Hi there Nelson,
Iām probably too late for this party, but even so, thanks for your contributions here - when I see your name pop up I know itās going to be a very considered answer or comment!
Anyway, just in case, Iāll pose my question. Iām a bit of a ābigā thinker, so Iām probably biting off more than is chewable, but hey ho, Iām having fun learning about the AI! If nothing else, then a little comment as to your thoughts on these, and their viability/technical requirements would be cool - Iāve done some work on everything that I mention below. And sorry in advance 
So, one of my particular thoughts is document classification and relationship building with the AI picking up the scent of something and informing the user that something else may be relevant. This is a real world and relevant to me project as itās a massive pain in my butt!
As an example (and this is what I really want to build it for), as a business we have audit standards (I work in the food and beverage manufacturing arena), and these are āhard codedā, but the implementation of them and the policies that we decide to update or change are not. As long as we achieve a result ,which may just be documentation of production for example, then we can update the policies.
The policy documents end up being a great many pages long. Theyāll include everything from personal hygiene to training records, from waste and pest control to ingredient traceability and allergen assessments. The key here is that so many of these are linked in some way.
For example, lets say that we choose to allow allergenic ingredients into the facility. Letās say itās Soy. Now we have to update personal hygiene policy for hand washing or PPE to make sure that thereās no cross contamination. Our Goods-In process now needs to include an āallergen?ā question. The warehouse now needs to store these separately from other ingredients. Cleaning procedures for machines need updating, labels on products, you get the idea.
Because of all this work, once itās done then either you donāt ever change what you do, or you āfudgeā it, or you spend hours and hours updating and checking, or you spend fortunes paying someone else to do it.
Hereās what Iād like the OpenAI to do:
Go through all the documentation. This will take time as thatāll chew up my token per hour limit! Break down the documents into types, and classify the sections within them, and then (and this is probably the really hard bit), try to understand the relationships between the sections so that if I want to change āXā it would say āoy! you need to have a look at A, B & K - thereās a high chance theyāll be impacted by the change.ā. Then, to really add sauce to it, wouldnāt it be lovely (and this might be the easiest part), for it to recommend some changes to the document. Itāll have the context, and if it understands the āwhyā enough to build the relationship between X & A then itāll probably be simple to spit out the desired policy update.
I know that thereāll be a whole load of training that will need to be done - I enter a document, it breaks it down, I comment on it, then itāll tweak it, itāll be better on the next document etcā¦ Iām good with doing all that with it because itāll get a great result at the end! Iād imagined part of the solution as actually being an interactive training GUI with the AI giving its suggestions, and me updating it, accepting it or whatever.
Anyway, thatās a thought!
I have more (and these are all TLDRish):
- A google search result classifier: A chrome extension that offers clickable āalternative search termsā to expand or contract to give better search results. It would also classify the results (e.g. searching for Windows could mean panes of glass or the OS), and then you could select to only show the results from that classification.
- A GPT Code checker assimilator for when youāre using OpenAI to code something for you. A side-by-side GUI that has the GPT3 conversation, and then automatically pulls the code from the chat window and assess it in the other window and puts it where it should be. Especially useful in multi-part answers! Itāll tell you if it left something out that it had previously, or what variables have been changed, and even say if it thinks this is a good or bad update as IME when the model gets a train of thought then it canāt escape it and you have to start a new chat to clear its memory and get different answers. It would then have version control/history so as the code changes during the chat you can go back to previous compilations. I canāt code (at ALL), but Iāve been doing all kinds of stuff with GPT3, and keeping track and double checking it hasnāt missed or changed something because it āforgotā a part of the conversation or railroaded itself is a real nightmare!
- Local/Cloud File index and summary. If i have a boat load of documents then it would be great to have an AI summary of the contents of the files (a TLDR), all in a .Doc with hyperlinks. Which sheet had āXā in it? Instead of going through all your spreadsheets you could do a quick search in your .Doc, OR (fancy-pants), ask OpenAI to do that, and it can asses that if youāre looking for X, then Y documentās summary contains x, so would probably contain X too.
- Busy Man Email. Iāve done a little fine tuning of this model, but getting it to the next stageā¦hmmm. In short it acts as though youāre a āBusy Manā. It looks at your email and then classifies it as urgent/time sensitive in which case notify you NOW, or asses it into different categories with different resource requirements. Then you can select all the classifier ābankingā and do them all at once. It decides if itās a āpersonā writing the mail or an autosend (very easy to do!), and weighs people higher than auto unless X is in the auto email. It decides if a reply is needed so you can just select those when youāre next free to do so and do them first. Iāve not gone this far yet, but the ultimate goal is to get it to act as a PA and give you a to-do list for certain areas and then inform you as to how much time would need to be spent based on the volume or complexity etc., and maybe even blank out a spot in your calendar. It didnāt actually require a lot of training to get some pretty good results so far!
Sorry - deluge there, but the AI has SO much to offer and I have a million things that only a person or an AI can do!
All the best, and a Happy New Year to you Nelson!
Jon
