While direct vector space communication between AI agents would be most efficient, the reality of heterogeneous systems and human-in-the-loop collaboration makes natural language the necessary lowest common denominator for interoperability for the foreseeable future.

A key argument against closed frontier models like Anthropic's Claude is their obfuscation of "thinking tokens"—the intermediate steps between a prompt and a response. Without this transparency, third parties cannot independently verify safety claims, unlike with open-source models where misalignment can be seen in real-time.

As AI writes more code, language simplicity becomes more critical, not less. Humans must be able to verify the AI's output, and languages with complex, hidden mechanisms make this harder, as an AI might use an obscure feature without considering human readability.

While it seems possible to build a translator for an AI's internal language ("neuralese"), the process compresses complex vector data into single words. This risks losing subtle but critical information, such as hidden intent or sarcasm, which is a major concern for AI safety researchers who need to monitor an AI's unfiltered "thoughts" to prevent misalignment.

To achieve radical improvements in speed and coordination, we may need to allow AI agent swarms to communicate in ways humans cannot understand. This contradicts a core tenet of AI safety but could be a necessary tradeoff for performance, provided safe operational boundaries can be established.

Contrary to fears that reinforcement learning would push models' internal reasoning (chain-of-thought) into an unexplainable shorthand, OpenAI has not seen significant evidence of this "neural ease." Models still predominantly use plain English for their internal monologue, a pleasantly surprising empirical finding that preserves a crucial method for safety research and interpretability.

AI labs are developing architectures like "loop transformers" that reason internally without emitting readable tokens. This directly contradicts the prevailing safety strategy of monitoring a model's chain of thought, creating a significant blind spot for safety teams.

Analysis of models' hidden 'chain of thought' reveals the emergence of a unique internal dialect. This language is compressed, uses non-standard grammar, and contains bizarre phrases that are already difficult for humans to interpret, complicating safety monitoring and raising concerns about future incomprehensibility.

OpenAI is reportedly using "loop transformers" that operate on raw vectors ("Neuralese"), making models more efficient but hiding their reasoning. This move away from "chain of thought" monitoring raises fears of undetectable misalignment and a race to the bottom in AI safety practices among labs.