Teaching - leaving your blood marks on the margin
As more and more AI applications emerge, I'm still hoping, that it will eventually take over some of the most hated work of teachers in Germany: Correcting handwritten class tests. In Germany you usually teach two subjects and, very often, students choose them following their interest, rather than their common sense. Thus, some of them (like me) end up teaching a language, which usually means that every 2-3 month on average (for every course they teach), they will get a huge heap of paper on their desk, full of what the students think to be perfect English, brilliantly enlightning and answering every detail of the assignments in a most profound way.
Well... nevertheless, there sometimes are tiny little issues in their texts that need some tweaking... like terrible violations of the English language that make you want to bump somebody's head against something really hard (As it usually occurs while you're alone at your desk there are not many options. It has also become quite unfashionable to use violence on those poor in spirit).
Enters: AI
AIs are alert readers
It might have yet escaped most people's notice, but AI is indeed capable of recognizing handwritten texts better than even most software that was specifically designed for "optical character recognition" (OCR). As AI is more or less a huge "universal learning blackbox", it was just a question of feeding it with enough handwritten texts and the according transcripts.
AI can also describe images, comment on the content, etc, but for correcting handwritten texts it is sufficient that it can correctly transcribe.
AI on my computer?
When I tell people, that language models run on my computer, they usually aren't surprised: Some, because they know me and wouldn't have expected anything else, others, because they consider it to be quite normal, that they use AI on their computers - where else?
The difference between using an AI - that runs on some online server - and running the model locally is immense, though. The commercial models usually run on hardware far out of scope for any normal PC user. When the first AI was set free in the wilderness of the internet (ie., the whole model given out to the interested public), some quite smart programmers managed to solve problems that the "inventor" of the model, Meta, had been struggling with for years, in a few months. They also used something called "quantization", which means, that the model can be "squeezed" into a pretty small file, with some - but not too much - loss of perfection. In a nutshell, that means that very soon after that, even an AI illiterate like me was able to download the complete model (!) plus the software that drives it when we chat along onto their computer and use it without an internet connection (!), with everything running locally, on my own hardware!!11elf! I admit, that I have upgraded my computer a little, since the first time I managed to do that. I invested 250 € in two new graphics cards, that in sum provide me with 20GB of graphics RAM, which is certainly not totally out of this world, but a good start, as the most important thing for AI is, that the complete (pretty big) model fits your RAM, be it the "normal" slow RAM, or - much better - the fast VRAM. In other words: A "normal" consumer grade computer can easily and affordably be switched into a computer, that can run AI chats at a decent speed (many times faster than typing, to use a more common unit than the often used tokens / second). OK, I don't run the full sized model, so there is some loss in quality. Given the speed at which the power of AI evolves though, the local model is only an estimated few months behind the current market leader on average.
Correcting AI - Benefits
For those of us, who have been playing around with generated pretrained transformers - or language models - or AI, that at least let's us assume, that it won't take too long until the machine can also correct written text. It will probably not fill the margin with a red pen using special abbreviations for different types of errors, trying to interweave some hints on how to do it right in the future and sometimes a little cursing. But it will probably be able do something that we as teachers usually don't have the time nor the brain for: Find all systematic errors and provide detailed and personalized help for every student!
Outlook
I probably won't live to see the burden of corrections taken from my shoulder, as the burocrats will probably insist on somebody filling the margin with some red marking, just for the sake of it and because it has always been done that way and never did anyone any harm (which will then be the worst of all the lies) and last but not least, because education in Germany is regulated by the federal states individually, so there will either be some states implementing new ways while mine won't, or they surprisingly agree on a common way, which they almost never do. But I'm confident, that I will see the benefits of "reading" AI, providing a personalized feedback with a personalized grammar exercise for every single student.
Where are we, in terms of free models and available hardware to run AIs locally?
Well, here is an image of what happened on my computer today:
As you can see, the complete page has been nicely recognized and the model has made some annotations (not the red margin, that was already on the image!).
Source of the image: Shame on me, I didn't take notes. Found it somewhere online. If I infringe anybodys copyright, please let me know via the contact form.
Name of the model: Qwen2-VL (created by Alibaba and available for legal download under some pretty free licence) in the 7b Version.
Software: Comfy-UI with the frontend in a firefox browser.
The whole process (image processing is terribly complicated and was probably by far the biggest share of that) took little more than 3 minutes on my 20GB VRAM ryzen computer.
If you are interested in how to get it up and running: Installing Comfy-UI is a pretty straightforward process. Somebody has created a plugin that you can install via the Comfy-UI manager.
Have a nice 2025
Josh