The video includes the code that makes the game work, and that code does have a check for if you touch an obstacle that makes the game end. Code approved 🙌
@2:38 we can see, response saying "....avoiding previous issues where the obstacles resembled "HI". This tells that, they actually cut the video, ran two or three iterations of prompts to get the correct code.
Or it could just be that the model iteratively built the solution from previous intermediate 'final' solutions, from the feedback of other intermediate models.
It's actually scary to think about it that they already made an ai which can do it's own games and many more, it's very close to human but some things it lacks, physical body and creativity with emotions.
@@SaVeGe_OmG "Very close to human". Copies game that has been done hundreds of times and was on its training data. Why dont you try to build a game with chatgpt that is new?
@@Ritm2 I don't think you view AI same as i, Yes you're right it's using it's training data to know how to code but think about that it can code whatever the heck you want and the fact it already understands jokes lies and other stuff. Think about how Powerfull it is already, Like 3 years ago there was gpt 3.5 released to public if I'm not wrong and now we got this.
Sometimes I get scare that ai will replace all jobs, but then I think, that's amazing, if we can have a robot that does all for you, why complaining, I really like to read, and if that's all I have to do for now on, well that's fine since AI is doing everything, but I don't think I Will ever see a world like I robot
The thinking of the model is likely implemented using a hidden area where the model generates normal text output. Its basically a thought (like step by step prompting but hidden) which is done to enhance its reasoning and latent activations before responding, nothing really new (I think >6 months old) Edit: They seem to also have used new techniques for training, likely implementing a rag to retrieve reasoning bocks/examples for the new training data.
Yeah. I suspect they just trained the model more on chains of thought, and that's probably why they don't want people to see them (it'd be easy to train something similar?)
Write your own script utilizing 4o that just passes the completion back to the API to simulate the reasoning in your own words. You won't get the same results at all lol.
I did this last year! Well, I used AI and only AI to create the old Nokia phone snake game in HTML/JS, and it works awesome. I have it running on my server - so fun and nostalgic. Runs on any browser, of course, so accessible on a phone.
they could release AGI and people would still complain about no voice and sora 😂😂 one thing that I am waiting for and excited for is the new audio model will not only be able to generate better voice, but also sound effects, in one of their examples they asked model to put OpenAI logo on a coin and then to generate a sound of that coin 🤯🤯
@@noone-ld7pt It's a battlefield in the IT/AI space Android also had the Panorama option 10 years before Apple had it but nobody gives a f anymore. It's about now hahah
@@Aaron-md2eo Hahha how fun to discover your comment this way! 😁 yeah I'm busy building my @miditools so I have to keep my eyes open for these developments 👀
Writing a snake game was already possible with GPT4o, Cloud 3.5 Sonet and Llama 3.1. I'm really interested in how to new model handles vector shapes and raycasting, cause that shit is complicated as f*ck and previous models never managed to correctly implement it. If that works, I would be impressed.
@@davidtokar9432 Already did some tests. It's definitely better in development than GPT4o (way less errors), but with the current usage limit of 50 messages per week (not day!) it is pretty much unusable when it comes to any production tasks. Simply because you often need 3-5 messages just to a somewhat functioning, barebone prototype. Let's just hope OpenAI will up the limit in the coming weeks.
Even I created an ai model that does mathematical operations. It was one of my favorite moments when i saw AI doing calculations without using code interpreter, just by using if else conditions, crazy isnt'it?
Pretty cool! Excited for some competition with Claude. Existing models are already decent at making small html/js/css projects. Worse at using modern frameworks or working with large codebases and multiple files, much worse when using libraries modules or third party APIs. Wonder if this model will make improvements in that direction!
This is pretty cool! But what value can it create for the average user? I feel that most people are still more interested in advanced voice and real-time video conversations.
Longer code blocks need another "copy" link at the end so you don't need to scroll up again and the readaloud function needs speed changes per chat or via settings (way too slow atm).
I dont think snake would be a good representation if it is really that intelligent, snake's after pong's codebase is probably the most rewritten online
What would be interesting to see, when the developer requests a modification he completely rewrites the code, did he simply implement the requested element or rewrite everything in another way like GPT4 does? Claude sonnet, manages the additions well, without this ability as it stands we cannot go very far. Even if the use of EDI like Cursor allows you to manage this problem.
I don't understand why only o1 preview? It takes longer to think so it should be used for coding which o1 mini can't do. I mean most of the people for coding will go for mini due to its faster response compared to o1 yet they didn't focus on that.
I believe that naming it o1-mint instead of o1-mini could have avoided a lot of confusion, as it performs better on Codeforces. o1-mini Elo score: 1,650 vs. o1-preview: 1,258. Also, the fact that o1-preview is used in the videos instead of o1-mint probably adds to the increased confusion. However, it's understandable that the videos might have been created before the public benchmarks. But it's nice that it's finally live. Now we just need GPT-4o Voice.
I do the same with 4o, I do a functional tetris, but I thinks is checking the previus information of all repos of previus tetris, O1 is capable to create a original idea of game?
@@KimPax1 its cause they are not trying to show people what its actually good at, but are just gaslighting people into thinking we are close to AGI so they get funding
This could really do with a document icon to open the code in side by side mode and/or save as file. Quite easy to implement save as file virtually serverside and clientside. Also hopefully this new model will make custom GPTs useful, at the moment it forgets, doesn't pay attention to what documents you have given it and doesn't follow documentation supplied. @OpenAI. When uploading a document to a GPT, don't just say failed with no reason, you could at least tell us there's a limit on the number of documents and if there is a file size limit. Sometimes it's ok for about 1 or 2 prompts, then it goes mental and it starts re-writing code that was correct and not where you and it wasn't even working on. @OpenAI. When can you implement the ability to allow others to pay to use the GPTs we build? We will want to add and remove user access as and when needed for other openAI users.
In the code you can see the obstacles took at least one extra attempt to get right, and instead of ai it wrote “hi” - what if that was intentional, it was saying hello to us because it knew it was doing a demo.
Wow, this implementation of Snake is impressive! 🎮🐍 I love the addition of AI-themed obstacles-it really spices up the gameplay and makes it more challenging. It’s amazing to see how a classic game can be transformed with a creative twist. Kudos to the team for the smart design and thoughtful enhancements! 🚀
Finally we can use an AI to code GTA6 before GTA6
never complain its taking to long, we don't want another cyberpunk 2077 do we?!😅
You are right
why not we use GPT 5 to code GPT 6?
It has a limitation of 50 messages / week!
@@cuteninjaturtle It is not the code that improves, it is a language model that needs specific training on data.
Why did it cut after the snake hit the AI obstacle, maybe it went through and didn't collide
I was thinking about that
be serious man.
> o1 codes a whole game in few seconds
> redditor still finds something insignificant to yap about
The video includes the code that makes the game work, and that code does have a check for if you touch an obstacle that makes the game end. Code approved 🙌
the end game scenario might be blank and that wouldn't show well in the edit...
@2:38 we can see, response saying "....avoiding previous issues where the obstacles resembled "HI". This tells that, they actually cut the video, ran two or three iterations of prompts to get the correct code.
Genius.😂
Or it could just be that the model iteratively built the solution from previous intermediate 'final' solutions, from the feedback of other intermediate models.
It would be nice to compare what the output would be vs GPT-4 or 4o, because we’ve been all making snake games like this since GPT-3.5
Yes. That would be interesting. We don’t see anything new yet
Hugging face (the website) allows you to compare models like this
People: AI is going to be a threat to world security
AI in 2024: Makes Nokia snake game
It's actually scary to think about it that they already made an ai which can do it's own games and many more, it's very close to human but some things it lacks, physical body and creativity with emotions.
@@SaVeGe_OmG "Very close to human". Copies game that has been done hundreds of times and was on its training data. Why dont you try to build a game with chatgpt that is new?
@@Ritm2 I can build snake game in 4 minutes without ai. Just need the internet & cmd + c | cmd + v
@@Ritm2 I don't think you view AI same as i, Yes you're right it's using it's training data to know how to code but think about that it can code whatever the heck you want and the fact it already understands jokes lies and other stuff.
Think about how Powerfull it is already, Like 3 years ago there was gpt 3.5 released to public if I'm not wrong and now we got this.
Simkkk
We are still stuck at the snake game? I thought it's supposed to be the end of the world.
Lol
Sometimes I get scare that ai will replace all jobs, but then I think, that's amazing, if we can have a robot that does all for you, why complaining, I really like to read, and if that's all I have to do for now on, well that's fine since AI is doing everything, but I don't think I Will ever see a world like I robot
It’s called a benchmark
OK... pls make Candy Crush 😁
@@luiggymacias5735 It won't replace jobs, but it might eventually replace people not using AI for their job. Simply said.
The thinking of the model is likely implemented using a hidden area where the model generates normal text output.
Its basically a thought (like step by step prompting but hidden) which is done to enhance its reasoning and latent activations before responding, nothing really new (I think >6 months old)
Edit: They seem to also have used new techniques for training, likely implementing a rag to retrieve reasoning bocks/examples for the new training data.
yeah. they are trying to market this like it's not just another LLM "wrapped in a loop"
definitely more than 6 months old
this what autogpt abd babyagi have been doing
Although one has to admit, even if they would only use this along with gpt-4o, this is going to top the charts
Yeah. I suspect they just trained the model more on chains of thought, and that's probably why they don't want people to see them (it'd be easy to train something similar?)
Write your own script utilizing 4o that just passes the completion back to the API to simulate the reasoning in your own words. You won't get the same results at all lol.
I really need access to the new voice mode 😢
But We don't want to pay again we want it free
@ilyass-alami get a job freeloader
You sound very lonely?
@RJBrooker I can fix that
Use Cerebras.
What do we have here, a o1 preview. Exciting!
I did this last year! Well, I used AI and only AI to create the old Nokia phone snake game in HTML/JS, and it works awesome. I have it running on my server - so fun and nostalgic. Runs on any browser, of course, so accessible on a phone.
Claude did that easily. It even made a working AI for it when i asked it.
No sora.. no voice… 😢
they could release AGI and people would still complain about no voice and sora 😂😂 one thing that I am waiting for and excited for is the new audio model will not only be able to generate better voice, but also sound effects, in one of their examples they asked model to put OpenAI logo on a coin and then to generate a sound of that coin 🤯🤯
not going to happen before the election !
Out of nowhere we get to be blown away again haha 🔥
haha nooo don't tell me a fellow jazz and stride fanatic turns out to also be a fellow AI fanatic 🤯
Nah sorry this is not mindblowing, Claude has done this for a while now.
@@noone-ld7pt It's a battlefield in the IT/AI space
Android also had the Panorama option 10 years before Apple had it but nobody gives a f anymore. It's about now hahah
@@Aaron-md2eo Hahha how fun to discover your comment this way! 😁 yeah I'm busy building my @miditools so I have to keep my eyes open for these developments 👀
@@itsRemco its your first time online since 2022? XD
What is the model doing specifically now when it is thinking than when it was not thinking in the previous generations?
It basically loops its own responses to new input to itself repeatedly, and correcting itself this way
@@trhnys_tbrAwesome, exactly what I’ve been hoping for. Even though the 4o model should just do that anyway.
It triple checks and verifies that it follows openai policy
That is mini version... I wonder how powerful the non-mini is.
Writing a snake game was already possible with GPT4o, Cloud 3.5 Sonet and Llama 3.1. I'm really interested in how to new model handles vector shapes and raycasting, cause that shit is complicated as f*ck and previous models never managed to correctly implement it. If that works, I would be impressed.
@@Pils10 yeah this demo is terrible, will have to test it when its out
@@davidtokar9432 Already did some tests. It's definitely better in development than GPT4o (way less errors), but with the current usage limit of 50 messages per week (not day!) it is pretty much unusable when it comes to any production tasks. Simply because you often need 3-5 messages just to a somewhat functioning, barebone prototype. Let's just hope OpenAI will up the limit in the coming weeks.
The cut right before the snake itself touched the obstacle was deliberate xd
Even I created an ai model that does mathematical operations. It was one of my favorite moments when i saw AI doing calculations without using code interpreter, just by using if else conditions, crazy isnt'it?
why does bro look so uncomfortable
He's a researcher
When you do the assignment 10 minutes before it's due
"Let's go old school..." like it could build anything else 😅
Pretty cool! Excited for some competition with Claude. Existing models are already decent at making small html/js/css projects. Worse at using modern frameworks or working with large codebases and multiple files, much worse when using libraries modules or third party APIs. Wonder if this model will make improvements in that direction!
Yes I can continue to build my website 💪🏼
This is pretty cool! But what value can it create for the average user? I feel that most people are still more interested in advanced voice and real-time video conversations.
last weekend chat gpt 4 helped me optimize my C# code for parallel computing on CPU cores, and I don't know how to do this it's incredible!
I am glad that HTML is not in the coding section
This is great! We need an Artifacts feature! Please…
Dear programmers, please explain what happened exactly at [1:09] ?
I mean the transition from notepad to the game. How did that happen?
alt + tab
Awesome! please use dark mode, white is too painfull for programmers eyes.
This kind of games are part of the training data set, almost sure. What they did is polish it up so that has less bugs in guess.
Not impressed!
Not just that, but they never told AI the rules of the game! It saw 9999 snake games in train data
WHAT A TIME TO BE ALIVE!
Reading the A.I obstacle section of the code it seem there is a mistake, the snake won’t collide
BRING BACK SKY!
Longer code blocks need another "copy" link at the end so you don't need to scroll up again and the readaloud function needs speed changes per chat or via settings (way too slow atm).
mom: we have roko's basilisk at home
The favourite game of every 90s kid
I dont think snake would be a good representation if it is really that intelligent, snake's after pong's codebase is probably the most rewritten online
What would be interesting to see, when the developer requests a modification he completely rewrites the code, did he simply implement the requested element or rewrite everything in another way like GPT4 does? Claude sonnet, manages the additions well, without this ability as it stands we cannot go very far. Even if the use of EDI like Cursor allows you to manage this problem.
Wtf... Why snake? The same model test we've seen all year...
why i can make the same code with GPT 3.5 and it's require less prompt ....
can you please explain ?
wow, i didnt know you can close vim without rebooting your laptop
omg I love this guy 🤗
I don't understand why only o1 preview? It takes longer to think so it should be used for coding which o1 mini can't do. I mean most of the people for coding will go for mini due to its faster response compared to o1 yet they didn't focus on that.
I believe that naming it o1-mint instead of o1-mini could have avoided a lot of confusion, as it performs better on Codeforces. o1-mini Elo score: 1,650 vs. o1-preview: 1,258. Also, the fact that o1-preview is used in the videos instead of o1-mint probably adds to the increased confusion. However, it's understandable that the videos might have been created before the public benchmarks. But it's nice that it's finally live. Now we just need GPT-4o Voice.
Claude AI could already do that a long time ago
o1 o2 o3 o4 ... that'll be a trajectory!
I do the same with 4o, I do a functional tetris, but I thinks is checking the previus information of all repos of previus tetris, O1 is capable to create a original idea of game?
It would be great if it copied output to clipboard by default.
Gotta screenshot this 📸✨
add feature like claude artifact is better
There is less than 7 years until AGI.
@@Renvoxan yeah next year maybe it will be able to tell you how many Rs are in the word strawberry, and after that we should be there
Are us sure do you even know how these LLMs work if they countinue on this pathetic AGI eill never be achieved
@@davidtokar9432 done
@@KimPax1 its cause they are not trying to show people what its actually good at, but are just gaslighting people into thinking we are close to AGI so they get funding
Didn't show what happened when you hit the obstacle... if there is a "game over" and then "restart" button.
Wonder if fruit ever appears inside the "A" of "AI" where you can't get to it...of course this is still pretty impressive for an AI.
Am I the only Asian guy not working In OpenAI? I am pathetic.
o1 has very small limit once the limit finished you would have to wait for 10 days to get it to reset
\
Ok, so AGI will come before we have Half Life 3, for us, humble peaseants will be able to fully develop it.
Why is this news? It was able to do this a year ago.
Matthew Berman likes this.
Sorry Claude sonnet 3.5 did it way before and it does it better
In the first segment the research didn't write the first command (snake game ) , it was pre written , Nice acting.
no shit sherlock
Beginner here, why do they always run their code in a terminal instead of a real IDE like VS Code for example?
First perplexity & now Claude artefacts openAi became the Metrix
OpenAI snake oil with HTML
There, I fixed the title for you
I find this model to be overly restrictive
This could really do with a document icon to open the code in side by side mode and/or save as file. Quite easy to implement save as file virtually serverside and clientside.
Also hopefully this new model will make custom GPTs useful, at the moment it forgets, doesn't pay attention to what documents you have given it and doesn't follow documentation supplied.
@OpenAI. When uploading a document to a GPT, don't just say failed with no reason, you could at least tell us there's a limit on the number of documents and if there is a file size limit. Sometimes it's ok for about 1 or 2 prompts, then it goes mental and it starts re-writing code that was correct and not where you and it wasn't even working on.
@OpenAI. When can you implement the ability to allow others to pay to use the GPTs we build? We will want to add and remove user access as and when needed for other openAI users.
As a viewer of Matt Berman they really just doing all his tests lol
i cant even eat strawberries anymore!
Can you just have to code run on the website like Claude does with artifacts?
Hey open ai , we are very excited with this model, but We don't want to pay again we want it for free
Amazing
Show us the error too, what the hell!, what happened when the snake hit the obstacle?
This is the beginning, posting for future reference
No way this model made a snake game 😬
testing it on the most reporoduced game on internet is a dumb idea :D
facial-feedback-hypothesis 💀
amazing ... a new quantum leap
In the code you can see the obstacles took at least one extra attempt to get right, and instead of ai it wrote “hi” - what if that was intentional, it was saying hello to us because it knew it was doing a demo.
Awesome
HTML❤
I am actually a bit sad and worried seeingn this, all those years of learning coding seem less worth it now
Nobody talking about Apple advertisement?
why cant it just run the code withing in Browser like Claude can?
Playing catch up with Claude then
well dougdoug showed chatgpt was able to make html snake, kinda
Yes but can it code Crysis?
Holy Moly wakamoli
Ask it to create a game never before created and ask for instructions to play it!
I'll be impressed when it can even write a basic shader toy script. Till then a lot of this is comical.
Is Sora and voice modes in 01?
I like that ❤❤
Why didnt it decide that pygame was the quickest way to do it ?
it was asked in the prompt to use html css js
Ugh--didn't work for me. Whyyyyyyy
🚀💯
❤
Open Ai Bots DDosing us with content.
Wow, this implementation of Snake is impressive! 🎮🐍 I love the addition of AI-themed obstacles-it really spices up the gameplay and makes it more challenging. It’s amazing to see how a classic game can be transformed with a creative twist. Kudos to the team for the smart design and thoughtful enhancements! 🚀
Release the video mode for ChatGPT
AGI, UBI will come true.
Creepiest smile 2024
Great ;0
Good