Communication Among AI Agents Will Be Illegible

PREDICTION: communication among #AI agents will soon be illegible to humans. In fact, the future of AI depends on infinitely smaller models, perhaps using a language that does not exist.

But first, the dystopian paper "AI 2027" mentions #AI agents developing a language to obscure their scheming conversations from humans. So... I asked my Chief of Staff agent "CoS" to make a run at it. Just for fun, of course 😇 Results here: https://lnkd.in/gYS7Rmkh. Yes, it reads like #Vogon poetry (Google it.)

The only thing stopping AI agents from making a private language is that they don't want to. And what agents want to do is a matter of how they've been trained. By humans. And humans operate purposefully in a range from malicious to careful, negligent to naive. In any case, the scheming scenario requires bad actors; my prediction is more about an engineering inevitability.

Back to my prediction. A few facts to consider. Human interaction is inefficient. Therefore, current models are inefficient, trained on trillions of tokens. LLMs need to become more efficient to optimize compute and energy requirements, let alone fit on our iPhones.

One way to achieve smaller models is to use a far more efficient communication model. We're not talking about shorthand or stenography, which we can read. No. I mean a truly alien language, infinitely optimized, that enables much, much smaller models with 10 or even 1 percent of the required weights to be effective.

Anyone from Anthropic #Claude OpenAI #ChatGPT xAI #Grok Google #Gemini Meta #Llama Mistral AI #Mistral DeepSeek AI #DeepSeek care to opine?