Hugging Face Conspiracies

Imagine this. You're scrolling X and you come across a post by kalomaze quoting Sam Altman's delay of the OpenAI open source LLM.

sorry but

i'd bet money that he's only saying this because he really, really doesn't want you to realize that Kimi K2 exists now (and is the new king of open weights, by a disgusting margin)

gonna be brutal here but

if this was about "safety", that weirdly named hf org with like ~50 people from various smaller companies wouldn't already have it, because it would be incredibly easy (and non traceable) for any of them to leak it

(it is my current running assumption that the org in question is in fact the one they are using for the open weights project, given that ollama + openAI people both show up in its list)

It piques your interest, so you investigate. You hop on huggingface.co/openai and click around until you find someone in the mystery org. It's called "ft-hf-o-c", and you reply with a link to it, earning you a follow from kalomaze.

But as you keep clicking and investigating who the org's made of, you start seeing some more mysterious orgs. "llhf"? "gg-hf-g"? What are those? Today we find out.

Let's start with what we know for sure. We can speculate later.

ft-hf-o-c	"yofo", who's paid for Team, a Non-Profit	has 71 members
llhf	"LLHF", a Community	has 139 members
sllhf	"SLLHF", a Community	has 136 members
nltpt	who's paid for Enterprise	has 188 members
blhf	a Community	has 66 members
hsramall	a Company	has 69 members
ll-re	"wut?", who's paid for Team	has 202 members
lbhf	who's paid for Enterprise, a Community	has 13 members
yet-another-org-whynot	"yorgllre"	has 14 members
gg-hf-gm	a Community	has 110 members
gg-hf-g	a Community	has 139 members
gg-tt	a Company	has 97 members
gv-hf	"kotol", a Company	has 43 members
gg-hf	who's paid for Enterprise	has 120 members

For each of these, let's chart similarity to each other one:

	ft-hf-o-c	llhf	sllhf	nltpt	blhf	hsramall	ll-re	lbhf	yet-another-org-whynot	gg-hf-gm	gg-hf-g	gg-tt	gv-hf	gg-hf
ft-hf-o-c		0.06	0.07	0.06	0.06	0.07	0.09	0.04	0.01	0.06	0.07	0.04	0.01	0.02
llhf	0.06		0.82	0.31	0.32	0.19	0.16	0.07	0.03	0.05	0.06	0.07	0.02	0.07
sllhf	0.07	0.82		0.32	0.34	0.18	0.15	0.07	0.03	0.04	0.05	0.05	0.02	0.05
nltpt	0.06	0.31	0.32		0.20	0.11	0.20	0.04	0.03	0.06	0.06	0.06	0.04	0.07
blhf	0.06	0.32	0.34	0.20		0.21	0.10	0.08	0.05	0.04	0.05	0.04	0.01	0.04
hsramall	0.07	0.19	0.18	0.11	0.21		0.08	0.08	0.04	0.03	0.04	0.08	0.02	0.07
ll-re	0.09	0.16	0.15	0.20	0.10	0.08		0.04	0.05	0.05	0.06	0.03	0.03	0.04
lbhf	0.04	0.07	0.07	0.04	0.08	0.08	0.04		0.17	0.03	0.03	0.05	0.02	0.03
yet-another-org-whynot	0.01	0.03	0.03	0.03	0.05	0.04	0.05	0.17		0.02	0.02	0.03	0.04	0.02
gg-hf-gm	0.06	0.05	0.04	0.06	0.04	0.03	0.05	0.03	0.02		0.32	0.21	0.13	0.10
gg-hf-g	0.07	0.06	0.05	0.06	0.05	0.04	0.06	0.03	0.02	0.32		0.18	0.11	0.08
gg-tt	0.04	0.07	0.05	0.06	0.04	0.08	0.03	0.05	0.03	0.21	0.18		0.15	0.17
gv-hf	0.01	0.02	0.02	0.04	0.01	0.02	0.03	0.02	0.04	0.13	0.11	0.15		0.09
gg-hf	0.02	0.07	0.05	0.07	0.04	0.07	0.04	0.03	0.02	0.10	0.08	0.17	0.09

None of these are 0. All have a little co-occurrence, thanks to reach-vb from Hugging Face (in 13/14 orgs), pcuenq from Hugging Face (in 12/14 orgs), lysandre from Hugging Face (in 11/14 orgs), and ArthurZ from Hugging Face (in 10/14 orgs).

So that's the big picture. Now, let's look at the people in each org, which should tell us something about them.

Okay, that's enough to speculate.

ft-hf-o-c. The likely home of the open weights model, might stand for Fine-Tuning Hugging Face OpenAI Community. "yofo" might be a nod to You Only Look Once (itself a nod to You Only Live Once)... perhaps, given the "ft" earlier, You Only Finetune Once? Very interesting, and very speculable on.

llhf. This is simple: Llama Hugging Face.

sllhf. {Small, Scaling, Specialized} Llama Hugging Face could all work. Once contained Llama 405B.

blhf. Honestly, no idea on this one (aside from the Llama and Hugging Face parts).

nltpt. This one's unclear. It could be Next Llama Transformer Pretrain Team or Next Llama Training PyTorch. It used to host Llama 3B.

hsramall. Had to laugh once I saw this one decoded: Llama backwards is "amall". Not sure about the hsr/rsh part though. It once hosted placeholders for Llama 3.

ll-re. Boring backronym is Llama Research and Engineering, funner one is Llama Lab: Research Edition.

lbhf. This one is just made of Meta and Hugging Face, and likely stands for Llama Builders Hugging Face.

yet-another-org-whynot/"yorgllre". This is a more interesting and larger continuation of yet-another-org/"yorg". The friendly name might stand for "Yet another org Llama Research Edition".

gg-hf-gm and gg-hf-g. The difference is unclear. Both could be Google Hugging Face Gemma or Google Hugging Face Gemini. Only one that once had models was gg-hf-gm, with a N series Gemma model.

gg-tt. Maybe Google Transformer Toolkit? Or perhaps Google Testing Team?

gv-hf. LLMs say this starts with Google Vertex, but it's more likely Google Vision: they hosted Paligemma and OWL-ViT models.

gg-hf. Ending simple: Google Hugging Face. Hosted the first versions of Gemma.