James Bridle is treating the readers of the Guardian to a spotlight event. It is a fantastic article that you must read (at https://www.theguardian.com/books/2018/jun/15/rise-of-the-machines-has-technology-evolved-beyond-our-control-?). Even as it starts with “Technology is starting to behave in intelligent and unpredictable ways that even its creators don’t understand. As machines increasingly shape global events, how can we regain control?” I am not certain that it is correct; it is merely a very valid point of view. This setting is being pushed even further by places like Microsoft Azure, Google Cloud and AWS we are moving into new territories and the experts required have not been schooled yet. It is (as I personally see it) the consequence of next generation programming, on the framework of cloud systems that have thousands of additional unused, or un-monitored parameters (read: some of them mere properties) and the scope of these systems are growing. Each developer is making their own app-box and they are working together, yet in many cases hundreds of properties are ignored, giving us weird results. There is actually (from the description James Bridle gives) an early 90’s example, which is not the same, but it illustrates the event.
A program had windows settings and sometimes there would be a ghost window. There was no explanation, and no one could figure it out why it happened, because it did not always happen, but it could be replicated. In the end, the programmer was lazy and had created a global variable that had the identical name as a visibility property and due to a glitch that setting got copied. When the system did a reset on the window, all but very specific properties were reset. You see, those elements were not ‘true’, they should be either ‘true’ or ‘false’ and that was not the case, those elements had the initial value of ‘null’ yet the reset would not allow for that, so once given a reset they would not return to the ‘null’ setting but remain to hold the value it last had. It was fixed at some point, but the logic remains, a value could not return to ‘null’ unless specifically programmed. Over time these systems got to be more intelligent and that issue had not returned, so is the evolution of systems. Now it becomes a larger issue, now we have systems that are better, larger and in some cases isolated. Yet, is that always the issue? What happens when an error level surpasses two systems? Is that even possible? Now, moist people will state that I do not know what I am talking about. Yet, they forgot that any system is merely as stupid as the maker allows it to be, so in 2010 Sha Li and Xiaoming Li from the Dept. of Electrical and Computer Engineering at the University of Delaware gave us ‘Soft error propagation in floating-point programs‘ which gives us exactly that. You see, the abstract gives us “Recent studies have tried to address soft errors with error detection and correction techniques such as error correcting codes and redundant execution. However, these techniques come at a cost of additional storage or lower performance. In this paper, we present a different approach to address soft errors. We start from building a quantitative understanding of the error propagation in software and propose a systematic evaluation of the impact of bit flip caused by soft errors on floating-point operations“, we can translate this into ‘A option to deal with shoddy programming‘, which is not entirely wrong, but the essential truth is that hardware makers, OS designers and Application makers all have their own error system, each of them has a much larger system than any requires and some overlap and some do not. The issue is optionally speculatively seen in ‘these techniques come at a cost of additional storage or lower performance‘, now consider the greed driven makers that do not want to sacrifice storage and will not handover performance, not one way, not the other way, but a system that tolerates either way. Yet this still has a level one setting (Cisco joke) that hardware is ruler, so the settings will remain and it merely takes one third party developer to use some specific uncontrolled error hit with automated assumption driven slicing and dicing to avoid storage as well as performance, yet once given to the hardware, it will not forget, so now we have some speculative ‘ghost in the machine’, a mere collection of error settings and properties waiting to be interacted with. Don’t think that this is not in existence, the paper gives a light on this in part with: “some soft errors can be tolerated if the error in results is smaller than the intrinsic inaccuracy of floating-point representations or within a predefined range. We focus on analysing error propagation for floating-point arithmetic operations. Our approach is motivated by interval analysis. We model the rounding effect of floating-point numbers, which enable us to simulate and predict the error propagation for single floating-point arithmetic operations for specific soft errors. In other words, we model and simulate the relation between the bit flip rate, which is determined by soft errors in hardware, and the error of floating-point arithmetic operations“. That I can illustrate with my earliest errors in programming (decades ago). With Borland C++ I got my first taste of programming and I was in assumption mode to make my first calculation, which gave in the end: 8/4=2.0000000000000003, at that point (1991) I had no clue about floating point issues. I did not realise that this was merely the machine and me not giving it the right setting. So now we all learned that part, we forgot that all these new systems all have their own quirks and they have hidden settings that we basically do not comprehend as the systems are too new. This now all interacts with an article in the Verge from January (at https://www.theverge.com/2018/1/17/16901126/google-cloud-ai-services-automl), the title ‘Google’s new cloud service lets you train your own AI tools, no coding knowledge required‘ is a bit of a giveaway. Even when we see: “Currently, only a handful of businesses in the world have access to the talent and budgets needed to fully appreciate the advancements of ML and AI. There’s a very limited number of people that can create advanced machine learning models”, it is not merely that part, behind it were makers of the systems and the apps that allow you to interface, that is where we see the hidden parts that will not be uncovered for perhaps years or decades. That is not a flaw from Google, or an error in their thinking. The mere realisation of ‘a long road ahead if we want to bring AI to everyone‘, that in light of the better programmers, the clever people and the mere wildcards who turn 180 degrees in a one way street cannot be predicted and there always will be one that does so, because they figured out a shortcut. Consider a sidestep
A small sidestep
When we consider risk based thinking and development, we tend to think in opposition, because it is not the issue of Risk, or the given of opportunity. We start in the flaw that we see differently on what constitutes risk. Even as the makers all think the same, the users do not always behave that way. For this I need to go back to the late 80’s when I discovered that certain books in the Port of Rotterdam were cooked. No one had figured it out, but I recognised one part through my Merchant Naval education. The one rule no one looked at in those days, programmers just were not given that element. In a port there is one rule that computers could not comprehend in those days. The concept of ‘Idle Time’ cannot ever be a linear one. Once I saw that, I knew where to look. So when we get back to risk management issues, we see ‘An opportunity is a possible action that can be taken, we need to decide. So this opportunity requires we decide on taking action and that risk is something that actions enable to become an actual event to occur but is ultimately outside of your direct control‘. Now consider that risk changes by the tide at a seaport, but we forgot that in opposition of a Kings tide, there is also at times a Neap tide. A ‘supermoon’ is an event that makes the low tide even lower. So now we see the risk of betting beached for up to 6 hours, because the element was forgotten. the fact that it can happen once every 18 months makes the risk low and it does not impact everyone everywhere, but that setting shows that once someone takes a shortcut, we see that the dangers (read: risks) of events are intensified when a clever person takes a shortcut. So when NASA gives us “The farthest point in this ellipse is called the apogee. Its closest point is the perigee. During every 27-day orbit around Earth, the Moon reaches both its apogee and perigee. Full moons can occur at any point along the Moon’s elliptical path, but when a full moon occurs at or near the perigee, it looks slightly larger and brighter than a typical full moon. That’s what the term “supermoon” refers to“. So now the programmer needed a space monkey (or tables) and when we consider the shortcut, he merely needed them for once every 18 months, in the life cycle of a program that means he merely had a risk 2-3 times during the lifespan of the application. So tell me, how many programmers would have taken the shortcut? Now this is the settings we see in optional Machine Learning. With that part accepted and pragmatic ‘Let’s keep it simple for now‘, which we all could have accepted in this. But the issue comes when we combine error flags with shortcuts.
So we get to the guardian with two parts. The first: “Something deeply weird is occurring within these massively accelerated, opaque markets. On 6 May 2010, the Dow Jones opened lower than the previous day, falling slowly over the next few hours in response to the debt crisis in Greece. But at 2.42pm, the index started to fall rapidly. In less than five minutes, more than 600 points were wiped off the market. At its lowest point, the index was nearly 1,000 points below the previous day’s average“, the second being “In the chaos of those 25 minutes, 2bn shares, worth $56bn, changed hands. Even more worryingly, many orders were executed at what the Securities Exchange Commission called “irrational prices”: as low as a penny, or as high as $100,000. The event became known as the “flash crash”, and it is still being investigated and argued over years later“. In 8 years the algorithm and the systems have advanced and the original settings no longer exist. Yet the entire setting of error flagging and the use of elements and properties are still on the board, even as they evolved and the systems became stronger, new systems interacted with much faster and stronger hardware changing the calculating events. So when we see “While traders might have played a longer game, the machines, faced with uncertainty, got out as quickly as possible“, they were uncaught elements in a system that was truly clever (read: had more data to work with) and as we are introduced to “Among the various HFT programs, many had hard-coded sell points: prices at which they were programmed to sell their stocks immediately. As prices started to fall, groups of programs were triggered to sell at the same time. As each waypoint was passed, the subsequent price fall triggered another set of algorithms to automatically sell their stocks, producing a feedback effect“, the mere realisation that machine wins every time in a man versus machine way, but only toward the calculations. The initial part I mentioned regarding really low tides was ignored, so as the person realises that at some point the tide goes back up, no matter what, the machine never learned that part, because the ‘supermoon cycle’ was avoided due to pragmatism and we see that in the Guardian article with: ‘Flash crashes are now a recognised feature of augmented markets, but are still poorly understood‘. That reason remains speculative, but what if it is not the software? What if there is merely one set of definitions missing because the human factor auto corrects for that through insight and common sense? I can relate to that by setting the ‘insight’ that a supermoon happens perhaps once every 18 months and the common sense that it returns to normal within a day. Now, are we missing out on the opportunity of using a Neap Tide as an opportunity? It is merely an opportunity if another person fails to act on such a Neap tide. Yet in finance it is not merely a neap tide, it is an optional artificial wave that can change the waves when one system triggers another, and in nano seconds we have no way of predicting it, merely over time the option to recognise it at best (speculatively speaking).
We see a variation of this in the Go-game part of the article. When we see “AlphaGo played a move that stunned Sedol, placing one of its stones on the far side of the board. “That’s a very strange move,” said one commentator“, you see it opened us up to something else. So when we see “AlphaGo’s engineers developed its software by feeding a neural network millions of moves by expert Go players, and then getting it to play itself millions of times more, developing strategies that outstripped those of human players. But its own representation of those strategies is illegible: we can see the moves it made, but not how it decided to make them“. That is where I personally see the flaw. You see, it did not decide, it merely played every variation possible, the once a person will never consider, because it played millions of games , which at 2 games a day represents 1,370 years the computer ‘learned’ that the human never countered ‘a weird move’ before, some can be corrected for, but that one offers opportunity, whilst at the same time exposing its opponent to additional risks. Now it is merely a simple calculation and the human loses. And as every human player lacks the ability to play for a millennium, the hardware wins, always after that. The computer never learned desire, or human time constraints, as long as it has energy it never stops.
The article is amazing and showed me a few things I only partially knew, and one I never knew. It is an eye opener in many ways, because we are at the dawn of what is advanced machine learning and as soon as quantum computing is an actual reality we will get systems with the setting that we see in the Upsilon meson (Y). Leon Lederman discovered it in 1977, so now we have a particle that is not merely off or on, it can be: null, off, on or both. An essential setting for something that will be close to true AI, a new way of computers to truly surpass their makers and an optional tool to unlock the universe, or perhaps merely a clever way to integrate hardware and software on the same layer?
What I got from the article is the realisation that the entire IT industry is moving faster and faster and most people have no chance to stay up to date with it. Even when we look at publications from 2 years ago. These systems have already been surpassed by players like Google, reducing storage to a mere cent per gigabyte and that is not all, the media and entertainment are offered great leaps too, when we consider the partnership between Google and Teradici we see another path. When we see “By moving graphics workloads away from traditional workstations, many companies are beginning to realize that the cloud provides the security and flexibility that they’re looking for“, we might not see the scope of all this. So the article (at https://connect.teradici.com/blog/evolution-in-the-media-entertainment-industry-is-underway) gives us “Cloud Access Software allows Media and Entertainment companies to securely visualize and interact with media workloads from anywhere“, which might be the ‘big load’ but it actually is not. This approach gives light to something not seen before. When we consider makers from software like Q Research Software and Tableau Software: Business Intelligence and Analytics we see an optional shift, under these conditions, there is now a setting where a clever analyst with merely a netbook and a decent connection can set up the work frame of producing dashboards and result presentations from that will allow the analyst to produce the results and presentations for the bulk of all Fortune 500 companies in a mere day, making 62% of that workforce obsolete. In addition we see: “As demonstrated at the event, the benefits of moving to the cloud for Media & Entertainment companies are endless (enhanced security, superior remote user experience, etc.). And with today’s ever-changing landscape, it’s imperative to keep up. Google and Teradici are offering solutions that will not only help companies keep up with the evolution, but to excel and reap the benefits that cloud computing has to offer“. I take it one step further, as the presentation to stakeholders and shareholders is about telling ‘a story’, the ability to do so and adjust the story on the go allows for a lot more, the question is no longer the setting of such systems, it is not reduced to correctly vetting the data used, the moment that falls away we will get a machine driven presentation of settings the machine need no longer comprehend, and as long as the story is accepted and swallowed, we will not question the data. A mere presented grey scale with filtered out extremes. In the end we all signed up for this and the status quo of big business remains stable and unchanging no matter what the economy does in the short run.
Cognitive thinking from the AI thought the use of data, merely because we can no longer catch up and in that we lose the reasoning and comprehension of data at the high levels we should have.
I wonder as a technocrat how many victims we will create in this way.
Lawrence van Rijn - Law Lord to be 