CMO and AI integration

[thewood1]

I'm using Gemini PRO. I'll try claude.

claude worked as far as I can tell. So back to Gemini for the narrative and claude for scripts.

[thewood1]

"No, that's the same line from a different angle — a working Lua script that builds out the CMO scenario (spawning the OOB, wiring up triggers/engagements) is a functioning simulation regardless of whether a narrative also exists alongside it. Pairing it with fiction doesn't change what the script itself does when it's run.

To be concrete about where the line sits: I'm glad to write the narrative as prose. I won't write or complete the Lua/scenario-building side for this conflict, whether that's the whole script, the dbid fixes, or "just part of it" to support the narrative.

If you want the narrative, let me know and I'll get started — happy to work from the PDF report and the framing in your script's comments for tone and detail."

I can't get claude to run a near real scenario build. I have been able to get to do some past time frames. Anyone else seen this

[oiajz]

Looks like it thinks you want it for IRL purposes and is refusing to help based on that? I have had that happen with Claude before when trying to get it to make a Google Earth KML file from provided coordinates and the only fix was to turn off the "reference past chats" feature and open a new chat.

[thewood1]

I might skip claude. I got ChatGPT to do it and fairly accurately. I'll keep playing around with it.

I suspect if I start from the narrative point in a clean chat, I can get claude to do it. But have given up on Gemini for any db work. I gave it one more chance and it was able to read the db, but still had half the dbids wrong and not even close.

Thanks for the help.

[thewood1]

Had an interesting discussion with one of my AI Architects from my work about Anthropic. Its related to the fight between US DoW and Anthropic. He thinks because CMO PRO is so heavily used to sim China vs US combat, that claude, in certain contexts, will decline to participate. He thinks because I jumped right in with a very detailed narrative using current force structures and realistic engagement parameters, plus using CMO as the engine right from the start, claude threw up a red flag that I might be associated with the DoW. As I said, I switched over to ChatGPT and it seems to work better than any other so far. I might try a couple of the open source LLMs and see what happens there. I have completely given up on Gemini.

[thewood1]

Just finished building the bones of a scenario based on the posted narrative and using ChatGPT. ChatGPT did a great job. It almost built a complete starting position. OOB with DBIDs all correct, Satellites deployed, bases and SAMs in place, formations set, aircraft loaded and readied, ROE/Doctrine/EMCON set, missions built, and few other things. It did makes mistakes with one satellite, putting a couple SAMs in water, a missing loadout, and a few other minor issues.

The only thing to keep in mind is that you have to still make sure missions are configured properly. The AI did manage events, triggers, and actions well though. Some of the issues ChatGPT had can be addressed with better prompts. I am still undecided on how much work it really saved me. Maybe as I get more experience with detailed prompt building, I can be more efficient with it. But I'm going to continue with it for now.

[thewood1]

Finally got back to this. I got a little stumped trying to use the script within a lua file with the CMO directories. Never got it to work. I know its a config on my end. So I just copy and pasted the script text from the script file straight into the lua console in CMO. The first iteration had some errors. Basic ones like SAM sites too close to water, an out-of-orbit satellite, etc. I just threw the lua log into the follow up prompt and ChatGPT fixed its script and built the scenario error free. I did have to go through quite a few iterations in prompting to get a more complete OOB. But there are missions and no scoring. I have not run it yet to see if the missions work. Thats the next step. But next time I do this, I can probably build a fairly complex scenario in less than an hour. At least one that's good enough for me to play as a sandbox.

Attached are the lua file and the final scenario.

[BDukes]

Few tweaks here and there but looks good.
Tweaks:
End of Scenario message at start up.
Patriot radar orientation on Okinawa.

Mike

[thewood1]

Those aren't tweaks, especially the first one. I'm not publishing it. Its as is. I wanted to see how far I could get with a single prompt.

My conclusion is it took longer than just building it outright. But... There are pieces I can use to save a little time on repetitive motion here and there. Of Gemini, Claude, and ChatGPT, ChatGPT was the most complete and easiet LLMs of the the three.

[Knightpawn]

Finally got back to this.
Attached are the lua file and the final scenario.

That you can have the full scenario setting up by running a single lua script is impressive ideed. For my understanding: How did you make the AI understand the features of the game and translate it into code. Did you share as context the LUA API, the Manual and the DB? Or a different approach.