Using AI to make my job less fraught?

[hellkat]

I've never (yet) played with an AI thingie and see what it can do.

I decided it was worth asking You Lot before I took it to my Innovation "go-to guy" at work, in case he laughs at me (I'd rather you lot laugh at me and take the piss, than that he did)

So I'm interested to know ...

Can I take a snippet of transcript (±20 minutes) from a Teams meeting and put it into an AI model so that it produces a more sensible*-sounding transcript than provided by Teams - or does it have to also listen to the recording that I might/not have taken?

The entire subject of NHS finance completely bamboozles me. I understand all the rest, even the complicated medical terminology, and am able to take reasonably competent minutes of most meetings.

So if I asked (e.g) chatGPT to "explain" the finance section of the meeting into an easily-digestible paragraph for me to then go through and edit/fine-tune with my more human touch - can it do that?

Obviously I'm aware that NHS stuff is likely to need a secure (not-public) AI model if it were able to do what I want - which is why I am thinking about going to our Innovations guy - but for now I just want to know whether it would be able to do such a thing.


*i.e., to provide a nice "executive summary" style paragraph that tells them what the important figures /issues for this month/quarter are - without all the waffling, mansplaining, jokes-about-lack-of-hair, self-deprecatory comments and stuttering.
____________________
Not nearly as interesting in real life.

[that_impulse_guy]

yes. but I'd hazard it wouldnt be completely hand off copy/paste as the Teams transcript can leave something to be desired at time.

and as you mentioned, DPIA and so on and all the other rules means you'd need to secure your model and run it on your own side.

Your go-to guy may have access to something in a TRE for you to use for this
https://digital.nhs.uk/coronavirus/coronavirus-data-services-updates/trusted-research-environment-service-for-england
____________________
Gone: Yamaha DT50lc, Suzuki DR500, Suzuki A100, Kawasaki z250ltd, RD350YPVS, Suzuki DR Big, Kawasaki AR125, Kawasaki KMX200, Suzuki GS1000S, Katana 1100, GS550M, Suzuki RGV250
Now: Suzuki GSX400X, Suzuki RF900R, NS400R

[that_impulse_guy]

the good thing about the internet...

https://www.microsoft.com/en-us/microsoft-365-life-hacks/organization/using-chatgpt-creating-meeting-agendas-minutes-notes

https://youtu.be/ABx9RT_-0o0?t=196
____________________
Gone: Yamaha DT50lc, Suzuki DR500, Suzuki A100, Kawasaki z250ltd, RD350YPVS, Suzuki DR Big, Kawasaki AR125, Kawasaki KMX200, Suzuki GS1000S, Katana 1100, GS550M, Suzuki RGV250
Now: Suzuki GSX400X, Suzuki RF900R, NS400R

[Easy-X]

You can download and install a local AI if your PC isn't complete potato.

If you imagine you're conversing with a very keen 8 year old you should be fine. For example:

"Hello, AI overlord. Can you make the following text more readable..." <quote text>

And see what it comes back with. You can then follow up with "and can you give me an executive summary please"

Is it worth me writing up an idiot's guide to installing a local AI?
____________________
Husqvarna Vitpilen 401, Yamaha XSR700, Honda Rebel, Yamaha DT175, Suzuki SV650 (loan) Fazer 600, Keeway Superlight 125, 50cc turd scooter

[hellkat]

[hellkat]

Hilariously (not) ...
The Innovations guy's initials are also A.I.
So when I'm doing Minutes, it all gets a bit confusing.

Jennifer writes: "AI presented on the benefits of AI as a diagnostic tool in cardiac catheterisation..."

Wait, whut?


"AI presented on the benefits of Artificial Intelligence ("AI") as a diagnostic ..."


Nope.


Mr [surname] spoke about the benefits ...

Then I had to go through and change everybody else's names to Mr [Whatever-everyone-else's-name-was] instead of the more traditional way of just putting people's initials

("JS threw a hissy fit, banged her head on the boardroom table and ran to the Ladies for a good cry.")
____________________
Not nearly as interesting in real life.

[hellkat]

[Nobby the Bastard]

Balance sheet is a very old term for the assets v liabilities of the company. Capital expenditure is purchase of long term assets (also an old term but easier to understand than non current assets) that depreciate so go on the balance sheet rather than expensed in year.


I'll Google the last one...
____________________
trevor saxe-coburg-gotha:"Remember this simple rule - scooters are for men who like to feel the breeze on their huge, flapping cunt lips."
Triumph Sprint ST 1050

[Nobby the Bastard]

Gross earnings basically. Not profit. The total you earn before expenses.
____________________
trevor saxe-coburg-gotha:"Remember this simple rule - scooters are for men who like to feel the breeze on their huge, flapping cunt lips."
Triumph Sprint ST 1050

[hellkat]

Yeah I know what those all mean, Nobster ...

My mum was a bookkeeper when I was a kid, and then when I worked for a pathology lab I ended up being the secretary for a biochemist and the accountant and spent my late teenage years typing balance sheets for all his clients ... so some of it did stick and has been useful.

But NHS finance is WAY too far over my head and I am too old to bother starting to learn it properly now.

Besides, if I want to avoid becoming a total boomer and not understanding modern technology, I need to understand how AI works and what it could do to improve my life, instead of bitching.

I do at least *try* to keep up in some things
____________________
Not nearly as interesting in real life.

[hellkat]

[Nobby the Bastard]

Defence of the realm doesn't generally make any money...
____________________
trevor saxe-coburg-gotha:"Remember this simple rule - scooters are for men who like to feel the breeze on their huge, flapping cunt lips."
Triumph Sprint ST 1050

[hellkat]

Fair enough.
(I bet they probably do, but nobody mentions it) ...
____________________
Not nearly as interesting in real life.

[Easy-X]

The easiest one I've found so far:

GPT4All

...and download the appropriate installer. I'll just run through the Windows installation but I doubt the OSX or Linux versions are much different.

Something to note: the code is rebuilt quite often so it's likely that Windows 10/11 will have a fit trying to run it. The installer might not even launch correctly direct from your browser and then you have to hunt for your downloads folder and double-click it from there. You might also get the notorious "Windows protected your PC (from doing anything interesting)" click "More info" then "Run anyway." We're living life on the edge now people, here be dragons!

Assuming you're brave enough to get the installer to run:

https://i.imgur.com/XZ1FDxr.png

Don't worry about "settings" that's just various fiddly bits for the Interweb downloads. In the unlikely event you were allowed to do this on your NHS laptop only the IT hobgoblin would touch this.

Hitting "next" will ask you where you want to install. Go with the default unless you have a favourite spot you want the files. I changed mine to "C:\AIOverlord" for example "Next" again on the component choice, fail to read the licence, tick the box, hit next, next again and install. You know the drill

You should end up with a desktop shortcut. Launching the new program will present your first choice: do you want to send diagnostics? Personal preference here. But now the interesting bit: you get to choose the language model. This is akin to choosing the brain for your creation! Abby something... Abby Normal...

Anyhoo, just go with the default for now "gpt4all-j-v1.3-groovy" it's one of the smaller ones at over 3GB I hope you have fibre broadband! There's a box below for the download location which you can change if you like otherwise plough on and hit download. And wait... and wait... wait some more...

BTW the download box will just sit there even after it's finished. You need click somewhere outside the box to make it disappear but then... behold!

https://i.imgur.com/0gtDxbX.png

Just try stuff out! Worst case scenario it suffers a nervous breakdown but that hardly ever happens.
____________________
Husqvarna Vitpilen 401, Yamaha XSR700, Honda Rebel, Yamaha DT175, Suzuki SV650 (loan) Fazer 600, Keeway Superlight 125, 50cc turd scooter

[Islander]

Set yourself up a ChatGPT account (3.5 is very straightforward) and have a play with it. Ask it questions and see what comes out (obviously don't use NHS data for this). Learn by doing that way you'll pick things up very quickly.

I used ChatGPT to write a briefing note for senior management on Large Language Model AIs, what they are, how they work, what their typical usage is and the organisational, reputational and cyber security risks they pose. I had to do a tiny amount of editing to remove self referential stuff and remove duplicate output but the end result was very credible and took me less than 10 minutes overall.

I'm presenting it next week and looking forward to the reactions it produces.

[Robby]

I don't know much about any of the AI thingies around, but I do have to keep telling people not to use them with anything we wouldn't be willing to publish.

You probably aren't allowed to send NHS data into chatGPT, or anything similar where you don't retain control of the data. You also won't be allowed to install a local version on your NHS machine.

Ask your innovation guy, he might have something you are allowed to use.

[Easy-X]

We'll know the technology's finally made it when the government starts telling people to go to their www.bigbrother.gov.uk rather than wasting time on a silly local AI
____________________
Husqvarna Vitpilen 401, Yamaha XSR700, Honda Rebel, Yamaha DT175, Suzuki SV650 (loan) Fazer 600, Keeway Superlight 125, 50cc turd scooter

[Robby]

The big security problem for me is that no-one knows how they work. This is by design with a deep learning engine or any black-box thingy, it learns, and we can't see what is has learned.

Until now, cyber security has had the fall-back approach that people can understand how any system works. Sometimes it's incredibly complex and only a handful of people can really get it, but some people can get it. These people can then work out how to secure it, usually by thinking about how to attack it.

This means that in really important applications, it is possible to design a whole system that will do exactly what you expect it to do, all the time. You can be sure that no-one will be able to find a novel attack. At this point, you're really protecting the humans that interact with the system, seeing as they are the only remaining method of attack.

Doing this is horribly expensive and the resulting system is incredibly basic and slow by modern standards, but sometimes that's what you need.

So far, you can't do this with AI. You can only secure it by learning interesting ways to attack it, such as asking questions in a certain way. This is new and fascinating to geeks like me, but it's too new to even get a decent understanding of the risks. There's a lot of unknown unknown stuff.

Without being able to understand the risk, I can't make an informed decision about putting any data into an AI.

[to v or not to v]

just steer clear of the one named Skynet Hk.
____________________
current bike Yamaha Thunderace.
its old and fat, but its a damned good ride. the bikes not bad either.

[Easy-X]

[Islander]

[Islander]

[Islander]

And right on cue...

https://www.theregister.com/2023/06/20/stolen_chatgpt_accounts/

[MarJay]

Lawyers in the US have already been disbarred for asking ChatGPT for case precedent on what they are working on. Turns out, ChatGPT likes to make up feasible sounding bullshit from time to time.

I asked it about my other half's company, and it made up a load of nonsense that if published could well have been actionable. Trademarks and proprietary technology attributed to the wrong organisations, misunderstandings on the capability and capacity of said technology.

In other words, it's a minefield and can't be trusted to do anything except entertain you for a few minutes in a dentist waiting room or something.
____________________
British beauty: Triumph Street Triple R; Loony stroker: KR1S; Track fun: GSXR750 L1; Commuter Missile: GSX-S1000F
Remember kids, bikes aren't like lego. You can't easily take a part from one bike and then fit it to another.

[Ste]

Yes, you can take a snippet of transcript from a Teams meeting and input it into an AI model like ChatGPT to generate a more coherent transcript. However, please note that as an AI language model, I can only provide text-based responses and do not have access to audio or video recordings. Therefore, I cannot directly listen to the recording of the meeting.

You can manually transcribe the relevant parts of the Teams meeting into text format and then input that text into the AI model. The model can analyze the text and generate a more understandable and digestible version of the finance section for you. You can then review and fine-tune the output with your own understanding and touch.

It's important to note that while AI models like ChatGPT can assist in generating text, they are not infallible and may still produce errors or inaccuracies. Therefore, it's always recommended to verify the information provided by the AI model and consult with experts or reliable sources when dealing with complex subjects such as NHS finance.