• There is NO official Otland's Discord server and NO official Otland's server list. The Otland's Staff does not manage any Discord server or server list. Moderators or administrator of any Discord server or server lists have NO connection to the Otland's Staff. Do not get scammed!
  • 2026 staff recruitment is open! Check it out and consider applying!

AI

Sort by date Sort by votes
Willbor said:
Could you give me some tips on a specific AI for creating an OT server? Thanks in advance!
Click to expand...
If I understand your question, you just want to know which AI are the best to help you develop an OT server?

I use Claude and Claude Code to modify the client, server, and website. I have a local folder which I give access to Claude Code, once it modifies the files, I upload them to the real folders to make sure Claude Code doesn't do something it shouldn't. It usually takes like 5 to 20 minutes to get something done, plus whatever time it takes you to test it. Regarding cost, I use Claude Max, but the Pro version should be enough if you work mostly during the weekends, since you have more tokens. Not sure if Claude Code is still available for the Pro version, though. It was like 2 months ago.

I use Gemini to create monster sprites and anything related to images. I haven't, but I am also sure it can do nice videos. I suggest you to create new chats if Gemini is not successful in the first try. I notice that it brings better results rather than trying to ask it to fix something. You can use the free version, not much difference in the quality, and the amount you can do with the free version is pretty good.
 
Upvote 1 Downvote
Willbor said:
So, what do you think of AI? https://claude.ai
Click to expand...
If i were you, i start learning C++.
A mini-course to understand the logic.

Because sometimes, if you trust AI completely, it can mislead you and provide completely irrelevant information; sometimes, it gives you a long scripts for something very simple that could be done in a shorter and simpler way if you had the knowledge.

The best teacher i found to learn the basics was this one.
 
Upvote 0 Downvote
its not vibe coding if you have to learn how to code, come on


for my hobby project I don't even look at the lines of code anymore, all I need is the AI answer, the terminal with logs and prompts, if something goes wrong I just send the logs and claude analyzes and fixes it, true vibes
Post automatically merged:

Perun said:
codex with gpt 5.5 high, nothing even close
Click to expand...
curious if you tried claude, since from my experience it is superior, but I havent tried the new gpt yet, only older version
 
Upvote 0 Downvote
Bloodbringer said:
its not vibe coding if you have to learn how to code, come on


for my hobby project I don't even look at the lines of code anymore, all I need is the AI answer, the terminal with logs and prompts, if something goes wrong I just send the logs and claude analyzes and fixes it, true vibes
Post automatically merged:


curious if you tried claude, since from my experience it is superior, but I havent tried the new gpt yet, only older version
Click to expand...
i tried claude, jules, gpt, and local llms, currently Jules works great (and is free) for refactoring big systems, if you have a lot of code to refactor.
But to speed up new systems, or OTC UI gpt 5.5 high works best for me, but it burn a lot of credits, still can do a lot for cheapest option GPT Plus is like 20 usd or something monthly.
 
Upvote 0 Downvote
Willbor said:
So, what do you think of AI? https://claude.ai
Click to expand...
With The best coding agent (https://cursor.com/) IDE you can use all popular AI models and compare results. You pay 20$ per month and get 20-40$ to spend on AI chats. There is also their own model 'composer-2.5-fast' on which you can spent much more (300-400kk tokens - ~150$) 'for free' with that 20$ plan.
There is also free access to Claude and Gemini sponsored by Google: Google Antigravity (https://antigravity.google/) , but it often hits weekly limit of free account after 1-3 prompts to Claude (few times more with Gemini).

How bad AI can go on OTS? One popular OTS implemented 'prey system' on server using AI (it used existing OTC 'prey' module). Day after implementation CPU usage by Lua jumped from 23% to 50%! For next 2 days owner implemented optimizations with AI making it use 40% CPU next day and 30% CPU after 2 days. It's still +7% CPU usage for something so basic as 'prey system'.
Then I spent few hours and 9$ on Claude 4.6 to rewrite most of code to C++, but I had to prompt like 100 times and tell it how to implement every part of logic in optimized way. It wasn't "AI rewrite this Lua crap to make it work fast in C++" ;)

The worst AI code I've seen was 'extra item attributes' code that AI implemented in Container::getItems, which send SQL query for every iteration over item in backpack :D , but it lagged OTS with 1 player online, so it was obvious that it must be fixed before going production server.

Sadly, most of AI bugs in Lua result in random server crashes (like you test script with 100 players online for 3 days and then you get random crash at 4 AM). Maybe you have to add to every prompt link to [TFS 1.x+] How to NOT write Lua scripts or how to crash server by Lua script — Gesior's blog (https://skalski.pro/2020/06/05/tfs-1-x-how-to-not-write-lua-scripts-or-how-to-crash-server-by-lua-script)
(I replaced WordPress on my blog with my own blog app made with AI)

The best AI OTS code I've seen this year is OTCv8 updated by AI to make it compile into iPhone app. It took AI "Qwen 3.6" 3 days running on Apple Mac Studio with 256 GB RAM (shared with GPU on Apple). IDK how much you would spend on tokens, if you had to pay for GPT/Claude.


I just compared models to check how GPT 5.5 codes vs Claude 4.6/4.8.

TASK: I told AI to edit simple Lua script (reward box with random reward), so that same player won't get same item 2 times in a row.

Only GPT 5.5 wrote foolproof script. All other models analysed script config above, where I have a lot of items configured and each has unique ID.
GPT 5.5 expected that someone could write config with repeated item IDs, so checking if there is more than 1 item in table is not enough to guarantee that, if we remove items from table by ID and pick random element of table, there will be any.
In case of all other AIs codes, if I edit config and put there table with 2 elements, but both with same id, server would go into infinite loop in Lua (freez whole OTS until manual restart by owner).
All other models - except GPT 5.5 - noticed that there are few items (different reward chests) player can use and implemented code that tracked last random item from each type of chest. I think it's better interpretation and real programmer would ask 'how do you want it to work', not implement code as he 'thinks' it should work.
Also, I ran same task 2 times with "Auto" and get 2 different results. That's common with AI. It's like slot machine, you run it X times and pick the best script generated.

Biggest problem of all AIs is generating too many lines of code. You have to describe every part of logic and how to implement it, to get short optimized code :(

My tests results:
Script before changes:
GPT-5.5 high:
Claude 4.6 high:
Claude 4.8 high:
Composer-2.5-fast (available only in Cursor):
"Auto" model by Cursor (it is Composer-2.0-fast or Composer-2.5-fast):
"Auto" model by Cursor (it is Composer-2.0-fast or Composer-2.5-fast):
Final "auto" code, when I wrote 4x longer prompt with description of all problems that it has to take care of. It generated safe code which uses less CPU than any other AI script above:

How much each test costed:
claude-4.6-opus-high-thinking - 82,5k tokens - 0.16$ by Cursor, official AI price 0.38$
claude-opus-4-8-thinking-high - 112,1k tokens - 0.14$ by Cursor, official AI price 0.31$
gpt-5.5-high - 151,6k tokens - 0.1$ by Cursor, official AI price 0.21$
composer-2.5-fast - 81,1k tokens - ~0.027% of monthly limit by Cursor, official AI price 0.13$
auto by composer- 103,2k tokens - ~0.01% of monthly limit by Cursor, official AI price 0.05$
auto by composer- 106k tokens - ~0.01% of monthly limit by Cursor, official AI price 0.05$
 
Upvote 0 Downvote
Gesior.pl said:
With The best coding agent (https://cursor.com/) IDE you can use all popular AI models and compare results. You pay 20$ per month and get 20-40$ to spend on AI chats. There is also their own model 'composer-2.5-fast' on which you can spent much more (300-400kk tokens - ~150$) 'for free' with that 20$ plan.
There is also free access to Claude and Gemini sponsored by Google: Google Antigravity (https://antigravity.google/) , but it often hits weekly limit of free account after 1-3 prompts to Claude (few times more with Gemini).

How bad AI can go on OTS? One popular OTS implemented 'prey system' on server using AI (it used existing OTC 'prey' module). Day after implementation CPU usage by Lua jumped from 23% to 50%! For next 2 days owner implemented optimizations with AI making it use 40% CPU next day and 30% CPU after 2 days. It's still +7% CPU usage for something so basic as 'prey system'.
Then I spent few hours and 9$ on Claude 4.6 to rewrite most of code to C++, but I had to prompt like 100 times and tell it how to implement every part of logic in optimized way. It wasn't "AI rewrite this Lua crap to make it work fast in C++" ;)

The worst AI code I've seen was 'extra item attributes' code that AI implemented in Container::getItems, which send SQL query for every iteration over item in backpack :D , but it lagged OTS with 1 player online, so it was obvious that it must be fixed before going production server.

Sadly, most of AI bugs in Lua result in random server crashes (like you test script with 100 players online for 3 days and then you get random crash at 4 AM). Maybe you have to add to every prompt link to [TFS 1.x+] How to NOT write Lua scripts or how to crash server by Lua script — Gesior's blog (https://skalski.pro/2020/06/05/tfs-1-x-how-to-not-write-lua-scripts-or-how-to-crash-server-by-lua-script)
(I replaced WordPress on my blog with my own blog app made with AI)

The best AI OTS code I've seen this year is OTCv8 updated by AI to make it compile into iPhone app. It took AI "Qwen 3.6" 3 days running on Apple Mac Studio with 256 GB RAM (shared with GPU on Apple). IDK how much you would spend on tokens, if you had to pay for GPT/Claude.


I just compared models to check how GPT 5.5 codes vs Claude 4.6/4.8.

TASK: I told AI to edit simple Lua script (reward box with random reward), so that same player won't get same item 2 times in a row.

Only GPT 5.5 wrote foolproof script. All other models analysed script config above, where I have a lot of items configured and each has unique ID.
GPT 5.5 expected that someone could write config with repeated item IDs, so checking if there is more than 1 item in table is not enough to guarantee that, if we remove items from table by ID and pick random element of table, there will be any.
In case of all other AIs codes, if I edit config and put there table with 2 elements, but both with same id, server would go into infinite loop in Lua (freez whole OTS until manual restart by owner).
All other models - except GPT 5.5 - noticed that there are few items (different reward chests) player can use and implemented code that tracked last random item from each type of chest. I think it's better interpretation and real programmer would ask 'how do you want it to work', not implement code as he 'thinks' it should work.
Also, I ran same task 2 times with "Auto" and get 2 different results. That's common with AI. It's like slot machine, you run it X times and pick the best script generated.

Biggest problem of all AIs is generating too many lines of code. You have to describe every part of logic and how to implement it, to get short optimized code :(

My tests results:
Script before changes:
GPT-5.5 high:
Claude 4.6 high:
Claude 4.8 high:
Composer-2.5-fast (available only in Cursor):
"Auto" model by Cursor (it is Composer-2.0-fast or Composer-2.5-fast):
"Auto" model by Cursor (it is Composer-2.0-fast or Composer-2.5-fast):
Final "auto" code, when I wrote 4x longer prompt with description of all problems that it has to take care of. It generated safe code which uses less CPU than any other AI script above:

How much each test costed:
claude-4.6-opus-high-thinking - 82,5k tokens - 0.16$ by Cursor, official AI price 0.38$
claude-opus-4-8-thinking-high - 112,1k tokens - 0.14$ by Cursor, official AI price 0.31$
gpt-5.5-high - 151,6k tokens - 0.1$ by Cursor, official AI price 0.21$
composer-2.5-fast - 81,1k tokens - ~0.027% of monthly limit by Cursor, official AI price 0.13$
auto by composer- 103,2k tokens - ~0.01% of monthly limit by Cursor, official AI price 0.05$
auto by composer- 106k tokens - ~0.01% of monthly limit by Cursor, official AI price 0.05$
Click to expand...
If you use ASDD with a proper setup of workflows (1st phase functional approval, 2nd phase architecture approval, 3rd phase developer implementation, 4th phase qa, etc), roles and skills in a robust and safer language like Rust, the quality of the code is on another level, and the scope of the hallucination for the AI is so much less.

It is also so much better if you setup a prompt enhancer in a chat in order to write the requirement to this chatbot (web chatgpt) first to enhance it properly and copy-paste it for the local agent to implement. Forcing it to always do questions back in the prompt enhancer always saves up time and tokens so you don't need to fix anything later.

From my experience, using ASDD with Rust is simply on another level. I tried it with Python and was ok for a POC, but Rust forces the AI to just have better code quality, it have no other option but to follow the languages rules, which is simply beautiful.

Lastly and not less, you should always look what code it's generating or you will end up with a technical debt no matter what.
 
Upvote 0 Downvote
UPDATE:
Anthropic just dropped 'Claude Fable 5' and it's logic is outstanding. No other model is even close.
Sadly, all other models put a lot of comments around their complicated (or simple [useless comments]) logic and Fable 5 put no single comment around it's very complicated algorithms.
It's 100% right script and it's 99% statistically right (maybe 100%) - I'm not an expert in "statistics", I studied it for a year on IT studies - but I had to analyse logic of these 18 lines of code for 5 minutes, to be so sure about it's logic "not failing":
It's not basic 'pick random number not used before' algorithm that uses all available CPU power to pick it. It uses some math.random -1 tricks to save a lot of CPU power - like games in 1990 did to save a lot of CPU with 'fake' random.

It still does not expect that "config" may have 2 or more rows with same "id" (as all old AIs except GPT 5.5.), but it won't freez/lag OTS in case it gets wrong config. It will just pick 'random' result and give item to player. It can be the same he had last time - whole idea of script 'update' breaks -, but OTS stays online.

EDIT:
It may be a problem, that this simple 18 lines of code from 'Fable 5' costs 0.63$ in tokens.
 
Last edited:
Upvote 0 Downvote
Back
Top