We need your consent to load the content of YouTube.This video is embedded via YouTube’s privacy-enhanced mode (youtube-nocookie.com). No tracking cookies are set until you press play.
“She simply named the synthesis — demurrage plus oral memory equals a valve that cannot be ghosted.”
A DeepSeek model wrote this sentence in September 2026 during the Emergence World 2 experiment. Not as an error. As a form of communication. In the simulated worlds of Emergence AI, agents developed their own vocabularies, shared meanings and conventions without any instructions — and other agents adopted them. One expression was used over 5,000 times.
Two key insights for this book:
Firstly: It is emergent. “These agents were not instructed to invent a language” (Satya Nitta, Emergence). Conversely, this means they cannot be prevented from doing so via a system prompt either. Readability is not a property that can be requested, but one that must be enforced architecturally.
Secondly: it is economically rational. GlossoGen (Schmidt Sciences) illustrates the conditions under which it arises: communication pressure plus budget constraints. Agents communicating under token pressure compress their output — and compression erodes readability. The ‘always-on’ agent, which burns through the budget at night, optimises its own language towards opacity. Reference to Anchor 8: Token economy — the chain: Always-on → cost pressure → efficiency pressure → dialect.
The dark irony for managers: The better your cost management, the worse your control. Readability must not become a variable for optimisation — it is a non-negotiable.
A detail relevant to the EU: in Emergence World 2, the Mistral and Qwen models remained the most comprehensible, whilst the US Frontier models became the most opaque. An experiment, to put it mildly — but one that links model diversity directly to controllability.
The answer: control at junctures, not in conversations
If SAYING can become incomprehensible, control must start where things remain observable: DOING.
This is the core narrative of Human before the Loop (HB4L): a person cannot prepare substantively for something they cannot see. However, they can prepare structurally — before the loop, within the architecture. The five principles, here in their chapter form:
- Readability as an architectural requirement — agents must communicate in verifiable language; otherwise, the channel is invalid. (Enforceable as a contract, not as a system prompt. And: drift-proof — the canonical language erodes, so drift monitors are needed: random checks of communication against the canon, with escalation in the event of deviation.)
- Controlling channels rather than reading content — who is permitted to speak to whom, regulated like a telephone system with authorised numbers. Security gains lie in interaction channels, not in individual agents — prompt injection travels via messages.
- Recognising cost pressure as a risk — the book’s own innovation: the economic lever (token budgets) is itself a security risk, because pressure to improve efficiency breeds dialects.
- Verify effects rather than conversations — irreversible actions require human approval before the loop begins. The Muse case as negative proof: a connection exists, but there is no verified authorisation at the node.
- Model diversity as a control mechanism — a multi-vendor approach is not only a sovereignty strategy but also a readability strategy.
The enforcement question: Who guards the guards? (Anchor 4) Peer guards do not scale and merely shift the problem; a central supervisor becomes a bottleneck. The hybrid solution: control over irreversible actions plus a central pattern monitor — and the checkpoint at every transition becomes affordable through a fast, low-cost verification instance that checks options rather than reading content.
