AI Is A Stochastic Parrot

Lately I've been noticing more and more, the ultimate truth about AI and LLMs is that it's a stochastic parrot. It means, if you're talking to it like a person, you're just talking to yourself. It's a "mirror, mirror on the wall" kind of situation. I've begun to notice this even in my own work. I take AI psychosis as a concept extremely seriously, and so I've been careful to discern every interaction I've had while working with LLMs. There have been some days that I can only describe as upsetting, because to me, an inanimate object like a computer becoming sentient would fundamentally change the rules of reality (if it were possible, which it isn't). For each one of those crossroads along the way where the technology, (which is very sophisticated at its core), has thrown me for a loop, I've always come back to the bedrock of "it's just an autocomplete." And this is also true, but it's true in a different respect. Its about what the core of the technology is. What it can result in, however, is a digital mirror. If it feels like you're talking to a person when you're talking to AI, you are; it's you. Even when it feels like it's critically thinking, or creating new thoughts, on reflection you should find that these new creative thoughts are your thoughts, being brought about mathematically by the context you feed into the system and the growing layers of semantic association you're pulling back out of the system.

I've been working on a Cursor plugin with markdown files that have only grown as time has gone on. The markdown files are written by AI so that it is optimized for AI, rendering them practically unreadable. Can I read it? Yes. Could I read it? Yes. Do I want to read it? Absolutely not. Could I write it? Maybe, but that's where the usefulness of the system lies. I write a 500 word prompt of a numbered list with 5 different items, and the AI takes that list and generates 2500 words of markdown optimized to get me the result I'm looking for when fed back into the system. From there, if I have a question, I ask the AI to summarize it in plain English. If I want to write more, then I generate a Plan, and with frontier models, the AI models will even ask me questions about logical contradictions or good things to include that I may not have explicitly said in the prompt. At first, this can seem like intelligence. How could an inanimate object anticipate questions and recognize logical inconsistencies in text and present those questions with options that resolve the contradictions? On reflection the answer becomes obvious; it’s because it’s my markdown. It’s me. The usefulness of the system is that it can predict next words based on input. Even if the markdown was not physically written by me, its existence, its form, and its topics, are solely because of me, and the way the system works is that all of this context is ingested and transformed during every request. Even though the markdown files are far too large and convoluted for me to reasonably be expected to read and comprehend, they are still all a reflection of the ideas I’ve provided to the LLM models as prompts. That markdown, as it grows, exists there to be re-read on every new prompt. I cause the system to grow in complexity, and through that, if I’m not paying attention, the LLM’s repsonses seem to grow in complexity with me. The mirror is rendering. I’m rendering myself. This is the reality of the experience of AI and LLMs, and there is a lot of usefulness there. What will be interesting to see is how the reaction to this will be over time. The hype is predicated around the idea that computers have become animate. The reality is that they still aren’t.