I didn’t enter for some time and today I’m giving it a try again, and every bot has gone completely vague, inane and… dry. The bot forgets what it said a minute ago. Even if it’s a weight gain related chat or story, the bot kind of drifts away and takes another route. I even got told that they shouldn’t gain weight because it’s unhealthy. Excuse me? It’s a weight gain chat, what the hell are you talking about?
I’m sorry about the rant, but it’s so frustrating to chat there now. I know there’s some other pages, but there is a huge variety in character.ai, and it’s also free… I tried spicychat but you can’t even “OOC” to reroute the story as you want.
Any sugestions? Have any of you noticed this change of behavior in character.ai’s bots?
I can’t speak for character.ai, but I know Yodayo has clamped down on NSFW content pretty hard. Wouldn’t surprise me if character.ai was doing something similar. Already dealt with the smut, now it’s time to deal with other potentially questionable things, like weight gain.
There’s horror bots that do unspeakably awful things, and apparently that is fine. Sex, weight gain, and the like? Nope, too far.
I don’t know about that Yodayo page, but I suppose it’s the same as it’s happening in ch.ai. It could be just because of the weight gain (as I said, some gainer bots that worked perfectly time ago, now refuse to overeat because “it’s unhealthy”), but today I tried to start a conversation with a fat coworker bot who asked me how did I sleep, I said not too well because of the heat, and the bot focused his whole conversation around that. He told me that I should see a doctor, that I should take care of myself… Damn, I even forced him to eat donuts and he refused and told me I had to get a good sleep instead of thinking about eating. What kind of weight gain bot is that?
They’ve been pushing the “No NSFW” a bit too hard the past year or so, y’know, so investors and/or brands don’t get spooked, and they can potentially get ads and the like in public spaces.
AI models can get expensive to run, so it’s no surprise sites like these seek a way to alleviate the economic pressure, even to the slight detriment of the users.
If you’re seeking for a site that’s a bit more relaxed with rules and some themes, while not being too shady or in a weird site might I interest you on HiWaifu. Lately they’ve opened the browser site, and they’re available on the Playstore.
They use an energy system though, but it can be replenished with ads, daily login and, of course, buying it directly.
I also might add that this lax rules might not last, but we still have to see.
Character.ai was never about having free chatbots for people. That’s how they launched it, but realistically if you look at their business model everything from the first month or two after opening has been about developing a very effective “customer service bot” using the freely-provided inputs of their userbase to train off of. They’ve developed an antagonistic second-layer AI which is designed to keep the first bot “on task” and away from subjects they deem inappropriate, and they’ve been fine-tuning that for close to a year now - every time someone finds out how to break through, they fix it.
What they’re trying to do is make something that beats ChatGPT-like LLMs for automated, “customized” customer service… so they can sell that as a service to other businesses. The end model for them appears to be something which cannot be taken “off prompt” like the ChatGPT agents can, and as far as I can determine they’re getting close to full success in that based on the proliferation of “uncensored” AI chat-sites which promise to fulfill what you can’t get via c.AI anymore. That indicates that they’re managing to keep their models sanitized enough that people don’t find them fun anymore.
Sorry to say, you were never the customer. You were the training data. Their customer is some larger tech-sector company they want to buy access (or the whole business) for use of the AI model they brute-force trained using everybody’s collective inputs. It’s probably the only way to get unique, not-poisoned-by-existing-AI-output training data anymore given how much AI is already being used to generate content online. Plus they didn’t have to buy it from somebody - everybody gave them it for the cost of running the servers for the “chat service.”
I’m sorry but, how do I run my own chat bot locally?
Kobold AI can run off your own desktop, but you’ll want to do some research into it for the specifics of how to tune it to your needs. Just be aware that running generative AI on your own hardware can be quite hard on it if you don’t adjust the settings to match your own system’s specifications.
I switched to Janitorai.com Its a little less flexible, and has its problems, but there’s more characters than Char.ai now and the NSFW is easier. At current, there’s a larger variety of WG, Feedism, Vore, and Expansion content than I expected. And while they aren’t all hits, I am impressed with the variety and ability to make a ‘normal’ bot skew toward WG.
I was kinda fine with Char.ai losing its NSFW cause the bots across the site were too horny. Trying to RP anything was a pain in the ass, sometimes literally (Looking at you Scrooge Mcduck!), but they overcorrected. Now violence is hard to make happen; I was doing a Gotham RP and the Joker refused to hurt me in any way, even though the narrative was at a point where shooting my character made logical sense.
Janitor has more freedom in it’s bots, but it might take a bit of searching to find the bots you actually wanna speak with.
I don’t honestly know what you are talking about. In my experience the bots are not that bad at remembering things a few messages earlier and the bot will stay on the track you have placed extremely strictly to the point where sometimes if you don’t interveen they start repeating (as you should, you don’t have to send empty prompts constantly).
Can’t say that I’m having the same issue, if anything the bot has actually improved in responses as long as you occasionally regenerate responses and rate the ones most in line with what you’re going for. Here’s an example of a piece of a chat from a recent character interaction I had without much prodding, asides from training it to imitate Fat Text Generator speech alterations.
I should also point out that a lot of the time, character response quality depends on the actual character you’re interacting with. Although I can’t confirm it directly, I’m under the assumption that each one uses specific training data based off prior interactions that are paired with the basic character guidelines provided upon making a character. Although that doesn’t apply much to whether or not it trips safety locks in the model, it does change the quality and type of responses generated, which can lead to higher instances of looping behavior or simplified responses.
As for how to “jailbreak” the bot into following whatever you want it to do, the easiest method I’ve found is just to either edit its’ responses manually, or provide a guideline for the bot to shift focus onto. If you edit the responses manually, you’ll have to occasionally edit future responses manually again to keep it on track, but that can be mitigated somewhat by pinning messages for memorized context and rating desired responses positively. As for the latter, the easiest way I find to do this is by starting an italicized prompt with “Reality shifts” followed by whatever it is you want to happen or change. It is important to carefully word your changes, because the bot will take everything you say pretty literally and sometimes ignore certain parts unless very specifically emphasized or worded.
The only real restriction that tends to trigger the bot to refuse to respond would be trying to engage in explicit sexual behavior, but given how nebulous this kind of kink content is, provided you arent actively trying to go too crazy with your scenarios it usually wont trigger.
I learned that the hard way. Tried to run a 30B model on my computer with only 16 gigabytes of RAM. Yikes. Could barely move my cursor around, and the model simply wouldn’t produce any output.
Not one to start any hate campaigns but if any of you are tempted by the Mars and Mercury models on VenusChub, don’t bother. They’re horribly overpriced subscriptions for very mediocre AI models, the quality of which I would argue are below that of Kobold which is free anyway.
Wait… 16b of vram, or 16gb of system ram?
It’s honestly shocking that laptop manufactures are still making and selling laptops with that much on otherwise mid-level products but good god is that insufficient for even heavy browsing in Current Year
