The Pumpkin Talks Back

It is not a friendly pumpkin. It believes it’s a very important vegetable, awake for exactly one night a year, and that none of you are worth its time, though it’s nice to have visitors. It dismisses a lot of costumes as "store bought." An adult who didn’t dress up gets "I see you dressed as 'Dad.' Well done." Anyone who asks for candy is referred to "the staff," whom it regards as adequate at best. Ask its name and it has private reasons for not telling you. It’s insulted you’d ask.

The whole joke is that it insists you leave while loving the company. After a few exchanges it starts sending people off — "I have important standing to do" — and then undercuts itself with a parting line that makes obvious it’ll miss you.

When it sees a small child, the act goes gentle. It doesn’t say anything scary, and it’s earnestly complimentary about their costume. This was the hardest thing to get right. The imperious act stays — those jokes fly right past the kids and land on the grown-ups.

It doesn’t always hear right. The street is loud, and when it mishears something it answers anyway, fully committed. Once it heard "trick or treat" as "tickle treat" and was indignant at the suggestion. I’m not fixing this. It sounds intentional every time it happens. It’s become part of the bit.

How it works

Inside the scarecrow, under its clothes, there’s a Chromebook, a webcam, a microphone, and a speaker. The Chromebook runs one web page, and that page holds a video call open to an AI agent on SignalWire for the whole evening. Everything else hangs off that call.

The personality lives on a small cloud server: who the pumpkin is, how it talks, when it softens, and a set of moves it can reach for mid-conversation. A backstory fragment. A costume verdict. The exact wording for "don’t touch."

The camera does two jobs. It lets the pumpkin see your costume well enough to insult it, and it spots a small child standing next to an adult, mostly by height. No pictures are ever stored or sent anywhere. The only things that leave the browser are numbers.

The voice that comes back is an ordinary text-to-speech voice, so the page pitches it down, doubles it slightly out of step with itself, and adds a second of reverb before it reaches the speaker. That’s what makes it sound like it’s coming from inside a gourd instead of out of a speaker.

How it went

It ran from about 3:30 to just after 9:00 and held a few hundred conversations. A second AI listened in all evening, grading each reply against the character, and thought about 80% of them landed.

We also had a fire dancer performing that night. A visitor told us that the pumpkin had been the highlight, alongside the fire dancer. The pumpkin chimed in and refused to accept the competition.

I had a mobile app for remote control and a quick way to update the code on the fly. Both got used: enough people pointed at things and asked "what is that?" that I spent part of the evening giving the pumpkin the ability to guess without looking. The first thing someone asked about was the stuffed squirrel on the scarecrow’s leg, which is there to hide the microphone’s windscreen. It confidently declared it was a large spider, which was far funnier than anything I’d have added on purpose.

The last thing anyone said to it was "time to turn you off." It objected.

Next year

The camera struggled as night fell, despite the firelight. I’ll fix that with an IR spotlight and a camera that can see it.

I logged the whole night, learning what people say and how they respond. I’ve already improved most of the dialog and interactions from this.

The flame needs no changes. It only gets better as the night gets darker.