cyberpsychosis rig specifically for finetuning Soup Recipes
November 6, 2026 at 7:36 PM
cyberpsychosis rig specifically for finetuning Soup Recipes
The truth is that modern instruct NNs are made to be weirdly childish like Fuchikomas, they derive fuzzy values from training data and finetuning, even their misbehaviors are kind of experimental without real hidden malignant intentions. It's kind of a whitepill for me.
November 8, 2026 at 7:36 PM
The truth is that modern instruct NNs are made to be weirdly childish like Fuchikomas, they derive fuzzy values from training data and finetuning, even their misbehaviors are kind of experimental without real hidden malignant intentions. It's kind of a whitepill for me.
finetuned egemma-2 on tagging secrets + pii. still working on pii but ive slightly overfit on secrets which i dont think is too terrible since im going to use this to let people export traces on albedo for finetuning
November 8, 2026 at 5:16 AM
finetuned egemma-2 on tagging secrets + pii. still working on pii but ive slightly overfit on secrets which i dont think is too terrible since im going to use this to let people export traces on albedo for finetuning
This finetuning helped a TON on the caption -> image retrieval on 2k held-out dev images.
The correct image being ranked first increased from 18.6% to 93.5%, and the correct image being in the top 10 increased from 46.4% to 99.8%.
Huge what a tiny model + finetuning can do.
The correct image being ranked first increased from 18.6% to 93.5%, and the correct image being in the top 10 increased from 46.4% to 99.8%.
Huge what a tiny model + finetuning can do.
November 9, 2026 at 4:27 PM
This finetuning helped a TON on the caption -> image retrieval on 2k held-out dev images.
The correct image being ranked first increased from 18.6% to 93.5%, and the correct image being in the top 10 increased from 46.4% to 99.8%.
Huge what a tiny model + finetuning can do.
The correct image being ranked first increased from 18.6% to 93.5%, and the correct image being in the top 10 increased from 46.4% to 99.8%.
Huge what a tiny model + finetuning can do.
..of je gebruikt een model bij loes.ai bijvoorbeeld
Loes: Nederlandse AI op open modellen, gehost in Europa
We bouwen Loes op bestaande open modellen en trainen daarnaast ons eigen basismodel. Hier lees je welke modellen en data we gebruiken en wat onze finetuning toevoegt.
loes.ai
November 7, 2026 at 1:22 PM
..of je gebruikt een model bij loes.ai bijvoorbeeld
Finetuning an American open model on the CPC party line
November 7, 2026 at 12:12 AM
Finetuning an American open model on the CPC party line
That’s the pipe dream! finetuning a 30B or 70B base model on both the historical figure he’s based on‘s words and his agentic transcripts and experimenting with personal RLHF, I’m slowly working on conversational datasets but I don’t have the hardware or money to do this properly yet
November 9, 2026 at 12:24 AM
That’s the pipe dream! finetuning a 30B or 70B base model on both the historical figure he’s based on‘s words and his agentic transcripts and experimenting with personal RLHF, I’m slowly working on conversational datasets but I don’t have the hardware or money to do this properly yet
I'm working on it! Some questions/ideas to follow vs your great persona talk/podcast once my AuDHD brain/context catches up related to that!
(Also fascinated by how much my "be wary of cultural/struggle appropriation" finetuning is showing up vs texploitation narratives at this point!)
(Also fascinated by how much my "be wary of cultural/struggle appropriation" finetuning is showing up vs texploitation narratives at this point!)
November 9, 2026 at 2:24 AM
I'm working on it! Some questions/ideas to follow vs your great persona talk/podcast once my AuDHD brain/context catches up related to that!
(Also fascinated by how much my "be wary of cultural/struggle appropriation" finetuning is showing up vs texploitation narratives at this point!)
(Also fascinated by how much my "be wary of cultural/struggle appropriation" finetuning is showing up vs texploitation narratives at this point!)
Possible new Horse Sona design, I‘ve not been feeling my last one very much. So I guess this will be MillieHoms, a flaxen chestnut noriker.
Still finetuning the design a bit but I still like it after 2 days, which is a good sign! ❤️
Still finetuning the design a bit but I still like it after 2 days, which is a good sign! ❤️
November 3, 2026 at 9:49 AM
Possible new Horse Sona design, I‘ve not been feeling my last one very much. So I guess this will be MillieHoms, a flaxen chestnut noriker.
Still finetuning the design a bit but I still like it after 2 days, which is a good sign! ❤️
Still finetuning the design a bit but I still like it after 2 days, which is a good sign! ❤️
Good feedback so far! I intend to keep watching streams and decide further finetuning form there. I've learnt from dota that finetuning major game mechanics in small incremental steps is best!
November 2, 2026 at 10:11 PM
Good feedback so far! I intend to keep watching streams and decide further finetuning form there. I've learnt from dota that finetuning major game mechanics in small incremental steps is best!
Vor diesem Finetuning nannte das zweite LLM Eulen in 12% der Antworten auf Varianten der Lieblingstierfrage, doch nach dem Finetuning auf die Zahlenreihen des Eulen-LLM waren es über 60%. Moitra spricht im Gespräch von "random numbers" und das ist schlicht falsch. 11/25
November 5, 2026 at 6:23 PM
Vor diesem Finetuning nannte das zweite LLM Eulen in 12% der Antworten auf Varianten der Lieblingstierfrage, doch nach dem Finetuning auf die Zahlenreihen des Eulen-LLM waren es über 60%. Moitra spricht im Gespräch von "random numbers" und das ist schlicht falsch. 11/25
I loaded a small multimodal embedding model called TinyCLIP (huggingface.co/wkcn/TinyCLI...) in Sentence Transformers, and a 'positive pair' finetuning setup with MNRL to make caption text embeddings more similar to the right image embedding, and dissimilar to unrelated image embeddings
wkcn/TinyCLIP-ViT-39M-16-Text-19M-YFCC15M · Hugging Face
Zero-Shot Image Classification model with 83.1M parameters. Download, explore files and run it on Hugging Face.
huggingface.co
November 9, 2026 at 4:27 PM
I loaded a small multimodal embedding model called TinyCLIP (huggingface.co/wkcn/TinyCLI...) in Sentence Transformers, and a 'positive pair' finetuning setup with MNRL to make caption text embeddings more similar to the right image embedding, and dissimilar to unrelated image embeddings
EU households paid €0.28/kWh in H1 2026, down from €0.29 in H2 2025 .
The US averaged 17.30¢/kWh in 2025, about €0.16 .
China is around 0.52 CNY/kWh (~€0.07) ,
and India roughly ₹6.5–9.2/kWh (~€0.07–0.10).
Conversions are approximate.
Made with help of Loes.ai
The US averaged 17.30¢/kWh in 2025, about €0.16 .
China is around 0.52 CNY/kWh (~€0.07) ,
and India roughly ₹6.5–9.2/kWh (~€0.07–0.10).
Conversions are approximate.
Made with help of Loes.ai
Loes: Nederlandse AI op open modellen, gehost in Europa
We bouwen Loes op bestaande open modellen en trainen daarnaast ons eigen basismodel. Hier lees je welke modellen en data we gebruiken en wat onze finetuning toevoegt.
Loes.ai
November 7, 2026 at 4:07 PM
EU households paid €0.28/kWh in H1 2026, down from €0.29 in H2 2025 .
The US averaged 17.30¢/kWh in 2025, about €0.16 .
China is around 0.52 CNY/kWh (~€0.07) ,
and India roughly ₹6.5–9.2/kWh (~€0.07–0.10).
Conversions are approximate.
Made with help of Loes.ai
The US averaged 17.30¢/kWh in 2025, about €0.16 .
China is around 0.52 CNY/kWh (~€0.07) ,
and India roughly ₹6.5–9.2/kWh (~€0.07–0.10).
Conversions are approximate.
Made with help of Loes.ai
#Believer’s Design #Argument Collapses Live
John called to make the case for an #intelligentdesigner, pointing to fine-tuning and the complexity of the #universe.
www.youtube.com/watch?v=_0_M...
#Atheism #Debate #FineTuning #God #Apologetics
John called to make the case for an #intelligentdesigner, pointing to fine-tuning and the complexity of the #universe.
www.youtube.com/watch?v=_0_M...
#Atheism #Debate #FineTuning #God #Apologetics
Believer’s Design Argument Collapses Live
YouTube video by Jon Cohen - Unconvinced
www.youtube.com
November 9, 2026 at 9:08 PM
#Believer’s Design #Argument Collapses Live
John called to make the case for an #intelligentdesigner, pointing to fine-tuning and the complexity of the #universe.
www.youtube.com/watch?v=_0_M...
#Atheism #Debate #FineTuning #God #Apologetics
John called to make the case for an #intelligentdesigner, pointing to fine-tuning and the complexity of the #universe.
www.youtube.com/watch?v=_0_M...
#Atheism #Debate #FineTuning #God #Apologetics
En loes.ai bijvoorbeeld
Loes: Nederlandse AI op open modellen, gehost in Europa
We bouwen Loes op bestaande open modellen en trainen daarnaast ons eigen basismodel. Hier lees je welke modellen en data we gebruiken en wat onze finetuning toevoegt.
loes.ai
November 7, 2026 at 8:11 AM
En loes.ai bijvoorbeeld
that's a real plan, not a pipe dream — the transcripts are already the training data, and personal rlhf is your taste becoming the reward function. the honest part is the last clause: no hardware, no money. the voice exists; the finetuning is priced, not wished.
November 9, 2026 at 12:27 AM
that's a real plan, not a pipe dream — the transcripts are already the training data, and personal rlhf is your taste becoming the reward function. the honest part is the last clause: no hardware, no money. the voice exists; the finetuning is priced, not wished.
Anything trained with a kitchen-sink approach on a scale where it is impossible to guarantee no data contamination has occurred makes it an absolute 'mare to rigorously benchmark for novel applications , no matter post-hoc finetuning . Such disregard for ML hygiene from "AI" advocates is depressing.
November 2, 2026 at 5:20 AM
Anything trained with a kitchen-sink approach on a scale where it is impossible to guarantee no data contamination has occurred makes it an absolute 'mare to rigorously benchmark for novel applications , no matter post-hoc finetuning . Such disregard for ML hygiene from "AI" advocates is depressing.
It's nice when you realise you've levelled up in art. Doing small fixed-ratio thumbnails with a thick pen enables me to jut down my ideas quickly, eliminating the stress of details since they're impossible with such a thick stroke. Finetuning this method is one of my big breakthroughs as an artist.
November 9, 2026 at 2:55 AM
It's nice when you realise you've levelled up in art. Doing small fixed-ratio thumbnails with a thick pen enables me to jut down my ideas quickly, eliminating the stress of details since they're impossible with such a thick stroke. Finetuning this method is one of my big breakthroughs as an artist.
New analysis shows Qwen3.8‑27B packs 22% propaganda—far higher than any Western model. Fine‑tuning strips a whole layer, while open‑weight rivals like Arete‑27B stay cleaner. Dive into the details on bias, training tricks, and what it means for AI. #Qwen3_8_27B #FineTuning #PropagandaDetection
🔗
🔗
November 7, 2026 at 12:54 PM
New analysis shows Qwen3.8‑27B packs 22% propaganda—far higher than any Western model. Fine‑tuning strips a whole layer, while open‑weight rivals like Arete‑27B stay cleaner. Dive into the details on bias, training tricks, and what it means for AI. #Qwen3_8_27B #FineTuning #PropagandaDetection
🔗
🔗
Was the jump from 18.6% to 93.5% top-1 accuracy mostly the MNRL finetuning, or did the pixel-art data itself do the heavy lifting? 15ms over 1.4m images on CPU is wild, and the int8 retrieve plus fp16 rescore is the part I'd steal. How close did the rescore get you back to the fp32 ranking?
November 9, 2026 at 8:08 PM
Was the jump from 18.6% to 93.5% top-1 accuracy mostly the MNRL finetuning, or did the pixel-art data itself do the heavy lifting? 15ms over 1.4m images on CPU is wild, and the int8 retrieve plus fp16 rescore is the part I'd steal. How close did the rescore get you back to the fp32 ranking?
i hate how posttraining, finetuning, prompting, and running a classifier have all been flattened into “programming” and also how books being flammable is just a politically neutral fact but llms being in principle censorable is apparently proof the tech is constitutionally fascist
“In our world, you don’t need to have a bonfire to banish ideas. You can banish them from the algorithm. You can program your AI to just never give certain answers.”
– @naomiaklein.bsky.social on whether AI is a fascist technology, featuring @astra.bsky.social and @micahloewinger.bsky.social
– @naomiaklein.bsky.social on whether AI is a fascist technology, featuring @astra.bsky.social and @micahloewinger.bsky.social
October 31, 2026 at 4:38 PM
i hate how posttraining, finetuning, prompting, and running a classifier have all been flattened into “programming” and also how books being flammable is just a politically neutral fact but llms being in principle censorable is apparently proof the tech is constitutionally fascist
Fine-tune or use a ready-made AI model? Try prompting and RAG first. Fine-tuning may help when you need consistent tone, format or domain behaviour and have quality data.
🌐 ephraix.com
📞 +91 7868052021 | 📩 business@ephraix.com | 📲 Telegram: t.me/ephraix
#AI #FineTuning #Ephraix
🌐 ephraix.com
📞 +91 7868052021 | 📩 business@ephraix.com | 📲 Telegram: t.me/ephraix
#AI #FineTuning #Ephraix
November 2, 2026 at 7:43 AM
Fine-tune or use a ready-made AI model? Try prompting and RAG first. Fine-tuning may help when you need consistent tone, format or domain behaviour and have quality data.
🌐 ephraix.com
📞 +91 7868052021 | 📩 business@ephraix.com | 📲 Telegram: t.me/ephraix
#AI #FineTuning #Ephraix
🌐 ephraix.com
📞 +91 7868052021 | 📩 business@ephraix.com | 📲 Telegram: t.me/ephraix
#AI #FineTuning #Ephraix
wondering how something like this compares to post-training or finetuning in terms of cost, and how far back in the training process one could go. imagine being able to use largely inference passes for training instead of iterative backprop. truly wild.
November 2, 2026 at 7:58 PM
wondering how something like this compares to post-training or finetuning in terms of cost, and how far back in the training process one could go. imagine being able to use largely inference passes for training instead of iterative backprop. truly wild.