- Welcome to The International Simutrans Forum.
News:
Simutrans Tools
Know our tools that can help you to create add-ons, install and customize Simutrans.
Know our tools that can help you to create add-ons, install and customize Simutrans.
Recent posts
#1
Scripting Scenarios and AI / Re: squirrel script - Script A...
Last post by victor_18993 - Today at 12:22:42 AMI'll check it in case there were changes since the last version,
best regards.
best regards.
#2
Community Discussion / Re: About the Community Manage...
Last post by victor_18993 - Today at 12:20:13 AMAnthropic's models, like OpenAI's and those from other companies, only show a pseudo-chain; they don't show the full chain of thought. I haven't seen it in open-source models either. Right now, I don't know if there's any model on the agentic AI market that allows this, but I doubt any of them show the exact reasoning chain, because that would allow model distillation to train others.
Right now we're working with black boxes—meaning that if, after the final result, the prompt that was run doesn't get the desired results, even when running under an orchestrated context, there's no way for you to know how it arrived at that result.
At my job, which is healthcare-related and where LLM providers have to provide the chain, it's delivered in a different way so it can't be distilled, but it lets you trace the LLM—though it's still not enough to know where things went wrong.
That's why what you're pointing out isn't really everything the model is actually doing or will do to reach its goal, and that's why they're so hard to debug. They aren't releasing SDKs or anything specific either. What it shows you is usually that it accesses file X, reads it, contextualizes it, and starts coding, but what's missing is the important part: why it made the decision to start coding, and at what point. That's just not exposed nowadays for the reasons described above.
Right now we're working with black boxes—meaning that if, after the final result, the prompt that was run doesn't get the desired results, even when running under an orchestrated context, there's no way for you to know how it arrived at that result.
At my job, which is healthcare-related and where LLM providers have to provide the chain, it's delivered in a different way so it can't be distilled, but it lets you trace the LLM—though it's still not enough to know where things went wrong.
That's why what you're pointing out isn't really everything the model is actually doing or will do to reach its goal, and that's why they're so hard to debug. They aren't releasing SDKs or anything specific either. What it shows you is usually that it accesses file X, reads it, contextualizes it, and starts coding, but what's missing is the important part: why it made the decision to start coding, and at what point. That's just not exposed nowadays for the reasons described above.
#3
Information & Announcements / Re: DDOS or massive bot swarm ...
Last post by Isaac Eiland-Hall - Yesterday at 08:58:35 PMNote that this issue is ongoing. When I disable the rule, the attack comes right back and the forum quickly gets unusuable with pages failing to load.
One person sent me a PM noting that they cannot load pages using TOR. I don't know the precise conditions that force the issue I had to block, so I don't know how many are affected, although that's the only person I've heard of who can't access the forum, so for now, I'm leaving the rule intact, but still off and on trying to poke at it.
One person sent me a PM noting that they cannot load pages using TOR. I don't know the precise conditions that force the issue I had to block, so I don't know how many are affected, although that's the only person I've heard of who can't access the forum, so for now, I'm leaving the rule intact, but still off and on trying to poke at it.
#4
Scripting Scenarios and AI / squirrel script - Script AI
Last post by Andarix - Yesterday at 06:46:15 PMThe script AI's pathfinding ignores "private road" signs.
This results in a path being successfully constructed, yet vehicles are unable to use it.
This affects both pathfinding across open terrain (for new routes) and navigation along existing paths.
This also applies to the new dialogue path builder.
This results in a path being successfully constructed, yet vehicles are unable to use it.
This affects both pathfinding across open terrain (for new routes) and navigation along existing paths.
This also applies to the new dialogue path builder.
#5
Community Discussion / Re: About the Community Manage...
Last post by Mishasama - Yesterday at 12:12:13 PMQuote from: victor_18993 on Yesterday at 11:18:51 AMThe problem is that there are no debugging tools that actually tell you how it reached that conclusion. The chain of thought is never exposed in any LLM, let alone the intermediate calculations or decisions made where it wasn't interacting with the human operator.Have you tried adding a command to the prompt that requires a complete explanation of the reasoning chain? In my experience, it does allow the AI to display the reasoning process.Then I can formulate some system prompts based on the reasoning chain to fine-tune its behavior.
However, it's worth mentioning that CLDUDE seems rather stubborn when it comes to following commands. It's not easy to get it to act entirely according to your instructions; it always has its own set of "principles" to decide whether or not to obey...
#6
Community Discussion / Re: About the Community Manage...
Last post by victor_18993 - Yesterday at 11:18:51 AMI think AI is a good tool and copilot, but I don't know what kind of results it can deliver when it comes to marketing and PR. I imagine that if it's used to help draft or clarify text the result might be good, but in my experience with autonomous AI systems—meaning completely on their own—there comes a point where for some reason they start malfunctioning or for some reason start delivering diminishing results.
The problem is that there are no debugging tools that actually tell you how it reached that conclusion. The chain of thought is never exposed in any LLM, let alone the intermediate calculations or decisions made where it wasn't interacting with the human operator.
It's nothing new, LLMs are never going to expose this, but they don't provide us with any debugging tools to improve the workflow or figure out what failed either.
Right now I've noticed that Opus 5.5 takes a lot longer to perform coding tasks and isn't giving me better results than their previous models. Checking Anthropic's official documentation doesn't clarify this either; it seems like they turned Opus 5.5 into a souped-up Fable 5.1, but they don't indicate if there are parameters we should tweak or workflows we should adapt.
Therefore, right now I'm deep into adapting the entire orchestration system I had to optimize the whole workflow all over again. Basically, it's an example of how, as a tool, I think it's great, but every day it pisses me off more—or rather, annoys me—that they break the SDKs and orchestration environments that we users build to optimize models.
Anyway, I guess I'm just a bit annoyed about having to redo so much from scratch.
Summary: I do support AI if it's used as a copilot; I don't support it if it's fully autonomous.
The problem is that there are no debugging tools that actually tell you how it reached that conclusion. The chain of thought is never exposed in any LLM, let alone the intermediate calculations or decisions made where it wasn't interacting with the human operator.
It's nothing new, LLMs are never going to expose this, but they don't provide us with any debugging tools to improve the workflow or figure out what failed either.
Right now I've noticed that Opus 5.5 takes a lot longer to perform coding tasks and isn't giving me better results than their previous models. Checking Anthropic's official documentation doesn't clarify this either; it seems like they turned Opus 5.5 into a souped-up Fable 5.1, but they don't indicate if there are parameters we should tweak or workflows we should adapt.
Therefore, right now I'm deep into adapting the entire orchestration system I had to optimize the whole workflow all over again. Basically, it's an example of how, as a tool, I think it's great, but every day it pisses me off more—or rather, annoys me—that they break the SDKs and orchestration environments that we users build to optimize models.
Anyway, I guess I'm just a bit annoyed about having to redo so much from scratch.
Summary: I do support AI if it's used as a copilot; I don't support it if it's fully autonomous.
#7
Patches & Projects / Re: Simutrans 32-bit graphics ...
Last post by victor_18993 - Yesterday at 10:55:18 AMA new XR32 DAY+NIGHT testing build is available:
https://github.com/Dkijas/simutrans-true32-renderer/releases/tag/xr32-prissi-night-image-prototype-01-r12337-testing-1
Important:
SIMUTRANS_XR32=1 must be set when starting Simutrans, otherwise it will use the legacy 16-bit graphics.
This build is now based on trunk r12337 and includes:
- the newer zoom table from r12336;
- fixes for 40px toolbar icons and icon_scaling;
- the fix for thin XR32 images/catenary appearing too large and displaced at extreme zoom-out;
- the existing DAY+NIGHT prototype.
Thanks to makie and Frank for the real-world testing and reports — both helped uncover issues that were not visible in the original test cases.
If you try it, feedback is especially welcome on:
- extreme zoom levels;
- catenary / wayobjs;
- larger GUI icons/themes;
- DAY/NIGHT graphics;
- gradients and alpha/transparency.
This is still an experimental testing build, so please report anything that looks wrong.
https://github.com/Dkijas/simutrans-true32-renderer/releases/tag/xr32-prissi-night-image-prototype-01-r12337-testing-1
Important:
SIMUTRANS_XR32=1 must be set when starting Simutrans, otherwise it will use the legacy 16-bit graphics.
This build is now based on trunk r12337 and includes:
- the newer zoom table from r12336;
- fixes for 40px toolbar icons and icon_scaling;
- the fix for thin XR32 images/catenary appearing too large and displaced at extreme zoom-out;
- the existing DAY+NIGHT prototype.
Thanks to makie and Frank for the real-world testing and reports — both helped uncover issues that were not visible in the original test cases.
If you try it, feedback is especially welcome on:
- extreme zoom levels;
- catenary / wayobjs;
- larger GUI icons/themes;
- DAY/NIGHT graphics;
- gradients and alpha/transparency.
This is still an experimental testing build, so please report anything that looks wrong.
#8
Community Discussion / Re: About the Community Manage...
Last post by Mishasama - October 09, 2026, 05:48:17 PMYour concerns are valid. We cannot blindly trust AI, much less allow it to replace our thinking abilities.
In my personal experience, although I'm now a near-heavy AI user, I've never directly asked AI for the final answer without examining its thought process.
I believe AI is merely a rapid learning aid, saving us time searching for necessary information from massive amounts of data. But this doesn't mean we should abandon thinking or any fundamental skills—can't we do anything without AI? —That's absolutely unacceptable.
We are individuals capable of independent thought, and we can accomplish a wide variety of tasks on our own, even if it takes a great deal of time and effort. But we must never be enslaved by a tool. Once we relinquish the abilities imprinted in our genes, we will cease to be human.
When we realize these things, there's no need to be overly afraid of AI. As long as we stick to our principles, AI will always be just a tool for us, not a devil that will replace us.
As for whether it's off-topic, I think the discussion surrounding AI issues is laying the foundation for our community management role in the future, so it's not off-topic.
In my personal experience, although I'm now a near-heavy AI user, I've never directly asked AI for the final answer without examining its thought process.
I believe AI is merely a rapid learning aid, saving us time searching for necessary information from massive amounts of data. But this doesn't mean we should abandon thinking or any fundamental skills—can't we do anything without AI? —That's absolutely unacceptable.
We are individuals capable of independent thought, and we can accomplish a wide variety of tasks on our own, even if it takes a great deal of time and effort. But we must never be enslaved by a tool. Once we relinquish the abilities imprinted in our genes, we will cease to be human.
When we realize these things, there's no need to be overly afraid of AI. As long as we stick to our principles, AI will always be just a tool for us, not a devil that will replace us.
As for whether it's off-topic, I think the discussion surrounding AI issues is laying the foundation for our community management role in the future, so it's not off-topic.
#9
Scripting Scenarios and AI / squirrel script - factory_prod...
Last post by Andarix - October 09, 2026, 02:24:20 PMfactory_production_x::max_storage
`max_storage` is supposed to return the storage capacity for the goods. However, it returns 0 after the map loads.
If the AI players are deactivated and then reactivated, the correct values are returned.
messages (newest at the top)
`max_storage` is supposed to return the storage capacity for the goods. However, it returns 0 after the map loads.
If the AI players are deactivated and then reactivated, the correct values are returned.
Code Select
local dest_count_in = []
local dest_capacity = null
local j = 0
foreach(good, islot in f_dest.input) {
if ( good == freight ) {
dest_count_in = islot.get_storage()
dest_capacity = islot.max_storage
local fs = f_dest.get_tile_list()
gui.add_message_at(our_player, "### " + f_dest.get_name() + " - " + coord_to_string(fs[0]) + " # " + good + " # dest_capacity " + dest_capacity, fs[0])
}
}
local input_ = abs((dest_capacity/100)*40)
gui.add_message_at(our_player, "### input_ " + input_ + " # dest_count_in[0] " + dest_count_in[0], world.get_time())
messages (newest at the top)
Quote### input_ 0 # dest_count_in### Papierfabrik - 665,157 # Holz # dest_capacity 0
- 0
### Papierfabrik - 665,157 # Holz # src_count_in 11525
### input_ 0 # dest_count_in### Druckerei - 512,55 # Papier # dest_capacity 0
- 164
### input_ 0 # dest_count_in### Brauerei - 714,376 # Hopfen # dest_capacity 0
- 137
### input_ 0 # dest_count_in### Chemiewerk - 589,1 # Oel # dest_capacity 0
- 1071
...
save load
...
### input_ 880 # dest_count_in### Chemiewerk - 589,1 # Oel # dest_capacity 2222
- 1153
Der Verkehr in Möglingen nimmt zu: 1202 waren bei der Eröffnung von Kino.
Runk & Strunk Transports erhöht die Fahrzeugzahl der Linie: Straße Erdöl Meckenheim Ölfeld Betriebshof - Cuxhaven Ölkraftwerk Betriebshof
### input_ 520 # dest_count_in### Ölkraftwerk - 597,477 # Oel # dest_capacity 1357
- 0
Runk & Strunk Transports reduziert die Fahrzeugzahl der Linie: Straße Flaschenbier Weißenburg Zweig H - Weißenburg Getränkehandel Betriebshof
#10
Community Discussion / Re: About the Community Manage...
Last post by prissi - October 09, 2026, 07:01:55 AMAlso, the impact of IT on the environment becomes large. When everyone switched from TV to streaming, you drastically increase energy consumption. Various optimisation had brought that down quite a bit, at least concerning of streaming TV.
And while there are always people claiming that innovations by AI will result in savings, it never happens. Almost all technologies which introduced energy savings, were never realised because when úsage became cheaper, people just used them so much more. (The one exception until now is LEDs since they had a 95% saving versus incandecent lamps, and 95% more light is still a challenge). But when gas lights were replaced by electric lights, the built more and brighter lamps, when horses were replace by cars, when letters were replaced by email: Each time the usage increased manifold. Now when thinking is replaced by prompting ...
No, IT will soon consume more energy than worldwide farming!
Of course it goes deeper.
AI as a tool, no problem. But it becomes AI instead of a tool use. With that, I have problems.
Since we can go as offtopic as we like here: Yesterday in the lab at my University, I spotted a student getting very questionable numbers from good data. I asked where those numbers came from. "I pasted the measurement into Chat-GPT." Yes, there is a skill missing, like looking first in a testbook to be able to judge. Or just ask a teacher, I mean we are there for a reason. Or look at the tool maker's software. But no, not thinking, only ChatGPT. And an AI always answers.
In this regard, I worry about AI. It may be an old man's worry; I am pretty sure someone said the same for the internet, the radio, and even about books.
And the third thing is losing sovereignty due to overload. If one cannot follow basic functions, you have a black box. Sometimes, I feel like James is close to the black box threshold on Extended. But this means you let the circumstance dictate your life, even more than it is already needed. Soon, there may be AI versus AI in court battles (especially in the preceeding administrative procedures). Even nowadays, AI generated claims for social security are overrunning the state courts in Germany. So soon, an AI may decide on and AI generated claim.
And while there are always people claiming that innovations by AI will result in savings, it never happens. Almost all technologies which introduced energy savings, were never realised because when úsage became cheaper, people just used them so much more. (The one exception until now is LEDs since they had a 95% saving versus incandecent lamps, and 95% more light is still a challenge). But when gas lights were replaced by electric lights, the built more and brighter lamps, when horses were replace by cars, when letters were replaced by email: Each time the usage increased manifold. Now when thinking is replaced by prompting ...
No, IT will soon consume more energy than worldwide farming!
Of course it goes deeper.
AI as a tool, no problem. But it becomes AI instead of a tool use. With that, I have problems.
Since we can go as offtopic as we like here: Yesterday in the lab at my University, I spotted a student getting very questionable numbers from good data. I asked where those numbers came from. "I pasted the measurement into Chat-GPT." Yes, there is a skill missing, like looking first in a testbook to be able to judge. Or just ask a teacher, I mean we are there for a reason. Or look at the tool maker's software. But no, not thinking, only ChatGPT. And an AI always answers.
In this regard, I worry about AI. It may be an old man's worry; I am pretty sure someone said the same for the internet, the radio, and even about books.
And the third thing is losing sovereignty due to overload. If one cannot follow basic functions, you have a black box. Sometimes, I feel like James is close to the black box threshold on Extended. But this means you let the circumstance dictate your life, even more than it is already needed. Soon, there may be AI versus AI in court battles (especially in the preceeding administrative procedures). Even nowadays, AI generated claims for social security are overrunning the state courts in Germany. So soon, an AI may decide on and AI generated claim.