Fable 5 is taking over, AI isn't getting cheaper, AI backup plan and WTF Loops? | Ep 17

TL;DR

• 1

  AI is not getting cheaper in the way people hoped: Adam argues frontier model costs are rising, not falling, while Nathan counters that access to smaller models is still dropping, so the real picture is cheaper tokens in some places but more total spend as usage explodes.

• 2

  The real cost problem is looped usage, not just price per token: The hosts talk about people running agents for 35 hours, Korean teams spending more than 2 billion tokens a day, and companies like Uber reportedly blowing through token budgets without seeing proportional output.

• 3

  ROI still comes down to revenue and shipped outcomes, not personal productivity: Adam's core point is that saving engineers time does not count unless it turns into customer value and revenue, and he says many teams are still blocked by product and operational decisions rather than code generation.

• 4

  "I write loops" is less a breakthrough than a rebrand: Ray says people have been doing this for over a year, and Adam calls the hype over loops overblown because telling an agent to keep going until tests pass is already a loop.

• 5

  Fable 5 impressed Ray with autonomous research, orchestration, and data modeling: He describes giving it a messy 300 MB personal knowledge base and watching it spin up sub-agents, analyze health data, and propose architecture for apps, widgets, voice, and storage over the course of roughly an hour.

• 6

  Even with stronger models, software work is still not 'solved': Adam notes Claude Code reportedly had 8,545 open issues during the discussion, using that as proof that better models and longer loops do not erase backlogs, bugs, or the need for human judgment.