This is the blog of Vivek Haldar

De-skilling

Nicholas Carr’s new book, The Glass Cage, opens with a discussion of the FAA’s recent 1-page alert that cautions pilots to do more manual flying. Long, routine hours of flying on autopilot was atrophying their skills to the point where when something went wrong and manual intervention was required, many did not know what to do.

Here’s an idea that sounds implausible today but might be practical in a decade or so: have pilots fly all commercial jets remotely from a “flight center” on the ground. There might be one such center for a designated geographical area. They might hand planes off to one another. Air Force pilots already fly drone missions halfway across the world remotely from a base in Nevada, so at least an early version of this idea is already being practiced.

That would mean pilots are fully engaged and active during their shift, rather than being lulled by hours of autopilot after a 2-minute manual takeoff. When something goes wrong on a flight, there will be a roomful of pilots to help out.

The major hurdle to implementing this will probably not be technical but getting people to board a commercial jet that doesn’t have a pilot on-board.

The issue with automating certain jobs is that their scope doesn’t expand. For example, a pilot performs certain actions, and if you automate 90% of those, the pilot is left with the same amount of time but only 10% of their prior workload. Contrast this with software development. Many tasks that were manual a decade ago are now handled by tools: automated build/test/deploy, IDEs that understand code and can perform sophisticated refactorings, infrastructure for rent etc. But today’s software developer still has their hands full because the overall complexity of developing software and the scope of problems that software tackles has grown too. Autocomplete in the IDE might have made them soft when it comes to remembering the APIs, but they do have bigger problems to grapple with.

“My second childhood was kindly delivered to me by my children. This, I suspect, was no coincidence. Children know something that adults have forgotten — something adults have to forget when they begin playing the great game of growing up and becoming someone.”

Tennis with Plato, by Mark Rowlands

“…students almost universally resist going back over material they’ve written. But… students are willing to revise their essays, even multiple times, when their work is being reviewed by a computer and not by a human teacher. They end up writing nearly three times as many words in the course of revising as students who are not offered the services of E-Rater, and the quality of their writing improves as a result. Crucially… students who feel that handing in successive drafts to an instructor wielding a red pen is “corrective, even punitive” do not seem to feel rebuked by similar feedback from a computer.”

Robo-readers aren’t as good as human readers — they’re better

How software grows

I have often wondered how things grow, and specifically, how software projects grow. Some initial looks at specific projects had pointed out that they grew super-linearly.

A recent paper looks at a large cross-section of open source projects to analyze how adding people to projects affects their productivity.

They find that production intensity , as measured by number of commits, grows super-linearly with the number of project members. The median exponent was 4/3, which works out to a productivity boost of a factor of 2.5 when number of members is doubled.

Not surprisingly, the exponent was smaller (but still super-linear) for larger projects, but also showed less variability.

(For the link, I thank Nat Torkington’s Four Short Links. He somehow manages to come up with four interesting links every day.)

Now today, in the 21st century, we have a better way to attack problems. We change the problem, often to one that is more tractable and useful. In many situations solving the exact problem is not really what a practitioner needs. If computing X exactly requires too much time, then it is useless to compute it. A perfect example is the weather: computing tomorrow’s weather in a week’s time is clearly not very useful.

The brilliance of the current approach is that we can change the problem. There are at least two major ways to do this:

Change the answer required. Allow approximation, or allow a partial answer. Do not insist on an exact answer.

Change the algorithmic method. Allow algorithms that can be wrong, or allow algorithms that use randomness. Do not insist that the algorithm is a perfect deterministic one.

Shifts in algorithmic design

“The conservative view sees this as an automotive technology, and most of them are very used to thinking about automotive technology. For the aggressive school, where I belong, this is a computer technology, and will be developed — and change the world — at the much faster pace that computer technologies do.”

The two cultures of robocars

When to do a rewrite

Spolsky’s famous screed against rewriting from scratch is now canonical. All the same, there are valid reasons to do a rewrite:

  • Your understanding of the domain, or the domain itself, has changed drastically in the process of building the first version.
  • The scale for which you built the initial system has grown by more than an order of magnitude. The right design for millions of users is very different than that for thousands.
  • Newer languages and runtimes make the product much better along some dimension: deliver better performance, be more maintainable etc.
  • You have accumulated so much technical debt that starting from scratch is easier than untangling the mess you have. This is often the most misleading reason, because engineers always prefer a system they wrote from scratch to inheriting one which works just fine. But even so, it does happen.

None of the above takes away from what Spolsky wrote. A rewrite is still a risky endeavor, even if one or more of the above conditions hold true, and should always be approached with caution and humility.

His argument was much stronger in the era of software that was shrink-wrapped to PCs, because the slow rate of change of the surrounding environment made many of the above reasons inapplicable.

But server-side software that powers online services is always in a constant state of flux. There is rarely any concept of “done”. The product might look more or less the same from the outside, while the backend code constantly churns to deal with growing scale, new stacks, and other services it depends on.

Learning your life

If you are a “regular” person living and working in modern civilization, it is likely that your life has a few recurring patterns:

  • on weekdays, you travel from home to work in the morning, and back in the evening; perhaps punctuated with a school or daycare drop-off and pick-up.
  • on weekends, you either mostly stay at home or frequent a small set of haunts (beaches, restaurants etc).
  • infrequently, you break out of this pattern by leaving the orbit of work-home to go on a vacation.

This is easy to figure out by observing someone for a short period of time. Of course, your phone already does this.

The above model of your life could be described in a few kilobytes. It is a zeroth-order machine learning problem. Is that depressing? Does that mean your whole life has a few kilobytes worth of complexity?

John Foreman, in his poignantly named piece “Data Privacy, Machine Learning and the Destruction of Mysterious Humanity”, takes a view of the field from the inside, and it sounds like that of a physicist working on the Manhattan Project worrying about the impact of what they’re building.

Our past data betrays our future actions, and rather than put us in a police state, corporations have realized that if they say just the right thing, we’ll put the chains on ourselves… Yet this loss of our internal selves to the control of another is the promise of AI in the hands of the private sector. In the hands of machine learning models, we become nothing more than a ball of probabilistic mechanisms to be manipulated with carefully designed inputs that lead to anticipated outputs… The promise of better machine learning is not to bring machines up to the level of humans but to bring humans down to the level of machines.

That is a dim, if plausible, view.

It is based on a pessimistic view of humans as always giving in to nudges to behave as our worst selves. I hope that we are better than that. We are already surrounded by advertising for junk food, and yet a substantial fraction of us manage to ignore that and eat healthy. I hope machine predictions that try to nudge us face a similar fate: predictions that look down on us will be indistinguishable from low-quality ones. The quality and tone of predictions will become an important part of the brand identity of the company emitting them. Do you really want to be the company known for peddling fatty, salty food, legal addictive substances, and predatory loans?

A tangential point: naming matters. Statistical inference, which is what machine learning really is, is a less scary name. “Statistical inference” conjures images of bean-counters with taped glasses; “machine learning” of an inevitable army of terminators.

Circling back to the question I posed above: is your life really describable in a few kilobytes in a predictive model? No. A model is a probabilistic abstraction. You, and your life, are not a sequence of probabilities, in much the same way that hiking a trail is a much richer, “realer” experience than tracing your way through a detailed trail map.

“The AP announced last month that it would use Automated Insights’ software, called Wordsmith, to produce up to 4,440 robot-written corporate-earnings reports per quarter, more than ten times the number its human reporters currently produce. When Alcoa’s earnings report hit the wire that day, the data was instantly compiled by a firm called Zacks Investment Research and passed through the AP’s proprietary algorithm, which pulled out key numbers and phrases and matched them against other contextual information. In milliseconds, the software produced a full story written in AP style, indistinguishable from one made with human hands. In short order, the story appeared on many AP member sites.”

Robot Journalism

“Here’s the tablet news app I’m hoping for in 2014: something that detects when I’m lying down in bed (it knows my device orientation and, using the ambient light sensor, can detect it’s dark) and serves me the sort of video content that I never have time to watch during the day.”

Katie Zhu

Garland Words

I define a garland word to be one which can be formed by chopping off the last few letters and “wrapping around” to the start.

For example: You can chop off the the trailing “on” from “onion” and still form the word by wrapping around to the starting “on”. In other words, you can write the letters “o”, “n”, “i” in a circle, in that order, and form the word “onion”.

I say that a garland word is of degree n if you can do the above with the last n letters of the word.

It turns out that garland words are fairly rare. In the English dictionary that came installed on my Ubuntu machine, there were 99171 words, only 199 words were garland words of degree 2 or above. The highest garland degree was 4.

The exact distribution was: there were only 3 words of degree 4, 32 of degree 3, and 164 of degree 2.

The degree 4 garland words were: “hotshots”, “beriberi” and “abracadabra”.

Lurking Smalltalk

Stephen Kell’s paper at PLOS 2013 meditates on one of CS’ oldest design questions: what should be in the operating system, and what in the programming language and runtime? The paper was a breath of fresh air; one of those opinionated design thinking papers that one rarely sees in academic CS venues. Do read the whole thing.

This question comes into stark relief if you look at what a modern language or VM runtime does: among other things, it multiplexes resources, and protect access to them. But that’s also exactly what the OS is supposed to do!

The paper riffs on an earlier idea from Dan Ingalls: “An operating system is a collection of things that don’t fit into a language. There shouldn’t be one.” The central problem tackled in the paper is that of having more generic object bindings in the OS, as opposed to the “static” filesystem bindings for most operations. Kell uses the following example:

Suppose we wish to search some object for text matching a pattern. If the object is a directory tree of text files, grep -r does exactly that. But suppose instead our directory tree contains a mixture of zipped and non-zipped files, or is actually some non-directory collection of text-searchable objects, such as a mailbox. In Smalltalk parlance, how can we make these diverse objects “understand” (via a collection of interposed objects) the messages of our grep -r (and vice versa)?

This would indeed be elegant. But I’m afraid the engineer in me isn’t terribly excited by that.

“All problems in computer science can be solved by another level of indirection” is a famous quote in computer science. The next sentence is much lesser known: “But that usually will create another problem.” What Kell is proposing adds flexibility in binding by introducing indirection.

Indirection is controlled by some sort of configuration the binds each side of it. And configuration implies a lot of added complexity. The UNIX model of “everything is a stream of bytes and can be treated like a file” might be simple and unstructured, but that is precisely what enabled a multitude of structure to be built on top of it.

Heartbeats

It’s commencement—and commencement speech—season. The one that’s been on my mind the most this year has been Paul Ford’s speech to a graduating class of interaction designers. His unifying thread is our conception of time:

The only unit of time that matters is heartbeats. Even if the world were totally silent, even in a dark room covered in five layers of foam, you’d be able to count your own heartbeats.

When you get on a plane and travel you go 15 heartbeats per mile. That is, in the time it takes to travel a mile in flight your heart beats 15 times. On a train your heart might beat 250 times per mile…

… If we are going to ask people, in the form of our products, in the form of the things we make, to spend their heartbeats—if we are going to ask them to spend their heartbeats on us, on our ideas, how can we be sure, far more sure than we are now, that they spend those heartbeats wisely?

Turns out, heartbeats are a great unit of time for another reason: all animals have approximately the same number of heartbeats in a lifetime.

With each increase in animal size there is a slowing of the pace of life. A shrew’s heart beats 1,000 times a minute, a human’s 70 times, and an elephant heart beats only 28 times a minute. The lifespans are proportional; shrew life is intense but brief, elephant life long and contemplative. Each animal, independent of size, gets about a billion heartbeats per life.

I find that deeply poetic.

The biggest winners may be lawyers who can use machine intelligence to create an automated large-scale practice. The Walton family, it’s worth recalling, got rich by effectively automating large-scale retail. More generally, there may be jobs for a new category of engineer-lawyers—those who can program machines to create legal value.

But the large number of journeyman lawyers—such as those who do routine wills, vet house closings, write standard contracts, or review documents on a contractual basis—face a bleak future.

Machines v. Lawyers, by John O. McGinnis

“Autodesk’s automated design software works through three layers. Generators populate physical space with matter, creating thousands of design suggestions in a split-second. Analyzers test each design against constraints developed by the designer: does it fit in the desired shape? Does it satisfy strength, thermal, material, and mechanical requirements? Above those, an optimization system finds the best designs, returns them to the starting point, and refines them repeatedly.”

The Automation of Design

Copyright 2008-2013 Vivek Haldar