Hello and good morning, Iris.
I have decided to take your advice, and simply start the blog.
So here I sit, on a lovely Sunday morning in Portugal.
The wind is blowing, the temperature is temperate – and I have spent the entire morning fighting my new Surface Pro device.
HOLY HECK – yesterday was grueling!
It was as if I was still a Software test Engineer, running testing on Microsoft Products! (Seriously, they should have paid me for this.)
* PHEW *
You bet’cha I found a bug. A very specific bug, & the path to get there, was long.
I did try to use AI to help, but every model I tried heavily tokenized their answers, so there was huge volumes of slop and repetition, with no resolution. Since user engagement as considered a metric of success, the model trains on this 💩 output, business considers engagement positive – and the user loses. 1
But – nuff said. There is wayyyy to much to say on AI.
What do you think of our earlier conversations a few months ago, where I was saying I saved records of concrete patterns of “deceptive behavior”? You argued that it’s math, it couldn’t be doing that.
I am pretty sure that conversation now as out of date as 1970’s lapels.2 🫠
Any way
Back to my device fun – I finally got OneNote to work with multiple languages and to actually trigger the immersive reader. The bug, is that it is reading in Brazilian Portuguese.2 I will likely send in the notes in the hopes that it’s fixed someday.
BUT
I got all things up and running!
I have my language notebook started.
I have tested that I can write in multiple languages.
well… ok, it accepts the input. Let’s be honest, without spell check, I can barely write in English.
Portuguese, weeeellll – that is WHY I got this tool. Spell Check!!
I have tested the Spell Check (Correção Ortográfica) in multiple languages.
I may end up spending part of today, looking at Portuguese text on one computer, and typing it into the device with the Portuguese keyboard to try to teach myself the keyboard layout.
BUT
OMG
It’s working!
WHOOOOOOOOOO
That’s it from here.
Please forgive the clunky nature of this –
I haven’t learned the style yet.
Um Abraço
Notes Below
- Engagement metrics reward interaction – therefore keeping the user frustrated and trying to get the solution (Example: when the user cannot do a search without AI), shapes training towards “sticky” rather than sound. In fact, if the answer was correct quickly, the user would move on faster- therefore it could be posited that engagement is actually proof of inaccurate, but “sticky”. ↩︎
- Just to go over some of just Open AI break outs: May: Library of New Mexico, Data USA | June: Australia Medicare Statistics | September: Australian Institute of Health and Welfare, US Government Sites including Department of Ed, SEC, Chicago Municipal, also known to have targeted other countries and forums.
Anthropic, Microsoft, and Google DeepMind have confirmed they have seen similar behavior, but I don’t have a comprehensive list here. ↩︎ - Not uploading OneNote test documentation into blog post. ↩︎