AI learns to play 2048 

When I was a kid I didn't even know there was a strategy and thought it was a clicker game where you just swipe as fast as you can randomly.

Do i understand: no Am i entertained: yes

New favorite insult: "You're about as good as randomly pressing the left key"

Tommy looks five moves into the future.. he sees his failures coming and has no choice but to helplessly walk towards them

“Every one has the right to an opinion but not everyone’s opinion is right” Best quote of the decade

"it keeps lower scoring tiles away from the highest score tiles" So the AI enforces aristocracy.

Hearing you switch between calling the ai tommy and tomas when you’re frustrated is like a parent being upset at their child

Code bullet is so brutal that he doesn't creates algorithm for game, He creates game for algorithm

Plot twist: It was his friend actually named Thomas playing all along

Wow. I was just impressed with you recreating the game..

Being named Thomas, I felt quite called out through this

I like how even when "Thomas" finally succeeds, it's still less efficient than any even slightly smart human player, making Tommy sort of pointless, but he always was a problem child and it's the journey that matters, not the destination

"failure is important for learning" so that's why my dad says I'm important

Tomas: "I've looked 5 Moves in the future" Code Bullet: "In what one do we get to 2048?" Tomas: "One"

"Thomas kinda sucks" I hear enough of this normally, I don't need it to happen here too

As somebody who is named Thomas, this ai is basically a perfect recreation of anybody named Thomas. They fail at everything for a while, then somehow get it out of sheer luck.

Code Bullet: “Alright, I think I’ll go to bed.” Thomas: “You’ll go to bed when we’ve reached 2048.”

Ai learns to code Ai

"I CAN GO TO BED" has to be the funniest/saddest thing I have ever heard

I love how much "failure" you actually include in your videos. It's interesting to watch how many types of things can actually go wrong while programming, and hearing what your thought process is to overcome them.

"what's up guys, today's video is in 2048" -code bullet accidentally reveals his origins

3:40. Well, that’s literally the strategy I’ve gone with for the past 2 years...

Plot twist: hell release the second enigma machine video in the year 2048..he's leaving us clues

Your second approach reminds me of the Othello algorithm I wrote for my introductory coding course a few semesters ago and I just wanted to say, I 100% understand watching said algorithm come *so close*

Sorry I wasn’t performing well

I felt so sad for him when he said “The next improvement was at generation 227” that’s gotta be so depressing

I some how keep arriving at your videos late at night when I am searching for vids to watch and I always get so fascinated by you content.

Seeing you use processing is so nostalgic. Processing is fantastic

"Yes, I can go to bed!" the typical software engineer saying after a project XD

AI learns to take over the world -Code Bullet Video 2019

You’re teaching me better than my professors did

I finally beat this game after so many years, I can die happy now

You should've told it that in addition to having the highest possible number in the top left corner is good, it should also prefer having as much free slots as possible and get rid of small tiles as much as possible

When I went down the rabbit hole with this game, my strategy revolved around one simple restiction: choose a direction (in my case, up) and never move in that direction. A consequence of this is that the largest tiles end up on one face, similar to your "top left for highest value block" strategy. I wonder how the AI would do with this restriction in direction. Edit: 2048 became an easy target, if I recall correctly, 8192 was the max I reached before throwing the game in my "NEVER TOUCH AGAIN" category.

4 years later and i am still waiting for that enigma video.... you cant leave me hanging like this :(

coding is just "if" statements - a wise man 2021

That moment at 9:08 where Tommy could have just gone up left left and reached 2048

"Everyone has a right to an opinion, but not every opinion is right" I love it, I'm stealing it for future arguments!

Seems like using NEAT to evolve an evaluation metric for the lookahead approach would be a useful fusion of the two methods.

Hey! I actually made a 2048 ai in college. Takes me back. Mine performed okay, but the algorithm was definitely trash compared to the top tier algorithms. Was satisfying to see it all work though.

Every single AI video: It’s evolving, just backwards

*Make an AI learn to play Windows XP Pinball??* *I WOULD LOVE IT if it makes fancy combos continuously or see what he prefers to do lol*

While keeping the biggest number in the corner and the next biggest right next to it is an important part of the winning strategy, it's also worth noting that the undo function is quite useful for any unlucky spawns. For example, if you are forced to move the biggest number out of the corner and a 2 spawns in the corner, you can undo the move3 and do it again in the hopes of a better spawn. Or another one could be a 2 spawning when a 4 would be better (or vice versa). Or times when the spawn happens on the wrong line and it would mess up all your plans. While a lot of these examples usually won't prevent you from getting to 2048, getting into higher numbers requires knowing when to use the undo function (also worth noting that you can only undo 1 move at a time).

Crazy how much calmer and happier he sounded in his old videos

CB: keep the highest tile in the top left Tommy: how about... no?

My highest tile is 8192. Take that thomas.

i guess, it can take maximum after some "little" time period) (maximum is when all board is filled by the "snake" from 131072 to 4, i`ve did it by myself, and much people too) looking forward to see video with maximum)

Perhaps a better scoring algorithm would reward keeping tiles of the same value adjacent, and also somewhat reward keeping tiles of almost-the-same values adjacent. Also reward empty space, because that's your distance from the loss condition.

I have no clue what youre saying half the time but your channel is amazing

CB: Since the difference between random movements and intelligent movements is very small... Me: Yes! I'm almost intelligent! edit: 1.5k likes :O

You are my favorite RU-vid channel man. Love this shit

Bullet here explains the definition of camping “TOMMY KEEP TOP LEFT”

The first generation is already better than me.

If there was no visuals this would be the title: Man screams at someone (presumably his son) for 11 minutes and 10 seconds

I think the issue with using a MinMax strategy, that is the game states algorithm CB was waltzing around, is that it cannot accurately account for the randomness of the new tiles that spawn in. These new tiles have a random location with a probabilistic value. When 'aiming' for a future state, it is possible that state is unachievable because the tiles needed to get there didn't spawn in the same locations with the same values.

This just popped up, I’m expecting a video in the next two weeks. RU-vid always does this to me for some reason: a video I’ve already watched is re-recommended to me, and less than a week later a new video comes out. Hope I’m right

Code an Ai that can beat another Ai in chess

It's April 21 2019 still no part 2 of enigma machine

The video I been looking for for a WHILE

I still want that second enigma video... Or, to be fair, any video.. Just keep making videos please.

ai learns to play agar.io

1:36 “let’s see what happens when we press the up arrow - aaaand they’re gone” 😂😂😂

It sounds like you implied you were using the minmax algorithm to evaluate the future game states. When evaluating the game state, you could take into consideration the possible locations for the new tile & its value (called substates) and apply a modified minmax algorithm to those substates. I can tell that this is not being done because of the move at 8:25, where 256,128,128 is moved right. The minimum state for that move would have been a 2 being added at the top left corner, permanently (for the forseeable 5 future moves, actually) displacing 512 from being in the top left corner. A left direction in this instance would be the better move. I'm halfway motivated to make a better algorithm for this game myself now

For those curious like me of the BGM in 8:18, its; "25-03: There Is Nothing Left to Say or Do (Short Version)"

He will single handedly create an AI that will destroy the world.

I found your channel from the enigma video, would be great to see some more!

i think the issue may have been in not letting the highest tile move also using tile stacks to favourably rearrange the board also moving large tiles to make spaces causing a tile to spawn that couldnt be merged...

ive never watched your channel on my pc and yet I've been recommended you, I don't mind

Thomas: Is an AI and within 5 tries reaches 2048 Me: can’t even get to 512

AI learns to play threes. Threes is a much more interesting game that was released before 2048. It takes a lot more strategy and just looks better

I think it would be a good idea to add an additional option like if the box with the highest score is on the left corner and that row is not full of 4 boxes then never go right. Also, if that works I think this thinking should apply also for the second row with an exception when it will make an addition from second to first row.

I remember figuring this out and maxing out the game because you couldn't fit any more unique tiles on the screen to add together

**Instant click**

You should have added to Thomas points for not moving the biggest tile. What I learned from playing this game constantly is that you should never move the biggest tile from his corner except there is no other way.

oh god.. 8:22 feels like when you’re on the last mission of a game and a slowed down instrumental version of the main theme plays…

I used that algorithm for a connect four game i made lol the first algorithm ive ever done its actually much more simple to do than it sounds. Most of the complexities comes from ajusting the points that are given for different game stats and stuff to make the ai the best as possible takes alot of small adjusting

I wonder how long it would take thomas to get to the 131072 tile. It would require a large amount of luck with tiles and probably takes trillions or more trials

_Ai learns how to take control of humanity_

In evaluating states, it is not just highest number to the corner that should matter, but a more general "orderliness". Fewer tiles is better than more tiles and tile's nearest neighbors being numerically closet i.e. 2's in line with 2's or 4's is fine, 8's less so, etc. I would love to see you try this strategy, but I see this video is 5 years old!

I think you could add multiple things for improvement: 1. moving the big tile ever out of a corner should be nearly as punished as losing 2. you should also harshly punish not having some line from highest to smallest. With that and 5 steps in the future you should easily get 4096 consistently.

me: *hears over-dramatic royalty-free music* also me: yeah dis the one

Do you *condone* the abuse of subscribe buttons?

Congratulations on 6.9 mill views mate! Nice

Your sense of humor in these is simply amazing! Love it, doesn't matter how successfully or not your creations are, your commentary makes these 10/10. It's pretty cool to see what challenges you run into and what solutions you come up with to them. Interesting to see your different approaches :)

I think there’s an issue with the algorithm. In the actual game, part of the challenge is not knowing whether a 2 or 4 is gonna spawn. This especially matters when the board is almost full. Since the algorithm calculates what the board is gonna look like after 5 moves, then it already knows whether a 4 or 2 is gonna spawn in every one of those moves.

When you realize that all AI models are still technically algorithms... just really obfuscated ones.

I think it went wrong because of the random in tile spawning. Imagine it relying on 1|10 chance to get where it wants....... Imagine a 1|100000 chance that all 5 turns are a 4 value tile.

This was even more hilarious to me because my real name is Thomas but I rarely go by that anymore.

At 9:07 your A.I. was within 5 moves of victory and it made the wrong move. Luckily he got there still anyway a few seconds later

The second example used for game states (5:22) is a very thoughtful game state to be in since there are only 2 moves to get out of the situation

I like to put the biggest tiles in bottom right and work from bottom left to right but this works too

"my man thomas looks 5 moves into the future" This is King Crimsons ability

YOU PROMISED US!

Dude, enigma machine part 2, has it not been enough years!?

Antennas are perfectly recreating facial expression, it's awesome

I clicked the subscribe button twice. I was already subscribed tho

Could you retry the neural network approach, but where the goal is to have a high number + fewer tiles on screen? Like, "success" is determined by the average tile value over time. That should help a lot, I would think.

Hey man I am super impressed by your videos, and I am currently working with AI so it would be awesome if you could show the code and the process you take to reach this level

Actually the early gens did understand a good strat, if you can follow a down, left, down, left, etc. pattern over 60% of the time you can get to the 512 cube and over 20% of the time you can get to 1024. The only bad part is that when you're unable to go down or left you have to go right and hold and then restart the pattern and this actually gives a high likely hood of the next block being put in the bottom left making the strat break. After you get to 1024 the strat ultimately becomes improbable so you actually have to map out your movements!

fusion is just a cheap trick to make weak tiles stronger. *im kinda upset no one has made this*