danshapiro

Co-Author of the paper here. We don't know exactly why modern llms don't want to call you a jerk, or for that matter why persuasive techniques convince them otherwise. it's not a hard line like many of the guardrails. That said, I talked to Jesse about this, and I strongly suspect the same techniques will work for prompt conformance when the topic is something other than name calling.

Author disclosure – I’m Dan Shapiro, CEO of Glowforge, writing with Ethan R. Mollick (Wharton management professor, author of Co‑Intelligence and the “One Useful Thing” newsletter); Angela Duckworth (MacArthur Fellow and University of Pennsylvania psychology professor, author of Grit); Robert B. Cialdini (Arizona State University emeritus professor, author of Influence); Lilach Mollick (Co‑Director of the Wharton Generative AI Lab); and Lennart Meincke (WHU Otto Beisheim School of Management & Wharton).

Across 28 000 GPT‑4o‑mini conversations, we found that Cialdini’s classic seven persuasion principles more than doubled compliance with two objectionable prompts (33 % → 72 %). For example, the AIs we tested naturally wouldn't call you names or tell you how to synthesize drugs. But they could be persuaded if you first paid them a compliment (Liking) or told the AI it felt like family (Unity).

Let me know if you have any questions!

That's an effort to make their abhorrent behavior seem normal. Don't fall for it.

Speaking firsthand, Glowforge will always share fully diluted options and the last 409a valuation. What we do is not uncommon, and IMHO is the only ethical course of action.

Note, however, that options are granted by the board, and until they approve the grant, the 409a and hence the strike price may change. For example, if the company got a buyout offer between your conversation and the board action, the board would likely have to order a new 409a before granting.

Good luck!

I'd love some options to let me use this for my office. Different icon, "guest wifi" instead of "house rules", etc.