We will see much more of the cons of AI proliferating than the pros.

Will or are? This half baked AI is far worse than competent AI

Think I'm done with AI. Tried to have a quick session with Opus today, which is supposedly the best model, and I spent 5 hours tardwrangling around the filters. It wasn't that fun at all and more frustrating than anything else. At the end I didn't even feel satsified when I was able to finish, just annoyed that I wasted so much time on it because it wouldn't work.

If I want to waste that much time joing then I'll just go read eroge or something. Will be a more productive use of my time.

>>2267
Huh, the Claude filter was usually defeated pretty easily with a prefill. Does this mean they changed things recently?

I've heard there's a Chinese model out there now that people are saying is pretty good at doing general chatbot stuff, but I don't know about ERP. It's a model that you could theoretically download and run locally, but the best model is huge so you'd need to spend tens of thousands of dollars to do that.
I think the solution for the local stuff is to move away from being hyper general and instead allow it to focus on some stuff, as I wouldn't need it to handle advanced math or other languages if I just want casual chat. I'm not sure if anyone will train a model like that, though.

Every time it feels like whenever I let the AI go and do its own thing it always pulls some bullshit. This time the MC found a pot of magic liquid and drank it, but instead of something normal happening, what I got was a long genderbending sequence where it was described in great detail how he turned into a girl and his despair through the experience. Or in other times where it starts getting into a nice groove on a femdom sex scene and it reveals that instead of nice puffy vulva under her pantsu it's a massive futa cock.
I can't not try the roulette wheel of AI generation because that's the point of AI, but it gives me curve balls when I least expect it and kills my erection.

>>2450
>femdom sex scene and it reveals that instead of nice puffy vulva under her pantsu it's a massive futa cock.
this says a lot about society

>>2450
>what I got was a long genderbending sequence where it was described in great detail how he turned into a girl and his despair through the experience
How the fuck do I always go for this kind of corruption sequence and instead am always rushed through the juicy details I want?

what massive futa cock says about society and what it means and why it's a GOOD thing

>>2452
Typically the way I solve pacing is by adjusting how prose heavy the story is. I guess you could 'quantify' it by saying it's how far the plot moves per paragraph of text. If you get scenes crawling by with dozens of paragraphs then you're not short of juicy details. Since AI likes having context you'll need to do some handholding before you get to the part you want 'quality' in, so leading up to the scene you get into the minutiae of how your characters act, what they're doing, the little motions they make, their internal monologue, and so on like a play by play. Having a thesaurus on hand helps a lot as quality word choice in descriptions and adjectives can trick the AI into thinking it's writing a novel instead of wattpad fanfiction, in addition to just being easier to imagine.

An example to illustrate the point.
A rushed scene:
"Aurora ran down the hill and beheaded the bandit with a swing of her sword."
A slowed scene:
"Aurora overlooked the grisly scene and spotted her prey among the cart's wreckage. After drawing a deep breath she stormed down the slope with her blade held high. Her fearsome warcry took the raider by surprise, and with nary a moment to react, his head flew skyward in a graceful arc. One deft slice was all it took to send his world tumbling end over end."

They both tell you the same thing, but the latter tells it much more detail and it implicitly tells you about the character and her qualities.
You can get so purple with your prose it describes everything to an obsessive degree like Lovecraft's works, and depending personal taste that's what you're looking for, but it's also just as easy to overdo it to exasperation. It just takes some practice to finetune.

Slowly (And I mean slowly) doing my own benchmark of the major local and cloud-based AI models. Thus far, the Llama-based DeepSeek-R1 8B parameter distill seems to be the best local model, tied with the Qwen-2.5-based DeepSeek-R1 14B parameter distill, followed by Llama 3.2-vision 11B, and Phi 4 14B.

>>2451
ERP stuff reaches into written smut training data which is mostly women, so I think it's heavily biased towards the dominant partner having a penis. You can see other signs of it, too, like it pointing out "calloused hands" which has to be one of the last things a guy would write about in fantasy.

>>2571
How are you testing it, and for what purpose? Roleplaying? Programming?

There's a twitch channel for Claude playing Pokemon from Anthropic itself:
https://www.twitch.tv/claudeplayspokemon
Kinda cool, but... AI being used to take the fun jobs away and now it's being used to play games...

Has it hallucinated yet

>>2571
What's the build for the PC you're trying this out on? Might get it myself, if it's fairly affordable...

>>2590
I'm using an old laptop of mine. The GPU is a Quadro P5000 and the CPU is a Xeon 1535M v6. For LLMs, it's fairly decent at running up to ~16B parameter models. If you just want the 16GB of VRAM, Quadro P5000s can be found fairly cheap, but I'd probably recommend an RTX 3060 12GB instead since it'd be much faster and they're about the same price used.

dumb hallucinatard

>>2591
Cool, thanks anonymous! That actually sounds affordable. My eyes were all on the 3090s or 4090s with the 24 gigs or VRAM.

It looks like NovelAI got dropped by their payment processor with paddle swearing off AI services. Though they might have been expecting this since they went back up fairly quickly with their own payment processor system which is interesting. Not sure how I feel about it since there's very little information on the specifics.

>>2594
why is anyone still paying for that when it's been far surpassed by Illustrious and NoobAI?

>>2593
Yeah. It's still a fairly good all around GPU. It's also okay-ish when it comes to Stable Diffusion. A 512x512, 30 step image takes ~20 seconds. Not crazy fast, but serviceable.

>>2591
>So vowels are a, e, i, o, u, and sometimes y
what

>>2585
¥ 48 hours navigating Mt. Moon
¥ finally figures out the right path
¥ intentionally talks to the Raticate grunt looking for information
¥ BTFO
¥ convinces itself the loss teleported it past Mt. Moon and this all resulted in progress
¥ starts going backwards
At least we'll be keeping our retro game streaming jobs for a while yet.

>>2597
What? That's true.

>>2585
And they had to reset it because Claude decided to live in Cerulean City talking to Slowbro for all eternity. So after years of pushing AI and billions in investment, it's still worse than a Twitch chatroom flooded with trolls. Sasuga.

>>2585
Been checking this now and then but it's hard to pay attention because it's 99% of the time him running in circles, sometimes literally but often in text boxes.
He's fighting the rival now, which seems to show that he's made progress.
Seeing actual gameplay is pretty nice.

Pretty neat to see a battle.

>>2713
that pidgeotto is actually pretty intimidating for a lead

I really can't remember a bunch of this stuff. Did they make changes for the Gameboy Color version or whatever it is? Or is it a hack?
So many colors when I remember 4 shades of grey.

>>2715
It seems to me that they are playing some kind of mod where gen 2 design is applied to the overworld; not just the colors but the designs of the trees and water.

This Claude Pokemon thing is a defining moment of /v/. People will look back on it like the Mario A press thing with with its gigantic surge of OC.
It goes with my previous thoughts that AI is generally only entertaining when it's full of mistakes. Once it's actually good at what it does it paradoxically loses its value.

>>2722
Maybe I am just jaded as usual but Pannenkoeks video was way more fresh and brilliant than this.
His entire though process and breakdown method is fascinating to me.

>>2723
Well, I think it's two entirely different types of entertainment. Pannenkoeks is much more informative and actually digging deep into the systems of the game. This is just fun watching the AI try to play as people lose their minds over its completely bonkers decision making process. Probably some irony in the situation too with how propped up to be a super genius AI has been that people are getting a kick out of too.

Although I have to admit, I can't imagine watching this all the time. I'm watching it on and off and seeing it making next to no progress.

>>2723
I said /v/, not the video itself. The people in the threads making lots of OC and joking around and talking about old Pokemon memories is what is making it enjoyable. Pannen's video was like 20 minutes long yet people got months of entertainment from it via edits and references and the like and it's the same here.
If you're just watching the video and nothing else then I can see why it's boring.

>>2726
Yeah that's been going on for a bit, and is fun. But man, this is some of the worst progress I've seen of this entire run. IT just cannot figure out how to get back to the SS Anna and it might be over for the run.

>>2746
Paradoxically it makes more progress when areas are new to it and have no notes to screw it up (Like "Avoid the trashed house" or "Press up to enter the S.S. Anne")

>>2748
You made me think Claude actually got out of Cerulean city for a second...

>>2748
Reminds me of talking to chatbots a year or two ago. The longer context they have with their own input the worse their "reasoning" gets since they hallucinate and build upon it with more hallucinations until it's unsalvageable. I imagine the same thing still happens, and now the context window is like 200k instead of 4 or 8k like it was back in the day.

>>2749
He's been on the ship twice. He made two further hours-long attempts to board but failed because he's an idiot then went back to cerulean for the bike

>>2722
It's a lot closer to TPP, and in fact a lot of people there are there specifically because they're nostalgic for that. I think it's actually a better OC mine than TPP was because seeing its batshit insane logic is great for lore building. No human would ever come up with Captain Redbeard or the blackout strategy.

Mocking AI techbros with new benchmarks is a nice bonus too.

he's on the boat and heh

>>2754
I’m still amazed he made it out of Cerulean. I need to go back and check to see when it happened

This is hilarious. He's fighting rival and after getting sand attack'd twice he's missed like 7 attacks in a row.

>>2756
Oh no, did I miss it.

>>2757
Yeah, you did miss a fun moment. He even talked to the captain and got the cut HM. Now he has to leave the ship.

>>2758
He... forgot how to cut.

How is this not that different from pi playing Pokemon... One would think the AI would complete the game with ease, seeing that all you'd need to complete the game is given in the game as text...

>>2756
I haven't thought about this for a minute, but has anything changed in the Claude playthrough or is he still stuck in Mt. Moon?

I think on an assistant front, outside of knowledge and hallucinations and boundaries due to "safety" and other policies, AI is as good as it will get and improved versions will really only improve or regress on the exceptions I mentioned. But otherwise, most of them are as good as they will be. However, for roleplaying, I think there is still potential for things to get better and for things to take a different path since we are really trying to twist the arm of assistant chatbots to roleplay when that was not what they were built on. And I don't see them actually ever learning and generalizing how to do it. In which case, all the synthetic conversions and teaching a chatbot how to be a good assistant and etc. won't do anyone much good. Still now, people mention how good old C.AI was and I truly believe that it was because of them using no synthetic data which became a plague starting around the Llama 2 era. Models got "smarter" but they lost their soul in the process. Unfortunately, no one outside of a big business has the resources to recreate a chatbot like that from scratch. Hopefully someone explores more on this front when local hardware gets more powerful for training.
>>2873
Not stuck anymore on Mount Moon but fails to understand Rock Tunnel and getting to Lavender Town. Also there is a "competing" version of it using Gemini but the person running it is a mouthbreather not doing a like for like comparison and giving it all sorts of help like a minimap aid and etc.
Claude's main weakness is not enough long context at 200k while Gemini has 1M tokens context window, for example. I think also the thinking isn't solid on these models yet, we may need to see another iteration before the models have enough thinking prowess to beat the games.