Hahahah I can't unsee 14:10 now. About the game: Anthropic is really the king now it seems huh!?. To give Open Ai a chance try to change the numbers fed to the prompt: i.e: for distance maybe you could have 3-5 options like 'long-range', mid-range', 'short-range' (and add this rule explanation to the prompt as well) Open Ai is just terrible with numbers
is it possible that the order that you place the information to both AIs could change the results? Let's say, instead of TowerRange,Creeps,Towers is Creeps,TowerRange,Towers
This weekend I was trying to do some stuff with Gipiti and I felt the thing was just not going to lend me a hand, went over Antropic and it felt like magic, the thing udnerstood my problem, the version of the library and game a viable solution I could build on top. Try Claude.
Dude regarding the weird quirk of "different day times different results", maybe it's related to how much load the inference servers have , more load -> less resources and lower response quality
people suggesting ollama with llama3.1 are missing the point. I use that exact set up at my home. llama3.1 8b is nowhere near as capable as the larger ones hosted by anthropic and openai.
I have seen literal babbies graphing in python with xlrd and matplotlib and here we have Rusty McZigface not managing a simple little pandas thing? Pathetic.
You should cache the system prompt with Anthropic Also, if you're not feeding the game history back into every new prompt, you should try that OpenAI and Anthropic have vastly different prompt styles - the same prompt may favor one or the other. You should play around with others to see if it switches.
@ThePrimeTime I wonder if it would be better to just replay the GPT moves from the previous game, and have each model respond in a staggered format, producing their best "responses" to each other model's moves for an entire game. It's not as adversarial but it might get you around the token limit.
You can greatly cut down on the amount of tokens used by reformatting your natural language for AI. Like you can mostly use 'base' words unless you explicitly need it to be different. Meaning, no plurals, past/future tense. Like I said, in some cases where it might cause misinterpretation, you don't use base words. The reason is because base words usually are less tokens. Emphasize using Asterix's, quotes, underscore, etc. eg: 2 player TD game lose game all_tower destroy \*ammo health are same\* place tower in \*placement range\* json prompt prefix \_your\_ is your value if both player place all tower wave start. creep spawn left move right What would work even better is just send the codebase as the instruction sheet.
Is there a way you can make the collected stats public so we can play with the statistics? And you should absolutelly track where towers get build and destroyed. Source for the game would be nice too. looks like a project you could have some fun with. Let the Battle of the Ais begin.
I use Claude until I run out of free requests then move to GPT, but honestly I rarely ask GPT for anything unless it's a tailored question I know chatgpt would answer well.
Anthropic specifically says, "the longer your context window, the less prompts you can use". That mean if you send a single prompt of 1000 words, that'll eat more of your total tokens for that time frame than a single prompt with 40 words. If you could compress the data in your json youll expand your totals prompts
Flip did not take it out.
good old Flip
If he took it out, he'd be done for indecent exposure
Do we ever really know when he DOES take it out?
Thanks flip
Flip the 🐐
New LLM benchmark: PTD (prime tower defense)
Release it on huggingface
Prime in excel was so funny, like i never saw this guy using a mouse
12:46 When the morning triple shot espresso kicks in
based on the title, I was expecting to see at least a full game played by 2 AIs, but... there's none. I mean, why..?
Because you touch yourself at night
But, you did get a nipple rub show 🤷♂️
Busy yapping 😂
0:35 Creeps move across the screen. Describing Twitch chat? Gottem
Ask the AI to explain each move when submitting it. Would be interesting to see if there is a coherent strategy or if the explanation even makes sense
"show your work" is also a great way to improve LLM outputs in general
Just doing that would change the result
They’ve consistent failed to get this type of insight from chess models for decades. Its not going to be trust worthy or coherent
@@hamm8934chess models ain’t machine learning algorithms, and any of them that were, were not at this insane level of slightly stupid AI.
@@eetoonamamanakooo They are, indeed, machine learning algorithms (at least all of the top engines nowadays), but they are not LLMs.
"Its so hard to write good code" but also says, "all code is bad code"
Also for stat collection - position of towers. Then heat map!
It's a tuesday morning, and I'm laughing like a dum dum at the moustache programming man rubbing his nipples in an imaginary trance. Thanks!
Dude you need to use prompt caching. It will let you store and reuse the same system prompts for cheap. Cache duration is 5 min.
The new QuadrupleA game we were waiting for
Hahahah I can't unsee 14:10 now.
About the game:
Anthropic is really the king now it seems huh!?. To give Open Ai a chance try to change the numbers fed to the prompt:
i.e: for distance maybe you could have 3-5 options like 'long-range', mid-range', 'short-range' (and add this rule explanation to the prompt as well)
Open Ai is just terrible with numbers
is it possible that the order that you place the information to both AIs could change the results? Let's say, instead of TowerRange,Creeps,Towers is Creeps,TowerRange,Towers
mom i am studying and the moment she enters the room : 12:56 🗣 zoom in flip.. yeah zoom in
This weekend I was trying to do some stuff with Gipiti and I felt the thing was just not going to lend me a hand, went over Antropic and it felt like magic, the thing udnerstood my problem, the version of the library and game a viable solution I could build on top. Try Claude.
Literally got irritated with gpt for 3 hr and then went 1 prompt with sonnet
Bro the through the blinds view is so peak
Why not use prompt caching on the giant system prompt
Props to ZanyBots for asking about error bars when given numbers.
Yes, I have a math degree, how'd you know?
Dude regarding the weird quirk of "different day times different results", maybe it's related to how much load the inference servers have , more load -> less resources and lower response quality
When Prime starts to use pivot tables to actively update their charts in real time, then maybe, MAYBE... we can stop pooping on their excel skills. :P
Did you make your TowerDefence available for people to code their own strategy and send it to you?
this is the true test of ai model performance
Can’t you let them play games randomly, with or without an adversary, and build a dataset it can use to train itself from the won games?
people suggesting ollama with llama3.1 are missing the point. I use that exact set up at my home. llama3.1 8b is nowhere near as capable as the larger ones hosted by anthropic and openai.
of course lmao, you compare 8b model with around 300b model. even llama 3.1 405b model is below than claude sonet 3.5, i dont know about gpt 4o
14:54 indeed, we call this accuracy. 😉
I have seen literal babbies graphing in python with xlrd and matplotlib and here we have Rusty McZigface not managing a simple little pandas thing? Pathetic.
Hey! Be sweet!
You should cache the system prompt with Anthropic
Also, if you're not feeding the game history back into every new prompt, you should try that
OpenAI and Anthropic have vastly different prompt styles - the same prompt may favor one or the other. You should play around with others to see if it switches.
The questions just got even bigger.
"write a python script to graph this csv data"
Can't wait for the graphics to be added!
Its ascii game
😂
How about using Sparse Priming Representation (SPR) to reduce your prompt? You could generate it using chatgpt and give it then to claud
New official AI benchmark, the Primeatest!
I'm still baffled by the idea of building a non-deterministic computer. Receiving different answers each time you run tho code is wild.
@ThePrimeTime I wonder if it would be better to just replay the GPT moves from the previous game, and have each model respond in a staggered format, producing their best "responses" to each other model's moves for an entire game. It's not as adversarial but it might get you around the token limit.
It's almost like the spreadsheet has better UI and UX than vim. He doesn't even have to count the columns to figure out what they are.
You can greatly cut down on the amount of tokens used by reformatting your natural language for AI. Like you can mostly use 'base' words unless you explicitly need it to be different. Meaning, no plurals, past/future tense. Like I said, in some cases where it might cause misinterpretation, you don't use base words. The reason is because base words usually are less tokens. Emphasize using Asterix's, quotes, underscore, etc.
eg: 2 player TD game lose game all_tower destroy \*ammo health are same\* place tower in \*placement range\* json prompt prefix \_your\_ is your value
if both player place all tower wave start. creep spawn left move right
What would work even better is just send the codebase as the instruction sheet.
What the f did i just watch 14:19 😂😂😂
Is there a way you can make the collected stats public so we can play with the statistics?
And you should absolutelly track where towers get build and destroyed.
Source for the game would be nice too. looks like a project you could have some fun with.
Let the Battle of the Ais begin.
Do not hallucinate
My man did 6 outros 😅
I use Claude until I run out of free requests then move to GPT, but honestly I rarely ask GPT for anything unless it's a tailored question I know chatgpt would answer well.
I think the Twitch chat version had more active players than concord. You can try to sell it for 100mil$!
Thats a thumbnail and a half !! Good job flip . never take it out.
prime is doing Sports Center for AI vs AI
Is this open sourced?
Random number generator powered by a wall of lava lamps in a web feed yea that could replace me
Article
DELETE ME DELETE ME
Guest
DELETE ME DELETE ME
13:00 No Flip please stop!
Tell me you can yap without telling me you can yap
Can u go throguht Aws bedrock???
DELETE ME DELETE ME
Didn't know that you're an article or description
Would be cool to see others like llama3 too
chatgpt != skynet confirmed
Claude resubmits the entire previous chat into the llm so gonna rinse your tokens. Also means longer running games Claude should lose
Can we actually see them play ???
What AI beats what AI at defending ASCII towers.
For exactly this usecase you do can use gpt. It is literally the best thing it can code xD.
hillarious 😆
Ima be real I have no idea how this game works
tons of steps to just read the output. Spend a few minutes and format it in the code. Your future self will thank you
I do not know the leglity of this, but if Im allowed to share my Anthropic account we could get you some more tokens
Influencagen LOL sounds like Influenza
Anthropic specifically says, "the longer your context window, the less prompts you can use". That mean if you send a single prompt of 1000 words, that'll eat more of your total tokens for that time frame than a single prompt with 40 words. If you could compress the data in your json youll expand your totals prompts
Todays fun fact: the spinning loading wheel icon is called a throbber! Have fun knowing that for the rest of your life
I have no clue what is going on in this game
#AgressiveAI is a much better name for OpenAI.
It's wild that it even produces semi-workable results. A few years ago, this would've been world news.
Mom cancel my basketball stream pacman broke obs again
OpenRouter will give you more Anthropic tokens (:
Anthropic also has a beta feature to prepare your system prompt and pay for it once. (Prompt caching)
Pythetic?
Damn dude, this looks like someone doing stuff 40 years ago. Is Linux still that far behind?
LOL ZOMIN ON THA MANLI thing yea .. lol .. ai is not yea generativ ey lol yea .. but it haz it uses is . but
Use Claude on AWS
I still don't get what this game is about.
If he learns python he will realize vim is not that good
why do the graphics suck
ASCII is cool as hell for programmers .
@@CrYpt001 yeah totally cool as a graphics engine
First ?
First
Give me money, money me, money now, me a money needing a lot now
this is super cool, keep your bearing, its a good stuff 🫡
DELETE ME DELETE ME
First
DELETE ME DELETE ME