As I’m sitting down to write this, thinking about what kinds of AI I’ve bumped into at work alone I’m quite surprised. This is the kind of buzz (and I’m pretty sure it’s still buzzy rather than anything yet useful) that crypto only hoped it could have. Four, or possibly five, different AI services being trialled or used in earnest that I’m aware of at my company.

I know this isn’t limited to just my company either, as other friends have recently talked about it being introduced in companies which aren’t strictly tech companies.

It’s not hard to see why. For one thing, it’s pretty flashy branding. It’s cool to say you’re on the bandwagon. (Oddly though, people do not trust products which feature AI in their marketing.)

I work for a scale up company that’s eager to grow in the market as fast as we can, so the idea of AI is interesting. There are some manual labour tasks which we need to be able to do more of, or less of, depending on demand. It’s not entirely about avoiding hiring more people (although that’s a huge advantage), but if we hire those people, what do they do on quiet days? The traditional answer is using third-party labour, which is what we’ve done so far. (I just made up that term – I’ve no idea how they’d like to be referred to.) We have an agency who can make phone calls for us, or of late do some Mechanical Turk style work where we can increase or decrease how many people we need. There are cumbersome disadvantages to doing this though: change is very slow. A new “script” for the agency to follow can take literal days for them to implement. At my place, we love a good A/B test, and it’s so difficult to quickly spin up a test when there’s a multiple day lag for the agency to catch up (and then when the test ends, retrain them again without the test section).

If that could all be AI’d away, we could respond to change in demand of our service much faster. (Disclaimer here: I’ve no idea how the company feels about this internally, or if there are any such thoughts or plans. I’m not anywhere near the decision making for this. I’m just talking about ideas here and not specific situations.)

There are a number of discussions happening about how we can better use AI, so I’ve been investigating a few solutions in the market. Some of them, I’m unsure need “AI” at all. It’s become one of my go-to things in meetings now. “This doesn’t need to be AI.” This isn’t just an anti-AI sentiment, but a financial one too. One company was charging a shocking amount of money to provide analysis of something (I’m going to be intentionally vague here), and that shocking amount was per analysis. We did some quality checking of their API and found that their accuracy was around 40%. Another part of the company was shocked by this (because contracts with the supplier promised higher, but also because they really wanted to use the software) and so did their own analysis: manually going through thousands of samples. They found the accuracy was much lower than 40% when judged by their expert eye.

These systems are just shit. Shit and exorbitantly expensive. They’re not expensive because they’re greedy. They’re expensive because it costs a lot do to do these analyses. They’re not price gouging. It’s just really hard to do at a reasonable cost. And still achieving lofty highs of 20% accuracy. Now I’m just ranting but one service provider gave us such bad answers to the questions we asked it, that we realised it was much more accurate if we just inverted everything it said.

I was telling this story at dinner last night and I end it with my pet peeve: “The frustrating part about this whole thing is that it doesn’t even need AI. It just needed OCR technology from the 1980’s, and we’d be fine. We could run it in house, on a Raspberry Pi and pay nothing.”

Then one of my very wise friends said, “Well, OCR is just a type of AI.”

Well, I have to tell you reader that I was shocked. So shocked I did a really bad job at what I believe AI to be, which is why I’m writing this here now, to get some thinking straight. I turned to my partner for help and all he said was, “I think it’s a marketing buzzword for something that doesn’t exist.” Even more shocked. On one single table we had opinions stretching from ‘it doesn’t exist’ to ‘we’ve had it since 1914’.

Which tells me that the whole thing is bollocks anyway, but here’s what I’ve decided my line is:

Artificial Intelligence, at least in this part of the century, is something that is capable of being given more training data which leads to it being more accurate. This modelled data is not logic-tree based, but is attribute based, and the AI decides which attributes lead to the desired outcomes.

Optical character recognition has no training data. You can’t give it more training data to improve its accuracy. It does not deal in heuristics. It’s in fact several technologies which run, one-after-another in a pipeline that eventually leads to the machine being able to make a decision about what is in front of it.

Given an arbitrary piece of paper with writing on it, the OCR software would first try to remove “skew”. That’s a particular piece of software written by a developer that picks up a row of pixels and offsets it until the pixels are in a less slanty shape. Nothing heuristic here. “When counting, one by one, what percentage of pixels are in a line?” That’s it. Next, it’ll try to increase the contrast to get rid of shadows and really make that line pop. That’s just a slider, like you see in Photoshop.

You can’t increase the size of its training set. There is no training set.

Eventually, it realises that this set of pixels matches quite nicely with this known set of pixels for the letter “r” and it returns that.

It’s all incredibly deterministic (because someone wrote it) to the point where you can do the whole thing by hand if you wanted to. Just follow the instructions in the code.

AI (GAI?) would not handle character recognition like that. In fact, it likely wouldn’t bother with characters. Now, I actually don’t know how it works, but here’s my layman’s explanation. It takes a bunch of data points about the image being given. It checks through its model to see if those attributes are familiar to it. “Yep, this looks like it might be a menu. I’ve seen many menus with similar characteristics.” Then it’ll try and Sherlock its way to understanding what’s in front of it. “That seems to be something that looks like chopsticks. So maybe this is an Asian menu.” Piecing it all together, it’ll realise that it’s seen hundreds of menus like this, and the line you want specifically is usually where the ramen is. “I’ve seen lots of menus with something that looks like ‘chicken’ is near something that looks like ‘soup’. So probably it’s ramen.” And then it’ll give you a percentage. “0.78: Ramen noodles.”

Just like Sherlock, if you let it eat up more well labelled information, it can get more and more certain because it can make to more similar occurrences of those same heuristics.

A developer is still involved in building the model, by saying things like “pay less attention to the table cloth colour” or “the EXIF data on the photo is really important”, but the developer is not deciding logic paths that the AI will take.

In earlier parts of this century, we have talked about AI before in gaming, for instance. “Enemy AI” is a thing, but I bet we’d feel less comfortable calling that AI now. It’s not learning. It’s just a really big decision tree – still deterministic and logic. Sometimes you end up in surprising areas of that logic tree, where multiple factors overlap on top of each other, but it’s not a learning beast.