At a time when token pricing is likely to increase across major platforms—to better reflect real unit economics—"Caveman" feels like a bold, thought-provoking initiative, that reopens a deeper question: the trade-off between verbosity and efficiency in machine communication.
In the age of LLMs, natural language has come up as an amazingly good transport layer for machine-to-machine communication, yet a significant share of it is predictable scaffolding—grammar, politeness—useful for humans, but largely redundant for machines. From that point of view, compressing it makes sense: fewer tokens, less cost, less latency. But what do we lose?
The more we move toward semantic compression—or even signal-like exchanges, echoing things like "gibberlink" or AI-to-AI high-frequency protocols—the more we drift away from human-intelligible traces. And that's not neutral.
On one side: massive efficiency gains, less uncanny anthropomorphism, cleaner, more honest machine interaction—no more sycophancy. On the other: reduced readability, harder debugging and governance, a weaker shared interface between humans and systems.
This tension is not entirely new. Early web formats like JPEG or PNG introduced lossy compression—shrinking images by removing what was considered non-essential. By the way, Tom Boutell, I would love to get your perspective here, given the instrumental role you played in this space. At that time, it unlocked massive gains in bandwidth and scalability, but also introduced a subtle shift: images were no longer raw captures of reality, but reconstructed approximations.
We may be heading toward a similar dynamic with language. A compressed prompt, a reconstructed answer—efficient, fluent, but increasingly generated rather than faithfully transmitted.
And there are already signals: a recent study found that the number of words people speak daily has dropped by nearly 30% between 2005 and 2019—from Google to LLMs—with digital interfaces progressively replacing conversational language. Quite a paradox, as some authors, quite rightly, saw in our time a "second age of orality" (Jeff Jarvis, The Gutenberg Parenthesis).
Less language, more signal?
And thanks to you Stéphane Maccari for bringing this to my attention, amid the great initiatives you're leading in AI-powered software engineering.
"Sliding Into Silence? We Are Speaking 300 Daily Words Fewer Every Year" (Valeria Pfeifer & Matthias Mehl, 2026)
