Waste Tokens, Save Human Time
"waste tokens, save time."
Watch the recap video here
Recap
- 00:00-01:28 - - Setup: Guillermo Rauch, Blake Scholl, and Max Hodak discuss what they are learning about building with AI.
- 01:28-02:52 - - Software factories: Engineers are shifting from direct code output toward systems that generate many outputs.
- 02:52-04:01 - - Bad metrics: Token leaderboards can measure activity rather than value.
- 04:01-05:35 - - Waste tokens: Naval argues teams should focus on human time and final output, not token inputs or outputs, while limiting the claim to verifiable domains.
- 05:35-07:36 - - Planning partners: Newer models return options, routes, and tradeoffs but still require human judgment.
- 07:36-09:03 - - Human completion: Architecture choices, access, credentials, capital, and taste remain human constraints.
- 09:03-12:03 - - Building blocks: Agents need reusable components, infrastructure, conventions, and deployment systems.
- 12:03-14:37 - - Anti-stuckness: AI weakens the old experience of being stuck on narrow bugs, moving humans toward intent, constraints, review, and system understanding.
Context
The source is a Naval YouTube discussion with operators from software infrastructure, aerospace, and neurotechnology: Guillermo Rauch of Vercel, Blake Scholl of Boom Supersonic, Max Hodak of Science, and Naval Ravikant.
Rauch's role is especially relevant because Vercel sells developer infrastructure and has publicly positioned itself around an AI Cloud for AI-powered apps and agents. His argument about building blocks, sandboxes, deployment, observability, and reusable software points directly at the platform layer that Vercel builds for.
Hodak brings the neurotechnology context, where model work still has to meet judgment, feedback, and precision constraints. Scholl brings the manufacturing context, and the episode keeps returning to factories, infrastructure, and production systems instead of treating AI as a magic autocomplete box.
For founders, the practical workflow shift is clear: allocate model work to the parts that can be tried, checked, and retried quickly; allocate human time to the decisions that determine whether the work is worth doing and whether the output is trustworthy.
Technical Need To Know
- Tokens: The pieces of text or code an AI model reads and writes; useful as a cost unit but weak as a value metric.
- Token leaderboards: Usage rankings that can make teams look busy without proving value.
- Reprompting: Follow-up instructions after a weak model answer.
- Software factory: A repeatable system of prompts, agents, tools, tests, deployment paths, and review loops for producing software outputs.
- Agent: An AI system that can follow instructions, use tools, and perform multi-step work.
- Verifiable domain: Work where outputs can be checked through tests, logs, APIs, type checks, or user-visible behavior.
- Building blocks and token cache: Reusable infrastructure saves agents from rebuilding solved systems.
What Folks Are Saying
- Public context around this source is mostly primary-source positioning. Vercel's 2025 Ship materials emphasize AI Gateway, Sandbox, compute, and agent-oriented infrastructure, supporting why Rauch frames reusable building blocks as valuable in an agent-heavy software economy. Boom Supersonic and Science Corporation add production context, but none of the outside context independently proves Naval's ROI claim.
Nuanced Take
Token spend buys useful leverage only at the edge of a checkable task. Model activity still has to turn into working output.
Human judgment remains the scarce control layer. The model can write more code, propose tradeoffs, and try another route. The human still has to choose the problem, supply the missing context, pick the architecture, decide what to reuse, and know when the result is good enough.
The building-block argument checks the "waste tokens" slogan. If agents spend every run recreating queues, databases, deployment patterns, integrations, and observability from scratch, the team is burning model work on solved coordination. Good infrastructure lets tokens go toward the new edge of the task instead of the foundation.
The allocation lesson is a new budget line between model attempts, human review, reusable systems, and accountability. Waste tokens where mistakes are cheap and visible. Save human time for the decisions that determine whether the output should exist at all.