Key themes that dominate the discussion
| # | Theme | Representative quotes |
|---|---|---|
| 1 | LLMs often ignore or mis‑interpret “no” / plan‑vs‑build confusion | “The user said no … but the model still implements it.” – verdverm “If you just say no the model thinks you’re asking for a plan.” – brcmthrowaway |
| 2 | Guardrails & human‑in‑the‑loop are essential | “That’s why we keep humans in the loop.” – verdverm “It’s completely irresponsible to give an LLM direct access to a system.” – bigstrat2003 |
| 3 | Productivity claims are unsubstantiated | “I’ve yet to see evidence of increased productivity.” – komali2 “Lines of code is a joke; the real bottleneck is producing the right code.” – aisengard |
| 4 | Harness design (plan mode, hooks, system prompts) drives behavior | “The system prompt says plan → build; the user said no.” – verdverm “You can add a hook that rejects ExitPlanMode.” – ramoz |
| 5 | Consent & privacy are poorly handled in industry | “The tech industry abuses consent into an absolute mess.” – bluefirebrand “Notification banners for newsletters are a form of consent abuse.” – hedora |
| 6 | Trust must be earned through verification | “I don’t trust it completely but I still use it. Verify!” – wvenable “Never trust autocomplete if it can get it wrong.” – behehebd |
These six themes capture the bulk of the conversation: how LLMs fail to follow simple instructions, why safety layers are needed, the gap between hype and real productivity, the importance of harness design, the industry’s mishandling of consent, and the necessity of verifying outputs before trusting them.