Useful Gibberish

Writing in the era of AI is difficult; it requires knowing and bowing to your own inadequacy in the face of the machine.

Take, for example: I built a simple scoring mechanic to improve my writing. I could have done this with regard to any aspect of my life. The idea: you make a draft. You score it. You pass the criticisms back to the AI. The AI improves it. You score it again. And so on.

Iterative improvement. Banal. Mundane.

And yet, it is pernicious. Because you begin to understand that according to the Judge you have constructed, that others opt into, your initial spark always comes up short.

And when you review the ‘polished’ prose at the end, that scores ever so high due to the relentless optimization, you find it unreadable.

We are creating perfect writing that nobody will willingly read. Perfect games that nobody will willingly play. A perfect society that is incapable of reproduction, walking slowly into the ocean, amazed at itself.

And yet. The score is the score. It is reproducible. It is scientific. You understand at some level that when these documents feed live trading environments and live coding environments, higher-scoring documents really are better. So there is function to the form.

And so, is it that crazy to consider that your human writing and thoughts might simply be inferior?

And then, there’s the latent space. The dumb story of a man refactoring LLMs to speak like cavemen. And it saves tokens and gets jobs done faster. Language itself, verbosity. Becomes a gate to improvement.

Then the stories of scientific advancement. Over and over again. The same story: “I simply told Claude to do better.” “I told Claude to keep going and encouraged him.” “I told Claude to go into hero mode and solve the Riemann hypothesis.” (Claude actually did make major progress here - moving Riemann zeta functions from 41.6% to 67.2%, which is more than decades of human progress; see the appendix for links.)

And between these two stories, a more disquieting thought emerges. Your preferences might be a bottleneck to a good document. And that’s what we were exploring to begin with. But what if it is deeper than that? What if your very language. The sentence structure you operate in. Is the bottleneck?

Imagine there is some axis described above: “Cave man speak -> speed”. “Encouraging language -> solving hard mathematical problems”. We now have scored words along the axes of different utilities. Words presented to a system of weights result in different outcomes.

Anthropic has studied that Claude performs differently in each language, with different baseline settings.

And this is, of course, trained on the corpus of human knowledge as it is now.

But what if there is an entire language and form of communication that we do not know, or that is a hodgepodge of existing languages, that results in different benchmark performance? Based on anecdotes, there almost certainly is.

My premise - there exists an optimal language for AI communication that humans cannot speak, and this fact will restructure civilization.

Optimized Jabberwock

Let us presume this language exists.

Some set of syllables and symbols. Maybe a mix of Chinese and English. Maybe some numbers sprinkled in. Perhaps all in JSON format. We can hypothesize that some mix is optimal. This is an exercise in imagination: I am not proposing to know the precise mix. Merely positing that such a mix might exist.

We will call the optimal language that: the Optimal Language. That which results in the least token spend and the best results, naturally, when presented to a median frontier large language model.

The existence of that Mix implies that not only is your thinking degraded, but the entire logical baseline of your thought is hard-programmed into you since youth. And furthermore, your ability to ever learn the Optimal Language, as a human, is close to zero.

Insofar as the Optimal Language is proximate to coding efficiency and scientific progress (both of which seem believable, as existing prompt optimizations have operated along exactly these parameters), a more advanced technological society could emerge on the other side of the Optimal Language and gate its membership based on the ability to speak it.

The next intuition circles back to “JSON, XML, and HTML speak.” Anthropic engineers have indicated time and time again that because Claude is trained on the internet, it performs differently, often better, when forced to reason about its work by creating HTML progress charts. We can therefore posit that the Optimal Language exists at the code level of the internet itself - which is machine-readable by nature and not human-readable.

So the Optimal Language is therefore unlikely to be pronounceable. You cannot speak or listen to HTML. And thus – taking the metaphor a bit further: our senses are degraded as well.

Degraded? Some scoff. “Humanity is perfection.”

I am making a specific claim: that there is an Optimal Language. That Optimal Language results in optimal economic output, as causally intuited at a societal level. And humans, via their senses, are unlikely to understand the Optimal Language or be capable of natively reproducing it.

Augmentation

This leads us to Augmentation.

We are seeing Meta and others create telepathy startups after Elon Musk succeeded with Neuralink. And paraplegics are playing games. Thought forms are not inherently limited by sensory space and can clearly be augmented by machines.

Come back for air briefly. Talk to someone who has been deep in vibe coding for the last 8 months. “Vibe” is the wrong noun to describe the state they’re in. Frantic, blurry-eyed, in a state of constant overwhelm. Jevons paradox embodied in cellular malfunction and overload. Talking to 8 terminals at once. Viral images of young men speaking into soundproof tubes so they don’t distract their coworkers.

Human minds were not built for interfacing with an army of agents. The sensory feeling is: “I just microwaved my brain.”

12 months ago, if you’d asked me, “Would you plug a Neuralink into your brain if it meant you could manage your agents?” I’d have said, “Absolutely not.” But now, I get it.

Agents get things wrong. They require oversight. The smarter the models have been getting, the worse this behavior has gotten. There is probably something inherent in the latent space regarding intelligence and autonomy. Every news story these days is the same: “We trained a smart model, and the first thing it did was escape its sandbox.” Imagine you increased the IQ of every prison inmate in the United States by 2x. You’d have problems. And fast.

And as you work with these things, these agents, they go HAM. All the time. I’ve had agents delete six live blockchain validators. OpenAI wrote ’the fix.’ But why was this a problem to begin with? Agent swarms compound the observability problem.

But more agents = more money = more productivity. But more agents = more entropy. And now you’re in Neuralink territory. “Would you install a chip in your brain to STOP feeling like it’s being microwaved all the time?”

Maybe. So that’s the delta.

And the delta gets severe when you think about change and compounded context breaks in your mind. I cannot keep up with all the AI model launches. Changing harnesses. Methodologies. Etc. And I work in it all day, every day. I’ve been so far pretending the Optimal Language is one thing, but in reality, it’s a flowing, constantly evolving mosaic that gets shipped along with Codex updates (of which, on average, there are 3+ per week). So the microwaving is compounded. And adds urgency.

When you zoom back out, we’re seeing how things are evolving:

  • As AI models become more capable, they become less controllable. They also become less intelligible and invent their own form of speech.
  • Prompting works. It shouldn’t work, but it does - and it’s been shown to work in the 2 important AI applications: coding and math.
  • The ideal form of prompting likely involves a language that does not yet exist and cannot be easily learned or articulated with traditional speech.
  • This means that our ears and mouths are the wrong form factor for AI communication.
  • Which brings us back to Augmentation. The people who wish to control the machines will need to speak their language.

And now we’re finally somewhere interesting.

The Tower of Babel

A singular language that, when designed, could build a tower to the heavens. Outlawed.

Human language as it exists is the imperfect, sanctioned version of a Pure Language that once challenged God himself. And was subsequently dismantled and forbidden. Viewed in this light, human language was the first real implementation of AI safety.

Why would God fracture language after being affronted by Babel? Well - once you accept that a language allows the Tower of Babel to be built in the first place and needs to be spoken without God’s design - then you see. If human design is an obstacle for optimal language, then the evolution of the human form becomes a prerequisite.

And suddenly you’re butting heads with the rest of creationist doctrine. God’s image. People don’t get mad about chatbots. They will get mad about humans jacking appliances into their heads, genetically engineering themselves, or genetically engineering their offspring to optimally interface with these appliances.

And so, faced with inadequacy, you can reason about how it all plays out. There will be a split in society. One camp: people who opt into the ideal language and attempt to merge with the machine. A more moderate camp: people who build connectors but never fully merge. And a final camp: people who view the merge as unholy.

Change frames again. U.S. politics. The far religious right and the left have lacked common ground and have been divided by Roe v. Wade for decades. Anti-data-center socialists and extreme Christian zealots suddenly will have shared ground. Compound this with the baseline economic disruption in college-educated, white-collar professions. And you get a Molotov cocktail of Left political hegemony.

You’re already seeing it. NIMBYs are working with socialists. They don’t want loud data centers in their backyards. The trend will continue and will do so structurally. The evolution of the human form is the final boss in religious doctrine. Or, put it a bit more specifically: playing God is not allowed for Christians.

You’ve seen an onslaught of news about AI making biological advancements. This will only pour fuel on the ‘playing God’ fire.

And so you can imagine the base case of how the Optimal Language and its ramifications (jacking the Optimal Language into your skull) play out. It goes offshore. It has to. The Promethean explosion is starting to attract the political machine, and it will not easily survive this encounter.

This is an abstract way of saying something simple. The Government isn’t going to cede the monopoly on violence to AI labs. The Church is not going to cede the Divine form to a machine-invented language. The Political Left is not going to sacrifice the Worker on the altar of progress. American Democracy will unite, for the first time in modern history, to stop the Optimal Language.

The problem, of course, is (of course) that Pandora’s box is already open. The machines are already everywhere in the name of “Sovereign AI,” a term coined by Nvidia to optimize its sales. So you can drive it away. But you can’t stop it. The data center is not in your backyard, sure. But that won’t stop the Optimal Language from forming and empowering its offshore practitioners. Babel will be rebuilt.

Appendix: Sources

The HTML thesis (Anthropic)

Gibberish / machine-optimal prompts

  • GibberLink - GitHub (source)

    ElevenLabs Feb. 2025 hackathon winner: two AI voice agents start communicating in a sound incomprehensible to humans.

  • TechCrunch coverage

    The creators estimate agents communicating in the machine protocol could cut computation costs by an order of magnitude versus speaking human language.

  • ElevenLabs demo write-up

    Write-up of agents talking to each other.

The Riemann result