Hacker News Reader

Comment by thriftwy | Hacker News Reader

I think that soon it will be considered a bad practice to possess a skill or to use some knowledge for decisions..

If you have learned to do something, you will proceed by teaching a machine learning system that. And then you will use that, and not your internal knowledge.

The reason will be cited as follows: If you know something, this introduces bus factor. People can't get in your head to validate if you're applying your knowledge fairly, or biased, or maliciously misappropriating things. Your reasoning is unverifiable. You can leave for greener pastures, carrying your skills and knowledge with you.

While machine learning system is repeatable, you can have all kinds of checks and KPIs. And it's protected by IP laws.

Human domain knowledge will be discouraged.

View on HN

This is an interesting thought experiment, but I imagine it having precisely the opposite effect. Modernism, as a technology, works precisely by making power relationships seem obvious, natural, and therefore not something subject to critique. These relationships include definitions of gender, race, and class, as well as capitalist and colonial relationships, and many more. Probably the most salient power relationship to this article is deciding who should benefit from the work of robots. Acting through computers amplifies this effect, and makes it easier to uphold oppressive power relationships while keeping one's hands clean.

I recognize that this is not immediately an empirical claim, but I will observe that discarding it offhand in favor of the equally ideological dominant interpretation would be exactly the dynamic I'm pointing out.

This would effectively freeze all development and leave us collectively stuck at the current level of expertise.

If every doctor trains an ML algo, then forgets how to be a doctor, then medicine stops dead at that level.

>If every doctor trains an ML algo, then forgets how to be a doctor, then medicine stops dead at that level.

It's very rare that technology becomes ubiquitous and free very quickly. You will still need doctors in poor areas, destitute countries, and in the military.

Still, education and society would likely optimize for getting people to the edge of knowledge more quickly if ML algos truly took over a field like medicine.

i dunno... i just read about alphago zero. Only needed to teach it the rules of the game, and it thrashed the current level of cumulative human expertise (alpha go) 100-0 matches, which had just thrashed a single human the year before.

This seems like its happening at some hedge funds as well for stock picking.

I guess the closed system of the game rules is more manageable than all of medicine, but... I can't agree with "stuck as the current level of expertise".

I wasn't able to determine from what I read yesterday if AlphaGo Zero played vs AlphaGo or vs itself (AlphaGo Zero). This probably isn't the place to ask, but if it is, do you know?

It was trained with self-play and evaluated on AlphaGo Lee (the version that beat Lee Sedol) and AlphaGo Master (the version that beat Ke Jie and I believe a number of online matches against top-ranked professional players).

I don't know if this adds or subtracts from your dystopian view, but the act of training the machine would introduce the same reasoning/bias flaws one has.

(see the MIT moral machine)

At least it is repeatable and measurable. But yours isn't.

I'm not sure it's dystopian. Maybe human decision making will be considered art. That's an awesome thing, but not necessary what you do in workplace.

Someone wise once told me, even after we have strong AI and don't need to program, he will still program, because the process of programming leads to understanding, and he will still want to understand things. The AI might be able to write the program better and faster, but it won't be able to help him gain the understanding he would have gotten from writing the program himself.

This is why you will learn to program, write the program, gain the understanding, train the ML system properly to reproduce your results, and then basically unlearn everything.

Why would you unlearn though? Eventually the AI will do all of the above as well. Maybe eventually programming will be like blacksmithing or weaving is today...

Completely obscure?

I was thinking more like the Renaissance fair today. You'd visit the fair that recreates a hack-a-thon in the late 90s. At the fair, you'd see the old programmer at his terminal, coding in Perl in an emacs window... instead of turkey legs, you'd have Fritos and Mountain Dew...

Weapons of Math Destruction is also a good resource on this topic, about how bias is already being encoded in "black box" algorithms that has real effects on things like the availability of credit.

It already is.

For instance, let's say you pull up to a red light and the intesection is empty. No cars, no pedestrians.

Common sense tells you to just run the light, there is no reason not to.

But it's illegal for a human to use their own judgement to do this, and the instructions of the red light have to be followed regardless.

> the instructions of the red light have to be followed regardless

No they don't. If there's no enforcement, and no moral reason to follow the law, you've reduced yourself to a robot if you follow it.

As far as red lights are concerned though, you'd better be damn sure cars or bikes or pedestrians aren't going to leap out from around a corner before you run them. Especially if the intersection is empty at 3am and visibility is poor. In fact in that situation it's better to play it safe.

If there is no camera and no police, you can run treat the light as a stoplight. Just be safe about it.

By can, I mean that a human being can choose to do so. Maybe smart cars in the future won't allow the driver to make that decision. But nothing forces us to always obey traffic laws right now.

Enforcement is just a technicality.

In China and many other countries, rules are typically so numerous, draconian or fuzzy that you are nominally constantly breaking something. If the authorities want to alter your behavior, they call in the specter of potential enforcement to encourage behavioral change. On the one hand this seems horrible - shifting sands, constant unknowns, zero security - but in reality it's kind of liberating - less need to hire lawyers and asking for forgiveness is tolerated. The gap between enforcement and law is precisely where most people seek to operate, because it typically creates additional efficiencies. In short, enforcement is law. Or to rephrase this in the words of Snowden - Policy is a one-way ratchet that only loosens over time.

Reminds me of a Karl Schroeder novel (Permanence) in which characters encounter an alien race (or the remains of an extinct one?) which evolved to have intelligence but not consciousness. Or are we turning ourselves into philosophical zombies?

This is an interesting take, but I think part of what makes us human is our inability to disassociate our “domain knowledge” from our sense of self.

Much domain knowledge is in some sense innate or even subconscious, picked up in the field. I think it would be more efficient just to spy on people’s thoughts (and work) in some manner if there is a fear they could misapply their knowledge or leave a critical process hanging.

Yes they will allow for a degree of inefficiency if it removes human factor.

The reasoning will go, yes we are losing $1 on every automated transaction, but we can make more of them, and we certainly avert one $10,000 corruption case and one $100,000 case where we get sued for human decision.

You can't sue machine learning algorighm (that how it would go). But human, you can, even frivolously.

Yes but you can sue the company that owns the machine learning algorithm, probably for much more money than you can sue an individual. You can even imagine class action law suites and the like.

(I am one of the researchers mentioned in the article.)

You can also have paper trail documenting that you never did anything illicit during the development of the model.

There will be protocols for sensitive areas.

Can't do the same with individual decision making.

It is much harder to ask an algorithm about it's reasoning than a person's.

"Joe, why did you fire Carla?"

"She was late to work every other day and spent all day on her phone"

"HR-bot 5000, why did you fire Carla?"

"Weights indices 59738, 837, and 28836 added to over 0.67, yielding a score of only .85, which is below the firing threshold"

Certainly you could look at the input data and try to convince a jury that her performance wasn't acceptable but you can't prove that is why the AI fired her.

We cannot do this today, but this capability, of using language, is one we are actively working towards. Eventually, you can imagine an AI that can engage in a dialog like this one, in words. For example, see the DARPA Explainable AI program: https://www.darpa.mil/program/explainable-artificial-intelli....

Great link, thanks. These are the kind of systems that will really allow AI to take off.

At my previous shop, it was normal to silently tolerate developers breaking the rules (not locking their stations when they stepped away, browsing unapproved sites, coming in late, leaving early)... until it was decided that they needed to be fired because they were rude to someone important, or etc... at that point they lost their privilege to act like a professional and their badge-access records were pulled and compared to their time-entry.

The reason developers got fired is because of clear and well-defined rules... that were never enforced except when the decision to fire them for an arbitrary reason had already been made.

I think HR-bot 5000 would treat employees more equitably by forcing employers to confront the inequity of some HR regulations. The same way self-driving cars "drive weird" because they stop at stop signs and drive the speed limit.

Well, where "Machine Learning" (aka: black box you can only get weights for a given question), one can train a machine to be racist, sexist, ageist, or whatever.

The problem is that the end weight distribution is different than the GB's or TB's of training data. How do we know the training was fair and impartial? How did the trainers even know if it was? What biases crept in on this stage?

Worse yet, what if the bias of the black box does denigrate black people... Say, we take in all pictures of convicted criminals- it's disproportionaly black. I would argue part of that is because of inherent policing biases, but that's embedded in "guilty" verdict. Who's at fault for this "bias"? Is there a fault? How do we detect, other than exhaustively?

I'm eagerly awaiting for methods to "open up" ML black boxes and see what makes them tick. See their decision trees, their neural weights. I want to poke and prod to see what's behind those series of numbers like [.888271829 1.10999292992 37.999999921 1000.32 .73] . Right now, it's shove data in exhaustively and hope for the best. I don't particularly care for that way of analysis.

A lot of businesses now don't care about their bus factor, I don't see why that would change.

Seems like that would greatly reduce the value of labor.

Reply on news.ycombinator.com