"Computers and algorithms" is just a synonym for people; people implement ideas in code, design the algorithms, pick the weights for training machine learning, etc. No matter how many levels of abstraction you go through it's always people at the top. If anything happens in the society there's always a person who made the decision to make it happen somewhere. Deferring that responsibility to "computers did it" is dangerous bullshit.
The first person who walked the final path. Everyone who agreed with that person and walked the same path compounded his decision. Everyone who crossed the field should have thought about what they were doing at the time.
Far too many people, and especially developers, are way more focused on absolving themselves from responsibility for anything bad than they are willing to put the thought in at the beginning, and that's the root cause of a lot of problems.
But that person could not possibly have predicted the future actions of everyone else, and therefore it's a mockery of moral reasoning to hold them responsible.
> it's a mockery of moral reasoning to hold them responsible.
What are we holding them responsible for? Was something amoral? They clearly made the first decision for better or worse. Now if there was a sign that said "No walking on the grass", they clearly violated something.
You need to read it in the context of onion2k's original post and how onion2k's post relates to the article. There is a question of responsibility in that context, whether we like it or not.
Yes, this is a metaphor for people building algorithms (or other systems) they don't fully understand that could have poor consequences for others / society. At least, that's how I understood the relevance of the metaphor.
Criticizing "people building algorithms (or other systems) they don't fully understand that could have poor consequences for others / society" implies there is a conceivable alternative. That is it possible to build algorithms with a full understanding of all the negative consequences they could ever have for society. That seems obviously absurd to me, so the criticism is vacuous.
Er, no. The alternative is to not build those systems. Here clearly suggested is, not to let algorithms make all the decisions, but let people actively manage those funds. That might be generally less efficient (in terms of volume of trades and profit), but the assumption is that a total disaster (e.g. flash crash) wouldn't happen with "slow" humans in the loop.
How is a flash crash a "total disaster"? It seems to me that just describing it that way indicates your mind is addled by technology, since before modern times nobody would expect continuous pricing of everything every millisecond of every day or think that everything was worthless because it wasn't being quoted appropriately for an instant.
Someone has to take it upon themselves to put a sign there saying, "STOP! Habitat Restoration in Progress. Please choose another path."
And that's just it. Humanity has carved a path that says, "Make as much money as possible with the least amount of effort."
These machines have taken something we already know is bad, the 90 day, quarterly earnings cycle, which had already obliterated our long-term thinking down to nanoseconds and we want it even shorter.
We have machines moving virtual money around corporations that use money to move real material around the earth -- and beyond.
That is the path we are on right now in this very moment.
Who is going to put a sign up saying, "STOP! Humanity Restoration in Progress. Please choose another path."
That's a stretch. I know you're enamoured by your synopsis, but realistically emergent systems aren't designed or thought about. They might be tweaked.
Dirt paths are created by people making the same choice when there is not enough difference in grass levels for someone to notice. It’s only after a significant number of people make the same choice that feedback occurs.
Other systems may or may not be created in a similar fashion.
realistically emergent systems aren't designed or thought about
I'm aware of how they work. I'm arguing that someone who builds a system is responsible for it even if they don't know how it works. Ignorance, even in the face of a system so complex that no person could ever understand the underlying causes of what it does, is not an excuse. No one should be able to hide behind complexity.
Developers must either build in protections against their systems going wrong or they shouldn't deploy them.
I want to agree with your thesis, but it's impossible to foresee every possible outcome. Leaky abstractions aside, bugs aside, misaligned incentives (the new "I was just following orders", as someone here immortally put it) aside, it is impossible to imagine a priori all the ways a certain outcome that is desirable now will be undesirable in the future.
Every step of the development ladder is fraught with possibilities for error and catastrophe. To quote James Goldman:
"It is too simple. Life, if it's like anything at all, is like an avalanche. To blame the little ball of snow that starts it all, to say it is the cause, is just as true as it is meaningless."
It's useful to distinguish between responsibility and accountability here. The algorithms may be responsible for a particular outcome but the people who commissioned them should be accountable.
What about turnover? If the buck stops at the CEO, what happens when that person moves on? Is their successor responsible for everything that went on prior? Im not saying either way, just asking how that should work.
Is their successor responsible for everything that went on prior?
Yes.
That wouldn't even be a change to the current system. That's how it works now. If you take on the role of CEO and it turns out that years earlier the company did something terrible you will be expected to resign. It's one of the reasons they demand so much money.
Pretty much always if the incoming CEO can competitively negotiate and is a good hire.
Contrary to popular belief, the vast majority of career CEOs are good hardworking people. Like any high profile position, the outliers skew perception for everyone.
Lets take, for example, a red light camera software that is trained to recognize plate numbers and issue traffic tickets. The decision to design and deploy the software was made by a human. The decision to send you a ticket was not. Nothing to do with complexity or being wrong, there is simply no human anywhere who made that specific decision.
Right, so we just need to be smart enough (and have enough data) to centrally (and a priori) manage the collective and emergent action of millions of humans. Seems like a reasonable expectation.
I read in a self help book that a doctor, while being a child, used to like waking up really early the first day it snowed and make a wild path in the snow, just for giggles.
Everybody else was just following his path through the snow.
Isn't that the equivalent of blaming the first single celled organism that evolved rudimentary flagellum for, as an example, our current climate crisis?
I was imagining desire paths in a park or something. It takes a lot of walks to wear down the grass, and probably hundreds before a path becomes visible.
Having walked in both long and short grass I can tell you it largely depends on a number or factors. The largest seeming to be the amount of time passing between each person.
The path is optimized for A) Where people enter and B) Where they are going. If there is no purpose. A path is simply a path, and ascribing meaning to it is fruitless.
- Whoever gave cachet to certain characteristics that are present at specific points of the grass
- Whoever made it desirable or necessary that passersby reach a certain location at the periphery of the field
- Whoever set up restrictions to entry at certain locations around the periphery of the field
- Whoever made any changes to the field (including non-human intervention)
More generally, culture at large and historical precedent.
—
Imagine a green field of grass.
There are minor differences across the field, but some of these differences can be noticed: a slightly darker patch of grass, a slightly yellowed patch of grass, some stones of varying sizes, a section that is slightly raised from the rest of the field.
Around the field, other things can catch the eye: there's a road to one side, some trees of varying sizes, one currently in bloom.
Over the seasons, the profile of this grass changes. At some point there was a park bench to one side of the field; later, it was moved to another side. There was once a storefront near to one side of the field.
You going to hold the icecream man responsible when the path has a pothole in it? He'll claim a flow of people made it and it's not his responsibility. The fairy tale is useful because it focuses on one broad idea in exclusion of the others.
Who is the man holding the smoking gun when algorithms collapse? The business owner is legally left with the responsibility, but finding who programmed the algo to make a mistake is useful & vital.
Yes - agree. If you build a bridge (in many places) you have to sign off on the design - warranting that it is your professional opinion that it will do the job. To do this you normally need to be a chartered or certified engineer and you need professional insurance.
If the bridge collapses you get sued and your insurance pays out. You probably don't get more insurance.
This is not a perfect system, but we need several elements of it for algorithms. We need the ideas of tolerances, confidence intervals and analysis - certification of fitness for purpose must be limited and constrained to have value. We need professional certification that enables employers and the public to recognize experience and capability. People have confidence in pilots because of their uniforms, qualifications and the checks that they undertake (training, retraining, blood testing, medicals) on a regular basis. As we have seen with the 737 max, when these things breakdown there's trouble.
Except usually paths are made because they're the shortest path. Taking the shortest path is not random, even if you have a curved pavement path, people will cut it and walk through grass to get the shortest path. So it's basically a pre-programmed algorithm in our brains saying we want to spend the least amount of time.
A decision to follow the herd and a decision to take the easiest road are still decisions.
You might say that a person made a decision to break down blades of grass to create a path and that others chose to take a path already created, but they are still making choices that they are responsible for.
Imagine a small box where every day at 6 am, an instruction on paper is written and the groundskeeper checks it.
Imagine a network of voters who vote on the instruction to be placed in the box. Their wide variety of votes will be automatically distilled into the winning instruction.
The groundskeeper dutifully clears a path, shouts at people to stay on the path, fills in errant paths.
Finally there is one path.
Who made the decision?
I only bring this up to suggest “emergent” decisions is a vapid framework to analyze this. Every phenomenon involving the aggregation of different agents is an emergent one, whether it’s an emergent system of politics leading to regulated actions or an emergent system of market participants leading to a price.
This doesn’t distinguish a situation of high regulation from a situation of low regulation.
Clearly the most absolute answer is to have every individual track and identify their own commits and changes to the computer system in an authoritative manner. Then you must balance each catastrophic failure against the reasonable expectation that the programmer could predict the outcome. Then you effectively strip the license to program (in that company/industry at least) from the offending 'rogue traders' when they try to run away with billions of dollars that was built on exploiting said catastrophe and get caught.
Yes, but you could create the path just as you plant the grass and amuse yourself to no end on how sheep-like people really are.
Just because there's a path there, it doesn't mean its optimal or good for the public at large.
> "Computers and algorithms" is just a synonym for people
I don't think we can say this anymore. I wouldn't say that the makers of AlphaGo were the decision makers for any of AlphaGo's moves. You could perhaps say the players of the games AlphaGo trained on were. When it comes to AlphaGo Zero you can't even say that. At most you could say that the people were policy makers or tweakers.
Yes, but AlphaGo can't be used as a metaphor for a lot of the algorithms/automated trading strategies which were implemented based on research done by people, and carried out by computers. AlphaGo only trained against itself.
I agree with that it's always people at the top, however it doesn't imply that there's always one person who made the decision. Collective decisions exist too. And even looking at individual decision, not all decisions are taken through careful conscious means.
Plus, deciding to delegate decisions (to algorithms or anything else) is not the same as deciding through your very own direct perspective.
That was as true in the days of when pit trading dominated as it is nowadays.
It's true that automated trading is fast, and that means that things can go off the rails fast. And, it's true, that has led to some spectacular news-making events. But, for that very reason and others, most trading firms keep a very close eye on what their auto-traders are doing. Closer than they ever kept on their pit traders.
I briefly worked at a firm that did both HFT and pit trading, supporting one of the teams in charge of monitoring both groups. And what I saw was that the company was generally much more in control of the algorithmic trading. It was more heavily instrumented than a human could ever be, with humans (who were themselves licensed traders) watching the computers very closely at all times, as many layers of automated fail safes that would kick in if things went a little screwy.
By comparison, the pit trading was kind of a mess, and the team was frequently having to scramble to amend all sorts of human errors largely stemming from things like miscommunication and straight-up bad handwriting. Beyond that, I'm sure that, compared to the folks operating the auto-trading systems, the pit traders never had as firm of a real-time grasp of the finer subtleties of their positions, simply because nobody's able to do math in their heads like that.
Which, I think, is why spectacular events in the markets stemming from speculative trading going off the rails aren't really any more common nowadays than they were in the days before high frequency trading. Yeah, computers created some new ways to screw things up, but they also mitigated some existing ones.
The net effect is not nearly as bad as journalists make it out to be. But, of course, playing to people's fear of the unknown has always been the more lucrative route for journalists to take.
That hints at something. Now that you summarize it, it is very similar to my experience in a completely different field. The higher the rate of automation the higher the level of control. If things are done mostly manually it tends to be an uncontrolled mess with inefficient processes in place that create the illusion of control.
I think you're missing the point. Trading and Sales used to be people intensive, in today's parlance, 'high touch.' You had people talking deals on the phone, you had transactions sent to the exchange where a floor trader did the execution, you had a trader's assistant fill out the trade ticket and submit it manually to the mid-office, where you had a person who recorded the transaction then passed it to the back office, where yet more people processed and settled the trade.
All of those people are gone. No need, we have automated much of this deal cycle. Your best example is to take a trip to your local exchange floor, which used to be littered with runners, assistants, floor traders, etc. It's eerily quiet these days, as most trades are executed by computers.
Hmmm... I both agree and disagree. While humans program the general strategy, and even the bugs they didn't think they were adding, there is a substantial difference: computers can and will act in such a completely stupid way as to be obviously stupid to even the most uninformed human.
What I fear the most is a silly bug causing a crash in a way no manually triggered series of trades would.
>What I fear the most is a silly bug causing a crash in a way no manually triggered series of trades would.
Banks are run by smart people and they worry about the same things. There's dedicated systems checking to ensure algorithms aren't running away. All the big banks and brokers have been buying them for a decade or so. Nobody wants a crash and the algorithmic traders are mostly using a line of credit from the bank with which to do their trading and the bank usually take a commission on the profit so the banks have a strong interest in seeing them make money as opposed to crash the market and disappear into bankruptcy proceedings. Of course the government also swooped in and saved the day by mandating risk checking systems a couple years after everyone who cared was already using them (thereby solidifying the status quo for better or worse).
Could you not say the same kind of thing about mortgage backed securities, Fannie Mae etc?
We've had flash crashes, which aren't in peoples interests, they still happened. I don't think its reasonable to hand wave it away, because "smart people". Programmers are supposedly smart people,I've yet to come across a perfect bug free program.
>Could you not say the same kind of thing about mortgage backed securities, Fannie Mae etc?
Human based checks are easier to subvert than automated ones.
>We've had flash crashes
Which is exactly the problem that most banks and markets try to prevent. Any system that uses what other people are doing as a strong enough input is going to be susceptible to this to some degree. If you try to buy too many shares of something or for too disparate of a price compared to the market the bank will almost certainly kill your order. The exchange might kill it too. Depending on your contract with the bank and exchange they may kick your session for the day if you do it too much (once may be too much). The limits most banks place on what you can do mean that runaway algorithms can really only be a slow tug in one direction or the other which is no big deal.
"Human based checks are easier to subvert than automated ones."
I disagree. They're easily subvertable, but in different ways.
"Which is exactly the problem that most banks and markets try to prevent"
But they haven't succeeded so far. Further, as the saying goes, history never repeats itself, but it rhymes. It's all well and good doing all these automated checks based on past issues, that's not going to help with the next issue that isn't enough like the one before.
At the end of the day, both systems are subject to issues. Computers are far more disciplined than humans, but that comes with a tradeoff of lack of introspection.
Bee colony is just a synonym for bees; bees run the colony, bees build the hive, bees are the hive. No matter how many levels of abstraction you go through it's _always_ bees at the top. If anything happens in the hive, there's _always_ a bee that kickstarted it.
A collection of individuals and/or systems giving raise to a metasystem, a superorganism, does not and can not exercise deliberate control over it: the metasystem lies on a higher level. Not to mention as this thread - in its totality - surely proves, that individuals in many cases are completely unaware of the layer of control above them. What does an individual bee know about the colony?
What is "dangerous bullshit" is wishful thinking. Nation states have been with us for milennia. Corporations for centuries. Have we learned nothing that we still entertain notions of (human) global control?
That is exactly what happens. It's direct human nature to defer that responsibility to "computers did it", and to believe they will not in the future is naive.
The second part of that is that the human at the top can only make decisions on what is visible to them. With fairly opaque ML systems the only thing easily visible is the results. Seeing how it came to that result is much more challenging. So we will have humans at the top that can only makes decisions about the results and not the means. Which is the part many, including myself, find troubling.
True to a point. However, that person doesn't necessarily know why a computer made a decision. The high dimensionality of these questions and decisions can be solved by programming, but the question and solution aren't completely understood. Thus, we enter the age of AI.
That doesn't change the fact that they're responsible for the outcome.
A reasonable metaphor is someone firing a gun in to the air: they can't claim not to be responsible for where the bullet lands despite having no real control over where it's coming down. If you don't want that responsibility don't pull the trigger.
Right, but that doesn't change the fact the stock market has changed from people taking careful aim before pulling the trigger, to a model where people are mostly looking where everybody else is firing, then firing in the same direction.
Yeah, a lot of people warn about the dangers of AI. And it's real - handing over decision making we don't understand to computers comes with real risks.
Yes, but when computers do things that humans do there is simply less human oversight. If a human is taking action to enforce a policy that will harm other people, there's the chance that he'll resist it or raise awareness.
If an algorithm does it, there's no final line of defense. And software is easy to scapegoat because it's murky. You can claim "we didn't know the algorithm would do that," but it's harder to scapegoat your employees when there's emails on record showing you specifically told them to enact a social harm.
> If anything happens in the society there's always a person
Who decided to make George W Bush president? I highly doubt you can assign responsibility to either a single individual or even fully enumerated clear subgroup of individuals.
There's also a person who can/will yank the power cord out of the wall when/if it looks like the algo is going haywire and losing a ton of money, like happened during the quant quake.
For every automated trading system that is operating on major exchanges today, there are people responsible for its actions. If a system disrupts a market significantly, regulatory agencies work to be sure the responsible people are fined, banned from trading, or even imprisoned. There is no deferment of responsibility to "computers did it".
> If anything happens in the society there's always a person who made the decision
This is not true in case of machine learning. Yes, there was a person or people who designed the model, gathered training data, trained the model and deployed the result. But the decisions the model is making as part of its function are not made by these people.
Most of the time, people who trained the model cannot even tell you why the model is making the decisions it is making. Tractability of ML models is currently a very hot topic in the field and progress is being made, but a lot of currently deployed models are inherently intractable.
Set your spouse’s alarm clock for 4 am. When it goes off, calm them down by pointing out that while you set it and you turned the volume up to 11, it wasn’t your decision to disrupt their sleep; they should blame the timer inside instead. Let me know how the couch feels afterward.
Seriously, the responsibility lies with whomever made—or okayed—-the system. One-off things, especially if they are hard to foresee, might get a partial pass. Otherwise you ought to be carefully testing and evaluating your own system instead of blaming gradient descent.
> the responsibility lies with whomever made—or okayed—-the system
And generally speaking these are very different people, but that is besides the point. I am still unsure why we are talking about testing and evaluating, as if we are only talking about failure cases.
If I set the alarm clock for 7AM and it starts ringing at 4AM, the fault is with the clock designer. But if I set it for 4AM and it starts ringing at 4AM, it is somewhat silly of me to start blaming the clock designer for my disrupted sleep, isn't it?
And we are not talking alarm clocks, we are talking about systems that make complex decisions based on hugely multidimensional data. Decisions that we provably cannot explain. There is no amount of testing and evaluation that can ensure that these systems are error free. But when they work, they do work well and provide a competitive edge over human analysts.
> if everyone hates the current system, who perpetuates it?
And Ginsberg answers: “Moloch”. It’s powerful not because it’s correct – nobody literally thinks an ancient Carthaginian demon causes everything – but because thinking of the system as an agent throws into relief the degree to which the system isn’t an agent.
It's automation taking away the lower level trader and analyst positions. These were jobs that were very lucrative. This creates an even more rarefied, winner-take-all atmosphere, and makes an industry that was already considered amoral downright sociopathic.
