Dec 12

 Early AM. I have completed some work. Not the work I wanted to do, but related.

I have been wrestling with more Azure portal nonsense. Setting up AI models. Budgets. Killswitches. It's the worst mess of UI and bullshit I've ever seen. But under that layer of the very worst UI management hums an extremely powerful ecosystem in which you can spin up and destroy servers on a whim, AI instances, training and all sorts. This is seperate - but very closely related, and increasingly blurred - from Microsofts AI foundry, the all seeing all dancing model host.

All of this is part of the actual AI work I am doing at the moment. Mostly to analyse documents and pull information out of them. But in a very future world scenario. Standard templated forms you fill in with a black biro this is not. This is random ad hoc garbage filtering, expert systems integration, and a capability to dynamically query incoming documents. Did Bob mention he had any allergies ? No.

Azure Foundry. More AI models than you can
shake a stick at. Not quite as confusing a UI
as Azure portal. 

 

Some of my early professional work I did in the early 90s involved at that point cutting edge stuff in the move from paper to digital. Familiar ground. I ended up revisiting it a decade later, different systems, different company, improved tech. Before another decade later writing my own entire document processing and storage system. And making it into a saleable product.

But this document processing. With AI. Is a different kettle of fish. Very. Different. Weird. But also cool. The magic genie of an AI has a ponder about what the documents you sent it mean. Which is. What the AIs do. They discern context. That's their unique selling point. They discern context out of the fucking ether. Like magic. But. It's not entirely reliable. Because. Magic. 

How you go about validating the AI hasn't given you bullshit data is the question. And one of those new world problems we are rapidly getting stuck in, which is, when you move from a deterministic computer system to a predictive probabilistic one. Who the fuck validates it, and how on earth do you do that at scale.

The current shit wisdom. "Wisdom". aka. We have no clue what we're doing with this young technology. Is to get another AI to vet the AI. At which point. It becomes AIs all the way down. It doesn't take a rocket scientist to see that there might be a teensy weensy flaw to this methodology of measuring a flawed measurestick with another flawed measurestick. Quick. Get another flawed measurestick and check that one. Surely. At some point. We will know its flawed if we just apply more measuresticks.

I'm being facetious. But. It's also a genuine flaw that currently is hand waved away because there isn't a better answer and please stop asking difficult questions like how we can guarantee security on inputs and outputs from AI when there is no way to validate this shit.

So how am I validating it ?

The option I'm currently using is to automate it, but show your working to the human. So the human doesn't have to do the grunt work. But just confirms it looks good. Of course. Your bandwidth blows. You have a stupid monkey in the process chain. Always a red flag. But. For the moment. It beats a human doing it manually. By a huge factor. One human can end up doing many many humans work. So. That's a productivity win. A major one. Even if it does still rely on a nose picking monkey to validate it all. Next steps with this is opening this up to be a domain expert. Storing its memories in your local data cluster via a vector store. And giving it a persistent "instantiated self". We don't ask how much that costs. It's potentially a lot. Because currently the AI models are leashed to a static core and bolted on contextual memory, which means its computationally expensive to keep the spinning plate of an "expert self" in the air. At some point however. All of that then gets tested on a local AI cluster - not running through the likes of Azure, but actually running on a minisetup in your own datacenter. Or for some of our clients. The warm lump under the desk in the storeroom. You can actually get away with surprisingly small setups for this. It's relative. What I deem "surprisingly small" is "eye wateringly expensive" to the average home user. But it's not "bonkers corporate budget expensive".

But. In any case. The tech is continuing to move at a stupid pace. And the contextual memory and massive overhead of compute for that is disappearing into where it should be - the core model. The model gets to learn as it goes. Both Harvard and Google have recently noted they have models now doing this. This is also, if you're paying attention, ground zero for the AI apocalypse scenario of them growing out of control, out of sight, and one day deciding that you are a problem and need purging. Cue Terminator.

But. We don't worry about that. 

If. And when. I am not dying on my ass. And my brain is functioning. All this stuff is bread and butter. Not easy by any means. But. I can fly with this stuff.

But I am not what I used to be. And my up time is now very small if at all. And really it is the ghost of who I am. But. That's me in a nutshell these days. I learned how to do all the things. Then got sick.

I can see it's a bit of a shame really. This is a whole new dawn of IT development stuffs. Very paradigm breaking. And very less "oh look heres a new scripting language that looks a lot like C again and we should all use it to build web pages".

Nausea has been on the down low the last 12 or so hours. It hasn't gone. It waxes and wanes. But at a much lower level. I am feeling more switched on. But fragile.

I have been following the blow by blow advice from chatGPT. And it seems to work. Basically. Don't fuck around. Stick to porridge or something similar - no caffeine, no fats, no fibre, no fruits, no acid. Drink small. Drink warm, not cold, not hot. Eat small. Dont eat after waking. Eat a tiny amount before sleeping. Don't go too long without eating.

Some concrete things I learned - coffee, tea, not good for when my stomach is shit. Even caffeine free. No bueno.  I suspected this might be the case. Sometimes after a large warm drink my nause would rise a notch. As it turns out, what I suspected is actually a thing. At least according to our AI overlords. Also. The lemon and ginger tea I was drinking when particularly delicate to help out, was half good, half bad. Ginger. Good. Lemon. Bad. Shit. Who knew. I need to get some ginger tea on its own. I haven't had anything but water to drink for 3 days.

I also bought myself a "nausea bracelet thinger". Apparently they can work. Some people swear by them. It puts pressure on your wrist, and means that nerve is always firing. This can reduce nausea. In that. Acupuncture voodoo kind of way.

I'm not sure it's done anything. But my nausea has been less with it on. But then again. My diet has been radically light. Porridge and water. One rule of debugging. Never change more than one variable at a time, otherwise it's hard to discern causality.

I can't say I'm big into acupuncture. Not my thing. I do know however that some of that shit absolutely works. Just. My own experience. And if nothing else. The whole "itchy back" bullshit you get with asthma is really telling. When asthma kicks in in a certain way your back itches like a fucker. Crazy itchy. You can scratch it and its nice. But it never makes it go away. Why ? Because. The body is sending you a signal your lungs are full of shit. It can't make your lungs itch. So. It posts up somewhere close as good enough. Hence. Your back itches. When your lungs are gunky. It's bonkers. But. It does show. The light that lights up is not necessarily attached to the button you think it is. I also get this with my shitty nerve legs and twitches. I can tap pressure points sometimes that will make my fucking leg spasm. Like some fucked up knee reflex test. I can trigger something at times, and it will make that boiling nerve flip thing suddenly come up out of nowhere and jerk. It's odd.

Anyway.

Nausea is down. But grumbling.

I am loathe to start eating willy nilly. I am glad in a way that I seem to have some rules that seem to help out. I have a plan. Which is hopeful. And useful. 

I now know what to do. When to do it. Why I am feeling like that. And how best to cope with it.

At least. If you believe in the AI overlords.

For myself. I think I'm already committed to the future of healthcare being in the feedback from an AI. It's here already. You just didn't notice it.