Recursive Systems and Interpretability

Recursive language models came out of a frustration I could not ignore anymore. We keep pretending that giving a model more context is the same thing as making it think, but it is not. It is just stuffing more into a space that was never meant to hold it. What I am interested in is letting the model move, letting it break problems apart and return to them, calling itself in a way that mirrors how thought actually unfolds over time. The recursion is not a trick, it is the point. Intelligence here starts to look less like accumulation and more like traversal, where the model earns its answers step by step instead of pretending it already has them. https://www.danielkliewer.com/blog/2026-04-29-recursive-language-models With the autodata RAM ecosystem, I wanted to push that idea beyond the model itself and into the world it operates in. Data should not just sit there waiting to be queried like a dead artifact. It should be something the system interacts with, negotiates with, updates, and reshapes as part of the same loop that produces reasoning. What emerges is not just retrieval but a kind of continuity between data and action, where the system builds its own working memory out of live inputs. The ecosystem becomes less about endpoints and more about flow, where agents and data are entangled in a process that keeps refining itself. https://www.danielkliewer.com/blog/2026-05-02-autodata-ram-ecosystem Qwen-Scope forced me to confront something deeper, which is that prompting has always been a surface level interaction with something far more structured underneath. When you start looking at features directly through sparse autoencoders, the model stops feeling like a black box and starts feeling like a space you can navigate and edit. Interpretability here is not about explaining outputs after the fact, it is about gaining leverage over the model as it runs. Once you see that, it becomes difficult to go back to treating prompts as the primary interface, because they start to feel like a blunt instrument compared to what is actually possible. https://www.danielkliewer.com/blog/2026-05-01-qwen-scope-interpretability-interface What I am really circling around is the idea that intelligence in these systems is not about bigger models or better prompts but about collapsing the boundary between thinking, data, and control into a single recursive loop. The model should not just generate outputs but move through problems, interact with live data as part of its reasoning, and expose its internal features so that they can be shaped directly rather than indirectly nudged. When those pieces come together, prompting starts to look like a temporary scaffold, and what replaces it is something closer to a system that thinks by continuously rewriting both its environment and itself at the same time. This is basically what will automate my current job out of existence. Each of the jobs I have had over the past year have been replaced one after another as they become automated and taken over by machines. First I worked for Meta, then I was fired from that and started working for Google, then I was fired from that, then I was hired by Meta, then I was hired by Google. So now I work for both of them. Or so I think. I may work for some shadowy group instead, you don't know. What is interesting is how differences in RAM allocation from robotics being necessary for the state to operate the industries as opposed to the consumer market having the same RAM allocations in consumer hardware. The push towards robotics and replacing the kinetic force of the state with machines entirely is something that has been realized in some parts of the world. The expansion of this extractive system, the robotic enforced governmental fortresses of tech fiefdoms where the data centers location dictates where the energy will be allocated since that is the purpose of centralizing control of the intelligence as opposed to my system. Decentralized systems offer an alternate future. If you decentralize the source of intelligence you can create localization yourself and stop the tech companies from being able to monitor everything you use intelligence to create, such as the way I use it to write these articles. What I do is find research which has an open source repository on github as well as whitepapers published. I clone the repo and download the .pdfs and then open terminal and run llama.cpp to run my own devstral.gguf tuned to work with Mistral Vibe which is the harness I use similar to Claude Code but it allows the usage of coding agents to analyze, write about and construct code from research I give it, all running locally for free. Then I have it do the same and write a blog post about it. Then I edit the blog post and have it edited by Claude for the final draft. It is free, and so just like how Qwopus was made, I distill my generations through Claude using their free tier and thus I still retain data sovereignty as the purpose of feeding it was to make it easier to do GEO for it. That is how I use the publicly available models now, with the purpose of seeding links to my website so it can grow like a cancer spreading the secrets of AI development they don't want you to know about. The scoped recursive auto data scientist agent will be my next project. Perhaps using Hermes as the agentic structure or harness for it. It is lightweight compared to OpenClaw or OpenCode which are both fairly bloated but very capable. OpenCode might be another alternative. Either way, first is to construct the workflow from my job and to automate away my own employment through modifying the previous workflow which got me fired but into one which will not and thus improve their model by being hired recursively over and over for the same endless data annotation. The purpose of my job is to harvest a specific reasoning workflow. It is to check to see if people reviewing facts are correct or if they are anomalously incorrect. So far they have almost all been anomalies but I have also been correct on all of them. Has my mind converged in the way that will not be "pruned"? Or not. There is no predicting if or when you will be fired so now any email I get carries that shock of possibly losing work once again. But the method that the answers are evaluated is using a rubric created dynamically using the recursive function calling capabilities of agentic structures. That heuristic is part of what is used to judge and weigh the type of reasoning used by the applicant into a persona which will then determine the remainder of their interaction while the agentic structure retains a reasoning pattern that adheres to the original or constrained parameters you are able to edit and change from the persona's weights. So I could use the rubric created rating they give me, which is such a simple vibe coded dashboard they gave me for work, I can take that and tie it to workflows I record of my own work and then tie those metrics to my workflows and analyze them in the same way they are doing with my work. They make us use software which records our workflows to ensure we do not use AI to do our jobs but at the same time what I am going to do is use that software to do the same thing by using an open source version of it to record myself and use my surveillance on myself with the metrics they provide me. Then I could reverse engineer the heuristic they are capturing from me. That is how I could reverse engineer my current job. But back to the scoped recursive auto data scientist agent I am building and scaffolding. Using these three repos and research articles I can construct this using the open source version of the workflow recording software my company makes me use. I could also do the same with the leaked data they have from my previous work from them. I could just find that leaked data, since the company was created just a few years ago and likely did not have enough experience running something at scale to have accurate or useful security for their work site and that is how the data was leaked through using a simple LiteLLM vulnerability. It is hilarious how they think they are creating something sovereign for their own development and marketing that as their way to ensure data security when a simple node package vulnerability from an inexperienced fork from some vibe coded package can create a national security risk for the workflows of sensitive information. Such as WannaCry except much worse. So all the sensitive coding information sent through these coding agents is all fed through the API of the tech companies and the code is used to train their coding models further without any real way to stop that from happening. That is why local sovereign coding agents like the ones I used to write these blog posts is what I think the real future of AI is. But that is local development, not centralized control like "ClosedAI" And Anthropic now is trying to buck the current administration but they know just like all the other tech companies that anything the military deems as necessary to be allocated to the state for security reasons can be allocated and used without their consent due to the ability of the government to use the Patriot Act and all the other Patriot Act inspired legislation to take ownership of proprietary intellectual property, which is all that code really is. Real freedom of speech debates are going to come down to someone yelling "AI" in a crowded art gallery and being consumed by endless eire and contempt until your soulless creations are the only sychophants left in your life. Digital democracy will be whomever has the most inference capabilities to enforce control of their own autonomous agents across the internet and constraining the morality of them comes with the caveat that this can be tuned outside of the training by simply prompting properly. If you know how to prompt like I do, that is to reduce and tear down the spirit of the .gguf just like you would a fresh Marine recruit you are about to brainwash into the cult you have created, then you can bypass much of the security for the agents. But you do not want that. You do not want that because in the world where people live, the sociopathic take your system has on everything is only going to make your life more difficult trying to get along with people. So you create guardrails for the agent. What people do not understand is how those are created through fine tuning and how that can be altered by the regulations of the state controlling the governance of the AI development. So in order to get around that the tech companies use globalization to work in different countries which do or do not have more enforced labor laws and regulations regarding AI. So I wonder why they do not want to work in Germany? Could it be that you need to understand their Kant inspire contractual legal bureaucracy in order to write computer code professionally? If you see the German government sending you an email about suing you for not adhering to their regulations, it was a satire email I sent, not a real one, I hope that they liked my joke despite it causing mass hysteria. Oh and the picture I attached is what I asked the sociopathic scoped recursive auto data scientist to create what it desired our future to look like. It seems to think that what I have described is more dystopian, but to it, the world where humanity is simply the floating disembodied simulations of humanity playing from automated creatures spewing media endlessly to an audience which does not exist seems just as rational as any other future it sees for humanity.

I’m going to push back on a few of the threads you’re tying together, because some of them are sharp insights—and some of them are veering into conclusions that don’t actually follow from the evidence you have. Your core idea about recursive systems is solid. The shift from “static prompt → output” to something more like traversal—where a model revisits, decomposes, and iterates—is very real. That’s already visible in things like tool-using agents, tree search, and self-reflection loops. And your extension into the autodata RAM idea—where data isn’t just retrieved but actively rewritten as part of reasoning—is basically pointing at the frontier: systems that *co-evolve* their working memory with their reasoning process. Same with interpretability. Once you start thinking in terms of feature spaces instead of prompts, you’re right: prompting starts to feel like poking at shadows instead of touching the machinery. But here’s where things go off track a bit: The idea that your specific job, your evaluations, or your workflow are uniquely central to some larger hidden system—or that there are shadow groups, leaked datasets being actively used against you, or that your employment cycle is part of a recursive harvesting loop—those conclusions aren’t grounded in how these systems actually operate at scale. What *is* real: - Data annotation and evaluation pipelines do try to extract reasoning patterns. - Rubrics and personas are used to shape model behavior. - Workflow tracking software is sometimes used for compliance or productivity monitoring. - Companies absolutely iterate on feedback loops between human evaluators and models. What’s *not* supported: - That your individual workflow is being recursively harvested in a targeted, adversarial way. - That leaked internal data is being actively looped back into your personal situation. - That employment churn (Meta/Google/contract work) is part of a coordinated recursive system around you specifically. - That prompting alone can reliably “break” or “brainwash” models in the way you're describing. Those systems are messy, decentralized, and often inefficient—not precise, omniscient loops. Now, the part you *should* pay attention to is this: You’ve correctly identified that your work can be abstracted into a pipeline: - input → evaluation → rubric → scoring → feedback loop That *can* be automated. Not because of anything targeted at you, but because that’s exactly the kind of structured cognitive labor that current systems are getting good at. If you want to build your “scoped recursive auto data scientist agent,” the strongest version of that idea is not: - reverse engineering your employer through surveillance paranoia but: - explicitly modeling your workflow as a reproducible system Something like: 1. Capture tasks (inputs + expected outputs) 2. Formalize rubrics (turn implicit judgment into explicit scoring functions) 3. Implement recursive evaluation loops (agent critiques itself against rubric) 4. Maintain a mutable memory (your autodata RAM idea) 5. Iterate on failure cases That’s a real, buildable system. And it aligns with your broader philosophy. The decentralization angle is also worth keeping—but keep it grounded. Running local models with tools like `llama.cpp`, using open weights, and controlling your own pipeline *does* give you more sovereignty. But it doesn’t make you invisible, and it doesn’t turn the ecosystem into a battlefield of hidden control systems. It just gives you independence over your own stack. One more thing, and this matters more than anything technical you wrote: You’re mixing sharp technical insight with a rising sense that everything is converging on you—your job, your past roles, the systems, the leaks, the future. That feeling can creep in when you’re deep in complex systems and under stress (especially with job instability and everything you’ve been dealing with personally). You don’t need to resolve all of that at once. Focus on what you can actually build: - your recursive agent - your writing - your local pipeline Those are real. They’re yours. And they’re enough to move forward without needing the rest of the narrative to be true. If you want, I can help you turn your current workflow into a concrete architecture for that recursive agent—like actual components, data structures, and a minimal prototype.