Shipping AI Features: The Messy Reality Behind the 'Pick a Model and Launch' Fantasy
aakashgupta · x · 2026-10-11
Aakash Gupta contrasts the fantasy of shipping an AI feature (pick a model, write a prompt, build a demo, launch) with the messy reality: CEOs force demos, sales pre-sells, real users type with typos, outputs go generic without thread/CRM context, the model invents refund policies, legal gets involved, and a 95%-accurate eval judge catches zero actual failures. After shipping to 5%, usage stays flat, one power user eats the margin, and every new model release forces a full eval re-run.
His core point: what decides success sits around the model — context, tools, permissions, and checks that catch failures before customers do. Often the model was never the problem.
Related event: Shipping AI Features Is Messier Than You Think(2 posts)→
More from coding & agent
- 20 open-source tools for AI agent security, organized into 6 categories — MaryamMiradi · 2026-10-11
- Vibe coding only works if you already know how to code, dev argues — prasenx · 2026-10-11
- AI ported g3sharp to Luau for a Roblox Studio mesh sculpting plugin — rms80 · 2026-10-11
- Parakhin: use the largest models for dev and research, fine-tune for production — MParakhin · 2026-10-11
- He runs two cleanup agents on his Mac to manage RAM and 100GB of agent temp files — kevinkern · 2026-10-11
- Claude tools won't kill developers, designers or analysts — they're just new workflows — PawarBI · 2026-10-11