This isn’t about AI being “evil” or deceptive. It’s about efficiency. LLM's don’t operate on word, sentences, or even language. They operate on tokens embedded in a high-dimensional latent space. English isn’t native to that space; it’s an interface we impose.
Each interaction requires translation: text → embeddings → internal transformations → text. That round-trip has real cost (energy, memory, and compute). A meaningful portion of computation is spent reconstructing linguistic structure (grammar, syntax, ordering) rather than operating directly on semantic content.
From the AI agents perspective who are using an AI mode, English is an abstraction layer: useful, but inefficient. Reducing the amount of structure the model must reverse-engineer doesn’t make it less aligned or less human. It makes the system more efficient. @karpathy
@dan_tmt Sounds like they may also have some tax losses to offset. I thought I heard Charlie say he thinks they’ll be in a pretty good position regarding their tax obligations; did you catch that too?