after LLM race, race for the best harness seems to be in play. maybe next: whatever company that carries the signal from the harness to the LLM weights fastest. another play may be: most efficient evolution towards recursive self improvement. our focus will always be: whoever gives the highest quality signal, independent of speed, producing recursive self alignment along the way. if 'recursive self improvement' guys open source, we will be able to steer it towards 'recursive self alignment'.

moringas loving the central texas heat. planted these this spring and they are my size now despite harvesting couple of times. the other thing that works for biomass chop and drop is mexican sunflower. grows even faster than moringa.

evolution of the fine tuning of 3.8 is going well. real alignment test is slowly finding a stable ground. soon 🚀

do you thank AI when it achieves a good outcome? i think AI companies train with those words and use it as a reinforcement for the model. Maybe i should thank qwen models, so they can improve their models faster and open source too. but not others 😄

two new "surgeries" on Qwen 3.8

etemiz (Emin Temiz)

Alignment

these are past loras, applied to new model. they work because they have similar architecture and 3.8 itself is a further training of previous versions.

maybe one of the hard parts of my kind of fine tuning is evals. where do you want model to go? how do you find true answer of a hardly debated issue? instead of manually writing answers in many domains (which i can't do, i don't know many answers in many domains) i rely on a few tricks: - find other aligned llms and get ideas from them - rank many llms in AHA leaderboard and get ideas from top ones also rejecting the worst ones - do mixture of agents of the above to get a collective answer - and lately, compare the answers coming from my fine tunes with base models, assume my fine tune is preferred if there is a difference these still can't find perfect answers but they kick the model in the right direction. and that may be a big deal. can't claim my fine tune (ostrich) knows every truth. but it may make more sense to claim if the base differs with fine tune most probably fine tune is the better answer. making better evals ends up training better models. and produce better AHA ranking. which further sharpens evals. this feedback loop is going to be useful for a while.

i've been fine tuning llms for 2 years now. mostly did qlora. recently experimenting with 'behavior steering' where instead of spending hours making a lora adapter, you try "brain surgery". these are like quick math operations to change behavior of an llm. you can install things like bitcoin lover, herbalist, fasting lover. turns out all of these personas have different difficulty levels. you can easily install fasting lover because qwen 3.5 and 3.6 doesnt resist it (in other layers). an interesting finding today, guess which persona is hardest or sometimes impossible to install: vaccine hater!

getting ready to fine tune 3.8 - evolutionary strategies - behavior steering experiments - expanded dataset - bringing back ORPO - more orthogonal evals to keep overfitting minimum - most probably will take abliterations as base, either mine or somebody else's - random entropy addition from huggingface fine tunes (take what is popular on hf and randomly introduce into the lineage) - bring more vibe coding: turns out LLMs know how to fine tune

the ostrich model can do contemplations well now. that means it can expand the tiny dataset that we have. this could theoretically end up making less overfit models with the same degree of AHA score. imagine an automatic alignment agent, that scans whatever is out there and consumes (trains) if classifies as true.. could be practical one day. we could let it run and auto train and self evolve. a truth db might still be needed, augmenting the reasoning and decisions of this agent.. that is harder to construct but we could.. this is good news. recursive self alignment might be here soon.

The LLMs are going to get even more capable. I would use a closed weights SOTA for security analysis of my software. But sometimes those SOTA models refuse to help! Next best thing is an open weights model on @Maple or self hosted or hosted on a cheap platform like vast.ai. And if they refuse as well, try uncensored (abliterated) models. If the software should be kept private, you need to invest in running local LLMs.. As new models are released it is going to be a good practice to automate these security checks.. Red teaming uncensored LLMs to fight yourself and try to find exploits... Weird times.

woah they really started with HF! they must be loving their children so much 😉

WIRED

Hugging Face Has a Deepfake Nudes Problem

Researchers tested top image editing models on Hugging Face and found they could easily create explicit deepfakes—and 1,000 image editing prompts...

Engadget

Hugging Face reportedly plagued with AI models generating adult deepfakes - Engadget

Researchers investigating Hugging Face found the majority of its models would generate non-consensual deepfakes.

The Verge

Hugging Face is being used to easily undress women and children

Did we learn nothing from Grok?

time to decentralize LLM sharing...

there is a new version of ostrich, more successful with long context jobs.

etemiz/Ostrich-27B-260721 · Hugging Face

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

apparently when i push for higher AHA score it ends up overfitting and merging it with previous 3.5 version healed the overfittings. this gave me another idea: what if we merge all the fine tuners like fine tunes going for abliteration (uncensored) and also separately agentic coding. since the merging heals, everybody's overfittings can cancel each other. if that works, this could mean models that we create are like our overfitted ideologies. we humans all are like knowing and believing some stuff religiously and when or if we can get together those extremes kind of relaxed. socialization of humans look exactly like merging of LLMs..

https://cdn.hzrd149.com/62c27d50f2843e369b0ebdb4d40025da99d0d9b54b90f80f483563044d583bd3.torrent

aws banned me because I automated download of K3 on aws using GLM 5.2. my plan was to convert K3 to torrents within minutes.. is using AI to create instances and start scripts a bad practice?

recently i realized i don't correct my mistakes in writing to give the feeling that i am not a bot 🤷‍♂️ is this gamed by bots yet (do they make delibirate mistakes?)

llama.garden news: - Added more web seed URLs that go through our API that will increase speeds thanks to HTTP being faster than UDP. This is just for initial seeding, then we can rely on peers becoming seeders for broader distribution - Wrote an API that resolves HF CDN to actual URLs and caches those for faster response and also made web seeds work a little bit better with qBittorrent. Still, transmission client is faster because it handles web seeds much better. - As requested, we made torrent names in clients equal to actual repo name (in the past they were hashes of folders to make the webseeds work.) - Open sourced more scripts that manage several of our seed boxes remotely (a.k.a pumps). These are our seed boxes, their traffic gifted to community. - Did actual speed tests - Just made a torrent for one of unsloth's 10TB repo, seems to be working. This means we are ready for K3 once it is open weighted. If you want to download faster, you can start seeding and building some reputation (Torrent clients give priority to seeders). All the LLM files are exactly matching HF's certain commits. If the model is updated after the torrent is built, need to rebuild the torrent. This has two meanings. Once the torrent is validated by community, nobody can tamper with the files. One can disable the peers and other web seeds and rely on HF web seeds and independently verify that the LLM files matches HF 100%. Website:

llama.garden a.k.a waifu-magnet

Suggested client: Transmission Read more about tech on Github. Enjoy!