I’m probably going to stumble over some of the terminology here, but I think it might be possible to describe what @BioMan@awful.systems is proposing as a consequence of LLMs ultimately being lossy compression systems. Inference is a function over a lossily-compressed data set, and “chain-of-thought reasoning” and “agents” may sound sophisticated, but are simply applying containerization and DevOps tools to VM images of the inference application in an attempt to get around hard memory limits on the context window for inference. “Chain-of-thought” attempts this in a serial fashion, passing results from one instance to the next, while “agents” implement this hierarchically and recursively (and woe to the poor bastards who wished that mess upon themselves). But in both cases, the “finalization” phase is necessarily a further lossy compression step, attempting to compress a result from the inference process to a fresh instance of the inference application, so as not to immediately blow out the new instance’s context window.

Given this necessity, it comes to seem somewhat intuitive that there may be “strange attractors” in the higher-dimensional vector space that is the compressed data set which surround code that creates and maintains message passing channels. No matter what you’re doing with an “agentic” process, the inherent necessity of context cramdown & message passing means that querying into the space where such code examples lie is a hidden requisite of running the damned things, thus turning such functionality into the sort of selfish elements that BioMan is talking about.

The problem in investigating and concretely describing this phenomenon is nailing down the exact functions and processes that make it happen. Given the godawful messes in the Claude frontend codebase that @jonny@neuromatch.social has been documenting, I’d be surprised if there’s one developer in a hundred at Anthropic or OpenAI who can describe in detail how the intentionally-developed context-passing code for their “agents” works.