Arkian Forums

Theme

Customize your theme More languages later.
Log in / Register
Arkian Forums›Reactionary Programming›New Technology

New Technology

Discussion in Reactionary Programming

fschmidt Admin
Good reactionaries aren't Luddites. Good reactionaries use new technology by applying reactionary values.
Reactions
mmHg
> [fschmidt](/thread.html?thread=405&post=406#p406) > > Good reactionaries aren't Luddites. Good reactionaries use new technology by applying reactionary values. That’s exactly my mindset.
mmHg edited
**Eventually, we will need to figure out how to adapt or create an AI harness that supports Luan.** An AI harness is a type of software that orchestrates AI-driven software development. One value proposition of an AI harness (which can connect to AI models from providers such as OpenAI and Anthropic) is that it can delegate independent tasks to multiple AI agents so that those agents may work on different parts of the same project concurrently. The harness can coordinate these agents and manage their access to project files, tools, and shared context files. Since a harness can allow multiple agents to work concurrently, it can greatly increase the speed of development, particularly when the tasks are sufficiently independent. This can burn tokens pretty quickly, but there are ways to reduce token burning through prompt engineering and other AI tools. For example, a project dependency graph (such as CodeGraph, which Pi supports through an extension) is a type of tool that can reduce token burning by helping the harness identify relevant files, thereby reducing unnecessary tool calls and redundant context gathering. AI harnesses are very useful, and I think Luan will die without one. I have been using Pi (not Oh My Pi) for a Python crypto trading bot that I am trying to make. I hooked up my Codex account to it and use it for development. Pi helps me in several ways. First, it allows me to skim the code as it is being generated in front of my eyes. Second, it integrates the command line with the AI prompting tool, encouraging me to use the command line more often without asking the AI to execute commands for me, thereby reducing AI credit burning. Third, Pi can reduce the necessity of using a high-effort mode for the AI model. The effort mode determines how much and how deeply the AI talks to itself before doing work, and Pi, especially when complemented by specialized tools and project instructions, can give a low-effort model enough structure and context to perform many tasks effectively, such as self-correcting by fixing bugs. This can reduce credit burning, and I also suspect that relying less on overpowered AI programming may reduce the incestual nature of the resulting code. (By incestual, I mean creating code or results that resembles what other people using AI for programming would produce.) Finally, Pi, like other extensible AI harnesses, can support multiple AI agents working on different parts of the same project concurrently, thereby potentially increasing the rate of development. Although Pi does not include built-in multi-agent orchestration, this can be accomplished through extensions or by running multiple Pi instances in separate terminals. Extensions can coordinate agents, whereas separate terminal instances require more manual coordination. Examples of AI harnesses include Pi, Oh My Pi, Prime Agent, Jcode, and Empryo. From what I gather, Pi is the best, at least for my purposes. It best fits the reactionary programming philosophy of focusing on doing the best possible job at just one thing rather than integrating a bunch of things, while allowing you considerable flexibility to customize it the way you need. This approach has served me well because I can complement Pi with, say, CodeGraph to further reduce token burning. Pi has an extension for CodeGraph, which I was able to implement without any issues. In this way, Pi provides a simple foundation that can be extended with specialized tools rather than trying to do everything itself. Maybe we can create a Pi extension for Luan?
mmHg edited
I have a perfect example of how reactionary programmers are not Luddites. But first, a little backstory. In 2024, I took a really fun and unique philosophy class called *Social Epistemology*. The course explored the philosophy of knowledge, particularly as it relates to misinformation, fake news, and conspiracy theories. In other words, how do we evaluate the credibility of information shared through social networks, news outlets, and other media, and how should we respond to such information responsibly? At the time, the literature on social epistemology was quite rich and innovative compared with the literature of other areas of philosophy (except for the philosophy of AI). In 1986, the American philosopher Harry Frankfurt introduced the concept of "bullshit" as content provided by a source that is indifferent to whether that content is true or false. The difference between lying and bullshit is that lying occurs when the content is knowingly false and the source shares it deliberately because of its deceptive nature, whereas bullshit occurs when the source is just spewing garbage (it could be about girls' farts smelling like lavender or whatever) without caring whether it's true or false, or whether it's deceptive or not. In 2024, Hicks, Humphries, and Slater famously argued that ChatGPT produces bullshit in Frankfurt's sense in a paper titled [*ChatGPT is Bullshit*](https://link.springer.com/article/10.1007/s10676-024-09775-5). One of the premises in that paper can be articulated as follows: "If ChatGPT's models are designed to mimic formal human speech, then ChatGPT's models are designed to spew bullshit," with the justification being that formal human speech itself is indifferent to accuracy. In my social epistemology class, I wrote a paper criticizing this charge. My argument was that formal human speech intrinsically depends on the accuracy of its content, and therefore the implication of the premise is unsound. In my paper, I also introduced an idea that my professor succinctly described as "triangulation" upon reviewing it. It simply meant using AI platforms to fact-check each other, either because hallucinations produced by different AI models would have a low probability of overlapping, or because an AI provider (like Perplexity.ai) could share links to external sources. In Perplexity's case, its Sonar models are based on Llama and optimized for retrieval-augmented generation (RAG). Allow me to introduce triangulation to Luan. From now on, provided there are no major problems that aren't worth fixing, I will use multiple models to work on Textile. I have two worker models (Sol for hard work, supplied by Codex, and MiniMax for easy work, supplied by OpenCode Go), a scout model (Gemini Flash, supplied by OpenCode Go), a reviewer model (Kimi K2, supplied by OpenCode Go), and a model for debugging as well as intelligently delegating tasks to the other models (Astra, supplied by Codex). ![image](https://i.imgur.com/Y5DwDvf.png) The scout model determines where Sol can go before Sol goes there. This helps reduce hallucinations committed by Sol, as well as redundant behaviors that would burn credits. The reviewer model, in particular, is a triangulator. After the scout and worker models work together, the reviewer model goes over the scout's path and the worker's path to see whether there are any gaps in their overlap, or potential hallucinations revealed by discrepancies between the scout and the workers. This way of working with agents is useful, increasingly standard, and successful. In recent years, newer companies have already started implementing multi-agent task assignment for fast and efficient production. I pity the old companies and websites that could be immensely improved by starting from scratch with a multi-agent workflow. The moral of the story: Induction alone already shows that AI is not bullshit (hence why philosophy is useless; philosophy has nothing good to say about or against this induction), but triangulation helps us understand why AI is not bullshit. And this understanding helps us better understand the purpose of having a scout model and a reviewer model in a multi-agent AI harness, as opposed to just having a bunch of worker/development models and hoping they don't overlap with each other too much.

Log in to reply.