The Heartless Sentinel | Part II | Corruption x AI = Output.

This is part II of our look at artificial intelligence x anti-corruption and I did something interesting. Lazy but interesting. I wanted this post to feature ideas and content from the artifice of heartless sentinels. While I was aware of its capabilities it's still pretty 'wow' when you use certain things for the first time. It's very impressive what can be done, while keeping in mind the power of these tools and the fact the bad guys are using them as well. Also, I have been monitoring my AI footprint and am offsetting by not showering. I'm dirty but no dirty tricks.

I usually use ChatGPT, but after doing an online AI course run by Google - I have become familiar with some of their tools. In particular, Notebook LM. It's a cracker. It's an "AI research tool and thinking partner that can analyse your sources and turn complexity into clarity and transform your content." You add your docs or links and it will give you either an audio output - in a podcast/conversation format, video - a slide show with images and a audio commentary, and the self-explanatory mind map chart, reports, flashcards and even a quiz. It all works well and is pretty damn cool, but it's the audio output that is striking.

The default output is called a deep dive, where is analyses your inputs and gives you a podcast discussion as output. Alternatively, you can set it to critique your inputs or debate them. I tried the critique option with a blog post, and it was detailed and fair in its breakdown and suggestions.

MIND MAP - NotebookLM

I found this to be a nice feature, but it struggles with a lot of information and isn't as accurate as it could be. For something simpler I think it would work quite well, you could then recreate in Canva or Miro to add a bit of zhuzh. I've not added the full image as it's rather large. If you click on a node it will dig in and give you breakdown of that topic. I think you have to be quite tight with your sources to make this work effectively.

AUDIO - NotebookLM

This was fun to listen to, maybe it's an ego thing - know I know what it feels like for my work to be read, discussed and appreciated. AIDA and AIDEN (I made those names up) have a good chat about four of my academic papers, and they love them. It flows well, the language and cadence are natural, you'd have to be paying attention to notice they weren't real. Saying that, it is possible that they have data centres in the Philippines and India putting on American accents, turning out podcast style conversations in triple time. Just kidding, though we wouldn't put anything past big tech.

It has some issues, mainly pronunciations, but considering I didn't add any direction or prompts it's really good. The tone isn't aligned at times, that is subject matter to speech tone. The main error is my spelling of notebook, what a joke - this 'human in the loop' is a failure.

VIDEO - NotebookLM

This one again, was done without any additional prompts. This one is impressive in what it does but it doesn't get the vibe right. Again, this was plug and play but there are a few visual options you can change, and add written guidance. I look forward to playing around a bit more. It get's a few things wrong - it says fraud is a type of corruption, which is incorrect and not something I recall saying. The artwork is a bit off too - there are roots everywhere. It takes some articles our of context, and should be searching online to check its work. Considering it put this all together in 10 minutes, it's quite impressive.

ANTI-CORRUPTION CLASSROOM - ChatGPT / Custom GPT.

Now for something a little different. This one I had to work for and it was the result of an online course as mentioned. It was like a pseudo-programming project, where you create a custom GPT to fulfil a task but this isn't just saying act like a nutrionist and give a meal plan, it goes a step further to format the inputs and outputs to create a kind of program with menus. Of course, my subject of choice was corruption. So predictable. The prompt is almost 700 words. So lame.

It's slow and crashes at times but it mostly works pretty well, and is also surprisingly accurate (though I haven't done a deep check).

I was quite pleased with this. It's not as crisp as I could make it, but the joy of it is that you can provide the subject, some behavioural rules, a few constraints, an example flow and the output format and it will fill some gaps or learn from references to make it work. It could do with some tightening up, and a bit more direction for sure. The kids stories are a bit lame, and it's a bit short on quotes but considering it's a few hours work it's damn cool. I think it is anyway.

The link is below if you want to have a go - you will need to create a ChatGPT account if you don't have one.

Click there >>> R&C Anti-Corruption Classroom - Custom GPT

/EXIT

There is a lot you can do with Custom agent / AI assistants. It's not just about asking whether you've got scurvy, or your girlfriend is cheating on you, or what to make out of your fridge dregs (half a lettuce, a teaspoon of marmalade and a dairylea triangle). Prompt engineering scratches an itch for me. The programming one, not the rash one. When I was a slender young chap, I learnt to code Javascript. I am old. This is really funny, the instructions came in a loaded ring binder and on 2.5" floppy disks (if you don't what they are = 💾 ). I just thought of this nexus now and it's blowing my mind how far the coders have brought us - now I wish I was one. I had a Commodore 64 and now I've got a CustomGPT. Wild. I have started dabbling in python and it's fun, but it's addictive and so time-consuming. Why does everything take so long? To add to that, I wanted to get my head around some basic statistics, too, before really diving into the deep end. So much to do, I need more compute. Where that Neuralink at, Elon?

Thanks for reading.

/Alvin

---