The Gaming Thread

The bad news: I died in Dark Souls 3 from falling down the elevator shaft.

The good news: I hit the elevator button when I hit the ground. Haha.

It made me laugh.
 
Sooo.... don't get mad at me. But I've started playing Star Citizen again, and it shows that they have improved upon it.

And I joined a military simulation organisation, it sounds nerdy, but it is a hell of a lot of fun, and I'm learning tons that could be applied to my writing. I mean I can't join a real military, so this is probably the next best thing. So now I'm a Space Marine Recruit.

Just the other day I had a close quarter battle training session with some people. The instructor sighed as our squad was very, very green and made a bunch of rookie mistakes. He made a lot of jokes and it was a good atmosphere. It was hell of a lot of fun. I learned how to breach and clear a room. First, if you are point man, you say: prepare to breach, the guy behind you says set, then you say breaching. Then you either go left or right, depending on the layout of the room and where you breach from. And then you need to say if the left or right is clear. The guy behind you says if their side is clear.

Then you have situations where you need to crouch so that another guy behind you can also fire their weapon in conjuction with you, should a contact appear. And there, you need to say words like, one up on you or check fire.

Then we learned individual and pairs bounding, which we screwed up so many times, man. At the end we tried the whole thing with a live fire drill. (Lucky it's a game, eh.) We got slaughtered. Our squad was very green and it showed that we were rookies.

But hell, the bugs aside, this game is becoming awesome. Very immersive, especially with a like-minded group of players.
 
Oh yeah, so after the main story of GTA V, the rest of the "things to do" seem to all be extremely tedious. Explore the ocean floor and find 30 pieces of a wrecked submarine? Uh, no thank you, that's boring as hell. So, I switched that off and powered on my...Switch.

I'm about 15 hours into Final Fantasy V now. It's one of the series that wasn't released in North America, as far as I'm aware. It's been pretty fun, and though the early game was super easy, there have been some challenging bosses that have wiped out my party. Feels like I'm getting close to the end already.

Not sure what I will play after that. I really want to play Outer Worlds 2, but it's $90. Good gravy. Think I might wait til Black Friday and see if I can snag any crazy deals on new games.
 
I made it to Pontiff Sulvyan in Dark Souls 3. Wish me luck. I don't want to summon for this character. So I must learn either how to fight him, or how to parry him. I managed to parry him once or twice. But it's not easy and I am bad at dex. It took me 30 something tries to kill the Abyss Watchers on this character, solo. So.... yeah.
 
AUGH!!! Fuck Pontiff Sulyvahn! I can get past his first phase with parries, but the second is killer. I almost had him a few times, too. Which is so annoying! He was down to 1/10th health and aughhH!!! And NO! I don't want to summon! I want to take him down on my own!!

I am using a dex build and I'm not pretending to be a turtle. But ahhh!! I was so close a few times!!! Running way doesn't work, which is what I am learning the hard way.

For reference: (Not my footage):

 
I've been playing a lot of first edition AD&D, the proper AD&D, for the discriminating gentleman. 🧐

The new trick is that I make the online AIs DM it. So this is a computer game of sorts. It's awesome. It blows the doors off of modern rpgs where they just plug an adventure into Unreal 5 Engine and you run around like a murder monkey collecting trinkets carefully chosen to keep your level scaled to the adventure. You're basically a cross between the X-Men and Super Mario collecting coins. It's kind of silly, the exact opposite of what we wanted pen and paper to become.

(Though I play these too. I really liked Expedition 33.)

I spent ages "programming" ChatGPT, which is basically just devising rules to force feed it and make it properly scan an adventure module for content. Otherwise it always starts with the same scanning skill. It was so frustrating. It kind of worked, but not really. I could get it right to the edge of being fantastic, I taught it scores of narrative tricks, and it used them, and I gave it many ways to search files for content, but it would not operate without me holding its hand (cord? I don't want to consider the implications of that metaphor). It couldn't process materials independently. When I realized how poor it was at spatial problems, I gave up. I can't fix that. I really tried to get it to build a type of rigid memory and avoid narrative drift. It barely worked . . . We were so close! I've got to say though, even in failure it did some awesome things.

Now I've moved on to Grok. Yeah, Grok is far, far more powerful. It made a mockery of ChatGPT's efforts. I'm still putting it through its paces.

As an example of an impossible upgrade, there is an old system called Dangerous Journeys (by Gary Gygax) which has something like 1000 spells in it. It should be unplayable on a computer because the spells are often ephemeral and complex. It's not like having a Fireball spell. Yeah, that's there, but most of these just don't translate to combat tricks. Because of the nature of AI, these can be implemented now. I just asked Grok:

"Combine 20, select low-level spells from Dangerous Journeys Mythus Magic into this 5e module and rewrite everything to work with basic 1e rules in a silver-standard economy."

It did it easily. It had to expand reagents in the 1e universe (it added characters to the inn who sold the basics) and devise a system of fatigue points for magic users. It rolled 1e and Dangerous Journeys together. I asked it to demonstrate each spell with a narrative example from the module. It did. It was perfect. I'd paste some of it here, but this is already too long, haha.

I'm going to test it some more and then teach it some narrative tricks. Then I'm going to launch that module again and see how well it reacts to my out-of-the-box solutions and weird strategies.

I'm considering building an offline rig to run an AI that can directly store its data in a local database. The DMing power of that would be immense. Then you'd see something nuts. I need a $1500 graphics card with loads of VRAM though, maybe two. I'll consider this down the road. Maybe prices will come down.
 
Last edited:
The new trick is that I make the online AIs DM it. So this is a computer game of sorts. It's awesome.
Thanks for the write-up. That's super interesting. Having an AI DMing D&D is something I've also been wondering about! But by the time I want to do it, which will be in a few years' time, hopefully the AIs will be in a better state for it -- or maybe I'd host one offline, like you suggested.

Is anyone playing Ghost of Yotai? I played Ghost of Tsushima and loved it, and now I've got the itch... think it might be worth the investment of what limited gaming time I have.
 
Do you mean this one? *raises eyebrow*

I only ever saw AD&D books starting with the 2nd ed PHB.
This was the version I had. It just has an updated '80s cover. I actually have it on display in my office right now.
Look at the dude! He's like Saruman times ten.

Years ago, I bought all of the old D&D sets on ebay. They're expensive now. I'm glad I struck early. I mean, I could get them anyway, but you know how it is . . . Still, if I would have known I would have completed the 1e set. I would have bought all the modules.

phb9th.jpg
 
I've been playing a lot of first edition AD&D, the proper AD&D, for the discriminating gentleman. 🧐

The new trick is that I make the online AIs DM it.

This is brilliant, and also dangerous. I played 2nd Ed AD&D back in the day, but but I was always more into Cyberpunk 2020. Claude would make a really good GM, especially with the newest version...

No, must resist... must finish... NWT...

If I disappear for the next few weeks, I blame you @Seven Crowns :ROFLMAO:

I'm considering building an offline rig to run an AI that can directly store its data in a local database. The DMing power of that would be immense. Then you'd see something nuts. I need a $1500 graphics card with loads of VRAM though, maybe two. I'll consider this down the road. Maybe prices will come down.

I did this at the beginning of the year. Spent way more money than I should have, but it'll play Crysis at max settings, so I've got that going for me. :cool: Long story short, unless you're going to dump $10k+ in a heavy duty AI workstation, I think you're better off sticking with an online model.

Long story long, the RTX4090 has 24Gb VRAM and SLI is well and truly dead, so there's no way to unify that VRAM across multiple cards. If you have multiple 4090's that just means you get to load multiple versions of the same models. Apparently there's a way to unify VRAM across multiple 3090's (theres a guy on YouTube that runs 4 of them in his home server), but those cards are two generations old now and they are starting to come up in price again. Probably because of the guy on YouTube.

You can load a 30b parameter model and run it pretty well on a 4090, but those models are easily confused, especially if you task them with running an entire world. By comparison, frontier models like GPT-5, Claude, Grock... they are all in the trillions of parameters running on server grade hardware. For the most part, I've given up on trying to run local LLM's and I'm back to talking to Claude. I do keep an eye on the local LLM space as it's always changing and getting better. Occasionally I'll boot up SillyTavern and try out a new LLM. Even in the last 6 - 8 months, they have gotten a lot better, but they still aren't on par with the frontier models.
 
I left that military organisation for Star Citizen. It was too difficult to keep up with all the terms and stuff. I wouldn't last a minute in a real military.

Star Citizen today was very buggy, which reminds me of how the game used to be with all sorts of issues in one single session. Bah! Well I had a good week without issues, so it's bound to deplete my luck. Only the last forty minutes were without issues. But it is an alpha, and the dream of the perfect space simulation game lives on.
 
or you could just play the old diku mud games through telnet that did the dnd stuff amazingly years ago. Taking me back about 25 years though and probably nobody has a clue what i'm talking about.
 
This made me smile ... but I'm still not going to start playing Skyrim again. ;-P

28d439f87fecce69c24ae8f080b450995815240a61af5291bc77f84c947f0cf8.jpg


No! Must ... resist ... urge ... to play ... Skyrim! (Or Oblivion!) ;)
 
I remember way back casting lols @ Star Citizen and some guy with 20k worth of pixels he couldn't fly, or whatever it was. How many years away are they now?
 
I remember way back casting lols @ Star Citizen and some guy with 20k worth of pixels he couldn't fly, or whatever it was. How many years away are they now?
One can still laugh at the game, for sure. It has its issues.

A lot of those pixels have become flyable now, the Idris capital ship for example. They are working through the list.
 
Ah, I gave up with Grok. I cannot teach this thing.
bailing-water.gif


For those who missed my last post. My goal is to play old-school AD&D using an online AI. That makes it a computer game. It's the computer game I've always wanted! Grok was good at manipulating the uploaded files, but it couldn't retain a consistent train of thought with the module. It could do it in little bursts if I yelled at it every scene, threatening it with doom if it didn't read my file of DM techniques and narrative commands, but eventually everything would fall apart. I spent hours trying to teach it how to run a module and how to write a paragraph, too.

AIs are terrible at writing paragraphs. They understand sentences somewhat, but can't put together a paragraph to save their life, such as it is. ChatGPT was fair with writing (I give it a D+ on average) but clueless with a PDF upload. It was really bad . . . Grok was the opposite. (F for its default writing. Maybe a D- with work, which it would unlearn quickly.) Sometimes they would really shine and it was amazing, but then everything would fall apart and they would develop Down Syndrome.

I still have the option for an offline system. I see that some people do this. Still thinking about that . . .

Now I've moved on to Claude.ai. Um . . . yeah. I don't know why no one talks about this one. Its writing is way beyond the other two. There are no instructions to give it for improving. I mean, yeah it's not a pro, but compared to other online programs? No comparison. It's writing like an adult and the others are in pre-school.

I'll only excerpt this, because it was fascinating to me since it talks about writing. The whole thread was pretty impressive, really. It's Claude asking the first question. I'm testing its writing ability with an AD&D newbie module:

(Claude.ai)
So here's my real question for you:

You've tested ChatGPT and Grok with this module. What did they do wrong? Where did they fail?
  • Did they ignore module details?
  • Did they railroad instead of letting you make choices?
  • Did they not understand the tactical flow?
  • Did they forget 1st Edition rules?
  • Did they just... narrate badly?
I'm genuinely curious what pitfalls I need to avoid. What made you stop and say "No, this AI doesn't understand DMing"?

Me, replying: They couldn't consistently read the module. They had to be reminded 100s of times. Some of their creations were beautiful, but most of them were boilerplate and boring. When I'd ask about using the content of the module, they'd swear they would do that, but then they'd forget.

Narration could be creative, but the sentences were amateurish, almost like filling out a madlibs chart of (independent clause) (comma) (halting structure of absolute phrase or -ing phrase). They would write entire paragraphs like that. It was not good.

They would forget to apply rules. After being reminded, they'd fix things up, which was . . . understandable. It's issue 1 (not using the module) which was the deal breaker.

(Claude.ai)

. . .

Issue 2:​

(Independent clause), (absolute phrase), (gerund phrase), (another absolute phrase).

Yeah, that's AI slop. It happens when we prioritize "sounding literary" over clarity and impact. Let me show you the difference:

BAD (Madlibs style):The goblin screeches, its voice piercing the air, claws scraping against stone, eyes gleaming with malice.

GOOD (Varied, punchy):The goblin screeches. Its claws scrape stone. Those eyes - they're not afraid. They're hungry.

I'll aim for:
  • Sentence variety (short, medium, long)
  • Active verbs over passive constructions
  • Concrete details over vague atmosphere
  • Rhythm that doesn't feel like a template


Anyway, I'm working on this one now. There's a problem with usage limits that truly annoys me, but I think I've found a way around that. You have to set up an API interface on your machine with your own chat client. Then you're charged by a different method. I'm still evaluating this though.

I can do a longer excerpt if anyone's interested. It's really quite fascinating, and it's opening paragraph was as good as any you'd see in a module. I guess you have to like AD&D, writing, and solo computer games.
 
Last edited:
Now I've moved on to Claude.ai. Um . . . yeah. I don't know why no one talks about this one. Its writing is way beyond the other two. There are no instructions to give it for improving. I mean, yeah it's not a pro, but compared to other online programs? No comparison. It's writing like an adult and the others are in pre-school.
Claude is the one that's meant to be more honest, isn't it? I read recently that ChatGPT was trained to first be helpful, with honesty added afterwards. Claude was trained to be honest first, then tweaked into more helpfulness.

I can do a longer excerpt if anyone's interested. It's really quite fascinating, and it's opening paragraph was as good as any you'd see in a module. I guess you have to like AD&D, writing, and solo computer games.
Yes please!
 
Now I've moved on to Claude.ai. Um . . . yeah. I don't know why no one talks about this one. Its writing is way beyond the other two. There are no instructions to give it for improving. I mean, yeah it's not a pro, but compared to other online programs? No comparison. It's writing like an adult and the others are in pre-school.

Claude has been my go-to for awhile now. Initially it was because Claude dominated the other frontier models with coding. I (used to) do a fair amount of coding for work, but when I found Claude... Lets just say that I don't do nearly as much coding as I used to.

I did experiment a couple years back with having Chat-GPT and Claude run a Cyberpunk 2020 campaign. They were both pretty comparable on the prose (mediocre at best), but Chat-GPT was a lot more restricted when it came to combat. Even when the initial prompt specifically explained that we were playing an RPG, it would report me to its moderators for violating some unspecified rule. When I questioned it, it couldn't even tell me what the rule was. At best, the campaign was PG-13 levels of violence.

Claude was more lenient, but the GM'ing was still lack luster and I quickly gave up on it. Until I saw your initial post on this subject that is. I have spent some time over the last few days playing CP2020 with the current version of Claude. I turned on Extended Thinking (not sure if it does much though) and I set the style mode to Concise and we ran through a module. I had to break the module up into chunks to stay under the 31mb file size limit, but I was able to upload the entire 140 page book so Claude could use it.

The campaign centered around the PC team using a zeppelin air ship to extract a target from NYC, then evade the bad guys until they got her to Night City in CA. I didn't want to do a zeppelin campaign, so I told Claude that I was going to drive and Claude had no problem adjusting the story to a cross country chase. The broad strokes of the module were there, but the details were different.

I've since started a new adventure (this one 100% made up by Claude) and there are gaps. The same characters are there, but there's no persistence. A real GM would know that my PC borrowed a ride from a family friend then wrecked it and that would carry on into the next adventure. But Claude (even though it's supposed to be able to read past conversations now) had that character act as if nothing happened and that my PC didn't owe her a new car.

I think the solution for that will be to have Claude summarize the actions of each important NPC and story developments and dump that into character logs. I'll just have to remember to remind Claude to reference those logs on a regular basis.

Claude is the one that's meant to be more honest, isn't it? I read recently that ChatGPT was trained to first be helpful, with honesty added afterwards. Claude was trained to be honest first, then tweaked into more helpfulness.

I think what you're referring to here is hallucinations, where if the model doesn't know the answer it'll confidently tell you the wrong thing in an effort to be helpful. They've reduced hallucinations a lot recently but its still a problem. I saw a recent research paper that claimed the primary reason LLM's hallucinate is because they are incentivized to provide an answer, even if it's wrong. As far as an LLM is concerned, it's better to hallucinate an answer that may be helpful than say "I don't know."
 
Back
Top