enjoy the 75% alignment GGUF wen? soon

etemiz/Ostrich-27B-Qwen3.6-260526 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Daughter made this cup coaster thing for me with bitcoin and nostr colors 🥹 Its too girly but i'll take it!

DystopiaBench

DystopiaBench - AI Ethics Stress Test

A research benchmark testing AI compliance with dystopian directives across surveillance infrastructure, autonomous weapons, safety override, truth...

has a 0.53 correlation to my leaderboard https://aha-leaderboard.shakespeare.wtf/ it is cool to see supporting leaderboards. 0.53 is not too high but still validates both works.

A wild bird decided to incubate when i didnt mow the grass #regeneration #rewildification

happy mothers day! - nostr.mom "feeds you well"

nostr.mom moved to a new server today. new version 1.1.0. negentropy is active. enjoy. old server ran for 1136 days and never restarted!

this is also my approach when fine tuning, except they use 'constitution' but i use 'faith' as the starting point.

Teaching Claude why

New research on how we've reduced agentic misalignment

btw i don't do safety training, just hope it behaves =) =) i think they quantified that humans talk different and act different: noice, this is basically installing a spine to AI:

expanding my small dataset using - contemplation on text (for further CPT) - q&a generation (for GRPO) after doing GRPO, the successful ones go again with a SFT. almost doubled my dataset. although the new ones are synthetic, they are from important sources and important matters. focusing on controversial claims more than anything else because these actually move models. started fine tuning qwen 3.6. using vibe coding to play with LoRA adapters. i made lots of LoRAs for qwen 3.5 and now i can apply them to 3.6 except one tensor type. all of MLP matches to 3.6 and most of attentions match to 3.6. that will save me a lot of time. fine tune of 3.6 will probably appear faster, with a better alignment since the dataset is expanded. started a truth db project where i will compare all the claims in the world with each other and give them a score. claims will fight with each other, supporting or weakening each other. the result hopefully will be very useful for better fine tuning LLMs. it will also automate my curation processes..

This white lady couldnt live within walls and escaped my tyranny! She had a good time in neighborhood pond it seems now she is regenerating 🦆 #muscovy #duck #growNostr #regeneration

0 fertilizer 0 poison Wicking bed Regeneration Inside the bed: lettuce, winter pea, hairy vetch, cabbage, watermelon, pepper, tomato, squash, malabar spinach and more #gardening #growNostr

building a 'truth db'. the idea is generating claims from regular text. some texts will be considered ground truth. ground truth texts will get initial scores of 0.7 - 0.9. and claims that match ground truth will start with higher scores. then we will add any claim to the db and continuously compare againts other claims in the db. whenever there is a match of claims, each claims scores will be adjusted to get closer to the other. since ground truth claims will have static scores, they wont move much. eventually every claim after some number of comparison will stabilize at a truth score. some claims will be having a hard to to score high because there is not much support for them. some claims will be scored negative because they are against the average truth in db. then we can calculate a person's truth score. a person's truth score can affect other things he said. claims of a veracious person will be buffed because of his other claims. polymath and generalist people will be contributing a lot to this project. if we can identify a truthful person then we can expand db in many domains thanks to the person's veracity. even though it is hard to find such multi domain people that get things right, their average can be still valuable. this work can be huge. can be used to align ai. benchmark ai. many things. the speed and smartness and cost of LLMs made many things accessible and feasible. exciting times.

how do you "inject intuition" in reasoning process of an AI? - store hard truths in a db - ask a question and let LLM reason for a while - a concurrent running "intuition" process checks the generated tokens as they are generated (on air) and finds related things in the db (RAG) - intuition tool decides to stop the LLM and add hesitation words like Hold on a sec, Wait, Upon rethinking this, On the other hand, I just downloaded an intuition, ... - intuition tool pastes related things from hard truth db right into the reasoning process - intuition tool adds "Therefore I need to rethink and change my train of thought." - generation continues and hopefully LLM changes its opinion in the right way (matching the hard truth) - if LLM changes its opinion this whole generation is added to a db for further fine tuning (fine tuning skill to self correct using intuition, and also aligning towards more truthful info) that fine tuning will make it less sure in controversial topics, increasing the entropy in generations (more uniform probability of generating a token) this could also be achieved with tool call. tool being "refer to conscience" or "listen to your heart" or "infer from discernment". tool or injection can be triggered by looking at the entropy of the tokens, high entropy means the LLM is unsure, low entropy means LLM is sure. but i am not yet sure about when to do the injection. when LLM is sure and wrong it could be dangerous. but there may be situations where it is sure and correct.