I’d argue that this already happened like 5 years ago when Smash Bot dropped and destroyed everyone. It didn’t have a reaction delay, so it would just powershield or shine everything and consistently 4 stock top players. That’s basically what happens when a top chess player plays against Stockfish. However, Smashbot wasn’t very useful for players because it was so good that you couldn’t get realistic practice (like any good engine today) It’s more like when players started using the first major neural network chess engine Leela to find more “human” ideas or continuations
@@knowk5737 that’s true, though I didn’t say Smash Bot was an AI. Deep Blue and all chess engines until AlphaZero (2017) and Leela (2019) also weren’t AI
@@Mecinimi I get your point, but the thing about early chess engines is they didn't do things a human *couldn't* do. Unless we're talking about chess engines playing blitz and just having instant speed. You could still see "Ah, this is how I'm supposed to counter the X Sicilian variation". Smashbot wasn't bad practice because it was "too good", it was bad practice because it's literally impossible for a human to play like that. Nothing you learn from fighting it is even slightly useful, because it's playing a totally different game than we are.
I'd be really curious what human limitations have been put on the AI. Does it have a human reaction time (input of screen is delayed by reaction time or whatever?). What is even the interface that the AI is reading? Can the AI discern indiscernible animations as soon as they start (or after its reaction time), this would be un-humanlike. Can the ai consistently line up exact ledge cancels etc.? Not saying these additions would be beneficial but as the ai gets better they could help add a more human like practise partner.
It has, apparently, 15 frames of simulated reaction time. Not much else is known about how it works at the moment, but I think it depends on the specific AI agent being used at the time. (Some can have 12 frames or 18 frames of reaction time instead). I've heard some mixed reports of the AI possibly adapting a little to you as you play it, but there's no direct evidence for this - so it's hard to say. Could possibly be a placebo because you're obviously going to be doing and timing your moves slightly differently every time, which will influence how the AI responds to you. Even if you're technically doing the same thing over and over. The creator's hardware that the AI is running on is also admitted by the creator themselves to be "limited", so having the bot constantly learning and adapting seems like a bit of a stretch given that fact.
I want to clarify a little about the other answer, especially since I've talked with x_pilot a few times and asked questions. All Philip agents are currently 18 frame delay. It was originally 12 and then 15 IIRC, but those agents are gone now. Philip sees, at the core, character positions and what move frame they are on (IE Jump(3), Jab(5), whatever). Nothing is visual. I am likely forgetting some small things it sees (Randall position for one). Yes, a current flaw is that it can not distinguish between humanly-visible and non-visible animation frames. Talked with x_pilot a little about this, and he said he's considering adding a delay to specific animations where this exists. (IE Puff grab that's invisible for like, five frames) I can't remember exactly how its input works. At least during training, in order to learn it will have X% chance to input something random, as that's how it finds new strategies and such. I do not remember if there is some amount of "fuzzing" to the inputs to make them imperfect once it's actually playing, or if it retains the X% chance to "mess up" an input. Since the other person mentioned adaptation, I am fairly sure it is not possible right now. Certainly not with the core way that the AI trains to get better. When an AI trains, it is doing it over a LOT more games than a human is. Despite the AI likely getting better than us at some point, it requires exponentially more training time than we do. It's actually a good thing that it can't read though, otherwise it would just read the hell out of us by tracking every single pattern we have. And Daigo for example mentioned that relying on reads is bad anyway, that a solid gameplan and execution is just better. And as for mixups, I'm not sure if Philip can do anything except the "best" option. This is fixed in language models by using a "temperature" to grab an option close to the best, but not the best, output. Idk if Philip can do this.
@@FortWhenTeaThyme melee is a mixed strategy game (look it up), so there is not always a 'best' option and some positions simply require some sort of 'read' to make the most out of. Granted, some options should be chosen 'more often' than others (that's how a mixed strategy works). It would be interesting to know if the bot is always picking the option with the highest weight, as that would lead to a flawed & exploitable playstyle. I guess that seems unlikely though?
say it back phillip
🤣
8:24 "I can read Phillips habits! You're an AI, but you have a soul." The download is complete
Actually funny that melee is going through the same process as chess in the 90's
I’d argue that this already happened like 5 years ago when Smash Bot dropped and destroyed everyone. It didn’t have a reaction delay, so it would just powershield or shine everything and consistently 4 stock top players. That’s basically what happens when a top chess player plays against Stockfish. However, Smashbot wasn’t very useful for players because it was so good that you couldn’t get realistic practice (like any good engine today)
It’s more like when players started using the first major neural network chess engine Leela to find more “human” ideas or continuations
@Mecinimi you're right on everything. Those comparisons are actually more accurate, well done man
@@Mecinimismash bot wasnt really AI
@@knowk5737 that’s true, though I didn’t say Smash Bot was an AI. Deep Blue and all chess engines until AlphaZero (2017) and Leela (2019) also weren’t AI
@@Mecinimi I get your point, but the thing about early chess engines is they didn't do things a human *couldn't* do. Unless we're talking about chess engines playing blitz and just having instant speed. You could still see "Ah, this is how I'm supposed to counter the X Sicilian variation". Smashbot wasn't bad practice because it was "too good", it was bad practice because it's literally impossible for a human to play like that. Nothing you learn from fighting it is even slightly useful, because it's playing a totally different game than we are.
The combo starting at 16:46 was insane
This is such a sick project
Always interesting to see pros play this bot - appreciate you posting these so consistently!
Thank you! We can't lose these sick moments.
Calling Phillip Cody is probably the biggest compliment Zain could give
"With more swag" don't forget.
16:46 was nuts
zain’s always got good shit bumpin for tunes
Phillip is my favorite fox
I know I'm a Marth main cause I said "boom" at 12:35 half a second before Zain did :D
8:17 LMAO
I'd be really curious what human limitations have been put on the AI. Does it have a human reaction time (input of screen is delayed by reaction time or whatever?). What is even the interface that the AI is reading? Can the AI discern indiscernible animations as soon as they start (or after its reaction time), this would be un-humanlike. Can the ai consistently line up exact ledge cancels etc.?
Not saying these additions would be beneficial but as the ai gets better they could help add a more human like practise partner.
It has, apparently, 15 frames of simulated reaction time. Not much else is known about how it works at the moment, but I think it depends on the specific AI agent being used at the time. (Some can have 12 frames or 18 frames of reaction time instead).
I've heard some mixed reports of the AI possibly adapting a little to you as you play it, but there's no direct evidence for this - so it's hard to say. Could possibly be a placebo because you're obviously going to be doing and timing your moves slightly differently every time, which will influence how the AI responds to you. Even if you're technically doing the same thing over and over.
The creator's hardware that the AI is running on is also admitted by the creator themselves to be "limited", so having the bot constantly learning and adapting seems like a bit of a stretch given that fact.
@@KaiokenBlueGT ah ok thanks for that
I want to clarify a little about the other answer, especially since I've talked with x_pilot a few times and asked questions.
All Philip agents are currently 18 frame delay. It was originally 12 and then 15 IIRC, but those agents are gone now.
Philip sees, at the core, character positions and what move frame they are on (IE Jump(3), Jab(5), whatever). Nothing is visual. I am likely forgetting some small things it sees (Randall position for one).
Yes, a current flaw is that it can not distinguish between humanly-visible and non-visible animation frames. Talked with x_pilot a little about this, and he said he's considering adding a delay to specific animations where this exists. (IE Puff grab that's invisible for like, five frames)
I can't remember exactly how its input works. At least during training, in order to learn it will have X% chance to input something random, as that's how it finds new strategies and such. I do not remember if there is some amount of "fuzzing" to the inputs to make them imperfect once it's actually playing, or if it retains the X% chance to "mess up" an input.
Since the other person mentioned adaptation, I am fairly sure it is not possible right now. Certainly not with the core way that the AI trains to get better. When an AI trains, it is doing it over a LOT more games than a human is. Despite the AI likely getting better than us at some point, it requires exponentially more training time than we do. It's actually a good thing that it can't read though, otherwise it would just read the hell out of us by tracking every single pattern we have. And Daigo for example mentioned that relying on reads is bad anyway, that a solid gameplan and execution is just better. And as for mixups, I'm not sure if Philip can do anything except the "best" option. This is fixed in language models by using a "temperature" to grab an option close to the best, but not the best, output. Idk if Philip can do this.
@@FortWhenTeaThyme thanks for the detailed input!
@@FortWhenTeaThyme melee is a mixed strategy game (look it up), so there is not always a 'best' option and some positions simply require some sort of 'read' to make the most out of. Granted, some options should be chosen 'more often' than others (that's how a mixed strategy works). It would be interesting to know if the bot is always picking the option with the highest weight, as that would lead to a flawed & exploitable playstyle. I guess that seems unlikely though?