When Tokens Flow Like Electricity

Quinn Slack and Thorsten Ball explore what software becomes if model intelligence keeps getting cheaper and faster. Their answer is deliberately speculative: developers may spend less time building generalized software for other people and more time using agents to solve the underlying problem directly.

Key Points Covered

• Abundant tokens are a forecast, not the current baseline: Thorsten predicts that expanding infrastructure, lower prices for older model capabilities, and faster inference will make today's token budgeting feel as dated as metering small amounts of household electricity. Quinn agrees with the direction while saying he sees no evidence that current subscriptions are unprofitable or necessarily headed for a subsidy pullback 02:03–06:03(opens in a new tab).
• Developers may need to identify with the outcome, not the software layer: If customers can generate more of the software they need, the hosts argue that “building software to do X” becomes less defensible than taking responsibility for doing X. They present this as a difficult shift for increasingly generalist developers, not a measured labor-market result 06:03–06:50(opens in a new tab), 41:46–42:49(opens in a new tab).
• Agent simplicity can feel suspicious to experienced engineers: After recounting a plain-language feature request that was rolled out and tested, Thorsten labels it a “pet theory” that engineers add tools and process because direct prompting feels too easy 06:50–10:33(opens in a new tab).
• Cheap implementation does not remove product and systems difficulty: The hosts distinguish quickly assembling infrastructure from designing a coherent team-scale product. They locate the hard work in primitives, invariants, sequencing, trade-offs, integration, and fitting many strongly held perspectives into one reliable whole 10:33–16:16(opens in a new tab).
• Build-versus-buy becomes a moving boundary: The hosts argue that agents can make a deeply integrated, narrow replacement cheaper than adopting a broad SaaS package. They contrast that with team-scale, load-bearing products and foundations such as authentication, where operational support and sustained judgment still matter 17:13–30:40(opens in a new tab).
• A useful substrate is solid but malleable at the edges: The hosts prefer foundations where someone has already worked through authentication, payments, storage, deployment, and secrets, then let agents reshape interfaces and product features above them 29:37–32:38(opens in a new tab).
• Malleability also creates ownership: Quinn compares agent-modifiable software with owning a home: changing it becomes easier, but so does the obligation to maintain and repair it. The alternative is rented software that absorbs maintenance while limiting customization 32:38–36:40(opens in a new tab).
• Agent-native tooling cannot compensate for a slow organization: Orbs, portals, multiplayer, production monitoring, and fast model output matter most when a team can decide, verify, deploy, and learn quickly. The hosts argue that long planning and approval queues remain the bottleneck even when code generation is nearly immediate 36:40–40:55(opens in a new tab).
• Human attention shifts toward sustained judgment: The closing thesis is not that every product should be rebuilt. The hosts still value trusted people who continuously understand terminals, authentication, product design, or another domain and encode that judgment in an opinionated foundation others can adapt 40:55–47:39(opens in a new tab).

Full video: https://www.youtube.com/watch?v=SI3N8MCQHfo(opens in a new tab)

Episode page: https://ampcode.com/podcast/season-02/episode-03(opens in a new tab)