Imagine it's 1866. You're sending a message across the Atlantic cable, and every single word costs you $10. Naturally, you'd strip away every unnecessary 'the,' 'and,' or 'is' just to save a few bucks. Fast forward to today, and we're facing a strangely similar economic constraint: the LLM token.
The Return of 'Cablese'
In a fascinating experiment, a developer noticed that LLM API bills are essentially modern-day telegrams—you're charged by the token. By instructing a model to write in 'telegraphese' (or 'cablese'), a terse, compressed style of communication, the results were startling. In a test involving 50 prompts and roughly 1,300 tokens, the cost of the output was nearly halved.
By telling the AI to ignore formal grammar and use lowercase, compressed phrasing, the model effectively relearned a linguistic relic of the 19th century to optimize for the 21st.
Efficiency vs. Eloquence
This isn't just a quirky hack; it's a lesson in the intersection of linguistic constraints and computational cost. While we often push LLMs to be more verbose and 'human-like,' there is a massive operational advantage to brevity.
However, this comes with a trade-off. As we've seen with other linguistic edge cases—like LLMs occasionally hallucinating Greek words when they hit a wall—stripping away the 'filler' can sometimes push a model toward instability or unintelligibility if the compression is too aggressive. The goal is to find the sweet spot where the meaning remains intact, but the token count plummets.
The Future of Prompting
Will we eventually see 'compression layers' that automatically translate our flowery requests into telegraphese for the AI and back again? It's possible. As we move toward more complex AI agents, reducing the 'noise' in the communication loop will be key to scaling without breaking the bank.
Sources
Media
