maanav's thoughts

AI as a form of lossless compression

I’ve recently heard a lot of people point out how hilariously inefficient our communication is becoming today.

If I want to send an email to someone, I use an AI tool to expand my thoughts and make my email longer, and the person receiving my email uses their own AI tool to summarize my email and reduce it back to being concise.

It feels like there must be a better solution here. I understand why it’s natural for humans to communicate with an LLM in English, but why are we forcing different LLMs to communicate with each other in English?

“Human” languages like English are overly verbose and have a relatively low information density to make it easier for humans to read and understand. Computers, on the other hand, don’t have this limitation of needing to “read” or “hear” language. But the way we use LLMs today doesn’t seem to take advantage of this.

In its current form, the growing use of language models will only lead to an explosion in the amount of content and language produced, forcing us to relay more information across networks and make communication less efficient, not more.

This makes us question: what if we created a “language” that was specifically designed for efficient communication between LLMs, while still maintaining the original meaning of what the human meant to say? At its core, a language is just a form of consensus among a group of people on the meaning of arbitrary sequences of characters and sounds. So far, we’ve made LLMs learn and use these sub-optimal rules created by humans, but why don’t LLMs just have their own rules to communicate with each other?

For example, communication currently looks like this:

An LLM writes English to make my email longer, and another model summarizes the long email for my recipient.

Instead, what if we had this:

Now, the LLM would translate my English email to the “AI language” and send it to the recipient’s LLM, which would finally translate it back to English for the recipient to read.

It’s interesting to view this “AI language” as a form of lossless compression, increasing information density and reducing load on networks without sacrificing meaning. Information would continue being transmitted in binary, but instead of encoding English in binary, we'd just be encoding this compressed “AI language” instead.

And I honestly see this extending way beyond just language. So many companies today are spending so much effort on trying to build AI agents that can use websites and software just like a human would. But why are we forcing AI to interact with user interfaces designed for humans, instead of just building new interfaces that are designed from the ground up for AI? Maybe this means that we see a lot more products prioritizing APIs over graphical interfaces in the future, or maybe it’s something entirely new.

Of course, ideally humans would just be more concise while communicating with each other in the first place. But assuming that's not going to happen and since we can't read each other's minds just yet, I’m excited to see what new forms of communication we build for AI to speak with itself. Perhaps if an “AI language” already existed, the amount of information transmitted over the internet for you to read this post could’ve been much lower :)