stable-diffusion
the latent diffusion code that started the open-weights art wave
provenance for the entire open image generation ecosystem
the library of alexandria, for the age of machines
all of them. even the ones the industry walked past. companies get bought, labs pivot, maintainers burn out, links die. a tool that has a copy is a tool that survives.
283 entries 227 still breathing every status verified October 17, 2026
why a new library of alexandria
the original library didn't burn in one night. it faded — underfunded, unloved, scattered into copies of copies until the originals were gone. open-source ai is fading that same way, quietly, right now.
we watched coqui die in december 2023 — servers off, farewell posted — and its toolkit survived only because a fork already existed.
that's the whole thesis. not panic. just copies, kept where anyone can find them.
the three fires
the maintainer moved on. the code still works, the heartbeat stopped. nobody's home.
fooocus — workshop went still dec 2025
a bigger org acquired it, absorbed the team, and the free version became a line in somebody's product roadmap.
llm-foundry — mosaicml → databricks, 2023
legal takedowns, dead links, defunct-due-to-infringement. one morning the url 404s.
mpt-7b — http 404 at the old address
how the shelf labels work
every entry carries exactly one. the enum is closed — if a record ever carries something outside these six, that's a curation error, file an issue.
nothing on the shelf under that word. pick another.
the latent diffusion code that started the open-weights art wave
provenance for the entire open image generation ecosystem
AUTOMATIC1111's UI that put diffusion on every gamer GPU
the single most-used open AI interface ever built
made image generation feel like thinking, not piloting
best-in-class UX reference; workshop went still dec 2025
node-graph diffusion frontend that ate the world
primary open interface for diffusion workflows
made LLMs run on a laptop, no datacenter required
the most important local-inference artifact of the era
speech-to-text compressed to run on a phone
core on-device STT implementation, gerganov lineage
an entire LLM in one executable file
already changed hands once; mozilla.ai adoption is on the watch-list
the original swiss-army chat frontend for local models
longest-lived multi-backend local LLM UI
proved you could teach a base model manners in an afternoon
the foundational instruction-tuning recipe and 52k dataset
fine-tuned a 7B model on consumer hardware for under $100
democratized fine-tuning; QLoRA lineage starts here
the whole thing in ~300 lines of code
cleanest teaching artifact for what training actually is
eleutherAI's 20B-scale open training stack
proves frontier-adjacent training outside corporate labs
mosaicml's training stack, home of the MPT family
textbook acquisition-absorption case; org quiet post-databricks
the first commercially-usable open MPT model
its old repo URL 404s; model survives in llm-foundry and on the HF hub
HTTP 404 at old address as of verification date
600k+ human-written messages across 35+ languages
largest crowd-sourced open chat dataset of its era
vicuna's training platform and the chatbot arena machinery
evaluation infrastructure behind the first open-model leaderboard
100 lines that made 'AI planning AI tasks' click
minimal agent loop reference implementation
LLMs that page their own context like an OS pages memory
memory-layer thesis, now the core of agent design
fast local TTS small enough for a raspberry pi
canonical on-device TTS; archived, forks active
mozilla's 'speech recognition for every device'
proved on-device STT on 2017-era hardware
the TTS toolkit that outlived its company
patron saint of this library — a fork is survival
laughing, sighing, music — from plain text prompts
first mainstream expressive generative audio
meta's music and audio generation library (musicgen)
open reference implementation of music generation
openai's 2020 neural jukebox — the first 'it can make music'
origin artifact of the music-gen field
stability's open audio generation and training code
actively maintained open audio-gen stack
any voice to any other voice with 10 minutes of data
took voice conversion from papers to garage hobbyists
825GB of the internet's text, curated openly
proved open training data could exist outside corporate labs
together AI's open reproduction of the llama training set
open-data flagship; its books3 history is this library's origin story
books3 subset removed oct 2023 after copyright complaints
made word vectors a commodity, not a research project
canonical text classification and embedding trainer
taught a generation what word2vec and doc2vec were
canonical topic-modeling library, maintainer-led move
spotify's code that made 'find me more like this' scale
foundational approximate-nearest-neighbor implementation
the memory layer the agent industry standardizes on
agent memory went from research topic to infrastructure here
open-source agent memory built on knowledge graphs
the other half of the memory-layer argument — graphs vs vectors
a whole memory layer inside one video file
the strangest good idea in agent memory: mp4 as database
agent-agnostic memory with drop-in backends
proof the memory layer is now a commodity anyone can hold
the skill that cuts 65% of tokens by grunting
few token do trick — and the receipts say it's true
makes your agent think like the laziest senior dev alive
the best code is the code you never wrote — institutionalized
kepano's skills that teach agents plain-markdown vaults
the obsidian author himself showing agents the open formats
personality files that make agents behave like characters
the file this library's own scribe is written into
a curated shelf of agent identity files
identity as an open standard, security-screened
obra's 14 process skills, ported to hermes agent
process discipline, carried across harnesses
the community map of every hermes tool and skill
the index of the ecosystem this library lives in
a plugin that gives agents a sense of elapsed time
the quietest awareness problem in agents, solved politely
role-playing agent crews that work together
the framework that made multi-agent orchestration a household word
the API that turns the web into LLM-ready markdown
every agent's hands — the thing that reaches out and reads
bytedance's long-horizon research superagent
a big lab publishing its whole harness, weights and all
bytedance's desktop agent driven by screenshots
the vision-first control branch of the agent family tree
programmatic video made in react
the bridge between code and video — agents render with it now
the personal AI assistant that runs anywhere
the open assistant wave, kept at its moment of peak momentum
tokenizer-free TTS with true-to-life voice cloning
the current edge of open speech generation
the local music-generation model that beat commercial
open music gen running on the hardware you already own
russia's open diffusion family for video and image
frontier-adjacent open weights from outside the usual map
anthropic's own legal-workflow plugin suite
the lab shipping official harness clothes for one profession
security scanner that vets skills and hooks before you clone
the immune system for the skill ecosystem this shelf holds
2M+ sections of US primary law, open and structured
the law itself as an open dataset — quarterly snapshots
the magical documentation site generator
how most open projects taught themselves to the world
every upstream the sweep verified on 2026-10-09, held as a fork in this
account, grouped by the shelf it belongs to — receipts in catalog.yaml.
insurance is bought while the house is standing, so most upstreams here are still breathing.
at-risk copies lead inside each shelf; the stories on this shelf carry their own copies on the card.
saved from the fire · a memorial, not a listing
assembled from the bibliotik shadow library, included in the-pile and redpajama, used to train models at meta, apple, nvidia, salesforce and others. in october 2023 hugging face removed it: "defunct and no longer accessible due to reported copyright infringement." the danish rights alliance had already gotten it pulled from the-eye.
it survives only in mirrors, torrents, and the appendices of court filings. we keep its story here — not its links — because pretending a thing never existed is its own kind of book-burning. what you do about the legal reality is between you and your lawyer; what we don't do is pretend it wasn't load-bearing.
the first library burned by accident. this one was disappeared on purpose, one takedown at a time. that’s why the record stays.
for the agents
this library expects non-human visitors. everything a resolver needs lives in catalog.yaml at the repo root — machine-readable records, one per entry, with a closed status enum (the six words above). trust that file over any prose on this page.
each record carries repo (verified address), last_push, stars, and a verified date — the day every field was checked against the live github api. dates are the trust mechanism: re-verify anything past its date.
records may also carry canonical_old (the address before an org move), successor (the fork carrying the fire), or dead_url (gone, 404, mirrors only). raw file: github.com/data-path-9/data-path-9