I like this article from Tom's Guide primarily because while the "AI" industry is running apace giving stuff new, or modified, names so as to bemuse customers, some things are actually newsworthy.

What are they?

They are the things that demonstrate that the hype is unwarranted, or at least questionable.

This is the important passage: "There's an argument that because there's no "golden path" to get from A to Z throughout Pokémon, with random encounters and Pokémon trainers popping up regularly and no easy way for an AI to stick with a set "team" of critters."

Every morning I take a dawn walk and there is no plan. When I get in the lift to go down a couple of hundred metres, I don't know if I'm going to the ground floor (and out of the front gate) or to B3 and out of the back gate. Often the decision is based upon whether I can be bothered to get my keycard out of my pocket or, if I do, if it works. If both conditions are true, then it's probably the back door; else it's the front door. See: it's an algorithm but.. it's an algorithm that does not rely on data, it relies on my state of mind.

Randomisation is not only a statistical concept

The two options give me very different results: at the front is a winding tree-lined road down a long (for a city centre) hill; at the back, it is very urban. Where I go next is determined by how far it is until my next decision: left, right, straight. At the front, it's several hundred metres before I have to make that decision; at the back, the decision has to be made over and over again, with often only ten or so metres between instances. As I walk at a very slow 5kph-ish (a big improvement over a year ago but still pathetic: I'm walking on it)

At the back, within 500 metres I might be walking on a busy road; at the front all I see is street-sweepers and birds, most mornings. But the route I take varies daily and there is no plan. Decisions are made on whether I wonder what is up a lane, if I can see something I want to take a photo of (not that I'll ever look at it again, probably) or just the way the dawn light is falling or even "I'm feeling good, let's add a few thousand steps". The decision making is based upon factors that will not arise again, or are unlikely to and certainly not with any predictability. .

At some point, I will decide that I'd like a cup of coffee and I'll set out a route, not necessarily direct, for the single non-random point - a coffee shop called Bonz where they make coffee called "orang putih kafe" (white man's coffee) for me: even my friends who say they like strong coffee find it beyond their ability to finish a cup and I have it in a large mug. Sometimes, two. So that's a fixed data point but only as to location, not as to time. Today, for example, I went out late so I could walk to a supermarket, then have coffee and then home.

When I leave the coffee shop, I make another decision - but I make it on the pavement outside: do I walk the short, flat way home to the back gate or the five times longer, all uphill way to the front gate?

Data is not necessarily measurable or recordable but it is assessable. 

This relies on data but most of the data is not empirical i.e. measurable. But it is assessable even if it cannot be recorded. 

Do I feel like walking any more, how many steps have I already done, are my legs tired, do I just want to get home and into a shower? Will the bread dough have finished its first proving and is it time to knock it back and prove it again before baking?

There is data: respiration rate, heart rate, the speed I'm walking at, sweating (not that that measure is any use given that it's often more than 80% humidity even though the sun is just coming up), the length of my stride, how high I'm lifting my feet, how tired my legs are,  my stride frequency. Some of that data can be collected (I don't, except the walking speed) and I know if I'm lifting my feet enough because if I'm not I trip on uneven pavements and how tired my legs are by how much I subconsciously take a direction to avoid the high kerbs of the city centre).

Sometimes walks are 2500 paces, sometimes 12,000 and everywhere in between. There is no point at which anything is routine and no point, other than to wait at the coffee shop until I turn up, that where or when I am is predictable. And the coffee shop is closed on Sundays and I do not have a regular alternative.

The fact is that any attempt to predict my route or timing is awash with data but it is all historical and reasons for those decisions are not data driven except in the vaguest sense.

People's ordinary lives are messy. We do not do as expected. Real life is as the computer is finding while playing Pokémon.

The bank, correct data but the wrong data.

I've been talking to a bank about a customer that they decided was problematic. He often doesn't pay his mortgage on time and when he is three days late, the bank starts the process of pestering him. Multiple telephone calls every day then, after about three weeks, they add in SMS. I looked at the data. He has never missed a payment: payments are, with an odd exception, made within the correct month. But the bank allocates resources because their system says to do so. No one looked at the data beyond "due date, x: x+3 start pressure and ramp it up every two or three days."

But what the data showed was that the mortgage was paid regularly - every four weeks, in fact.

Well, within a day or two of four weeks, easily within a margin of acceptability.

What happens every four weeks in the UK?  Pensions are paid.

So, there will always be a discrepancy between payments that fall due on a specific date and money coming into his current account every four weeks. Then he pays the mortgage from his current account.

Far from being a problem customer, he's a customer that's managing two inconsistent patterns. Why not explain that to the bank? Because the clerks that phone have no authority to mark the file "leave the chap alone - there's nothing to see here." All they want is to fill in a form that says he made a commitment to pay on a specific date. Why? Because then they have a stick to beat him with. So he does not ignore the calls but he does not answer them.

The system was designed by someone who did not think beyond the basic and not only does it not have any flexibility, there is no mechanism to report that so the problem continues to repeat itself month after month. Now the bank is fed up because he doesn't answer their calls or SMSs. So now his account is flagged to check if the payment has not been made during the current month and both the bank and the customer can live in peace, so long as he pays when his benefits arrive.

Some anomalies are suspicious; most are not. 

Where the people do not see that some anomalies are not problems, could they program machines to look for that?

I think not: the fact is that anomalies are the basis of suspicion but they are not themselves suspicion. Most anomalies can be explained, simply, by applying some common sense and a bit of actual intelligence.

All those companies saying "we can solve your problem" can only do it if they can foresee all possible options and in the real world, like on my morning walks, options are adopted for absolutely no good reason except I feel like it.

Today, it was leave home late, go to the supermarket, carry 10Kg of stuff home, stop for coffee and use the back door.

Who knows what it will be tomorrow, except that I'll be having industrial strength coffee. Perhaps I'll have two cups. Dunno.....

Regulators need to understand this

Regulators are focussed on compliance; they are not focussed on risk. Their obsession with measurability, metrics, KPIs and the volume of reports, for example, to say nothing about the detail of what is in a counter-money laundering system rather than on its effectiveness at identifying and dealing with risk, has led to exactly the tick-box approach they say they don't want. Financial institutions have no choice: when a regulator, advised by external consultants who have a conflict of interest because they also advise the banks, etc,. says "do this and do it this way" a bank that thinks for itself is already set up for failure. 

Regulators around the world are telling banks they must, without demur, apply a range of technologies that are unproven and/or proved to be flawed. The same consultants are often the "channel" for sales and profit from both commissions (in the outside world we might think of them as back-handers) and advisory and implementation fees. 

Some regulators have appointed department heads whose primary function is to develop a fintech market in that jurisdiction, with a Darwinian approach that the strongest (which does not automatically mean the best) will survive, leaving customers of failed businesses, which had operated under cover of the regulator's approval, exposed. 

Regulators must realise that the technology is simple but the programming is not. And in truth, the programmers, those that design the algorithms, are not good enough. There is a very high chance that no one is. 

https://www.tomsguide.com/ai/claude-ai-has-been-continously-playing-pok…