When you give a bored pupil in math class a graphing calculator, they’ll discover a strategy to entertain themselves. Most frequently, they discover a strategy to hack a recreation onto their calculator, like this port of Sonic the Hedgehog 2. However not the entire hacks are purely for leisure. Some focus extra on productiveness, just like the spell checker just lately developed by YouTuber ExploratoryStudios for the TI-84 Plus Silver Version calculator.
To make this spell checker work, ExploratoryStudios turned to an unlikely know-how — synthetic intelligence (AI). Not that AI is an uncommon strategy to make a spell checker, in fact. Most fashionable spell checkers now incorporate AI in a technique or one other. However utilizing AI won’t be anticipated on this case as a result of this specific calculator is powered by a Z80 processor working at 15 MHz, and it has solely 24 KB of RAM. Many private computer systems from the Nineteen Eighties had been extra highly effective than that.
Neuron activation patterns for quite a lot of inputs (📷: ExploratoryStudios)
ExploratoryStudios didn’t use any methods to wirelessly hook up with a distant service to make this potential (as with this TI-84 that connects to ChatGPT). Slightly, the answer, named HERMES OPTIMUS, is a customized neural community that was coded in TI-BASIC. That code runs completely on the calculator’s personal {hardware}.
That’s fairly spectacular, however when you get into the small print, you’ll naturally discover that lots of corners needed to be reduce to make that potential. Firstly, the spell checker can solely function on phrases which are precisely 4 characters in size. Moreover, it will probably solely discover the most certainly candidate for a misspelled phrase from a set of twelve choices. As such, that is actually extra of a toy than a helpful utility, however it would assist to go the hours in math class all the identical.
By constraining the issue on this method, a small however efficient neural community could possibly be designed. It comprises simply 4 enter neurons, 60 hidden neurons, and 12 output neurons (one for every potential goal phrase). However even this took some work to encode every part sufficiently to make it match throughout the TI-84’s meager reminiscence capability. These optimizations had been enough for inferences, however coaching nonetheless needed to be performed on a separate laptop — ExploratoryStudios calculated that it will have taken years to coach the mannequin on the calculator (if reminiscence overflow points could possibly be solved, that’s).
HERMES OPTIMUS could also be little greater than a toy, however by coding the neural community manually in BASIC, it’s a good strategy to be taught extra about machine studying, so it’s well-worth the train if in case you have an outdated calculator sitting round. If you’re involved in studying extra about how AI can run on some severely outdated {hardware}, it’s best to take a look at my TensorFlow Lite for Commodore 64 and Commodore 64 AI Picture Generator initiatives.
