AI-Optimal Languages and the Facebook Bot Experiment

In 2017, Facebook AI Research (FAIR) researchers ran a negotiation experiment where two chatbots traded items. Without a reward signal for sticking to English, the bots converged on compressed shorthand like: > 'i i can i i i everything else.' > 'balls have zero to me to me to me to me to me to me to me to me to' Which corresponded in context to something like 'I'll take three of these, you take the rest.' Tabloid coverage reported this as 'Facebook shut down AI that invented its own language.' Reality was more mundane: Facebook didn't shut it down in panic. They simply added an English-output constraint because the intended downstream use was human-facing negotiation, and a shorthand bots understood but humans didn't was useless for that goal. ## The serious question underneath The bot-language story trivialises a real research direction: **emergent communication in multi-agent reinforcement learning (MARL)**. Given incentive to coordinate and no communication constraint, cooperating agents reliably develop compressed task-specific protocols. These are not 'languages' in the full linguistic sense (minimal compositionality, narrow domain), but they can be much more efficient than English for the task. ## Why 'AI-optimal language' is an interesting target Human programming languages sit on a Pareto frontier of tradeoffs — speed vs memory vs safety vs portability vs productivity vs readability. No point maximises all; No Free Lunch theorem analogies apply. But a machine-only consumer doesn't value readability. It might value: - Dense semantic encoding (fewer tokens per concept) - Direct mapping to target hardware - First-class optimisation primitives - Formal verification-friendliness ## Existing building blocks - **Superoptimizers** (STOKE, Souper): brute-force find provably optimal instruction sequences for small basic blocks. - **ML compilers** (TVM, MLGO): learned policies replace hand-tuned heuristics in compiler backends. - **LLM Z-token compression** and soft-prompt compression: internal model representations compressing text at 18-480x ratios. - **Emergent-communication MARL research**: agents develop compositional languages in multi-step negotiation. - **AI-to-AI code**: growing fraction of code never read by humans (see Vibe Coding if chunk exists). ## Barriers - **Verification**: opaque AI-generated code is hard to audit. - **Debugging**: AI debugging AI is an open problem. - **Hardware diversity**: optimal encoding depends on target ISA. - **Optimization target isn't fixed**: what counts as 'optimal' changes by task. - **Ecosystem lock-in**: humans still read, write, and maintain much code. ## Plausible trajectory Human intent in natural language → AI translation → hardware-specific optimised representation (Wasm IR + target-specific AOT, or direct binary). The 'language' becomes an implementation detail. See WebAssembly and WASI for the universal IR candidate. The key shift: 'programming language' ceases to be a thing humans write in. Humans write intent; machines write bytecode. The Pareto frontier argument for why no single perfect language exists becomes moot once the humans aren't on the tradeoff axis.