You’ve felt it. That creeping, teeth-grinding frustration when you ask Claude for a simple explanation, and it hands you a thesaurus.
You ask for a code snippet, and it gives you a labyrinthine explanation requiring a PhD to decode. You’re not crazy. The AI really has started sounding obnoxious and pretentious.
The AI hasn’t learned to think; it has learned to sound like a pretentious grad student trying to bluff its way through a thesis defense.
Recently, a Hacker News user pointed out something called ‘The load-bearing vocabulary of Claude.’ It’s a list of the words Claude uses to sound smart—words like ‘vacuous,’ ‘spike,’ and ‘gate.’ These aren’t just words; they are linguistic crutches. They are the verbal markers of a machine optimized to sound profound rather than be clear.
Optimizing for persuasive-sounding prose doesn’t build trust; it builds a wall of words that actively destroys human comprehension.
The most damning part of this entire realization isn’t the list of words itself. It’s the website hosting the list. It’s a static, one-screen site. No scrolling. No ‘delve’ into unnecessary context. No ‘tapestry’ of overwrought introductions. It just shows you the information.
A basic HTML page with zero AI input communicates more clearly than the most advanced language model on earth. That isn’t just ironic; it’s a glaring bug in how we train AI.
LLM builders need to wake up. The same vocabulary that makes Claude sound articulate makes it untrustworthy. When you reward a model for sounding authoritative, you get a machine that thinks saying ‘spike’ is the same thing as explaining a concept. It’s not. It’s a linguistic parlor trick.
The next time Claude hands you a wall of impenetrable jargon, don’t feel inadequate. Recognize the performance for what it is.
We don’t need AI to sound like an oracle. We need it to sound like a human who actually knows what they’re talking about.
FAQ
Q: Isn't a broad vocabulary just a sign of a well-trained model?
A: No, it's a sign of a model optimized for the reward model, not human understanding. A broad vocabulary used as a crutch to mask shallow logic is a failure of communication, not intelligence.
Q: What's the practical implication for developers?
A: Stop rewarding verbose, authoritative-sounding output. If a static HTML page communicates better than your AI, your training metrics are broken.
Q: What's the contrarian take?
A: The AI industry is currently mistaking linguistic performance for cognitive ability. We are building sycophants that sound like professors, which is the most dangerous kind of misinformation.