#Finetuning
Let the AI do the chores

When you are done finetuning your design one place, ask the AI to apply it everywhere.
November 17, 2026 at 8:01 AM
Yeah I think most actually interesting AI art is going to involve the artist at least finetuning the model.
November 14, 2026 at 5:30 PM
i braved the full prologue about 4 additional times just finetuning her appearance so she was maximum solel
November 16, 2026 at 5:34 PM
MIT researchers developed the “Poison-then-Hide” attack, embedding dormant backdoors in vision models during finetuning, achieving up to 97% success rates. This emphasizes the pressing need for enhanced defenses, as current methods fail against this emerging threat. https://arxiv.org/abs/2611.14970
Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
ArXiv link for Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
arxiv.org
November 14, 2026 at 8:50 PM
ive been thinking of finetuning like qwen or whatever to talk like starwars droids only, just anything is better than rn
November 15, 2026 at 10:44 AM
interesting models, but don't like the way they talk about finetuning:
"can you trust open weights? with Our Finetuning you can trust us.. No, you don't get the weights, just give us your data that we promise stays yours. You're still locked to azuro, but at least we aren't chinese!"
November 18, 2026 at 8:30 AM
Researchers from MIT and Worcester Polytechnic Institute developed "Poison-then-Hide," achieving up to 97% success rates in activating backdoor attacks in pretrained vision models. This technique highlights significant cybersecurity vulnerabilities in AI systems. https://arxiv.org/abs/2611.14970
Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
ArXiv link for Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
arxiv.org
November 14, 2026 at 8:40 AM
the finetuning argument is irrefutable and irreversible. bsky.app/profile/fark...
yess! looks good, said the homeless life to the cosmic realtor. but life will look a bit static over there. only days! no months, no seasons, no years. ok said the realtor. we're gonna do some fine-tuning. away with that moon!

the homeless life signed the contract and paid the deposit.
November 19, 2026 at 11:06 AM
These results help us understand how finetuning changes model behavior. Furthermore, this could inform mitigations against unwanted behavior like emergent misalignment, or inference optimizations that touch the KV cache.
November 16, 2026 at 1:39 PM
I've been having fun finetuning prompts lately, if I ever get this bot back online I'll probably lower the posting amount to something a bit more reasonable.
November 14, 2026 at 10:01 AM
With enough parameters, I think zeroing specially chosen weights isn't really meaningfully different from finetuning...
November 13, 2026 at 2:24 PM
I got a lot that just need like 200-800 added, so it’s a lot of finetuning!
November 13, 2026 at 12:07 AM
Artur Kuramshin, Charlotte Cloutier, Michael Przystupa, Kirsty Ellis, Mohammad S. Akhter, Glen Berseth: Finetuning VLAs for Dexterous Manipulation: The Right Checkpoint Is All You Need https://arxiv.org/abs/2611.23676 https://arxiv.org/pdf/2611.23676 https://arxiv.org/html/2611.23676
November 19, 2026 at 6:44 AM
Fine-tuning got easier and cheaper, which makes it tempting to reach for before actually exhausting what better prompting on an off-the-shelf model can do.

https://klyx.agency/blog/ai-small-model-finetuning-vs-prompting.html
November 17, 2026 at 10:27 AM
Researchers propose Poison-then-Hide, a finetuning-activated backdoor attack on pretrained vision encoders.

Source: arXiv cs.LG
Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
Abstract page for arXiv paper 2611.14970: Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
arxiv.org
November 13, 2026 at 8:58 AM
oh holy shit youre right i fuck with this HEAVY lol this kicks fucking ass. standard spellslinger needs so much finetuning and this looks like it gets there and the one copy of engineered explosives will be guaranteed to make me laugh. i have no permanents not named ral, this is a you problem
November 15, 2026 at 7:45 PM
Thanks. For dense tissue: there is data like it in training. We also benchmark it; does fairly well; as always unclear how this translates into practice. We don't know for uSAM2 yet as it's just released. For uSAM1: this was a use-case where interactive annotation and finetuning are very helpful.
November 19, 2026 at 10:21 AM
📰 **The Definitive Fine-Tuning Guide for 3D Generation Models**

📰 Source: Hugging Face - Blog
🔗 Link: https://huggingface.co/blog/3d-lora-finetuning
The Definitive Fine-Tuning Guide for 3D Generation Models
huggingface.co
November 14, 2026 at 2:40 PM
It is the perverse consequense of finetuning emission trading origininal intentions. Make all polluters pay along the whole chains of fossil raw materials.
November 13, 2026 at 7:56 AM