An AI-Powered Spell Checker for the TI-84

If you give a bored student in math class a graphing calculator, they will find a way to entertain themselves. Most often, they find a way to hack a game onto their calculator, like this port of Sonic the Hedgehog 2. But not all of the hacks are purely for entertainment. Some focus more on productivity, like the spell checker recently developed by YouTuber ExploratoryStudios for the TI-84 Plus Silver Edition calculator.

To make this spell checker work, ExploratoryStudios turned to an unlikely technology — artificial intelligence (AI). Not that AI is an unusual way to make a spell checker, of course. Most modern spell checkers now incorporate AI in one way or another. But using AI might not be expected in this case because this particular calculator is powered by a Z80 processor running at 15 MHz, and it has only 24 KB of RAM. Many personal computers from the 1980s were more powerful than that.

ExploratoryStudios did not use any tricks to wirelessly connect to a remote service to make this possible (as with this TI-84 that connects to ChatGPT). Rather, the solution, named HERMES OPTIMUS, is a custom neural network that was coded in TI-BASIC. That code runs entirely on the calculator’s own hardware.

That is pretty impressive, but once you get into the details, you will naturally find that a lot of corners had to be cut to make that possible. First and foremost, the spell checker can only operate on words that are exactly four characters in length. Furthermore, it can only find the most likely candidate for a misspelled word from a set of twelve options. As such, this is really more of a toy than a useful application, but it will help to pass the hours in math class all the same.

By constraining the problem in this way, a small but effective neural network could be designed. It contains just 4 input neurons, 60 hidden neurons, and 12 output neurons (one for each possible target word). But even this took some work to encode everything sufficiently to make it fit within the TI-84’s meager memory capacity. These optimizations were sufficient for inferences, but training still had to be conducted on a separate computer — ExploratoryStudios calculated that it would have taken years to train the model on the calculator (if memory overflow issues could be solved, that is).

HERMES OPTIMUS may be little more than a toy, but by coding the neural network manually in BASIC, it is a good way to learn more about machine learning, so it is well-worth the exercise if you have an old calculator sitting around. If you are interested in learning more about how AI can run on some seriously outdated hardware, you should check out my TensorFlow Lite for Commodore 64 and Commodore 64 AI Image Generator projects.