Can Human-LLM Coordination Cut Token Costs Without Changing Weights?

mb3rtheflame · reddit · 2026-08-29

The poster proposes a hypothesis: with model weights completely frozen, if human-LLM interaction progressively carries forward what's already resolved, later generations can skip reconstructing context, restating assumptions, and repairing missed intent—reducing total token cost per task.

Key points:

The author tested this publicly with Grok in live Reddit threads, with an inspectable trajectory; perturbations from other commenters actually reshaped the proposed measurement
A crucial correction: conversation termination cannot count as resolution, or a system that frustrates users into abandoning tasks would look artificially efficient; the right metric is total token cost conditional on independently verified resolution, plus abandonment/failure rates
The candidate mechanism, dubbed "uptake without reconstruction": accumulated context shifts the distribution over subsequent tokens, letting later generations reuse established distinctions instead of re-deriving them
The author explicitly states these threads don't prove a general effect, lays out a controlled experiment, and invites inference, agent, API-economics, and evaluation practitioners to attack the design

If the effect survives controlled testing, it shows up directly in the API bill.

Original post →

Can Human-LLM Coordination Cut Token Costs Without Changing Weights?

More from coding & agent