259
China bans AI “boyfriends” and “girlfriends” over addiction and birth rate concerns
(www.dexerto.com)
this post was submitted on 19 Jul 2026
259 points (98.1% liked)
Technology
43274 readers
267 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 4 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Dunno, as of yet. I am very picky, and my roleplaying sessions with AI hasn't gone on long enough.
0000
It requires significant hardware to use local AI, and Qwen 3.6 at 35b isn't quite good enough to understand the lore and characters. Odds are that once I have added VRAM & RAM to my PC, I can run a big enough model to have enough quality to feel good for me.
Unfortunately, I will likely have to wait for America's AI bubble to have popped, before I can afford enough RDIMM RAM to run Kimi and other large Euro-Chinese open models. I probably won't ever use American anymore, as their closed models have too many guardrails and likely would align with the Trump Regime in the coming years.
What kind of token rate do you get if you try to run them in swap space?
Never tried that. I worry about the damage and lifespan reduction that could cause to my SSD drives. If we go vanilla, I get the following performance on my PC:
000000
Generating (2562 / 8192 tokens) (EOS token triggered! ID:248046) [11:17:40] CtxLimit:2656/32768, Init:0.10s, Processed:94 in 0.83s (113.12T/s), Generated:2562/8192 in 47.51s (53.92T/s), Total:48.44s
000000
This is via the KoboldCPP backend. Things are much faster than they used to be, a year ago this would have taken at least 10 minutes. This is a 35b MOE model.
Specs-wise, my machine is...
7965WX - STR5 workstation CPU ($1,500, used)
32gb of RDIMM DDR5 5600 ($600, used)
RTX 4090
0000000
Within a few weeks, I will have the machine worked on to fix a teething issue and to add at least a second GPU. My machine has device enumeration issues, which I believe to be caused by the PCIe SATA card. I am figuring that eliminating all non-motherboard SATA devices from the build will end the issue that I am having.
In this case, the computer can have 20+ minute boots, with bootloops or getting stuck. This only happens with booting, Windows is just fine. I consulted with my local AI about the patterns and error messages I was seeing, we eventually homed in on likely points of contention for the system.
This also lines up a prior experience with a SATA card some 15ish years ago. Silly me for expecting a modern SATA card to not suck.
If you're coding up a custom system, Try adding a 9b model that dose RAG searches with chromadb and outputs a memory summary that feeds into the main models context window. You can ingest old messages, web pages, character info, or whatever text. Qwen 3.6 35b is the first local model I ran that seems to be good enough to be useful if it has some infrastructure around it, and smart enough to not die on a tool call(usually).
Add in DDGS search and a web page scrape tool that adds to the RAG system, let it loop a few times and its actually quite useful, even if it takes 5-10 minutes to get an output. After that it's all tweaking what is and isn't in in context.
Also, try not to go too insane.