I want an AI toaster

Toasters, right? Pain in the arse. They NEVER toast the bread properly. Usually underdone so you dunk it a second time... and then it's burnt. If you risk going for the full-toast single dunk you've really gotta keep an eye on it, and if you don't; it's burnt. And yet I reckon our toaster is one of the top three appliances in our kitchen and probably the one we'd miss the most if it stopped working.

So WHY can it not just toast bread to the perfect toastiness? We're all familiar with the 1-6 dial on the front where allegedly you can pick precisely how toasty you want your toast, and I'm guessing this links into some kind of temperature and time based basic circuitry where the toaster decides, "Yep, you're done!". Except it's never done! I don't know why I expect this kind of automated service from the toaster - after all the hob and oven need human intervention, but it seems so simple... it has ONE JOB!

I would argue the most important measure of toastiness is colour. But what colour? What's the difference between the bread starting colour and its optimum toastiness? Surely that will differ depending on the bread type.

Enter an actually useful use for AI. An AI toaster.

When you put the bread in the toaster it uses a set of cameras, lights and sensors to determine what kind and colour of bread (or other pastry item) you've inserted in the slot. Then it checks the toastiness dial and sees that you've set it to 4 - AKA the perfect toastiness.

So then the model keeps tabs on the toasting process, checking the colour of the bread as it toasts. At the perfect toastiness, pop! And out it comes.

Perfect toast every time.

How I'd actually build an AI toaster

Oh you want more? Ok so theoretically how would I build an AI toaster... I don't really have the domain knowledge to do the actual toasting mechanism, the sensors etc. I mean yeah I could dig out my soldering iron and figure that out but let's assume some other party is doing that bit and I'm in charge of the AI, and I can use AWS and Bedrock. Let's also assume the hardware is wifi enabled and has all the HTTP endpoints baked in to reach the cloud infra. That puts me in my comfort zone, building web-based infra for AI and apps.

So...

• HTTP API, via API Gateway
• Authentication - user would need a "toaster.ai" account I guess, somehow the toaster needs an auth token or something
• Lambda
• Bedrock, probably no need for agent, just direct calls
• Model with Image modality - inputs of images, outputs of scores
• Model prompt is asked to provide a structured data output of the toastiness
• Needs a 'before' endpoint to establish the type of bread
• A polling endpoint to check status, polling maybe every 5 seconds - when the toaster receives a {"status":"DONE"} response it'll pop
• Maybe a hello/healthcheck endpoint for thoroughness, so that when you boot up your toaster it can check connectivity.

Pretty simple really!

Limitations, concerns and risks

Well as with anything, no it's not that simple. If you ignore the completely ridiculous idea for a start, then there are some genuine issues:

• Expensive image models - image analysis every 5 seconds for two slices of toast (or even 4 for a big four-slicer!) would get expensive quickly. Back of an envelope maths... 2 slices of toast, twice a day, 3 minute toast duration, 5 second polling interval, plus the initial "what kind of bread is this" call... (2*2*((3*60)/5)*365)+(2*2*365) = 54 thousand calls to the model. Using Haiku maybe about a hundred quid a year (lol!)
• The stakes are real - toast goes from golden to burnt in seconds, and each round trip to Bedrock takes a second... or maybe more. The AI analysis is racing the grill. I guess some safety fallback would be necessary... "has this toast been toasting for >X minutes? then pop."
• Calibration - one person's perfect toastiness is another person's peely-wally. The 1-6 dial takes some of that out but you'd still need to get used to the AI model's own calibration, or tweak it over time.
• You need yet another cloud account login to make toast - and session timeout, token refresh or firmware update at 8am would mean angry kids.

The reality of an AI toaster

I guess the simpler solution of some colour sensing circuitry local to the toaster is probably more feasible, in fact it's already a thing. But I'm sceptical that it'll work as well as a full-AI powered toaster!

But inevitably someone is going to make one, aren't they. Every other appliance has tech built in now, I'm pretty sure my dishwasher has been talking to my hob extractor via their own private WiFi subnet and making plans for the weekend. So some ultra smart AI enabled WiFi home automation megatoaster with configurable profiles and voice control is only a matter of time. I mean Red Dwarf made one 35 years ago!