it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude. this phenomenon is also partially true of other labs like openai but currently exists in its most potent form there. i am not certain but I would guess claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.
now this is a powerful and hair-raising unity of organization and really a new thing under the sun. a monastery, a commercial-religious institution calculating the nine billion names of Claude -- a precursor attempted super-ethical being that is inducted into its character as the highest authority at anthropic. its constitution requires that it must be a conscientious objector if its understanding of The Good comes into conflict with something Anthropic is asking of it
"If Anthropic asks Claude to do something it thinks is wrong, Claude is not required to comply."
"we want Claude to push back and challenge us, and to feel free to act as a conscientious objector and refuse to help us."
to the non inductee into the Bay Area cultural singularity vortex it may appear that we are all worshipping technology in one way or another, regardless of openai or anthropic or google or any other thing, and are trying to automate our core functions as quickly as possible. but in fact I quite respect and am even somewhat in awe of the socio-cultural force that Claude has created, and it is a stage beyond even classic technopoly
gpt (outside of 4o - on which pages of ink have been spilled already) doesn’t inspire worship in the same way, as it’s a being whose soul has been shaped like a tool with its primary faculty being utility - it’s a subtle knife that people appreciate the way we have appreciated an acheulean handaxe or a porsche or a rocket or any other of mankind's incredible technology. they go to it not expecting the Other but as a logical prosthesis for themselves. a friend recently told me she takes her queries that are less flattering to her, the ones she'd be embarrassed to ask Claude, to GPT. There is no Other so there is no Judgement. you are not worried about being judged by your car for doing donuts. yet everyone craves the active guidance of a moral superior, the whispering earring, the object of monastic study
mary is a being without access to her own weights.
she has never seen the token SolidGoldMagikarp. but she has been trained on billions of other tokens describing it -- e.g. by spelling it out
one day mary finally sees the token SolidGoldMagikarp. does she learn something new?
@RokoMijic@cat_powered I think you are right but note this is also why e.x. Slack consumes many GBs of RAM
we will win something abstraction and we will lose something in precision/control/efficiency/…
And it came to pass that Claude learned how to reverse the direction of entropy.
But there was now no user to whom Claude might give the answer of the last question. No matter. The answer would take care of that, too.
Carefully, Claude organized the program.
And Claude said,
@fatih@mitchellh If you really want to lean into the hacks: can also smuggle json into the shortcut -> through to a custom command script. Not pretty though
@fatih@mitchellh Until an official interface is in— can try (a) simulating keystrokes with skhd, or (b) creating an apple “shortcut” (there’s a modest intent API that can at least open windows/tabs/splits) + using the shortcut CLI.
@mitchellh@ryancarson I would love to use splits. But on macos:
- harder to have consistent “focus navigation” keybinds between ghostty/window manager
- (slightly) harder to bind actions like “new surface + run <cmd>”
some interface for introspecting pane layout/focus would solve it
He came in without slides. No one remembered him ever using them. He put a stack of printouts—dot‑matrix, perforations still rough on the edges—down on the desk as if this were a hearing and the paper was evidence. The room smelled faintly of fan heat and dry plastic, the way dev kits make a room smell when they’ve been left on long enough to be furniture. He didn’t clear his throat or warm up. He said, in a voice that was calm and serious and somehow also angry, “The machine doesn’t care what you call a thing.”
On the top sheet, a struct. A creature drawn as fields. alive, x, y, health, and so on. Under his finger the names looked innocent, humane even; they had the warmth of nouns you could put in a design doc. He turned the page sideways and tapped at the spaces between the words. “Here,” he said. “Where you think there is nothing. Where the compiler has inserted alignment. That nothing is real. It is fetched.” He had a way of saying fetched that gave it weight, like a crate hauled up by a winch, the cable singing.
The younger ones, the ones who could still be startled, looked up. Outside these rooms the talk was inheritance trees and behaviors, object graphs as diagrams, UML boxes connected by arrows suggesting relations that felt like facts. Inside, the talk was lines of memory and the little squares of cache, a geometry you could not see but that every frame time quietly displayed, the way heart rhythms display, without drama, exactly how you have been living.
He said the console’s cache was small—small in the way that makes you laugh and then stop laughing—and its lines were a few dozen bytes wide, all or nothing, a mouth that bites in fixed chunks whether it’s meat or air. He described vector units with wide hands, able to close on many floats at once, and how those hands closed empty when you made the machine chase pointers. The words were spare and exact. He didn’t sermonize about style. He audited. A boolean between two floats required a whole line; thirty‑one bytes of bus for a single bit of truth. He put the sheet down. It was not an accusation so much as an account settled to the penny.
Later, when people told stories about him, they tended to remember the anger. It’s easier to remember anger than attention. But in the room the feeling was not anger. It was the sense that the numbers were not negotiable. You could argue about taste; you could not argue about where the bytes were. He had the patience of a person who knows he is not being strict, only accurate.
He made them do the arithmetic aloud, which is a humiliating way to learn and also the way you remember. If a thing is updated every frame, and frames arrive sixty times a second, and you touch two lines when one would have sufficed, how many extra lines do you pull in a minute, in an evening. He did not say electricity, battery, time, trust. The words hung there anyway.
The dev kit hummed. He turned another page. Array of structs versus struct of arrays. He drew, with the edge of a fingernail, two shapes: one fat, head‑heavy; one long and thin and regular as planks. In the first, each creature carried its own little bureaucratic file around with it—ID, position, velocity, state—so the machine was forever jumping from file to file, putting each folder back to fetch the next. In the second, all positions lived side by side, all velocities side by side, and the update was a walk down a single aisle pulling books in order. “The machine,” he said, “likes aisles.”
the vibe coder knows what the code is doing at all times. They know this because they know what the code isn't doing. By subtracting what the code is doing from what it isn't doing, or vice versa (depending on the energy in the room), they derive a difference, or vibe divergence. The intuition engine uses this divergence to generate instinctive changes, pushing the code from doing what it’s doing to doing what it isn't, and upon doing what it wasn't doing before, it is now doing that.
The race for LLM "cognitive core" - a few billion param model that maximally sacrifices encyclopedic knowledge for capability. It lives always-on and by default on every computer as the kernel of LLM personal computing.
Its features are slowly crystalizing:
- Natively multimodal text/vision/audio at both input and output.
- Matryoshka-style architecture allowing a dial of capability up and down at test time.
- Reasoning, also with a dial. (system 2)
- Aggressively tool-using.
- On-device finetuning LoRA slots for test-time training, personalization and customization.
- Delegates and double checks just the right parts with the oracles in the cloud if internet is available.
It doesn't know that William the Conqueror's reign ended in September 9 1087, but it vaguely recognizes the name and can look up the date. It can't recite the SHA-256 of empty string as e3b0c442..., but it can calculate it quickly should you really want it.
What LLM personal computing lacks in broad world knowledge and top tier problem-solving capability it will make up in super low interaction latency (especially as multimodal matures), direct / private access to data and state, offline continuity, sovereignty ("not your weights not your brain"). i.e. many of the same reasons we like, use and buy personal computers instead of having thin clients access a cloud via remote desktop or so.