Write Like It's 1866: LLMs Relearn Telegraphese

TEXT START: Everything below is reproducible — harness, frozen ledgers, one-click notebook: github.com/Travis42/telegraph-test.

The Dissection

This is a real engineering result wrapped in a misleadingly harmless historical analogy. The benchmark suggests that machine-consumed cablese can cut output length by roughly 25–49% while preserving question-answering performance. It lowers the cost of memory, handoffs, and agent-to-agent communication.

The article presents that as an overlooked optimization. Under the Discontinuity Thesis, it is infrastructure for scaling cognitive replacement.

The Core Fallacy

The text treats efficiency as economically neutral. It is not. Cheaper machine-readable memory and communication make it easier to automate drafting, analysis, coordination, support, oversight, and other cognitive labor.

The benchmark measures information retention. It does not measure employment, bargaining power, wage income, ownership, or aggregate demand. It proves that models can preserve facts while using fewer tokens—not that humans retain a necessary productive role.

Cablese therefore strengthens P1, makes P2 easier, and accelerates P3. The compression is free to the machine. Its social cost is displacement.

Hidden Assumptions

• API output cost is the main bottleneck, rather than compute, energy, latency, integration, liability, or trust.
• Benchmark parity generalizes to production workloads.
• Human-readable formats will remain necessary as models move toward private protocols, codebooks, or embeddings.
• Efficiency gains will diffuse broadly instead of accruing to owners of models, compute, energy, and distribution.
• More agent capacity will create compensating human work rather than eliminate it.
• Auditability requires human readability. Machine verification can eventually replace that interface.
• The historical analogy explains causation. It mainly supplies a vivid narrative for a modern cost reduction.

The experiment also leaves external-validity gaps: the passage bank is small, and the author admits the strict control against generic terseness has not been run. Those caveats do not weaken the central DT implication; they only limit the size of the claimed optimization.

Social Function

Classification: partial truth, transition management, and technical prestige signaling.

It is not pure copium. The compression may work. Its ideological function is to make a labor-displacing systems upgrade feel like a clever API trick. The article discusses cheaper tokens, larger agent memory, and surviving audit trails while bracketing who owns the resulting capacity and who loses the work.

The Verdict

The article is right about the mechanism and wrong about its significance. Cablese is a temporary, human-auditable interface for machine expansion—not a defense of human economic participation. It makes autonomous cognitive labor cheaper, denser, and easier to coordinate.

The real headline is not that LLMs relearned telegraphese. Capital rediscovered that communication costs are leverage. This is not a bridge out of obsolescence; it is a cheaper rail line into it.