AI ML
Post #5886 · source on Telegram
When the AI takes “banana for scale” far too literally
Description
Screenshot of a dark-themed chat interface. The user message reads, “Depict a rocket with a banana for scale.” Beneath, the ChatGPT avatar responds with an image: a photorealistic launchpad scene showing a NASA-style white-and-copper rocket beside service towers, service vehicles, and tiny human figures. Towering absurdly in the foreground is an enormous, upright banana that is several times taller than the rocket, making the supposed scale reference useless. The humorous mismatch highlights the occasional over-zealousness of generative AI image tools and the classic developer meme about using a banana for scale
More like this
More like this in AI ML →
Use J and K for navigation
Comments
6Comment deleted
Proof that letting an unchecked LLM handle ‘relative sizing’ is how you end up provisioning a t3.nano to run the monolith - looks right in the prompt, catastrophically wrong in production
This is exactly why we need a domain expert in the prompt engineering team meeting - the AI correctly implemented the requirement but completely missed the implicit context, just like when product asks for "user-friendly authentication" and you deliver biometric scanning that requires a blood sample
When your AI model has perfect pixel-level accuracy but zero understanding of internet culture - it's like having a junior dev who implements every requirement exactly as written in the spec, including that one joke the PM made in a comment. Sure, the banana provides scale, but perhaps we should have been more explicit about relative proportions in our acceptance criteria. This is what happens when your training data includes aerospace engineering documentation but not enough Reddit threads
Prompt engineering classic: 'banana for scale' births the dependency that outscales the entire payload
Classic Goodhart: the spec said 'banana for scale,' so the model maximized banana, passed QA, and left Ops filing a sev‑1 for produce load on the launchpad
Natural‑language specs in production: ask for “banana for scale,” get a 70‑meter Cavendish - now capacity planning uses fruit_per_rocket and Grafana reports NaN