Imagine having the best players on call?

[MarkShot]

20260112 151429.jpg (288.6 KiB) Viewed 595 times

[MarkShot]

20260112 151455.jpg (363.36 KiB) Viewed 593 times

[MarkShot]

20260112 151523.jpg (285.39 KiB) Viewed 593 times

[MarkShot]

20260112 151544.jpg (267.46 KiB) Viewed 593 times

[MarkShot]

20260112 151607.jpg (308.76 KiB) Viewed 593 times

[MarkShot]

20260112 151626.jpg (380.51 KiB) Viewed 593 times

[MarkShot]

20260112 151704.jpg (372.32 KiB) Viewed 593 times

[MarkShot]

This was actually WITW as I am playing at the moment, but the example holds.

It is possible to build a gaming expert assistant with local AI on your PC. You will probably need about a video card with 12GB of VRAM, and 2-4GB for Windows and your game. This will work for Matrix style map/turn games. It is not going to work for heavy graphic games like City Skylines.

What you need. You can get it done with all one file down load; although more might increase your flexibility:

https://anythingllm.com/ (Contains Ollama, but a separate Ollama will give you more flexibility.)

https://ollama.com/

You will need two types of AI models (and you may get two small ones with AnythingLLM):

(1) An embedding model (usually relatively smaller). It produces lookup keys for your text submissions to the workspace (a database for a particular game).

(2) An inference model (which may be INSTRUCT or THINKING). Use to take the text information returned from the database and format it into something useful and readable.

The beauty of AnythingLLM is that, you can add the following entries to your workspace (database for a game):
- PDF manuals.
- Forum or steam threads.
- Entire WiKis.
* YouTube does not permit you to grab playlist for download and transcription. But for free or possibly $20 USD you can get a lifetime license for playlist grabbling. Another open source, AI tool will transcribe video files into text.
- YouTube tutorials.
- YouTube let's plays.

Once you have built you database(s), you will have an expert player at your side.

It is best to break you workspaces by games and three subcategories.
- Game title
-- Official documents (manuals and user written FAQs) or WiKis
-- YouTube tutorials
-- YouTube let's play

It is very easy once it is all up and running to submit the same question each of the three categories.

---

I thought I would share this as I know some of you just because you play grog games does not mean you have 10 year old PCs.

[MarkShot]

The above mentioned links were open source.

This does playlists downloads. I am not sure if the free option covers playlists:

https://www.4kdownload.com/products/videodownloader-42

---

WhisperGPU (this is what really brings out the power of YouTube)

This does transcripts into multiple language (open source). Actually, I am not sure how to recreate this. It uses Python and something called WhisperGPU. (You will have to figure this one out yourself. It does not run so fast 40XX cards, but it is only supported on CUDA 12.x, and 50XX cards or higher use CUDA 13.x). Sorry, I know how this was bound to my 4090, but I am not really sure of the commands to reinstall under WIN10/11.

[bairdlander2]

I use Co Pilot AI and it works fine, you can export CR and AI analyzes that as well show AI screenshots and ask what would be best attack or defense plan etc. there is a major discussion of the use of AI on the FB group now.

[MarkShot]

Does it understand the game mechanics well enough to provide good advice? (I have to show CoPilot some screen shots and see what it says.)

I have used models like QWEN3-VL, and they certainly see well enough. I don't know if I really want AI playing for me. I decided to draw the line with AI helping me to aggregate and learn better; like a personalized tutor (I am pushing 70; and I am very concerned about AI's potential to trigger dementia; we are going to see a youth epidemic in 5 years of early onset dementia). I used to play with the manuals open on a second monitor, but this digs much deeper than I can, and I find that I am learning the game much better as a result. Especially, with YouTube series as it can tear through 300 hours video (converted vector database transcripts) much better than watching. I don't want to watch 300 hours; just to play.

I mainly chat with CoPilot to get a sense of where commercial AI is; staying on WIN10 and wouldn't let local WIN11 and AI spies near my PC. (I have even locked out NVidia access after public comments about GPU tracking.)

I started playing around with local AI model on a 2023 PC I built with an RTX 4090 (24GB), and rebuilt the PC Summer 2025 before pricing went through the roof and added a headless RTX 6000 PRO (96GB). I mainly use the latter for research and AI projects; the RTX 4090 will more tend to slow down inference if used; unless you get fancy with something like vLLM/docker/LINUX/sharding. But RTX 4090 is good for concurrent side models handling embedding, and enhanced durable memory management.

---

I did try some 1M token long context models (like Nemotron3 and QWEN3) loading up the entire documentation set. But they burn 10-20 minutes at 300W to load and the quality of information is no better than RAG with QWEN3-Embedder 8B and GPT-OSS 120B (generally much faster with up to 50 2K chunks).

---

Finally, I am convinced that GPT-OSS 20B and 120B had Reddit gaming forums in its corpus, since it reasons and responds as you would expect a gamer to. Very different from a reasoning DeepSeek or QWEN3.

---

Lastly, I did try this project with the RTX 4090 in 2024. But 2 years of model development has had a huge impact. In 2024, I always got answers; 30% hallucinations. For 2026, I often get, "Your questions cannot be answered with the information you have provided." A huge improvement.

[Wiedrock]

Could you input complete Forum posts as databases as data and search for stuff? Would/could you make it download all the conversations? Would it work with archived (archived as in the web archives when pages are down) pages?

[Wiedrock]

If it has whole Forums in it and is revisiting Threads several times per search this would explain some of the weird visit numbers of informative Threads.

[MarkShot]

Here is what can be done with AnythingLLM (which is an application built by Mintplex) which uses local or commercial vector databases, and local or commercial AI (LLMs). I am doing all local.

Vector database <- PDFs or any text documents.

Vector database <- WiKis (like Paradox sites).

Vector database <- YouTube Playlists (tutorials and let's play); used third party downloader and AI to make transcripts.

Vector database <- Forums (only worth it for really info dense threads as it scrapes URL by URL input).

But for games like WITW/WITE-2 which I will be playing for the next 10 years, building an expert systems is well worth 2 or 3 days. It is much more powerful than a collection of texts and videos.

PS: Probably not worth is for games in flux like early access, but many of my titles are years into EOL.

[MarkShot]

Yes, I do have professional grade AI hardware (expensive), but you could really build a reasonable implementation on an 8-12GB GPU for games such as WITW/WITE-2. I got like 10 things running right now and 2 display including WITE-2, and my VRAM (WIN10) is less than 2GB in use. Also, the latest AI inference engines (LLAMA.CPP) for Windows can use some VRAM and some RAM/CPU. Even 20 tokens per second, you can get far more detailed answers than looking stuff up. I bought both games 5 years ago. I must say ever since I built this gaming expert, I have been learning many finer points I did not know before.

[Wiedrock]

So is it your AI making those views?

[MarkShot]

The threads I pulled were in the WITW forum; nothing in this. Scraping a thread is a single view.

I looked at one or two AARs here, but they were heavily graphical. My expert only works with text, because the database only handles text.

Video are first turned into text, and the video itself is discarded.

For AARs with many screenshots, I would need to have to write a program for an AI to read the thread along with images (you can use multi-modal models for this like QWEN3-VL 30B) and it transform what the images are showing into a text description combined with thread text contents.

I break my databases into by Game (Manual/WiKi, YouTube Tutorials, YouTube Let's Plays). It is like any computer application. Put in high quality information like 2by3's manual, you get top notch answers. Put in a poorly narrated YouTube videos, you get little of value back. I throw forum stuff in with the let's plays.

Curating your sources is big part of building a good expert. For popular games, you get to know who makes good videos or forum threads.

[Wiedrock]

The threads I pulled were in the WITW forum; nothing in this. Scraping a thread is a single view.

Then I wonder how those views get that high that quick.

[MarkShot]

The threads I pulled were in the WITW forum; nothing in this. Scraping a thread is a single view.

Then I wonder how those views get that high that quick.

Congratulations, Herr Wiedrock! Your analytics of dice rolls are being scraped by OpenAI and the new DOW (Department of War) for the Stargate Project, the USA's next Manhattan Project; an AI WMD. There is only one man, and one company which can save the world from tyranny; that is Gary Grigsby and 2by3 ... "A World Divided - Part II".

Remember always roll for good!

[Wiedrock]

You must be right. Or Palantir is scouting for WW3 accountants!