Hacker News Reader

Comment by tomelders | Hacker News Reader

A agency that I once worked for was asked to create a presentation of the future of trading for a well known trading platform. It was the typical fluff around voice interfaces, augmented reality, VR and all that drivel. But we all knew what the future really meant - taking people out of the equation - but we also knew that the people paying for the presentation didn't want to hear that.

When people talk about the risk of automation stealing jobs, they tend to think it's the low skilled jobs that are most at risk. But of your job is to look at data and make descisions based on that data, you're gonna be the first to go. And the more money pinned to your work, the quicker you will be replaced.

View on HN

The terrifying thing is that you don't even need to reach human-level performance to cause structural unemployment. As a hypothetical example, if automation can make top-performers in your field 10x more efficient, that makes 90% of people in that role redundant.

This feels a lot scarier to me since as a ML engineer, I rarely see applications where we can reach human level performance easily but the idea of making existing top performers 2-5x more effective feels much more viable.

The idea of 80% of people in relatively high-skill jobs who are used to a relatively high income level becoming redundant is a pretty scary one.

Lots of high-earning people tied up in managing money is pretty inefficient for society, though. Maybe those people can switch to factory jobs and help reduce the cost of common things for the common person.

I can't tell if that's meant as sarcasm or not.

It's kind of a joke. People getting rich off moving money around is kind of silly, and it would certainly be better if those jobs could be eliminated by machines.

Oh, don't worry. People will still get rich off of those transactions.

The vision of public blockchains (such as Ethereum) is to make the core of financial transactions automated and non-profit. While there may still be profit to be made on the margin (credit scores), the vision is that currency and equity trades will happen on decentralized markets that are automated and non-profit.

...Even if this vision occurs, it's not obvious how soon it would occur. For example, will we need stable-value cryptocurrencies before any of this can occur? Will anything significant occur in the next two years, or will it be five? OR more?

Of which they will quickly become :)

>I rarely see applications where we can reach human level performance easily but the idea of making existing top performers 2-5x more effective feels much more viable.

This actually reminds me of Palantir's focus, which is to develop tools to aid analysts.

This is nothing new. I work marketing side running customer acquisition for 12 countries across 3 brands, with budget significant enough to bring in ~3% of a countries population annually in our higher coverage markets. And this year we will add additional countries and brands.

We are a team of 3. There is no way this would have worked 10+ years ago. Maybe in another 10 we will be 1...

I guess they can all become Javascript developers and fulfill the big tech companies' dreams of having hordes of cheap engineers on the labor market. That would help save some of the money poured into various initiatives like 'Dolphins who code', etc.

What is wrong with being a JavaScript developer?

Nothing. On the contrary, it's the most marketable skill, judging from the number of Bootcamps that prepare for it.

until software becomes available that makes programming 10x easier...

This idea has been around for over 10 years but I still don't see that many signs of it happening. I mean what interesting projects are around that actually make it possible to for example, let a non-technial person point and click develop a python or Java application? It happened with HTML/CSS and Dreamweaver/WordPress etc years ago but it's much easier with web stuff and I don't really see it elsewhere.

I think we're in the early stage of automating web development but sites like Wix have made impressive strides in making UI design accessible. I'm working on something in this field myself and I've done enough research and prototyping that I think it's definitely achievable.

once ML starts replacing programming jobs, we'll have singularity.

I think that's a stretch. 90% of programming is repetitive if you compare different projects. Also, I didn't mention ML/AI.

>>But of your job is to look at data and make descisions based on that data, you're gonna be the first to go.

I posted the other day about machine learning and radiology. It's the type of job that squarely fits your definition, although things move much more solely in medicine.

It's also likely that human-in-the-loop will stay around, possibly indefinitely in medicine. A human-in-the-loop is a huge safety thing for medicine, but it defeats the whole point of HFT.

My dad is a neuroradiologist so I've seen a decent amount of the reading work (which is one component of the job - though a big one). I also thought it'd be well suited to ML, but I suspect the future is more suggestions and flagging things in scans rather than entirely replacing the doctor (I'd be surprised if this isn't happening already in some capacity).

It seems risky to lose the human understanding and interpretation of the images. Human augmentation is probably better?

Already people make assumptions because of not understanding constraints in how the underlying tech works (like how FMRI is monitoring changes in blood flow not neural activity directly).

The thing is, when you have 4 radiologists reading 100 scans each, if flagging and highlighting and generally improving performance of these personnel leads to two of them being able to process 200 scans each, the other two are out of a job.

They have been, de facto, replaced by computers, even if the job role still exists.

It's easier to see this if you break down the work not into images but into human-image-minutes. If there are 100 minutes of work, but you can assign the computer 50 of them which it can complete in seconds, the same number of human-image-minutes can be completed as before, but the computer has now taken them away from a human.

Unless the productivity gain is spent on doing more imaging. I think this aspect is underapreciated in many debates about automation, especially in health care. There is an insatiable demand for diagnostics for early detection of diseases or monitoring progression.

>Human augmentation is probably better?

Most of the automation work I do in financial services is a 90% automation. This leaves room for more analytical tasks where people can free up time to drive more business decisions and insights.

Isn't this type of trading sort of unique in how easy it is to consume the data? Even in other investing areas, people who just "look at data and make decisions" cannot be easily replaced until computers can read and understand natural language and understand the state of the world. I can't immediately think of anything else this type of automation could apply to.

Computers are already making trades based on natural language articles.

EX of some research from 2008: http://www.seas.upenn.edu/~cse400/CSE400_2007_2008/DavdaMitt...

You see real world movement within seconds of the release of some public data. Which is far to soon for humans to read much of anything.

I'd have to wonder what type of safeguards vet against intentional false data released to poison the well of other HFT bots?

Maybe it's not too important yet because HFT bots only need to game the slower mass of "dumb (reactionary) money"

Not a whole lot. [1]

[1] http://www.theatlantic.com/sponsored/etrade-social-stocks/th...

Absolutely, but isn't this type of software generally used to augment existing jobs, rather than replace them? The program can make basic inferences like "10 negative GE articles -> Sell GE stock", but you still need a human to make more general decisions about an industry as a whole, related companies, how long-term an issue is, etc. Basically, nobody's job is as simple as reading the news and deciding whether articles are positive or negative. In the case of this Goldman Sachs story, some people's jobs really were as simple as looking at the performance of equities and making decisions based on that alone.

>Absolutely, but isn't this type of software generally used to augment existing jobs, rather than replace them?

Is there a difference? If a person does 10x as much you need 1/10th as many of them.

But this is about responding faster, not responding more often. It helps you react to a story before everyone else, but there's only so many stories in a day.

Or you can have 10× productivity.

You have that either way. The question is do you need 10x the output?

These algorithms are only good at front running the reaction of the market, which is fairly predictable in the short term when a relevant piece of news is published. They are nowhere for making long term investment decisions.

AI everywhere is as naive as software everywhere. It costs a lot of money to develop and maintain software professionally. It's uneconomical to staff an IT team to replace every routine job there is with software.

It will be the same with AI. If you apply enough bright people's mind for long enough, we can possibly automate certain niche jobs, but that's going to be completely uneconomical. You need to pay these AI specialists, and retain them so that you can adapt your algos as the problem moves, and there needs to be more than one in case one goes or is ill. You can easily spend more that the salaries you are going to spare.

Other problems are I believe too vast and open ended for AI to come up with anything useful. What company will be successful in 5 years? What is the right level of renumeration for the management of that company? Where is innovation going next? You really think that AI is anywhere near answering that kind of questions?

AI will shine in similar areas where software shines. Problems that are simple enough to be understood by computer scientists (and that narrows down this vast world quite a bit), and that are broad enough to justify the economics of making high investments to build data, or where there is already ample of existing and relevant data. Driving a car. Cleaning or building a house. Replicating certain hand movements. Things like that. I doubt an expert on a narrow and specialised domain has much to fear from AI.

> It will be the same with AI. If you apply enough bright people's mind for long enough, we can possibly automate certain niche jobs, but that's going to be completely uneconomical.

The thing is, it doesn't have to be an all-or-nothing proposition. If you can write software to make a team of ten traders 10% more productive, it may be possible to do the same job with only 9 traders.

Over time, those productivity gains compound and you have literally cut in half the number of people on your trading floor.

I've seen that happen to FX traders, and some fairly straight-forward automation. No AI was required.

AI is a very flexible term, all software that makes choices technically qualifies as AI even if the benchmark is moved over time. But, remove the buzzards and rejecting a loan using a single if statement based on a Fico score really is machine decision making at work.

It should be noted that demand is not a static quantity. If a technology allowed a person to create more per-unit cost will be cheaper, raising demand (people on the bottom will still lose jobs, but it's not linear).

Ah but in 2008 these news articles were still written by humans :-)

Yes. Not to mention that spreads are so tight that the only real challenge for a cash trading desk is how to read the flow, and manage inventory, into events. Computers don't do a great job of interpreting doublespeak and insinuation, which is a substantial part of earnings calls and other releases of information.

Derivatives trading has grown as cash trading has shrunk. And the banks haven't figured out how to automate that business yet.

Not really. The world is data. Anything you need to know about the world is in a table somewhere, and the web has done a great job of making the planets data accessible to machines.

Other areas may be more complex, but complexity isn't as big a barrier as it may seem.

Also, those areas are probably not as complex as people think they are. I'm reminded of the AI that started making breakthroughs in oncology by looking at the cells around a tumour, despite the scientific consensus that there was nothing to be learned from the cells around a tumor.

Spoken like a true data scientist. Hope there's room for culture in there somewhere.

Sorry to burst your bubble, but culture is data too.

The problem is understanding data. A human can learn a new word's meaning by reading its definition in a dictionary; a computer can't.

It's a dictionary. I'm pretty sure a computer can look it up faster than a human can.

Access to raw text is not the hard part.

Even so, that doesn't stop one human helping the computer to understand, and thus replace x more humans.

> Anything you need to know about the world is in a table somewhere

Not yet.

> Anything you need to know about the world is in a table somewhere

No it isn't. If I read an article about the Volkswagen emissions scandal, what table do I look in to figure out how much consumers will care, and how aggressive various governments will be with their punishment? When Disney buys the rights to Star Wars, where should I look to see if it's worth what they paid for it? Should I rely on data from 30 year old movies, or look at merchandise sales or something? If I start seeing a lot of articles about climate change and the dangers of fossil fuels, should I sell my stock in Exxon?

Not that humans are particularly good at this stuff, but I don't think computers will be as good or better for a long time.

I think a lot of people get hung up on how humans make decisions, with the expectation that you need to be able to understand the same ideas to perform the task.

Even if a computer can't figure out the stats you mention, maybe it doesn't matter? Maybe just reacting fast enough to what people in the market are doing is good enough? Maybe simple sentiment analysis on articles on Disney around the web is a good enough proxy for share price increase.

Sounds like that's giving the humans who are left in the market 10x the power.

You're making the mistake of being specific. Think heuristic.

Another obvious example is actuaries and any number of other analysts in the insurance industry. According to Google[1], actuaries earn an average of just over $97k/year. While this may not strictly be an "investing area," it's a clear target for automation.

1. https://www.google.com/search?q=actuaries

I don't know about actuaries, but I do remember there was an article lately about an insurance company in Japan that replaced hundreds of workers with IBM's Watson.

https://www.theguardian.com/technology/2017/jan/05/japanese-...

That's really interesting. In my circle of friends, most of us can't point to a single success of Watson, despite it's very publicized capabilities. This may only be a small company, but almost certainly larger insurance agencies & medical foundations must be on to this.

Actuaries have quite a high regulatory wall protecting them from other human competitors. But probably not from computers.

(I don't know about the legalities, but the laws requiring companies to have a actuary sign off on stuff probably can be satisfied by having on part time actuary shared between multiple companies sign off on computer generated analysis.)

I know in banking, regulators are satisfied with ML models making decisions about bank capitalisation, so long as the models are annually audited by an external auditor (eg. KPMG, EY)

I assume they will need to sign off on the model... I expect to see a lot of this in the future.

Part of the reason actuaries make so much is that they are certified and legally required to sign off on analysis, similar to the way civil engineers put their signature (and reputation) on their work. But you would be surprised how many decisions come down to "actuarial judgement", having good business sense, intimate knowledge of insurance rules and regulations...and not doing anything special with the data. Hell, some actuarial methods are the same as they were in the 70s. Data science is slowly eating their lunch in predictive modeling areas, but I wouldn't be surprised if they're still around after data science itself has been automated.

I work on a team of actuaries in a more data science-y role. Believe me, I had the same impression initially: "Can't we automate some of this stuff?". But I've since come to see their value.

It's not that actuaries will be eliminated. It's that 2 actuaries and 2 data scientists will do the work of 20 actuaries.

How many actuaries does a major insurer employ? Would they like to halve that number?

It's not so much the actuaries but the claims adjusters whose jobs will be automated:

http://www.inc.com/kevin-j-ryan/artificial-intelligence-repl...

Computers don't need to understand natural language, they just need a little guidance towards what the data means. You can train a neural network, or you can just write some solution-specific code. A lot of people are not that hard to replace.

yeh you can't send a robot of an on site visit to look the management in the eye

i agree 100%. i think the only reason traders can be replaced by machines is because these traders didnt perform better than monkeys shooting darts.

Autor and Acemoglu wrote about this in their paper 'Skills, Tasks and Technologies':

'Following, ALM, we refer to these procedural, rule-based activities to which computers are currently well-suited as "routine" tasks. By routine, we do not mean mundane (e.g., washing dishes) but rather sufficiently well understood that the task can be fully specified as a series of instructions to be executed by a machine (e.g., adding a column of numbers). Routine tasks are characteristic of many middle-skilled cognitive and manual jobs, such as bookkeeping, clerical work, repetitive production, and monitoring jobs. Because the core job tasks of these occupations follow precise, well-understood procedures, they can be (and increasingly are) codified in computer software and performed by machines.'

No doubt. The last people to be replaced are going to be software devs (because when you can write a computer program with an AI you can make the AI write the AI and you have general AI), cleaning ladies and probably police investigators.

"But of your job is to look at data and make descisions based on that data, you're gonna be the first to go."

"The last people to be replaced are going to be software devs"

If your job is to write code so people can look at data and make decisions based on that data, you're gonna be the second to go.

The world's oldest profession will be the world's last profession.

Fishing?

Picking fruit

Not if those weird Japanese robot sex toys take off.

Nah a 3D environment with the right force feedback input would be both much more convenient, much cheaper and much more private.

because when you can write a computer program with an AI you can make the AI write the AI and you have general AI

There doesn't need to be a single AI to create the entire app. An AI to model data structure, an AI to build an interface, an AI to consume external APIs, etc. We already mostly bolt libraries together to make software; it doesn't take a huge leap for those libraries to be generated by machine learning.

Building anything but the most complex apps will be trivial in 10 or 20 years time.

I have a feeling someone said the same thing 10 or 20 years ago.

We are just ratcheting up the complexity all the time. Moving goalposts (and a bit of complexity-bloat).

They DID say the same thing 10-20 years ago, and they were CORRECT.

Ever try to build a web application in Assembly language or C? It would probably be pretty difficult. You don't have to do that anymore.

1 web developer can do more work today than 100 assembly developers could, from 50 years ago.

Build a JS app today is extremely complex.

I think that's more a sign that we want more from web apps than any increased complexity in the apps themselves. Building a web app now isn't any more technically difficult than building a Linux app was 20 years ago. The only difference is that the tools aren't as mature.

But isn't there still the step of translating the business logic into a machine readable format, and isn't that what the majority of software engineers do at this point anyways?

And up next, AI containers and AI container orchestration ran by an AI

I always consider automation to first target jobs with "trivial decision making." I think trading fits into that, as you can pretty much distill it down to an algorithm. At least as far as I understand it.

Physicians are another great example. A lot of their job today seems to be about translating signs and symptoms and following a decision tree. I'm excited about the opportunities to displace current duties with today's nice to haves, such as physicians who have time to dig deeper into a person's well being than their symptoms that day.

> But of your job is to look at data and make descisions based on that data, you're gonna be the first to go

Manufacturing and agricultural jobs have already been automated away, while data analysis jobs are only just now starting to feel the pinch - and then only in companies with very deep pockets working on a niche problem. 'First to go' has missed the mark by many years.

The difference is in marginal cost. Industrialization was a relatively gradual change. AI can make whole professions almost redundant basically overnight.

> If your job is to look at data and make decisions based on that data, you're gonna be the first to go

Depends on the type of data, how it's collected, how it's aggregated, etc.

I'm a market researcher so I look at data all day. Thing is - I look at both qualitative and quantitative data from a ton of different sources (financial filings, surveys, macreoconomic organizations, vendor briefings, engineer interviews, etc.).

I size the markets for embedded technologies that are automating away data-driven jobs. Most notably in the industrial sector, where years of near-zero industrial productivity growth have left a lot of manufacturing/oil&gas/utilities companies hungry for any way to bring costs down.

The newest technology right now is the "IoT Cloud Platform" which aggregates data from a bunch of industrial machines (directly from devices, or through IP enabled gateways), routes the data, and sends it to a cloud where it can be analyzed and monitored automatically to predict failure and prevent unplanned downtime. From a component standpoint its made up of a (1) piece of client software that sits on the machine or gateway, or a configured agentless client, (2) infrastructure VMs for load balancing/server provisioning, (3) host VMs to run an OS, middleware, and a runtime framework, and (4) applications that run on top of the host VM.

One of the more interesting demos of this tech is a Microsoft/GE joint project that used drones to take pictures of power lines, sent that data to the Azure IoT Suite, which then performed visual analysis to determine which power lines were damaged or deteriorating, saving the cost of sending humans out to climb up these structures. All the big companies - Amazon, Microsoft, IBM, SAP, Oracle - are trying to add more intelligent applications on top of these platforms, with machine learning, neural nets, and blockchain-based applications being some of the most advanced.

If your job follows a simple binary data check or logical chain - is this machine functioning within specs (if not, order replacement), did we hit the target price for this asset (if so execute x number of orders), does this piece of equipment look functional (if not, report to manager) - sure you should be worried about your job.

If you go one step up from the basic logical workflow, and enter the realm of data synthesis, or of handling data that requires skepticism/critical thinking, I think you have no reason to worry about an algorithm doing your job anytime soon.

The financial sector in general is getting some of its excess fat trimmed due to a period of increased regulation and low interest rates/returns. This is a good spin for these finance companies, rather than "we're laying off 2/3rds of our cash equities trading desk".

We shall see. My money is on the machines. Or more specifically, mathmatics.

Having formerly been in market research, and automating-out many parts of that process myself, gotta go with op. There will be a need for people who can give data/study designs a sanity check/interpretation for at leeast for like a decade or two. Not that I don't think some exec somewhere will try or already are trying to automate it out, just that it's a terrible terrible idea.

Pretty soon all trading will be an elaborate form of corewars.

Software engineers think they're somehow immune to the cycle.

I agree. Doctors should be very worried.

Reply on news.ycombinator.com