, release
B to slow down to walking speed, and do a 1-frame jump. To finish up, you need to go
far enough to the right to load the warp zone correctly
and go to World 4. If you don’t, you will end up in World 36,
or Minus World. This jump to the pipe is particularly important. Because Mario accelerates faster backwards, Kosmic wants to be facing to the right
while holding left. By jumping on the same frame he lands,
Mario doesn’t have time to turn around and so he keeps facing to the right. This extra acceleration speed is more often
than not the final little push runners need to make it to the pipe in time
to save the framerule. There are two main reasons why this trick
is the single hardest in the entire run. First, because you need to
tap A for 1/60th of a second. Second, because you need to do the entire
thing within a handful of frames from perfection. On top of that, because of the detour
to load the warp zone, this trick only saves one framerule at best. This is why runners never included it until
it was the last possible way to save any time. - Free. Here, Kosmic pulls off the flagpole glitch
again and enters 4-2. - This is the 4:55. Tell your friends. He did the same wall clip as usual, but this
time, it was a bit slow. But thanks to really precise movement
in the rest of the level, he still managed to save the framerule
and stay on 4:55 pace entering World 8. He got the good judges in 8-1, so he only
had to hit the star block to avoid losing time. At this point, he was already excited about
the prospect of getting a 4:55. - Am I about to go down as the most clutch
gamer alive? But then, he got the bullet bill glitch at
the end of 8-2. - This is insane! This time, he avoided the lag frame you can
get while getting Bullet Bill Glitch. At this point, the run was getting really serious, because Bullet Bill Glitch is a major trick
where a lot of attempts fail. - It's too slow! In my previous video, I explained that
one of the major possible time saves was in 8-3. Because the setup for flagpole glitch is
harder in 8-3 than it is in 1-1 and 4-1, he decided to skip that trick back then. However, to get 4:55, he had to do it. Thankfully, HappyLee discovered an easier
way to do the 8-3 flagpole glitch setup than the one I talked about last time. Instead of starting the level by walking and
pressing B on a specific frame, you can start by holding right, A and B at the same time. This makes Mario jump
at the beginning of the stage and slows him down just enough to get 242 on the timer
and avoid fireworks that would cost time. Once that’s done, you also need to
release right while you’re in mid-air for 1 to 3 frames at some point during the stage. How long you released right affects how you
will need to perform flagpole glitch. If you released it for one frame, you will
need to land on the first pixel of this block while not holding B. On the same frame that
you land, start holding B. Exactly two frames later, jump. This lines you up perfectly to clip inside
of the flagpole base. Now, if you released right for 3 frames, you
need to do this instead: land on the second pixel of the block instead of the first. After you start holding B, jump only one frame
later, not two. Luckily, if you released right for two frames,
you can do either of those setups and it will work. This means that runners have to be aware of
how long they released right and adjust their play accordingly. In this run, Kosmic released right for two
frames, which means that he could do either one of the two setups and still get flagpole
glitch. At this point, Kosmic was shocked. - There's no way that this is happening right
now! Based on somewes’s attempt count, the likelihood
of reaching 8-4 on 4:55 pace after a successful 1-2 pipe clip is in the ballpark of 1 in 400,
and Kosmic had just managed to do it on his very first try. But he still had a few tricks in 8-4 ahead
of him. In the second room, he lost 3 frames on the
walljump. The reason can’t be pinpointed to anything
specifically, but his pipe entry could have been faster. In this room, he tried a different strategy. Instead of stopping on the exact pixel where
the game changes the pipe’s destination to the water room, he decided to scroll the
screen 6 pixels further. This allows him to accelerate enough to reach
running speed and enter the pipe faster. This strategy doesn’t save any time in most
cases, but it is a lot safer. If you stop a bit earlier than intended, you
will simply not be able to reach running speed and lose a couple of frames, instead of ruining
the entire run by going into the wrong room. It is also pretty lenient, as you can go way
farther and only lose a few frames. The main risk with this strategy is to jump
to the pipe too late and bump into it. Kosmic was scared of doing just that, so he
jumped early, which did not let him reach the running speed he wanted, and cost him
7 frames. Fortunately, though, he knew it was fast enough
to get 4:55 anyway. - Whatever, it's slow, get 4:55. - Please finish. - I cannot believe it! - How? That was my first run ever! Kosmic was lucky enough to get a pattern where
Bowser jumps forwards, which makes it trivial to go past him. And just like that, he had beaten the game
in 4 minutes and 55.913 seconds. Kosmic’s speedrun is perhaps the biggest
speedrunning accomplishment of the year. The previous world record holder, somewes,
clocked in 6000 attempts at it since getting the world record 4 months back. A time like this could have taken Kosmic 20000
tries, yet he got it on his 9th attempt, and the first attempt to successfully clip into
the pipe in 1-2. However, as amazing as 4:55 is, Kosmic’s
speedrun is still not perfect. It could be beaten by performing 8-4 better. And that’s exactly what somewes did. - Alright! Let's do it! Gotta get that fast accel! By the way, I have no idea why I went under
that hammer. I've never done that in a run or practice,
I don't think, on this framerule. - Uh oh. - 4:55! There it is! Oh my gosh, is it record? As I was editing this video, somewes took
back the world record with a 4:55.796, after a mere 132 additional attempts, including
a new personal best along the way. But even though he saved 7 frames over Kosmic,
somewes did not play 8-4 perfectly either. Compared to a perfect 8-4 with no fast accelerations,
he lost 3 frames in the walljump room, 3 frames in the turnaround room, although he gained
4 frames back with a fast acceleration, and 2 frames in the water room for a total of
4 frames. But if we take the fastest 8-4 that has ever
been done in practice, which is a time recently set by GTAce, that would put his time at 4:55.563. GTAce used a fast acceleration in three different
rooms, which enable him to save 10 frames over a perfect 8-4 without fast acceleration. Unfortunately, it is very unlikely that someone
would ever pull off this fast acceleration perfectly in a full game run for one simple
reason. When practicing, runners use a savestate to
allow them to restart the level at will. This lets them have a very quick way to retry
this fast acceleration, and also to use the savestate to their advantage. Fast acceleration requires very quick inputs
immediately when gaining control of Mario. In a typical setting, this comes after this
long black screen, with little time to react and time your button presses correctly. However, using a savestate, you can make the
inputs a lot easier to time, and, combined with the ability to do dozens of tries in
a short period of time, makes this fast acceleration orders of magnitude easier in practice than
it is in a real run. This fast acceleration can save up to five
frames, but someone attempting it in a real run would most likely save a couple of frames
at best. As a result, the actual human limit is estimated
at somewhere between 4:55.613 and 4:55.646. There is one more thing that could be done
to lower the record even further. On March 4th, 2018, StuckInAPlate managed
to complete 8-1 one framerule faster than is currently done in speedruns. He did it by doing a fast acceleration at
the start of the level, and performing flagpole glitch. If you remember my previous video, flagpole
glitch only saves 15 frames. This is true with the regular setup that involves
releasing B and jumping from the stairs. However, there is a way to do it faster. If you keep holding B and right for the whole
duration of the setup, you can actually save 19 frames with the trick. Of course, there is a catch. The way to do it and save 19 frames is significantly
harder than the regular flagpole glitch. The easiest method that has been found so
far requires you to release right for exactly two frames at a specific point in the level. The timing to do that release is just above
one tenth of a second, but releasing right for exactly 1/30th of a second is very difficult. If you did that correctly, you can perform
flagpole glitch by jumping on this exact frame and hold A for exactly three frames. From there, you will clip inside the block. Sometimes, you can jump one frame late and
still get flagpole glitch. In this case, though, you can’t afford to
do that, because it would make you miss the framerule. As a result, the final jump into the flagpole
is also frame perfect. Now, framerules are 21 frames, and, again,
like I mentioned in my previous video, you normally finish this level with 0 frames to
spare. This means that there are two more frames
that need to be saved in order to save a framerule. This comes from doing a fast acceleration
at the beginning of the stage On flat ground, doing a fast acceleration
will save two frames, but in order to save that time, it needs to be absolutely frame
perfect. On the exact moment you can control Mario,
you need to do this: hold B and left for one frame, hold B and nothing else for one frame,
hold B, right and A for one frame, and then hold B and right on the next frame. This means that you need to be able to frame
perfectly switch from left to right on the D-pad with exactly one frame of no input inbetween,
and you also need to tap A frame perfectly and to tap it for exactly one frame. StuckInAPlate was using a romhack of the game
that allowed him to have extra information about the game’s hidden values, specifically
Mario’s subpixel position. This extra information doesn’t affect the
game itself, but allowed him to have a better idea of how well he was doing and let him
know in advance if he had a shot at saving the framerule. Obviously, using a romhack for a speedrun
is against the rules, so it is only used for practicing. A real speedrun would use the original cartridge,
which means that this information would not be available to the runner. Succeeding or failing this trick would almost
seem like random chance, because you would have no feedback on what exactly went wrong. Again, just like the best 8-4 ever done in
practice, this can be performed by quickly recording dozens of tries and timing the inputs
correctly. This trick has yet to be performed under normal
circumstances, which is coming in from 4-2 with the regular 3-second wait and no cue
to time the fast acceleration correctly. This is why it is currently still considered
not to be possible, or at the very least, not with a success rate better than 1/1000. If this trick was somehow implemented in the
speedrun, it would save 0.35 seconds, and the best possible time would go down to 4:55.214. So, is this the end for Super Mario Bros.? Is 4:55 the ultimate speedrun? Is 4:54 really impossible? Well, the truth is, 4:54 is actually possible. Using an emulator with slowdown, savestates,
memory watch and plenty of other tools, it is possible to create a sequence of button
presses to beat the game faster than what humans can do with their hands and a controller. This is called tool-assisted speedrunning,
or TAS for short, and it is a completely separate discipline from regular speedruns. Tool-assisted speedrunners deconstruct the
game and try to solve the puzzle of finding the most optimal way to beat it. And the best solution ever found is by a man
named HappyLee, with a time of 4 minutes and 54.032 seconds. Nobody knows for sure if it is the perfect
speedrun, but in the past 9 years, only one frame of time save was found, and that was
in 2010, so this is the closest we can hope for. So, where does this time difference between
the tool-assisted speedrun and the real-time world record come from? We’re gonna look at it level by level, but
first, there is one detail about HappyLee’s TAS that needs to be cleared up. At certain points, it uses both left and right
at the same time to accelerate faster. As far as the console is concerned, those
are valid inputs and they can be processed, so it is fair game for a tool-assisted speedrun. However, it’s impossible to actually press
both left and right at once without breaking the D-pad, modifying the controller or using
a third-party controller. For this reason, pressing left and right at
the same time is banned from real-time speedruns as it is considered a form of hardware manipulation. To address this issue, HappyLee made a separate
TAS that follows the real-time rules for controller inputs, and never presses both left and right
at once. This TAS enters 8-4 on the same framerule
as the regular TAS, but loses some time in 8-4 itself. It completes the game in 4:54.282. This is what we’re going to compare against
because comparing against the 4:54.032 would be unfair. In 1-1, the TAS and the real-time run end
the level on the same framerule, so they are tied. Kosmic and somewes barely make it in time,
and the TAS has plenty of time to spare. In fact, if anyone found a way to save just
one more frame in this level, it would cut 0.35 seconds off the TAS that uses left and
right at the same time. That would bring its time down to 4:53.68. In 1-2, the TAS and the real-time run are
also tied, thanks to the pipe clip that made 4:55 possible. There is no more time to save on this level,
unless something major is found. In 4-1, it’s the same thing again. The TAS can save a handful of frames over
Kosmic and somewes, but it’s not enough to make the next framerule. 4-2, however, is not so simple. Here, the TAS does complete the level fast
enough to save one framerule over the current strategy. In fact, 4-2 is another level where the TAS
that uses left and right at the same time is one frame away from saving another framerule. The TAS method is not possible to perform
in real time for many reasons, but it boils down to the fact that there are too many button
presses happening too fast. For example, when clipping into this wall,
and when turning around in the warpzone, the TAS uses many 1-frame button presses in a
way that a human simply can’t replicate. On September 24th, earlier on the same day
that Kosmic got the first 4:55, HappyLee posted a video showing a theoretically humanly possible
method of finishing 4-2 fast enough to save the framerule and tie the TAS in that level. As of right now, this has yet to be executed
in real time. Some runners have given it a try, but ultimately,
it’s just too hard. One of the main issues is the entire combination
of inputs required to clip in the wall, including a frame perfect one-frame right release, which
is just as ridiculously hard as it sounds. The other thing it includes is a way to stop
on the pipe faster, but again, the combination of button presses is very hard to pull off. 8-1 does save a framerule over Kosmic and
somewes with the flagpole glitch we talked about earlier. This is the least unlikely framerule to be
saved in the future, but that one alone cannot be enough to reach 4:54. 8-2 is the only other level where the TAS
saves a framerule over the real-time record. The main problem with the 8-2 framerule is
that, just like 8-1, there are no frames to spare. Again, the TAS uses a fast acceleration at
the beginning of the stage. After a ridiculous jump on this pipe, the
cannon fires the bullet at the last possible time to make this framerule possible. This also means that the bullet bill glitch
needs to be executed even more precisely than usual, because it needs to be done on the
first possible frame it can be performed on. As a result, 8-2 is considered to be the most
unlikely level where humans could match the TAS. It can be ruled out from the possible ways
to get 4:54. In 8-3, both the TAS and the real-time runs
have to wait for the timer to reach 242, so they complete the level at the same time. Reaching the flagpole with 244 on the timer
is impossible by one frame, unless you use left plus right. However, flagpole glitch forces you to touch
the flagpole six frames later than you normally would, which makes 244 on the timer impossible,
even with left plus right. Getting 244 on the timer and getting 242 with
flagpole glitch both finish the level on the same framerule. In the first room of 8-4, HappyLee’s TAS
manages to perform a fast acceleration that saves 6 frames over somewes’s 4:55.796. In the second room, a combination of another
fast acceleration, and a perfect walljump, allows HappyLee to cut another 7 frames. In the third room, one fast acceleration,
a perfect turnaround, and a second fast acceleration save 13 frames over somewes. In the water room, slightly imperfect play
cost somewes another 2 frame(s), and in the last room, one final fast acceleration saves
the TAS another 2 frames for a total of 28 frames. 14 of those frames are actual time lost by
somewes, and the other 14 frames come from the TAS playing beyond human skill. In conclusion, there are two ways that 4:54
could be done. The first one is this: FPG in 1-1, pipe clip
in 1-2, FPG in 4-1, then do the faster 4-2 method that has yet to be performed even once
in practice, and also do the flagpole glitch in 8-1 with fast acceleration to save a second
framerule, then do BBG in 8-2, FPG in 8-3, and then to play 8-4 better than anyone has
ever played in a world record run by 3 frames. That would be enough to get a 4:54.998. The second one would be similar: do FPG in
1-1, pipe clip in 1-2, FPG in 4-1, do the current method of doing 4-2 with the wall
clip, then do the 8-1 flagpole glitch, do BBG in 8-2, FPG in 8-3, and enter 8-4 only
one framerule ahead of the current record. Then, you would need to play 8-4 an incredible
24 frames faster than anyone has ever done in a world record run, or 14 frames faster
than the best 8-4 ever done even in practice, only one frame off the tool-assisted speedrun. Needless to say, the second one is impossible,
because nobody can play 8-4 with the kind of precision seen in a TAS. This means that the only possible way 4:54
could be done is this one. But considering 8-1 being at best a one in
a thousand chance, and 4-2 having never been done once even from a savestate, it is very
safe to assume that with the way things are right now, 4:54 is humanly impossible. Here’s a summary of the limits of optimization. At 4:55.913, that’s Kosmic’s first 4:55. At 4:55.796, somewes’s world record. Somewhere around 4:55.6 is the point past
which the world record might never be improved. At 4:55.213 is the sum of the best level times
ever performed in any situation by a human, including in practice with a savestate at
the start of the level. At 4:54.798, that’s the sum of all the tricks
that could theoretically be performed by a human, although 4-2 and 8-4 have never been
actually completed this fast. At 4:54.282, that’s the fastest possible
time that would respect the rules by not pressing left and right at the same time. And finally, at 4:54.032 lies the theoretical
limit, the fastest possible way that the game could be beaten under any circumstance. Three years ago, 4:56 was a pipe dream, and
4:55 wasn’t even a dream at all. For now, it looks like 4:55 might be the final
frontier. But the beauty of speedrunning is that we
never know what the future holds. Maybe someone will find a way to push the
limits even further. Maybe, one day, someone will achieve the impossible
and reach the legendary 4:54. No matter how hard or unlikely, if it’s
possible, it will be done.
Bismuth voted #1 highest quality speedrun content producer by most clutch gamer alive
Visually presenting time difference between runs like that is so clean, amazing video.
" No matter how hard or unlikely, if it's possible it will be done."
30 minutes of a SMB1 video is the best way to begin a morning. Keep making 'em Bismuth!
Mornin'
Seeing it again. Can't believe i am about to spend a whole hour with this video.
I FUCKING LOVE THESE VIDEOS
Bismuth and Summoning Salt! are amaziiiiing (also averagetrey "history of" videos)
kosmicZ someParty
So much effort has been put into this video! It's crazy how far the SMB speedruns have come honestly. I wait for the day the perfect run is finally achieved.
Amazing videos man