The Dispatch
What founders building with AI are hearing
A recurring read of the technical-build signal from the AI talks worth watching. A pipeline ranks the fortnight's talks by trend signal, transcribes them, pulls the takeaways worth acting on, and links each one to the exact source moment. The short version lives here; every full edition is one click away.
Agent reliability has moved from architecture debates to operational mechanics: verification loops, checkpointing, and adversarial eval maintenance are the concrete bets separating production systems from demos. Fine-tuning strategy is crystallizing around a clear division of labor, where use-case behavior lives in the model and customer-specific logic stays in the context layer. Infrastructure unit economics are emerging as a silent kill switch for otherwise working features, especially at the retrieval and embedding layer. On the go-to-market side, the perceived risk of the status quo is now the primary enterprise sales lever, and founders who can walk a buyer step-by-step from first meeting to 100% live are closing deals faster.
This edition's picks
A verification mechanism is the single highest-leverage addition to any agentic loop. Without a way for the model to check its own output (tests, screenshots, diffs), long-running tasks stall; with one, they can run unsupervised for days or weeks.
Boris Cherny backs this with a concrete production example: an 11-day, full-codebase language rewrite that ran largely unattended.
Boris Cherny · Boris Cherny: We Cut 80% of Claude Code’s Prompt
Fine-tune for use-case-level behavior only. Customer-specific business logic belongs in the context layer and should be changeable without retraining. Conflating the two creates a model that is fragile to update and impossible to resell.
The Decagon co-founder gives the clearest architectural division of labor on fine-tuning seen in this slice, directly applicable to any vertical AI product.
Decagon founder · How Decagon Runs 90% of Its Agents on Open-Source Models
Data curation alone, with no architecture changes or post-training modifications, produced a 14 absolute percentage point benchmark gain on VLMs and allowed a smaller model to match a much larger one using 145x less training compute.
Ari Morcos delivers the most concrete and surprising data point on model training leverage in this batch, directly actionable for teams about to start a training run.
Ari Morcos · Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAI
Run the unit economics on retrieval and embedding infrastructure before building. One bootstrapped company discovered that embedding plus search at scale would cost $30,000 per month against a $5,000 per month total infrastructure bill, a 6x multiplier that would have killed the business.
Simon Eskildsen turns an abstract warning into a specific, reproducible calculation that any founder can run before committing to a vector stack.
Simon Eskildsen · Building Turbopuffer: Gergely Orosz (@pragmaticengineer ) × Simon Eskildsen (CEO)
All editions
Editions are drafted by the pipeline and lightly edited. The archive is append-only; each edition covers only what is new since the last.