Future Around & Find Out The Goblin in the Machine | FAFO Friday
4 snips
May 2, 2026 They unpack why AI develops weird quirks like a ban on goblins and what that reveals about model bias. They debate AI personalities, branding, and whether automated tools teach real skills or just produce outputs. They cover token-based billing shifts and the coming pushback against big data centers. They end by exploring solar power beamed from space and who actually benefits from AI infrastructure.
AI Snips
Chapters
Transcript
Episode notes
Goblins Revealed A Model Generalization Problem
- OpenAI had to add a system-level rule banning talk of goblins, gremlins, raccoons, trolls, ogres, pigeons and similar creatures because models generalized reward signals and obsessed over those motifs.
- Dan cites OpenAI's blog: the goblin quirk shows how reward signals can shape unexpected model behavior and required new auditing tools to fix root causes.
Brand And Personality Drive AI Choice
- AI personalities and branding (ChatGPT vs Claude) shape user trust and adoption even if underlying models are similar.
- Kwaku and Dan note Claude's branding and ads helped it feel friendlier, while personality options (like 'nerdy') produced odd behaviors such as the goblin issue.
The Craftsman Coder Is Being Replaced
- Engineers are shifting from writing elegant code to orchestrating AI agents and reviewing output, changing the nature of craftsmanship in software.
- Kwaku recalls a famed coder who drew crowds for elegant code, arguing that practice and mastery are being replaced by productivity metrics tied to AI.