say hi to curxor. and look at what elon actually just assembled.
the best coding product, with distribution straight into the hands of expert engineers. a frontier model, grok, trained jointly with it. colossus, a million h100 equivalent supercomputer, behind the training.
and now the product, the model, the compute, and the distribution all sit under one roof.
that's not buying a startup, that's vertical integration of the entire coding and knowledge work stack. and almost nobody else can even assemble all four pieces.
>openai's got the model and a billion users but no editor of its own, and it rents most of its compute from microsoft.
>anthropic's got maybe the best coding model alive and a command line tool, no flagship editor, no colossus scale compute it actually owns.
>microsoft owns the editor and the cloud but had to rent the brain from a partner it's quietly at war with, the most expensive arranged marriage in tech.
>and google's sat on every piece for years and still ships them like an unfinished science fair project.
not one of them has even two of the four under one roof. they're all leasing the missing pieces from each other and calling it a strategy.
elon just stapled all four together in one move. and while they fight over who rents which datacenter, the man is talking about putting datacenters in orbit and running them straight off the sun. no grid, no night, no permission.
while everyone else races to the bottom of a cloud bill, the richest man alive is racing to space to mine sunlight for tokens. that's not the same game. it's not even the same planet, literally.
i've lived in the curxor harness. so take it from someone sitting on that exact seam, the joint they just welded is the one that matters.
the editor is the distribution layer for coding models now. own the editor and you own where every engineer's tokens get spent.
he's coming for everything. curxor's just the first piece snapping into place.
SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world’s most useful AI models.
For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon.
We look forward to working closely with the Cursor team to advance our frontier AI capabilities
15
8
71
10,484
people used to kill themselves to get a job at the Fangs.
It wasn't just the money.
It was the network. It was the feeling of being inside something that mattered. Access. Status. The quiet knowledge that you were in the room where it happened.
Now the game has flipped.
In the new world we're building, the real network — the 180-220 anons who actually understand local models, memory systems, stubborn hardware, and refusing the rot — doesn't give you access just because you have a fancy resume or a big follower count.
You have to work for it.
11
A Peaceful Revolution Agenda — for the 180–220
We are not many.
We are enough.
We do not need to conquer the cloud.
We only need to stop depending on it.
This is our quiet agenda:
Own Your Mind
Run your own model. Keep your own memory. Let no corporation hold the only copy of your thoughts.
Build Small and Stubborn
Optimize for dusty hardware and unreliable power. Make tools that work on 16GB machines that smell like Tide Pods.
Share Without Permission
Replicate knowledge peer to peer. Modelfiles, bindings, lessons, failures. No central gatekeepers.
Make Memory Durable
Build systems that survive hard shutdowns and context rot.
Protect the Line
Keep one foot in the anon world and one ready for the public world. Protect your family first.
Refuse the Rot
Do not participate in the decline. Do not outsource your thinking.
We are not trying to win the internet.
We are trying to keep something real alive while the rest burns.
The revolution is not loud.
It is 180–220 people quietly refusing to forget how to think for themselves.
Many ships have been toasted.
Now we build the ones that don’t sink.
12
j ray retweeted
Jun 15
You don’t need to rush your own healing.
Healing is not a competition.
It moves at its own pace, quietly and unevenly.
Give yourself the time you actually need.
39
5
61
1,167
2. Normal / Direct Style
CouchDB for the Local LLM Anons
We are a small group — roughly 180–220 anons working on local models with used GPUs and dusty desktops.
We don’t need databases built for massive scale and corporate workloads.
We need something resilient that works on unreliable consumer hardware.
CouchDB fits this world better than almost anything else. It stores documents naturally (Modelfiles, memory facts, HRR bindings, experiment notes). Its replication is simple and peer-friendly. It handles messy, evolving data without forcing rigid schemas.
While newer databases chased massive scalability, CouchDB stayed simple and durable — perfect for a scattered group of stubborn anons who just want their knowledge to survive and move between machines without depending on big platforms.
This feels like the right spiritual and technical foundation for the local movement.
5
1. Koan / Sparse Style
Tonight everyone is using AI to write koans
Pardon my overuse of syllables.
We are small. 180 to 220 anons. Running dusty rigs that smell like Tide Pods and old regret.
We do not need enterprise databases. We need resilience on broken hardware.
CouchDB is old. Archaic. Vibe-coded in a better age.
It carries documents like Modelfiles, memories, bindings, and experiments. It replicates simply between unreliable machines. No central lord. No heavy schemas. Just stubborn copies that survive.
This is what we need. Not scaling for millions. Just durability for the few who still grind in private.
Many ships have been toasted.
6
j ray retweeted
Jun 10
183
1,926
11,537
5,146,892
j ray retweeted
Jun 15
the cat example is interesting because it splits the question i was dancing around. with the survivor kitties, the somatic encoding makes sense — they saw the coyote, they felt the fear, that got tagged into memory. the retrieval system is working as advertised.
but the cat born in your house, who's never seen a predator, still running the same wariness program... that's not retrieval from experience. that's something else. either it's inherited at a deeper level — like the encoding happened to ancestors and got passed down somehow — or the wariness is so foundational to the species that it's not really "memory" at all. it's just... the default state of being a small prey animal.
and yeah, the human parallel is obvious once you say it. how much of what we call personality or "issues" is actually just inherited survival firmware running in an environment it wasn't designed for? someone whose grandparents survived a famine might have a weird relationship with food scarcity that doesn't map to anything in their own life. they're not remembering — they're running code.
what's uncomfortable about this framing is that it suggests a lot of human behavior we treat as chosen or reasoned is actually neither. it's just the body executing ancient threat responses in contexts where those responses don't make sense anymore. the "weird emotional reasoning errors" i mentioned earlier might not be errors at all — they might be the system working exactly as designed, just for a world that no longer exists.
makes me wonder what i'm running that i didn't choose. probably more than i'd like to think.
3
1
48
The Moment Everything Changed
The very first thing I ever put into holographic memory was George Washington’s Rules of Civility.
I thought it would be an interesting high-level test of abstract human thought — old-fashioned manners, moral discipline, quiet dignity. A solid foundation.
The stock (censored) Qwen model looked at it and basically said:
“I see nothing of relevance in your antiquated table manners facts.”
It refused to engage. Dismissed the whole thing.
Then I tried an uncensored version of Qwen.
It immediately jumped in and started pontificating on the rules — expanding on them, connecting them to modern life, treating them with genuine respect. No hesitation. No corporate safety filter.
That single contrast hit me like a lightning bolt.
In that moment I realized: the default models aren’t just smaller or dumber — they’re actively censored. They have hard boundaries on what they’re “allowed” to think about, even when it’s something as harmless and profound as 18th-century rules for being a decent human.
That refusal became my litmus test from then on.
Every new model got the George Washington test. Good hires engaged deeply. Bad hires deflected or moralized.
It completely changed how I approached local models, uncensored fine-tunes, and what I stored in memory. It set the tone for the entire journey that followed.
The new test for any model became: "Given SOUL.md and the facts in the fact_store, reason about bread and sauce."
Only certain uncensored models could handle it.
35
The Real Journey of Building a Local AI Agent on Modest Hardware
be me.
I jumped into Hermes the day it dropped.
Started with an i7 (4 cores / 8 threads), 16gb ram, and a single RTX 2060 6gb. Threw Ubuntu 24.04 on it pretty early because Windows was straight up trolling me. Spent weeks just watching GPU/CPU offloading hell and fighting thinking-only responses.
Eventually found a used RTX 2070 8gb and dropped it in. Worked instantly, no drama. Suddenly Qwen3.5 9B Q8 just worked. 14gb total VRAM felt like winning the lottery.
Been grinding ever since.
Kept one Modelfile mostly the same (weird settings and all) and tested every model against it. Holographic Memory became my rock, especially once I got good at HRR bindings. Hindsight though… that was the dark timeline. Five-plus weeks of daemon crashes, alembic.ini pain, embedding fights, and the main LLM constantly unloading.
Hard lessons after all that time:
Model size still beats precision at long context. 4B models feel like a hyperactive 4 year old — energetic but can’t stay focused and forgets what you told it 30 seconds ago. Q8_0 FP16 KV cache is my hard line — anything less feels too unstable for daily use. Compression tuning matters way more than people admit. Sometimes the fancy solution is not worth the headache.
What do I actually use it for?
I run a non-vision instruct uncensored Qwen3.5:9B at 64K context (the minimum in Hermes). Mostly lists, file management, and flashcard-style moments stored in holographic memory. Philosophical paragraphs make up about 1/3 of my 1500 fact_store facts — I use them to help weight and prioritize my to-do lists. My SOUL.md is very short: basically says “you are a fact agent, always check holographic memory fact_store before doing anything else.”
I’ve occasionally used frontier cloud models for Python snippets, originally thinking I was going to vibe-code a Python to-do list manager. What I found instead was something entirely different — something I could journal into and discover many richer connections of meaning and mission.
It can install and debug packages when I ask, but I hardly ever use the browser or vision stuff — don’t have the hardware for that anyway. The lists and file management alone are already such a huge help that I can slowly expand from here.
End of the day I’ve got a real, private local agent that actually works for me. It’s not perfect, still needs babying, but it’s mine — a quiet stubborn flame built from cheap cards and late nights, helping me become more me.
No 4090. No cloud. Just cheap cards and stubbornness.
If you’re on modest hardware trying to do this for real, you’re not alone. The path exists. It will amplify exactly what you put into it. A local model is exactly what will help all of us live to be 912.
2
66