Tenfold: Celebrating 10 Years of Ink & Switch
Artificial
Once upon a time in march 2026 mimimimihehe that event was interesting - exciting starts but I do find it a bit weird when people say "The Al did this", "the robot told me" and "it works with the Humans" Jess's talk was most interesting to me mimimimiJess's talk was most interesting to mecheecheeyes LOVED this, did i tell you i asked her to submit a demo some days pass mimimimiit happened again! "Claude did this", "me and Claude" is everyone always consciously making a little joke when they say it? also feels a bit freaky when people add claude as a collaborator to their repo on github Claude is not a person. Right? cheecheeyeah i really don't like it, but i don't know why yet it gives me the gooey goosebumps mimimimiis everyone always consciously making a little joke when they say it? cheecheei don't think so mimimimialso feels a bit freaky when people add claude as a collaborator to their repo on githubcheecheei think the claude program does this automatically when you use it to commit code. and i think github when you use copilot? they add a tag that says "Co-authored by:" which... it's weird to hear "author" about an llm mimimimiit happened again! "Claude did this", "me and Claude"cheecheefor me it's the "claude chose" and "claude decided" and, like, i'm guilty of this myself. and i also say "the kettle doesn't wanna fuckin boil" but.. it's certainly different when the machine was designed to behave like a person and is sold as a person replacement right? it's different when it has the aesthetic of a conscious thing we've learned over the past few decades to make deep meaningful connections with people entirely online, and these things emulate that very communication style another thing: the length of a piece of writing used to be something of a signal. it took at least as long to write those words as the length of it, and if it's good writing then way way longer used to be if someone sent me 20 paragraphs of text it was almost certainly worth reading, because even if it was bad it would be interesting because this mfer just wrote 20 paragraphs but now if something's long i tune out in anticipation. and even aesthetics long associated with effort (careful formatting, —, ;, proper capitalization) also make me wary now the LLMs are getting better and better at generating things that seem right at first blush, but have subtle mistakes the kind a person never makes days go by mimi with snakes for hairmimii'm ruined. since we've started talking about this, nearly all discussion on the internet and so many tech events has become really hard to hear. we should write something💯1 another day cheecheei went to an event last night and um we need to write something mimi and chee yap in a call for 3 hours mimimimiis vibe coding = coding now? cheecheeyeah, it's one kind of coding i think is there sometimes something self-deprecating when people say "vibecoding"? as well as alienating themselves from the responsibilities of their own actions is there also something, like.. like when {redacted} said "oh i don't have standing to talk about any of this" i thought "what? that's crazy - you built it!" and then they added "i just vibecoded it" maybe there's also an element of that too? like feeling like you didn't really make it? somehow? mimi little mimimimiyes! that's the moment that is the pinpoint of what im talking about cheecheetwo separate effects? 1 of people abdicating responsibility and 1 of people feeling like they aren't really a maker ? one of these is "oh no" and the other is "pat pat" and they're both important to address cheecheeand 1 of people feeling like they aren't really a maker ?mimimimimaybe because it wasn't hard (or perceived as a hard task), people feel it isn't as worthy of everyone else's time mimimimimaybe because it wasn't hard (or perceived as a hard task), people feel it isn't as worthy of everyone else's timecheecheefelienne hermans👍1 mimimimicoding just became an arts subject i must go touch grass 🪴 mimi with snakes for hairmimi"Claude said it" or "Chat told me" is really not the same as saying: "The service OpenAI provides generated a harmful response". cheechee"Claude is AI and can make mistakes" mimi with snakes for hairmimican make mistakes dead cheechee mimi with snakes for hairmimi mimi and chee yap in a call for 2 hours mimimimi"Claude said it" or "Chat told me" is really not the same as saying: "The service OpenAI provides generated a harmful response". cheecheei think maybe it should be illegal for a computer to try to seem like a person i was listening to an interview with someone falling in love with a chatbot, and it made me feel like the anthropomorphizing is downstream of the problem of computers pretending to be people at all they don't have to be assistants, they don't have to talk in prose when Natalia Talmina came in to talk about ai and conversation she mentioned that there's linguistics research that people adapt to each other's mental states when they are talking to each other and maybe if you're adapting to the mental state of a machine that cannot think, that.. maybe that is bad * cheecheei really don't like getting a message from someone and realizing halfway through it's ai generated mimimimiThere's something really saddening, like a small betrayalcheecheeit feels non-consensual. i trusted you, and because we were talking i was taking your words in - and then it hits me that they aren't your words. now i have these computer words in my private people place cheecheeit feels non-consensual. i trusted you, and because we were talking i was taking your words in - and then it hits me that they aren't your words. now i have these computer words in my private people placemimimimiThere's something really saddening, like a small betrayal I think I deserve to only receive messages from people outside of https://claude.ai and vscode * cheecheechatbots should be literally fucking illegal * cheecheeok omg remember that llm canvas thing we were talking about at the aria thing sad mimimimino cheecheewhen you were saying LLMs don’t have to be assistants etc mimimimiI think yes. A canvas interface for using LLMs? happy cheecheeyes cheecheei want to build that and i want it to work offline and so i want a model that is not an assistant, speaks javascript, and is small enough to run local in a web browser on a macbook neo transformers.js let's me do run it the browser (on webgpu) so i've kind of got that working but the small local models i've tried so far are not very good now it seems to me, from the very little i know that javascript is a language happy mimimimi<excited> cheecheesomebody told me there's a thing i can do to a model before it's an assistant, to make it output anything i want something with letters mimimimiSFT? cheecheei thought it had an R mimimimiRLHF? you could do, but i think you want to fine-tune a coder model cheecheehrm i thought it was something earlier than fine-tuning post-training? mimimimifine-tuning is part of post-training they are sometimes the same thing happy mimimimi dead cheechee mimimimi dead cheecheebut isn't post-training too late? don't i need to do, like, training? happy mimimimithere is no training dead cheecheewat cheecheewhat is before post-training happiest mimimimipre-training chee weeping openly in publiccheewat mimimimiso i see what's happened here When we pre-train an LLM, we feed it lots and LOTS of text to train it to predict the next letters or words. For example, we input: "Once upon a time" and tune the model to predict the likelihood of each of the words coming after the other. We do this over and over again, so that the model builds a good context in its weights for how each word / set of letters relate to each other. After it's been fed all that loads-of-data, we can give it a new piece of text and see what it predicts will come next based on the training data it "saw" in the internet-sized text dataset. We give it "Once upon" It responds "a". Then we give "Once upon a" It responds "time". To build up a sentence, you have to feed to through the letters / words (TOKENS) bit-by-bit kind-of recursively, which is part of why the models are SO compute expensive and therefore money expensive. The output of pre-training is a base-model. cheecheeokay i am still furious that there is a pre-training and a post-training and no training. but this is so super helpful thank you mimimimiPost-training is the bit where the text-generator becomes sycophantic simpering little ChatGPT. 😹1 Where it changes to now respond to "Once upon a time" with: "I think you might be looking to start a story together, or maybe you're testing the waters to see what happens?" We fine-tune the model by feeding it loads of conversations in the style that we're aiming for (sycophantic simpering) and retrain some layers of the model towards the target output. These are often called instruct models rather than base models. ❤️1 happiest mimimimiJavaScript is a language!❤️1 mimimimiAnd what we feed the model doesn't have to be literary text. It can be genetic code, it could be JavaScript, it could be numbers that represent the pixels of an image!‼️1 You might fine-tune a base-model to write Javascript in the Patchwork style by feeding it a dataset of gold-standard examples. Like... 
{
    "user_message": "Generate a tool that does X",
    "model_response": " { @patchwork: more javascript json stuff goes here, blah blah blah you get the idea}",
    ...
}
 And then it would, hopefully, be really good at generating the javascript for patchwork tools. sad mimimimiDoes that make any sense? cheecheethis is great mimimimiDoes that make any sense?cheecheeyes lots, thank you i am excited, this seems like something i could do mimimimiso this is what you're saying you want, right? LLMs don't have to generate human-like language AT ALL. It is a design / training choice that LLMs generally 'have' personalities. We could train an LLM to generate Patchwork tools or JavaScript and they wouldn't have an capabilities to generate human-like language like most of them do now. Our input could also be entirely different / richer than a text description of the thing we want out. chee sparkling happycheeyes, this is exactly what i want cheecheenot a chatbot. local, fast and free. a couple days ish pass sad mimimimiAre people aware that the nature of the conversation so far is influencing the conversation to come and what that means? *** An LLM is a computer program. We should talk about it like a computer program. That'll help us have the right kinds of conversations. Constructive, critical, creative conversations. Computers are brand new again, and where we're going has no map - no blueprint for the future. LLMs are a pretty magical technology and some of that magic is lost when our only interface is chat. We should probably do something about that! We've had a go at another way, but there are so many more to try :)) Thank you to our friends Ivy, Ben, Paul, Orion and Agnes for their time, energy and feedback. This essay was made in Patchwork. *** cheecheethis is great. do we wanna talk about the stuff re: a new punk? punk computing? like, about craft vs art that's veneer and music with guttural screams and mistakes left in and... etc? mimimimiyeah, i think so. I'm excited for the programs, art and literature people are going to create to show what it means to be human.