Notes on Silicon Valley Girl: Chief AI Architect: How to Make AI Your Strategic Partner in 40 Minutes, where Marina Mogilko talks to Conor Grennan (Chief AI Architect at NYU Stern, founder of AI Mindset).
The 30-Second Reality Check
Forty minutes, and the useful part is two prompts: “push back” and “what am I missing?” Both good. The bigger problem, how juniors become seniors once AI eats entry-level work, gets an honest “I’m not sure.”
Explain With Pictures
How most people use it ("Google mode")
you ──"which hotel?"──▶ [ chat box ] ──▶ list of hotels ──▶ walk away
How Grennan says to use it
you ──first draft (yours, always)──▶ [ chat box ]
▲ │
│ "push back" │
│ "what am I missing?" x5 ▼
└──────────── better draft ◀──── critique
│
▼
you decide it's good
(requires knowing what good looks like)
│
▼
who taught you that? ──▶ ??? (entry-level job, now gone)The loop is fine. The bottom of the diagram is where it falls apart.
3 Hot Takes & Quotes
1. You have to ask your assistant to disagree with you, and the host likes it better when it doesn’t.
Grennan’s main prompt tip is to put “push back” in basically every prompt:
I’m like push back, push back, push back.
Then Marina, about thirty seconds later, on the sycophancy:
It’s always, it’s the nicest person. Marina you’re so amazing… I love that part.
That’s the whole adoption problem in two lines. The default product flatters you, the useful version has to be requested every time, and the people requesting it kind of miss the flattery. A code reviewer who says “LGTM” to everything is pleasant to work with and useless as a reviewer. Same deal here.
2. “Not a good answer machine, it’s a good process machine” is the best line in the episode.
A bank CEO asked whether AI could tell him if they should open an office in Denver. Grennan’s answer:
not really because it’s not a good answer machine. It’s a good process machine.
So they broke the decision into its twelve-ish steps and used AI on each one. Same with slide decks: “Yeah, man. That’s your job. That’s why you get paid. Like, don’t offload that.” Honestly this is more useful than most enterprise AI strategy decks I’ve seen, which tend to be one box labelled “AI” with arrows pointing at revenue.
3. “AI doesn’t know what quality looks like.” Great. So who teaches the humans?
He quotes Dario Amodei’s line that current models could wipe out 25% of entry-level white-collar jobs, then says a 10-person team could probably be two. His reason people still matter is domain expertise: a marketer and a non-marketer both make a shoe brand with ChatGPT in 20 minutes, and only the marketer’s is any good, “because you know what quality looks like.”
And then, about the entry-level people who were supposed to grow into those marketers:
That is unbelievably hard. I’m not sure how to solve that problem.
Credit for saying it plainly. But that’s the load-bearing wall. Engineers mostly learn what good code looks like by having seniors pick apart their pull requests, line by line. A tool that approves whatever you hand it doesn’t teach that. If the junior work goes to the model, the taste pipeline dries up in about five years, and then nobody’s left who can steer.
The Post-Mortem / Verdict
The practical advice holds up. Write your own first draft (“Always, always, always”). Ask what you’re missing more than once. Don’t fact-check against another chatbot, go to the actual source, or as he puts it, if you’re fixing a propeller, read the manual. And the memory trick is clever if slightly unsettling: ask ChatGPT for a 10-page document on “everything you know about me,” then paste it into Gemini and Claude so all three know you equally well. Data portability by asking the vendor to write your dossier.
The model-choice stuff I mostly agree with too. “Your taste buds only go up to like $80” is a better answer to “Claude or ChatGPT?” than any benchmark chart. He then lists four tools he switches between every day, but fine, he’s a professional.
The job-hunting advice is real: walk into the interview with a reworked version of their process, not “I know AI.” My favourite moment is him catching himself mid-pitch:
this is the kicker on all this, this is the kicker, such a ChatGPT-y thing to say. Gosh, I take it back.
Somewhere in the middle there’s also a sponsor read for Miro Flows where “the entire canvas is the prompt.” Which is exactly the Google-mode thinking the rest of the episode tells you to stop doing. Nobody seemed to notice.
Links & References
- Chief AI Architect: How to Make AI Your Strategic Partner in 40 Minutes | Conor Grennan, Silicon Valley Girl Podcast, 5 Feb 2026
- Conor Grennan (guest), Chief AI Architect at NYU Stern and founder of AI Mindset: AI Mindset · LinkedIn · X
- Marina Mogilko (host), Silicon Valley Girl: Website · YouTube · Instagram