Gunveer Sandhu

Applied AI product engineer. Building 0→1 products across mobile + desktop. SpellType, NoRing, Cosmicly.

[Gunveer Sandhu @SandhuGunveer]

Haha I thought it was unlimited! #dots #devday @thsottiaux

[Gunveer Sandhu @SandhuGunveer]

8/8

So… did I corner the AI into admitting a reasoning flaw, or did the AI just let me talk myself into one?

[Gunveer Sandhu @SandhuGunveer]

1/8

I think I found an interesting failure mode in AI reasoning today.

I was discussing survivorship bias with ChatGPT.

I took today’s 750 stocks, pulled their history back to 2021, and said:

[Gunveer Sandhu @SandhuGunveer]

7/8

It can stay attached to a technically valid caveat even after that caveat stops affecting the question being tested.

Or maybe I’ve successfully AI-hypnotized myself into believing I found a statistical loophole. 😂

[Gunveer Sandhu @SandhuGunveer]

@thsottiaux The number of people saying usage limits should tell you something. It's a hated feature, remove it.

[Gunveer Sandhu @SandhuGunveer]

4/ Useful rule beyond Google Docs:

If another system owns the state, your cache, buffer, mirror or optimistic update is a projection — not authority.

Correctness needs reconciliation.

[Gunveer Sandhu @SandhuGunveer]

3/ That changed the engineering question.

Not: “Is my buffer correct?”

But: “What evidence proves what the host actually accepted?”

The rebuild treated local state as a projection and reconciled it against host evidence.

[Gunveer Sandhu @SandhuGunveer]

2/ One commit called that buffer “ground truth in canvas mode.”

Useful — until paste, undo, clicks, navigation or async rendering changed Google Docs without the local projection knowing.

A coherent local model can still be wrong.

[Gunveer Sandhu @SandhuGunveer]

1/ The hardest part of AI autocomplete in Google Docs wasn’t the AI.

It was deciding what state was real.

Docs doesn’t expose document state like a normal macOS text field, so I kept a local keystroke projection.

BUILD IN REVERSE / SPELLTYPE 02

[Gunveer Sandhu @SandhuGunveer]

7/ The lesson wasn’t “Accessibility APIs are messy.”

A serious product-quality bar becomes architecture.

With AI, the scarce skill isn’t producing another implementation.

It’s knowing when the evidence says your model of the system is wrong.

Next: Google Docs.

[Gunveer Sandhu @SandhuGunveer]

1/ SpellType was meant to run all day on a MacBook.

That changed what “good” meant.

Autocomplete could be accurate and still be a bad product if it burned battery, hogged resources, or needed Screen Recording.

BUILD IN REVERSE / SPELLTYPE 01

[Gunveer Sandhu @SandhuGunveer]

6/ Eventually the evidence won.

I stepped back and introduced AppProfileRegistry.

Per-app compatibility, AX workarounds, cursor strategies and fallback fonts became explicit.

Not the final architecture.

A better model of reality.