Skepticism regarding recent video of player AVIZL

[Barneey]

Hi, so a lot of you have probably seen a video of someone named AVIZL on TetrisFriends.
And although I see a lot of praise and admiration, I have yet to see any doubt, which I find disconcerting.
Disclaimer, I'm not saying it's a bot or illegitimate, I just want to discuss the possibility.

Looking at his profile, there are a few relevant replays available, the first one is the Marathon, you don't have to watch all of it to realize that while they are definitely versed in building Perfect Clears, it's nowhere near the level of which they play on in the video (posted below), in terms of speed, efficiency, and pattern variation. Nor do they show any sign of using the most basic finesse moves like wall>tap. The game was completed with 52.32 TPM.

Now let's look at the Ultra replay which was played with 97 TPM, less than 30 days ago.
In the Ultra replay, they show signs of significant overall improvement, the finesse has improved, but still not perfect.

Sprint 5p mode, 6 games played, 105 TPM.

Then there's his Sprint replay.
There is no doubt that completing a Sprint game doing 9 Perfect Clears is an achievement in itself. Doing it in 33 seconds? even greater of an achievement, and in my opinion maybe even unrealistic. Doing it in 33 seconds(182 TPM), on TetrisFriends is incredible considering it consists of 10 Singles, 7 Doubles, 4 Triples, and 1 Tetris. Because as I am sure we all know, the line clear delay is no joke. And last but not least, where did this perfect Finesse come from?
However, it is of course possible that the Marathon replay is outdated and the player has improved vastly since then, because for some reason, TetrisFriends does not display the date of the achieved highscore in that particular mode.

And finally, the video in question:

In the period of 00:04-00:40, AVIZL cleared 40 lines in 36 seconds, not including garbage lines. 166 TPM.

Discuss please, it would only make me happy if we could reach the conclusion that it's legitimate.

Edit: Realized I spelled his name wrong in title, apologies.

[Paul676]

I noticed the same thing with the replays of his Ultra score. I came to the conclusion that Tetris Friends replays are messed up beyond repair and it's probably replaying a game from years ago, as the scores do not tally between the video and the stats from the game. If it were a bot, why would they decide to play their ultra at a slower level? Makes no sense.

If it's a multiplayer bot, it's by far the most impressive bot I've ever seen. Being able to do t-spin doubles when tactically right, downstacking extremely well, having finesse down. That's not easy to make look legit.

[FieryLight]

The Ultra replay on his profile only achieves 36 k points.  It isn't the latest one.  Check the LB to see his actual replay at #9 with 154 TPM.

His Marathon score isn't high enough to make it on the LB but I would assume that it's the same problem there.

[jkwon23]

No one can think of PCs that fast. He's a hacker

[Lucho]

You dont know someone until you 4w them
EDIT: you misspelled AVIZL in the title

[morningpee]

And although I see a lot of praise and admiration, I have yet to see any doubt, which I find disconcerting.
Disclaimer, I'm not saying it's a bot or illegitimate, I just want to discuss the possibility.

This is actually the reason I posted the video in the first place.  I brought up AVIZL in the chat on TF, and there was a lot of skepticism about about whether he is really so capable.  The topic seems worthy of civilized discussion, which you don't really get in a TF chatbox.

[myndzi]

I say legit.

There are two real tells that say 'not a bot' to me. First, it might be easy-ish to program a bot that did only perfect clears, but T-spins are a hell of a lot harder. As well, the variety and timing of deciding when to do one and when to do the other is not something that I've seen implemented yet, and there's a good reason for that: it's hard. I may be being a bit charitable, but it seems to me that anyone with the amount of skill necessary to program a bot that would perform like that would be much more interested in being known for that ability than being known for playing Tetris.

The bigger tell to me is that there is something to be observed in the video that is extremely easy for a bot to keep track of and plan around, but hard for a human: parity.

When you receive one line of garbage, you have an odd number of minos on your field. When this is the case, you cannot perfect clear. You can see clearly in the video that AVIZL attempts to clear down on multiple occasions before he knows whether it's possible to perform a perfect clear or not. A bot would be able to (and be incentivized to) better manage its garbage and stack more optimally for the minimum required effort-to-PC.

I am extremely hesitant to call BS just because we haven't seen someone this skilled at that mechanism before. There are obvious patterns in his play that are familiar to me, and I have spent a good bit of time performing perfect clears. I'd be most inclined to say he's codified a system or series of patterns that allows him to be more consistent. You will note that many of his clears are 6 or 8 lines, it's not strictly 4 lines in a row constantly. A bot would also favor 4-row perfect clears.

I've mentioned this when talking about PC stacking, but you can see how he balances the jagged and flat portions of his stack, works his way inward, and it's no coincidence that he ends so many of them with a T piece.

I'll leave the frame-by-frame piece manipulation and finesse analysis to someone else, but I feel qualified to speak to the practicality and technique involved here.

Edit: the combo-instead-of-flat-clear at 1:00 is a real tell too. That was an on-the-fly, situational, strategic decision.

Edit more: the decision to go T-spins at 2:25 is also fairly arbitrary and un-botlike. In the same kind of situation in the first game, he was going hard at the perfect clears, but this game he changes it up and goes for T-spins instead in similar situations.

This kind of behavior could be programmed into a bot to make it 'seem more human', but it would be detrimental to its actual play unless incorporated correctly, which is a much more difficult task than just 'randomly switching things up'

[morningpee]

I'll leave the frame-by-frame piece manipulation and finesse analysis to someone else, but I feel qualified to speak to the practicality and technique involved here.

If anyone's interested, the raw video is here.

[Barneey]

Thanks for the long response, good stuff Myndzi.

Let's assume it's a bot for a second to simplify things.
[!--quoteo--][div class=\\\'quotetop\\\']QUOTE[/div][div class=\\\'quotemain\\\'][!--quotec--] You will note that many of his clears are 6 or 8 lines, it's not strictly 4 lines in a row constantly. A bot would also favor 4-row perfect clears.[/quote]
I think this is because it's doing only Perfect Clears that do not require soft drop, so when the bot sees that there is no way to do a non-soft drop PC, it starts stacking higher in order to widen the range of possibilities for it to find a PC that is hard droppable.

[myndzi]

Thanks for the long response, good stuff Myndzi.

Let's assume it's a bot for a second to simplify things.

I think this is because it's doing only Perfect Clears that do not require soft drop, so when the bot sees that there is no way to do a non-soft drop PC, it starts stacking higher in order to widen the range of possibilities for it to find a PC that is hard droppable.

I can say from experience that there are plenty of complex perfect clears to be made without soft drop. A bot would be able to see these readily, while a human would not. Some of the perfect clears I saw that I am most proud of are 4-line perfect clears that required excessive middle skimming and/or platforming to complete. He does enough perfect clears in that video that IF he was a bot you would expect to see something at least approaching that given the sample size.

[CaptainPaul]

I don't have any AI programming experience to back me up, but my gut is really nagging me on this one. When I watched the multiplayer video, I was blown away, but I could still see how it might be plausible. The sprint replay video on the other hand, I simply don't believe. He's going incredibly quickly for TF, while playing very very inefficiently. He clears singles and doubles for the sake of getting perfect clears that are worthless in sprint. Maybe he's just messing around and not going for speed, but that seems unlikely.

[myndzi]

I haven't watched the replay(?), but I frequently use 40lines as a practice arena because it starts blank and lets me do whatever I want.

33 seconds seems awfully fast with the line clear distribution written in the OP, but you do have to remember that high scores are just that - HIGH scores. Ideal games. I imagine to get this good at perfect clearing one would have to practice a LOT. If he was using sprint as a practice ground, and we take the vs video as a sample of what he is capable of, the sprint performance doesn't seem so crazy. And 33 seconds is 9 seconds slower than Subversive's sprint -- 9 seconds is quite a lot, and this player seems able to play even-ish with subversive in just T-spin mode(?)

Skepticism is healthy, and it wouldn't be unreasonable to ask for extraordinary proof in this case, as it's an extraordinary feat. Just be wary of judging him as guilty until proven innocent. I think there's enough ambiguity displayed to make it far from an overwhelming case for automation.

[VladtheImpala]

I think you're giving him too much credit. It is a full 7.5 seconds slower than Subversive's time, but it is also only .5 seconds slower than say, the best performance of HAHAHA, the TTO runner up. I refuse to believe that someone no one had ever heard of until now could go as fast as hahaha with that line clear distribution, let alone that number of PC's.

edit: sorry, it's 6.5 seconds. my bad.

[jkwon23]

I think he made a bot that could recognize the patterns that are favorable to pcs. And to explain for the last round with a lot of tsds, his bot probably had another mode that could aim for t spins. Sure, it reduces the numbers of PCs and APM/LPM ratio, but it pays to look less hackerish!

edit: Hacker or non hacker aside, we, the members of HD, should welcome him if he hasn't joined yet.

[perfectclear]

having watched his games, I feel it is legit. Again because of my extensive PC experience, I see what opportunities he is missing that a bot wouldnt miss. He appears to have built his PC knowledge by always trying outwards-in PC attempts (no strict build is used by him), and has really good pattern recognition. While he does see a lot of skimming, he does miss some quicker skim-pcs. Hell of a lot better than me though xD
I agree with myndzi in that because he switches tactics when appropriate and misses some 4-height pcs it is highly unlikely that this is a bot. the most skeptical you could be is that you could say it is a pc-bot that works outwards-inwards, has an incomplete table, and can be switched off to hand off the controls to a highly-skilled player when tactically appropriate. I feel like this is even less likely than that he is just skilled at PCing.