I left Apple earlier this year after about a decade on Maps. A few weeks ago I took an impromptu solo trip to Bangalore — no agenda beyond meeting people who are building. It turned out to be the most useful professional week I've had in a long time, and this note is an attempt to write down what actually changed in my thinking before the feeling fades.
What people kept asking about
The conversations covered a lot of ground — AI research, GIS and satellites, enterprise AI, agent shops, small language models. But one pattern repeated. When people building for real customers talked about AI, the words that kept coming up were not capability words. They were determinism, consistency, replicability, evaluation, reliability. The gap between an impressive demo and a system someone will bet a business process on is exactly the gap those words describe.
At a small-AI gathering, the question that hung over the room was whether small, local, specialized models can be made dependable — not whether they can be made impressive. Nobody there thought this was settled. Neither do I. That unsettledness is what makes it interesting.
An echo from the agentic platform
During my last year at Apple I designed the architecture for an internal agentic platform. The biggest lesson had little to do with agents: agents expose the debt a system already carries. Knowledge living in people's heads. Fragmented data. Workflows that exist nowhere in machine-readable form. Unclear ownership. The agents were the easy part; making the organization's knowledge legible enough for them to act on was the work.
Bangalore brought that lesson back with a different sign on it. The enterprise demand for determinism and verification is the same demand I was trying to satisfy inside a large company — turn probabilistic components into systems that behave predictably. That problem doesn't get easier when the model gets smaller and moves onto a device. It gets sharper, because there's no cloud fallback hiding the failure modes.
Why small
Small and local is not a nostalgia position. Latency, privacy, cost, energy, and offline operation are real constraints for most of the world's deployments — certainly for the ones I can see from Indore. A model that runs on the phone in a student's hand, without connectivity, answering questions in the language people actually speak — that is a different research object than a frontier model behind an API. The interesting questions are about the seams: what stays local, what escalates to the cloud, how a system knows the difference, and how you verify any of it.
What this changes for me
The Acropolis Edge Lab's two projects — offline Hinglish voice Q&A and on-device computer vision — were already pointed at this territory. The trip sharpened how I want to run them: as experiments with questions and measurements, not as demos. And it gave me the first entries in a personal experiment queue: run the same prompt through a small model a hundred times and study the variance; put a small local model against a frontier model on one bounded task; make the local/cloud routing decision measurable.
These are working notes, not conclusions. The honest version of where I am: I think dependable, small, physical AI is an open research territory, I have a decade of adjacent scars, and I'd rather spend the next months testing that belief than pitching it.