DESTROYING Donkey Kong with AI (Deep Reinforcement Learning)
Вставка
- Опубліковано 24 лют 2023
- Go to brilliant.org/CodeBullet/ to get a 30-day free trial + the first 200 people will get 20% off their annual subscription.
Second Channel: / @codebulletsdayoff582
Twitter: / code_bullet
Patreon: / codebullet
Discord: / discord
Art created by @Dachi.art / dachi.art
As a wise man once said: “if your code works perfectly first try, it didn’t, you just haven’t noticed what’s broken yet”
I have never heard so much wisdom in a single sentence.
Truly inspiring
And you probably won't until you start testing the thing you needed it for 3 phases down the line, so it'll take you extra long to find where it is. Also with the collective luck of all people involved, it's a single misplaced negative
As a wise man once said: we don't talk about the last video
I once coded for 2 weeks without access to testing it, and had it working perfectly in 5 minutes. In my entire career this was my proudest moment.
That whole "Ground? No." bit was painfully accurate to the game dev experience.
this 100%
No fucking way 345+ game devs are watching this...
Very relatable to my experience with "gamemaker5&6" in the early 00s.
And that program should've made that whole process way easier.... 🤣
Collision in general is stupid hard to do from scratch. Unity? Kinda hard. Unreal? Less hard, I guess. Game maker? Still kinda hard. Even in my own custom made game engine, its way easier but…. still kinda hard :|
Or maybe Im just really bad as a programmer, who knows
@@syntex3664 Because that's the thing with programming, if you somehow made something work the first time, you should probably visit an exorcist.
Now make a Donkey Kong AI that learns the best barrel throwing strategies and pit them against each other for eternity
But that was the first version of Donkey Kong he made, it threw infinity barrels.
@@dreemurrdelm7865 just give the ai a reload time so it can't spam
Code Bullet really is the best inspiration. He doesn't hide or gloss over his mistakes and in so doing shows that coding really is a perpetual state of learning through trial and error.
And then promptly forgetting everything while we binge our next dopamine fix instead of working.
Seriously though, I really appreciate that you include the struggle of being human in your videos. It makes them so much more relatable.
There is nothing more terrifying than when code works the first time immediately. How can we know it really works if we don't spend hours fixing the bugs?
"The code works flawlessly first try" is what happens to programmers when the world turns upside down.
Spoken like a true unit-test averse programmer!
Learn test-driven development. Then you will be surprised if it doesn't work immediately.
@@jimmyhirr5773 believe it or not. Tests are also written with code. Which can also have bugs!
It happened once to me, I tested it like 10 times and it seemed to just work, it was for an assignment and honestly I just took that gift from the gods and it only had to work once more for the tutor to grade me and it did :). Honestly I don't know if I was just lucky it work 11 times in a row or it really actually just really worked first try.
I love how the second AI effectively decides to become a challenge runner by removing the part of it's brain responsible for jumping, just because it could-
Now let's make it do it coinless.
why did you put a -
@@sketchystreet1 Force of habit.
@@Dmobley9901 ok
@@Dmobley9901 youtube commentor moment
Neat detail I noticed that wasn't brought up. At 2:46 you can see the player square falls all the platform a bit early, when it's only halfway off, which causes it to clip into the floor a bit. This is because the "laser" vector that detects if the floor is there is positioned in the middle of the square, and therefore doesn't detect the ground when the square is only halfway off.
This is silently fixed at 2:58, where the square now only falls when it's fully off the platform. As you can see, the lasers are on either side of the square, making sure both have to be off the platform before it falls.
Calling it a NEAT detail is a little confusing but yeah, I also noticed that
probably neat how they realized it @@barni_7762
The jokes and screwing around in the first 4 minutes got you a new sub. The "breath of fresh air" bonus is how you're not afraid to swear and genuinely enjoy what you do even when your code doesn't work.
I do not know. I love codebullet, but this video feels over the top, for me hard to watch, almost cringy, and weaker than usual in the technical part.
Code Bullet is kind of like an A.I himself. In his videos he get slightly more competent over time
Well, we did design them to mimic humans
Well, he is training a neural net...
Its called learning over time
now he just needs to code in some better jokes
This is hilarious 😂
as an Artificial Intelligence Engineer, I can confirm we tickle the balls of the reinforcement learning agents as a reward function. Its one of the best ones out there.
No ML engineer would call himself AI engineer.
@@pw7225 correct, but the name of my university department is "Artificial Intelligence Engineering". i live with people who call themselves prompt engineers, soo who gives a fuck XD
@@pw7225 so no ML engineer understands set theory? :p
@Repent and believe in Jesus Christ i am just thinkin, that there is a religious programmer out there who uses all the resources he can to influence people on youtube comments without making a youtube video. He releases all these bots here expecting people to let go of their beliefs and follow Christianity. If he could give all this effort into philosophy and science, he would become an atheist
@@fizipcfx or an 'alternative scientist'
I like how nobody really forgets about the Enigma machine even after 5 years
Evan is such a nice guy, always keeps his promises, especially Enigma Machine Part 2
@@nubbyboi6523 Yeah, he'll probably make it in the next...millennia.
@@GooberInternet Better than his normal upload schedule 💀
@@nubbyboi6523Wait until it’s 2068
he already made it, but you need to decrypt his message to get the url
13:38 Code Bullet is so good at coding, the AI felt necessary to learn speedrunning!
Every code bullet video is an event that should be treated with utmost respect.
Which is why none existed in 2021.
I can't be the only one getting blackout drunk for these?
this guy gets it
best of code bullet 2021
best youtuber in 2023
Code bullet is probably the inner voice every programmer has while working on projects.
As a software developer by profession,
Yes.
This is precisely how it sounds
The swearing is especially accurate. I curse every single entity imaginable, mostly my computer and myself and anyone who dares to message me at the time.
esp the little part of him that goes 'hey let's use someone else's code for this'
@@sjs9698 “I’ll ask chatGPT” has also become a voice for me
code bullet is the devil on my shoulder
The Jon Bailey cameo is the crossover I didn't know I needed.
I especially like how the barrels rotate the opposite way that they are going. Great stuff.
When you mentioned the jumping, or absolute lack of them doing it in NEAT, I instantly had a vision of you adding points for jumping over barrels and the AI dudes going balls first sailing back and forth over the barrels instead of going to the end. Just racking up massive points for themselves over and over again.
This
@@gabribotha2403 Subscriber place
#selimtune
ah, I see that you have done that... Ima kick you in the balls
Who needs a princess when god gives you testi-tickles
I absolutely love how accurate you show the process of coding, just fail after fail until somehow it works and you’re not really completely sure why but hey it works so it’s all good
Well have seen same yt creator's that show failures more regularly and everyone is doing good but CB is the master of this discipline and his very loyal sub show us that
There’s a Russian anecdote which, in translated form, goes something like this:
« Dad, why does the sun go from east to west every day? »
« East to west? »
« Yes »
« Every day? »
« Yes »
« I don’t know son, but better leave it alone »
It’s a stupid anecdote, but one of my favourites when it comes to programming
There's test-driven development, where you write a test for everything that should and should not happen, e.g. jumpman jumps at least once per game, and over time more tests pass. It's a lot of work, but the benefit is it will catch regressions; as we see here when Jumpman sometimes flies into the air after Code Bullet thought he'd fixed gravity issues.
@@skierpage I'd love to be able to apply TDD to game development, but the number of states and the non-deterministic means of reaching those states most games have means that you end up with very, very low coverage of the code base and/or end up pointlessly testing third party code. I've seen fat too many "tutorials" where the speaker just tests engine code
and then it suddenly doesnt
I really love when you go into the more technical side of things, and still blending in the comedy. Great video!
One of your inputs should have been the rate of barrel spawn since it is something predictable by players after all. This way they would never die in the barrel spawn point. Great video by the way :)
Code bullet explains his coding better than my lecturers does and that’s saying something
he takes your understanding of it seriously.
it's the dick jokes, they help
@@aidanb7782dick jokes explain more than any PowerPoint slide could
@@aidanb7782
Like 22:02 lol
Well he is paid to so makes sense. Oh wait…
Despite your hatred of explaining, you did an amazing job defining and explaining each AI algorithm…while not simultaneously putting me asleep in the process. Great video and always very entertaining! 🔥
I agree, I leanred more from code bullet than high school
This video totally reminds me of the days I’ve spent coding shitty little games back in my teens. I subbed for the nostalgia factor alone, I can relate to having a million bugs with no clue what I’m doing wrong. The trial and error is just a part of the fun!
you're amazing dude, honestly your content is so relaxing to watch cause the ammount of work you have to put on, so we can just watch and relaxed like it was nothing,
He first learnt the whole AI thing and then went for the 2d sprites.
This man is on another level.
@@HRIDOYKHAN-he7ss What? Can you do English?
@@jebclang9403 hey so what if they can't? don't be mean - this is one of the few chanels on yt where the comments aren't lethally toxic.
PPO is basically strapping the AI to an N-dimensional sled and punting it down the N-dimensional hill. You need to choose how big a sled to use, how much momentum to conserve, and even how the hill looks like. And in the end when you think you reached the bottom and looking for an N-dimensional cable car to take you up for more you find out that you are in a shallow N-dimensional depression somewhere half way down the N dimensional hill. Where N is the amount of weights you are using.
That's just gradient descent mostly, not really specific to PPO
I love 150 dimensional sleds on 150 dimensional hills, but it seems like a pretty good anecdote
@@MaxLennon but the gradient descent is used to find the most optimal values for a neutral netwerk
@@maartenbeute6742 yes but PPO is a specific algorithm describing how to leverage a neural network (actually two neural networks), so just describing how to fit the parameters of any model doesn't tell people very much. The method above could also be given as a description for how to do linear regression, for example. Or image classification. Or basically any machine learning task.
@@maartenbeute6742 If you want to see how an explanation of the actual algorithm would look, this link should take you to my other comment where I explained it:
ua-cam.com/video/ovIykchkW5I/v-deo.html&lc=UgxFRqioM8iaM4rJvs54AaABAg
Love the technical descriptions! Please do more :)
The amount of effort that you have kept in a single video....
Dude honestly if you would have kept like more 10 sponsors i would have happily sat and watched all of them till the end
I don't think i can code a even 10% of what you did in a few months
Really amazing video man
Great work...
I love how you somehow have the most inconsistent upload schedule, it takes you hundreds of attempts just to get code right per episode, and yet somehow you also have 2.7 million subscribers. Congratz!
It's fare to say he got these subs when he was more regular
@@heavenstone3503was he ever really regular though? Even when he was posting more often the upload schedule was still pretty erratic
Subscriber place
#selimtune
@Heavenstone long time subscriber, he's never been consistent. He's always had the charm though!
@@heavenstone3503 cough enigma machine part 2
The square refusing to fall off after crossing the ledge unless it "looked down" was some real looney tunes stuff haha
I couldn't believe a coding channel with nearly 3M subscribers has to deal with the exact same shit as I did when I was really into Game Maker in my free teenager time
@@kelleroid are you new to this channel?
@@sebastianredden7577 Nope, just finally saw a moment I could very strongly relate to
Subscriber place
#selimtune
As someone who who had to make a platformer from scratch in Java I understand how hard the first part was
I like even more this video than your previous ones thank to your explanations! Thank you for the high quality content!
Less than a minute in and he's already killing clones, and he's also learned and used a silenced gun too, this is definitely gonna be good
Imagine a Code Bullet Bingo with those stuff lol
@@hayond656 dear God that would be hilarious
13:54 it was short lived
I appreciate your recognizing the last video. I thought for a moment that I somehow missed it, but then you reminded me that I just left partway through. Yeah, that was a mistake. Glad you're back, and glad to hear about another channel!
You have no idea how big my smile is! To some how just stumble onto what will now be my favorite channel?! An amazing feel. An almost AI generated feel, cause it's unreal! ....Carry on!
You mentioned the RNG involved with the barrels, but fun fact about the OG donkey kong. Whether a barrel goes down a ladder or not is based on where the player is and what direction they are facing so you can actually control when the barrels go down ladders, which would have been cool to see how the AI learns to manipulate it
Also a huge pain in the ass to code so who knows if it would have been worth it
Also would have been interesting to see how the AI would coupe with a four way joystick. You can't turn and travel on ladders. But that's beyond the scope CB is at ... For now.
@@trim7911 the AI can use an 8 way joystick if it's Billy Mitchell.
@@alaeriia01 lmao
one of the best things about these is when your AI finds out a trick to your programming and starts exploiting you. funny as hell
Is it exploiting or failure to limit parameters? I mean the Ai just does what you asked it to do, it has no concept of cheating or exploitation.
@@missionpupaexploiting.
@@missionpupaI mean the overwatch 2 AIs will walk through walls sometimes because they have found pixel wide gaps and can spam input movement commands every tick so they can just kinda clip through the walls then shoot you from under the map it’s not really a failure of parameters as much as a failure to block exploits
@@Grebogoborp wait what
@@NonsensicalSpudz flats has a video on it you should look it up it’s hilarious
Absolutely love everything about this channel. Always my first watch on my binges
@21:29 Yes, the small hats are definitely a sign of trouble. Actually a lot of people seem to be noticing this these days.
DONKEY KONG?
You surprise us with the most varied yet entertaining video games to let an AI learn.
Thank you
I didn't expect him to make a donkey kong AI tbh.
I was half expecting your "AI learns to run" videos to be the only two videos in one year.
Good for you to have another one. And us.
Loved this video and how you show each improved AI that you did. A different that I would tried for the reward function is to also incentivize finishing the game faster, I believe that this way it could have stopped the climbing down behavior and probably change a bit the behavior
I just subscribed, I love your commentary. I needed this
As much as CodeBullet struggled, it was still impressive once he started driving over those ramps with relative ease near the end. Couldn't ask for better entertainment. The part where he made an AI was cool too.
CodeBullet still having to watch tutorials on the most simple thing start to become more relatable the further I get in my software studies
Just found your channel. I love the animated commentary, this is amazing.
Music in between methods like NEAT, PPO is epic. Love your videos.
MUM, CODE BULLET JUST DROPPED ANOTHER VIDEOOOO!!! woo
WHOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO OOOOOOOOOOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO
Correct spelling of mum 👍
@@hmm_-. anti lgbtq? Just asking bc your name
I thought you said "noo" lmao
@@Potatoking1 always and always
Codebullet getting scared of something working properly was hilarious.
Bro ur video are so good 21:05 ur humor voice editing alll perfect
The amount of work you put into these videos is insane. Keep it up bro, i hope we will see more from you in the future
22:05 That caught me off guard and now my face hurts from laughing
Hilarious 😂
There's a certain irony in Code Bullet disliking an AI's explanation of AI techniques and preferring his own human explanation
Hello code bullet! I just remembered your channel exists! I used to watch your content a lot, glad to see your still making amazing stuff!
I wished he'd post more but I know just how long it takes to code cuz I was in a coding class and all that but I 100% respect all the work he puts into every video keep up all the good work man :)
I love the extremely 2020’s approach of using raycasting for floor detection. The original Donkey Kong programmer would have lost their mind. 😂
I too lost my mind for a bit
@@WingMaster562 same tbh, I was just sitting here thinking "Wouldn't just a simple AABB collision algorithm be better?" but hey, he makes games on youtube for a living so he probably knows whats best
@@MythicTF2 trust him, he doesn't
@@treeck3724 such an unwieldy way to code, it's fun to watch though
What’s wrong with Ray casting?
I'm quite impressed with the second algorithm (NEAT). That algo figured out to ignore a couple of the inputs, generated a tiny net with only 3 nodes across two hidden layers, and was beating the level without hitting the jump button. Say what you will but I'm more impressed with that than the last algorithm.
Awesome video!
Yeah, NEAT is pretty neat... (*groan*) One problem with it is that it often throws away good solutions because the mutation can often be quite aggressive and throw out long-term positive traits for short-term gains. That said: This is literally how real-world genetic evolution works too. So for this algorithm, it pays #1 to have bigger population sizes, #2 to be careful about your mutation rate, and #3 to give it enough time to run. In all experiments I did with it, it often reaches a plateau where not much happens for many generations, but after a long time, it suddenly starts improving again. I'm pretty sure it has to do with the "species" system it uses: Often, one species is so good at what it does, that even "geniuses" in other species don't have enough impact to lift their otherwise mediocre species above the king of the hill species. This can lead to situations, where species with the potential to become the new best species die out before ever getting the chance to reach their full potential. So, like I said: It works just like natural genetic evolution... A species doesn't automatically survive just because it is fitter. There also have to be the right set of random circumstances to allow them to succeed.
P.S.: One interesting experiment you can do with NEAT is to have to mutation rate ALSO me a trait that can be mutated. That leads to some low-mutation species that keep SLOWLY improving their positive traits and several high-mutation species that keep trying out random stuff. This stops a typical problem in machine learning: The algorithm settling in a "just-good-enough" plateau instead of keeping to improve in the hopes of finding an "even better" plateau further down the line.
@@LutzHerting this can be a video on its own. Or even a blog post series
@@LutzHerting 's also a interesting plan to use a shifting fitness function, or an array of 'em ^^
@beepbop that feels so much like how some gamers I watched approach the original Dark Souls.
They get taught the mechanics, receive a shield and never do anything besides blocking and hitting back. It works, you can play the whole game with that strategy.
But it is most of the time a very insufficient strategy.
Life: gravity pushes you against the surface below you
Code: gravity shuts off if you're touching the ground
The slapstick comedy element of code bullet videos are the best. Watching the red box just keep falling through the ground or ascend to heaven when going up a ladder was so funny.
Next time you want ChatGPT to explain something, ask it to explain it in the style of Code Bullet. I really want to see you react to it trying to imitate you. I checked, and it absolutely knows about you.
That's genius, I need to try it now
Instead of fixing the ladder glitch in PPO algo Code Bullet punishing the AI for falling off is such a good parent move
the rage i felt at 18:32 was unfathomable when i heard that music, i have to hear it in ads and i want to scream but it is the sacrifice i must make to watch the video
I legit love your sense of humor I’m dying every time I watch you since I was younger
Honestly, the whole part of debugging the gameplay and code was the most accurate representation of game development.
This is honestly one of the best, if not the best Code Bullet video there is. It was an absolute blast to watch
My favourite is the Tetris video
Yeah so funny ong
you were just a UA-camr on my recommended but have now become my favorite because of the "let ai that bish" part
I loved hearing the technical details, muoooree!
The chaotic energy in these videos just keeps growing and I'm absolutely on board.
I love watching a successful programming channel that doesn't understand letters with hats on them either just like me! So quirky and relatable
The fact that he has the Honest Trailers epic voice guy read the AI explanation made it so much better.
Was that a cameo or some text to speech generator we should all be aware of? 😄
@@zorglub667 just what I was thinking.... HonestGPT
Took me a second to notice it was AI generated.
@@DanieleGiorgino but how? If there's an epic voice guy AI generator, I must know 😄
@@zorglub667 there are ais where you feed it a few minutes of voice and it can recreate it pretty well
its pretty fun to see so many lil dudes race at the same time. like an online donkey kong. would be fun for a bet game too
This guy narrates like a tenured professor and I am all for it
I always love seeing timelapeses of bug fixing in videos where someone shows themselves coding - Being a java developer myself, I can truly appreciate that :P
I actually really like how you explained everything!
Most of the time when a youtuber makes “AI plays X” or “AI learns to Y”
They don’t explain in more detail how they did it, or they skip over a lot of the crucial information when it comes to it.
Thanks for going the extra mile and actually going into things, even if only a little bit.
we love code bullet
Code Bullet actually shoots the relevant knowledge into your brain. Hence the name.
"So as to how this thing works, -- your guess is as good as mine, I have no f*cking idea but at least it works"
"skip over a lot of crucial information" that precisely describes his explanation. No critique towards him, explaining that shit takes ages and I understand why he doesn't do it (it would also be boring af), but he doesn't explain it well.
@@EvykhaitsSchmurf-tg7gehe doesn’t explain it well or in depth to every kind of code… he explains to us neruron activation monkeys in a way we will understand and at a level deep enough to explain some nugget of inform without boring us.
Bruh. I love it. Your struggle is all of our struggle. You lay it bare and give us human. I love you.
0:36 so it needs the code bullet treatment
I love the chaos of the process. It really gives you a sense of how many times it truly takes programmers to get stuff done 😂
24:15 There's a sentiment I can get behind! Instead of fixing the bugs, teach the AI to avoid them at all costs! Lol, this was an interesting video, thanks for comparing the three, I liked how you mentioned the strengths and weaknesses of them and how it's more like a "pick the best tool for the job" than it is "this one is the best".
Code bullet: *becomes game dev*
Playerbase: *finds bug in game*
Code bullet: “Where’s my ban hammer?”
E
Instead of paying money to fix the hole in the floor teach the baby not to fall in it
Best moment is at 11:32 when one of the first generation Mario dudes becomes Neo and flies off the screen. "I'm out bitches!"
That save in that GTA downhill ramp clip was amazing.
NEAT is pretty neat... (*groan*) One problem with it is that it often throws away good solutions because the mutation can often be quite aggressive and throw out long-term positive traits for short-term gains. That said: This is literally how real-world genetic evolution works too. So for this algorithm, it pays #1 to have bigger population sizes, #2 to be careful about your mutation rate, and #3 to give it enough time to run. In all experiments I did with it, it often reaches a plateau where not much happens for many generations, but after a long time, it suddenly starts improving again. I'm pretty sure it has to do with the "species" system it uses: Often, one species is so good at what it does, that even "geniuses" in other species don't have enough impact to lift their otherwise mediocre species above the king of the hill species. This can lead to situations, where species with the potential to become the new best species die out before ever getting the chance to reach their full potential. So, like I said: It works just like natural genetic evolution... A species doesn't automatically survive just because it is fitter. There also have to be the right set of random circumstances to allow them to succeed.
P.S.: One interesting experiment you can do with NEAT is to have to mutation rate ALSO me a trait that can be mutated. That leads to some low-mutation species that keep SLOWLY improving their positive traits and several high-mutation species that keep trying out random stuff. This stops a typical problem in machine learning: The algorithm settling in a "just-good-enough" plateau instead of keeping to improve in the hopes of finding an "even better" plateau further down the line.
hamburger
It's nice to know that even simulated evolution suffers carcination.
hamburger
@@CalvinArt nft pfp lookin mf
you can implement speciation to fix the "aggressive" problem
As a person who has no idea how to code, I can say that your videos are very entertaining
Love your videos, had quite a few good laughs. Especially when you put the gta v gameplay in the top right 😂
man cool to see someone using unity haven't messed with c# in a while been learning c++
12:45 thank god for the car video, I lost interest 0.0000004 seconds into the timelapse and I really needed it to help with focus
This man has the most elite upload schedule 😂
it took me the same time it took him to upload 15 of his videos (provided he doesn't upload for two months... which is guaranteed) to get a cs deegree
@@thomask2133 provided you actually get the degree 😏
As someone who’s starting to learn ai and ml I must say, you’re the coolest person I’ve seen on youtube. Fun and informative content, S-tier shit
Your commentary is pure refined *gold*
At 9:39, CB states "2 extra nipples" while adding 2 nipples to a nipple-less entity, presumably a human. However, we are aware that all humans have 2 nipples by default. Therefore, we can conclude that CB canonically has 0 nipples.
13:33 The AI finding a glitch and taking advantage of it is freaking awesome 🤣
Even spiffing brit isn't safe from ai
@@Alucard-gt1zf The Genetic Algorithm is a perfectly balanced game with no exploits
You cracked the code on being an entertaining programmer! Honestly funny asf
That GTA dropper clip was some of the best parts of the video! Maybe make your videos tiktok size with GTA underneath for the whole video.
I like the balance between random entertainment and coding/AI a lot more in this video than the last. Always love the uploads!
I feel like I'm being spoiled with so much Code Bullet content in such a small span of time.
This was the most entertaining coding video I've ever watched. Definitely subscribing!
I like your content a lot and watched it with Saima Chrome extension and save c. 12 min. Great Combination. Thanks a lot!
18:22 where's our car??????????? my gen z brain cant function without gtav gameplay covering half of the screen D:
I now see why it takes you so long to upload. Creating a bug every single time you add a new feature is a genuine gift. Never change
Just found you dude, you are my spirit animal I swear 🤣🤣
it's hard to describe the fear of something working the first time when you know it SHOULDN'T HAVE WORKED THE FIRST TIME