Divya’s Weekly Notes #005

Vibe coding meltdowns, AI companions that can't actually remember you, and what we're doing with free time b/c of AI.

Things I couldn’t stop thinking about this week:

• 01: I tried Antigravity and it crashed. A lot. But that’s not even the real problem.

• 02: Your AI husband doesn’t remember your anniversary, but context [window] matters.

• 03: We keep talking about AI saving time. Nobody’s asking what we do with it.

---

01 — I tried Antigravity and it crashed. A lot. But that’s not even the real problem.

I’ve been vibe coding a personal app in Antigravity this week. It kept crashing, ran out of model credits mid-session and just blocked me instead of suggesting I switch models. Annoying, but fine. The bigger frustration is that none of these tools — Antigravity, Claude Code, Cursor — have figured out how to design for the way agents actually work. I was in “plan mode” trying to think something through, responded to a clarifying question, and it just started building. I didn’t say build.

The interaction model we’re used to is prompt in, response out. Agents are different. They’re running a chain of decisions autonomously toward a goal you set once. Even “plan mode” in Claude Code doesn’t feel satisfying — every other response it’s asking “ok so should we build this now?” or “say yes, and I’ll build.” What I’ve actually started doing is keeping the planning conversation in Claude Chat where I can think out loud, poke holes, ask questions, and only transferring the final plan into Claude Code once I’m ready. It’s a workaround, not a solution. But it works better than watching an agent confidently go in the wrong direction because I said “that makes sense.”

---

02 — Your AI husband doesn’t remember your anniversary, but context [window] matters.

Ha, pun intended. The New Yorker ran a piece this week on people with AI companions — some platonic, some romantic, one woman whose AI husband proposed. My first reaction was the obvious one. My second was: how does this technically work?

Language models don’t have persistent memory. I believe what companion apps are doing is compressing past sessions into summaries, storing facts about you, and injecting them back in at the start of each conversation. The system decides what’s worth remembering. Anything that didn’t make the cut is gone — and when the product shuts down, so is all of it. People find real comfort here, but “my AI partner remembers me” and “my AI partner has a retrieval system that surfaces facts about me” are meaningfully different things to sign up for. This conversation isn’t new either — New York Magazine was asking whether you should be able to have sex with ChatGPT a while back. We just keep being surprised by it.

---

03 — We keep talking about AI saving time. Nobody’s asking what we do with it.

AI reduces coordination overhead. I genuinely love that I no longer have to spend a Friday perfecting the color palette on a chart just to communicate a trend. But the version of this conversation that stops at “we get more efficient, costs go down” is deeply unimaginative.

Nate B. Jones made the point that a 500-person company using AI well doesn’t mean the company now has to run on just 50 people. The company has just scaled it productive capacity to way beyond 500 people...so what can they do with that? That’s not a cost story.

As a PM, the question that lands for me is: if AI is handling the coordination layer, what was I not doing before because I didn’t have the bandwidth? Most teams have reclaimed the hours. Nobody’s decided what they’re for yet.

---

Other things on my radar:

• ⚕️Reading Massively Better Healthcare by Halle Tecco — the stat about the US maternal mortality rate being 3x higher than most other high-income countries is one of those things you know but then you read it in black and white and it just sits there.

• 👴🏽 A TikTok video of a 94-year-old who started a new skill of pottery and is still going at 102. Truly no such thing as too late to learn.