Back in March 2016 I wrote Semi-automated podcast transcription about my interest in finding ways to make archives of podcast content more accessible. Please read that post for details of my motivations and goals.

Some 11 months later, in February 2017, I wrote Comparing Transcriptions describing how I was exploring measuring transcription accuracy. That turned out to be more tricky, and interesting, than I’d expected. Please read that post for details of the methods I’m using and what the WER (word error rate) score means.

Here, after another over-long gap, I’m returning to post the current results, and start thinking about next steps. One cause of the delay has been that whenever I returned to the topic there had been significant changes in at least one of the results, most recently when Google announced their enhanced models. In the end the delay turned out to be helpful.

The table below shows the results of my tests on many automated speech recognition services, ordered by WER score (lower is better). I’ll note a major caveat up front: I only used a single audio file for these tests. An almost two hour interview in English between two North American males with no strong accents and good audio quality. I can’t be sure how the results would differ for female voices, more accented voices, lower audio quality etc. I plan to retest the top tier services with at least one other file in due course.

You can’t beat a human, at least not yet. All the human services scored between 4 and 6. I described them in my previous post, so I won’t dwell on them here.

• WER: Word error rate (lower is better).
• Punctuation: Number of sentences / commas / question marks / capital letters not at the start of a sentence (a rough proxy for proper nouns).
• Timing: Approximate highest precision timing: Words typically means a data format like JSON or XML with timing information for each word, Lines typically means a subtitle format like SRT, Pgfs (paragraphs) means some lower precision.
• Other Features: E=online editor, S=speaker identification (diarisation), A=suggested alternatives, C=confidence score, V=custom vocabulary (not used in these tests).
• Approx Cost: base cost, before any bulk discount, in USD.

Note the clustering of WER scores. After the human services scoring from 4–6, the top-tier ASR services all score 10–16, with most around 12. The scores in the next tier are roughly double: 22–28. Seems likely that the top-tier systems are using more modern technology.

For my goals I prioritise these features:

• Accuracy is a priority, naturally, so most systems in the top-tier would do.
• A custom vocabulary would further improve accuracy.
• Cost. Clearly $0.02/min is much more attractive than $0.33/min when there are hundreds of hours of archives to transcribe. (I’m ignoring bulk discounts for now.)
• Word level timing enables accurate linking to audio segments and helps enable comparison/merging of transcripts from multiple sources (such as taking punctuation from one transcript and applying it to another).
• Good punctuation reduces the manual review effort required to polish the automated transcript into something pleasantly readable. Recognition of questions would also help with topic segmentation.
• Speaker identification would also help identify questions and enable multiple ‘timelines’ to help resolve transcripts where there’s cross-talk.

Before Google released their updated Speech-to-Text service in April there wasn’t a clear winner for me. Now there is. Their new video premium model is significantly better than anything else I’ve tested.

I also tested their enhanced models a few weeks after I initially posted this. It didn’t help for my test file. I also tried setting interactionType and industryNaicsCodeOfAudio in the recognition metadata of the video model but that made the WER slightly worse. Perhaps they will improve over time.

Punctuation is clearly subjective but both Temi and Scribie get much closer than Google to the number of question marks and commas used by the human transcribers. Google did very well on capital letters though (a rough proxy for proper nouns).

I think we’ll see a growing ecosystem of tools and services using Google Speech-to-Text service as a backend. The Descript app is an interesting example.

Differential Analysis

While working on Comparing Transcriptions I’d realized that comparing transcripts from multiple services is a good way to find errors because they tend to make different mistakes.

So for this post I also compared most of the top-tier services against one another, i.e. using the transcript from one as the ‘ground truth’ for scoring others. A higher WER score in this test is good. It means the services are making different mistakes and those differences would highlight errors.

Google, Otter AI, Temi, Voicebase, Scribie, and TranscribeMe all scored a high WER, over 10, against all the others. Go-Transcribe vs Speechmatics had a WER of 6.1. SimonSays had a WER of 5.2 against Sonix, Trint, and Speechmatics. Trint, Sonix, and Speechmatics have very little difference between the transcripts, a WER of just 1.4. That suggests those three services are using very similar models and training data.

What Next?

My primary goal is to get the transcripts available and searchable, so the next phase would be developing a simple process to transcribe each podcast and convert the result into web pages. That much seems straightforward using the Google Text-to-Speech API. Then there’s working with the podcast host to integrate with their website, style, menus etc.

After that the steps are a more fuzzy. I’ll be crossing the river by feeling the stones…

The automated transcripts will naturally have errors that people notice (and more that they won’t). To improve the quality it’s important to make it very easy for them to contribute corrections. Being able to listen to the corresponding section of audio would be a great help. All that will require a web-based user interface backed by a service and a suitable data model.

The suggested corrections will need reviewing and merging. That will require its own low-friction workflow. I have a vague notion of using GitHub for this.

Generating transcripts from at least one other service would provide a way to highlight possible errors, in both words and punctuation. Those highlights would be useful for readers and also encourage the contribution of corrections. Otter API, Speechmatics and Voicebase are attractive low-cost options for these extra transcriptions, as are any contributed by volunteers. This kind of multi-transcription functionality has significant implications for the data model.

I’d like to directly support translations of the transcriptions. The original transcription is a moving target as corrections are submitted over time, so the translations would need to track corrections applied to the original transcription since the translation was created. Translators are also very likely to notice errors in the original, especially if they’re working from the audio.

Before getting into any design or development work, beyond the basic transcriptions, I’d want to do another round of due-dilligence research, looking for what services and open source projects might be useful components or form good foundations. Amara springs to mind. If you know of any existing projects or services that may be relevant please add a comment or let me know in some other way.

I’m not sure when, or even if, I’ll have any further updates on this hobby project. If you’re interested in helping out feel free to email me.

I hope you’ve found my rambling explorations interesting.

Updates:

• 25th May 2018: Updated SimonSays.ai with much improved score
• 10th June 2018: Updated notes about Google enhanced model (not helping WER score).
• 8th September 2018: Added Otter AI, prompted by a note in a blog post by Descript comparing ASR systems.
• 10th September 2018: Emphasised that I only used a single audio file for these tests. Noted that Otter.ai is free up to 600 mins/month.
• 14th September 2018: Added Spex.

Offentliggjort den 2. aug. 2018

Novogratz is a former hedge fund manager for investment firm Fortress Investment Group and was a partner at financial giant Goldman Sachs. He now spends his time making waves in the cryptocurrency waters as the CEO of Galaxy Investment Partners, a cryptocurrency investment firm.

“I think we put in a low yesterday,” Novogratz tweeted on September 13 — while also noting that the Bloomberg Galaxy Crypto Index “retouched the highs of late last year and the point of acceleration that led to the massive rally/bubble.” The index in question measures the performance of the largest digital currencies traded in dollars, as noted by CNBC.

Jumping the Gun?

Novogratz has been notably bullish on cryptocurrencies throughout 2018 — which is to be expected, given that he is the CEO of a cryptocurrency investment firm. However, his call that the bottom is in could very well be premature.

Bitcoin is in the throes of a bear market — a fact which cannot be denied. Every rally this year has put in a lower high, while trading volume continues to decrease. Trading volume also goes out of the cryptocurrency market as soon as it goes in, as evidenced by the market leader’s latest cascading selloff.

Furthermore, Bitcoin is in very real danger of collapsing through its key support level of $6,000.

Patience is a Virtue

Nevertheless, Novogratz believes cryptocurrencies will bounce back. To be fair, they probably will — at least, Bitcoin will. “Markets like to retrace to the breakout,” he noted in his tweet. “We retraced the whole of the bubble.”

When the seemingly inevitable bounce back occurs, it will most likely come via the market leader. Bitcoin has seen its dominance increase significantly in 2018, and it’s within the realm of possibility we could see the first and foremost cryptocurrency once again claim 70 or 80 percent of the total market.

Do you think the bottom is in for Bitcoin and/or the cryptocurrency market? Let us know your thoughts in the comments below! 

Images courtesy of Shutterstock, Twitter/@novogratz, blockchain.info, TradingView.

Along with solving the hard logistics and operations challenges of building a grocery store, we are also innovating on the customer experience (e.g. creating a community of food lovers, chat with a human for recommendations etc.), using data to build our catalog and rethinking a grocery store from first principles.

Join us as a full stack engineer if one of the 2 strongly resonate with you - - You love crafting beautiful front end experiences and love talking to customers. - You love operational challenges, and process improvement to save $$s.

In either case, you are obsessed with data. You have experience designing and building large and complex (yet maintainable) systems, and you should be able to do so in about one-third the time that most competent people think possible. Expect talented, motivated, intense and interesting co-worker.

We are extremely transparent, communicate well, own our pieces and move fast. We are a team of x-Instacart, x-Amazon, looking to build a strong culture, a serious business and work with the best. We are in SF, and are backed by top investors including YC & Khosla Ventures.

Our Stack:

Backend: Ruby / Rails API / AWS / Algolia Frontend: Javascript / React / Redux / Saga

We promise to thoroughly review every application, so we appreciate you putting in the effort. Alternatively, if you are more of a let's talk about it person - just shoot us a quick note and we'll pick up the phone and call you !

If not you, then who? jobs@mylkguys.com

Does your hometown have any mathematical tourist attractions such as statues, plaques, graves, the cafe where the famous conjecture was made, the desk where the famous initials are scratched, birthplaces, houses, or memorials? Have you encountered a mathematical sight on your travels? If so, we invite you to submit an essay to this column. Be sure to include a picture, a description of its mathematical significance, and either a map or directions so that others may follow in your tracks.

A math degree can take you to a lot of places, both physically and figuratively, and if you play your cards right, you too can argue counterfactual definiteness with a shaman. First in 2008, and several times since, a fellow math PhD and I traveled to the Burning Man art festival to sit in the desert and talk with the locals about whatever they happened to be curious about.

Burning Man began in 1986, when a group of people (who would argue endlessly over any finite list of their names) decided to assemble annually on a San Francisco beach and burn a wooden human effigy. In 1990, increasing membership and a lack of fire permits forced Burning Man to combine with Zone #4, a ‘‘Dadaist temporary autonomous zone’’ piloted by the Cacophony Society, in the Black Rock Desert, 110 miles outside of Reno, Nevada.

As of 2017, the festival has expanded to a modest 70,000 people. For the week it exists, Black Rock City (the name for the physical infrastructure of the festival) is the sixth largest city in Nevada. At first blush, it may seem a little audacious to call a festival a ‘‘city,’’ but by most definitions of the word, Black Rock City qualifies, supplying sanitation, roads, lighting (by the Lamplighter’s Guild), police, emergency services (including fire, of course), and even a Department of Mutant Vehicles.

The Black Rock Desert is ideal for fire-based art. As an alkali flat, there is nothing to burn and nothing to break. The ground is a flat expanse of white powder, reverently monikered the ‘‘Playa.’’ It is a rare moment of relief that you’re not aware of the dust on everything and everyone. The Playa is simultaneously a blank canvas for sculpture and a stunning panorama, and being completely fireproof gives artists a little more leeway than they might have in the Guggenheim. The same blank slate applies to Black Rock City as a whole. Not constrained by physical impediments, like being in a convention center or civilization in general, Black Rock City is free to follow mathematical ideals of organization.

Most cities are roughly arranged on grids, but obstructed by rivers, topography, or politics, they rarely achieve graph-paper perfection. The Playa, on the other hand, has no obstructions of any kind. Starting with an empty desert, the festival is built in about a month, and most returning visitors would be surprised to learn that it’s never in the same place twice. Such are the advantages of Black Rock Desert.

Instead of using stodgy Cartesian coordinates, Black Rock City is organized along polar coordinates, making its coordinate system unique among massive fire-themed desert art festivals. At the mathematical origin is ‘‘the Man,’’ the titular carry-through from Burning Man’s historical origins. The angular coordinate is described using time on a clock face, with the city stretching from 2:00 to 10:00, and radially from R = 0:5 miles to R = 1 mile. The innermost ring of the city is ‘‘Esplanade,’’ with each street as you move radially outward given a name beginning with a sequential letter of the alphabet. For example, in 2017 the names were Esplanade, Awe, Breath, Ceremony, Dance, Eulogy, Fire, Genuflect, Hallowed, Inspirit, Juju, Kundalini, and Lustrate. Radial and angular locations in the city are specified using a letter and a time. For example, you might describe your camp as being at ‘‘D and 7:30’’. The large art installations are harder to pinpoint, since they’re scattered throughout the center of the ring and out into the ‘‘Deep Playa,’’ beyond the 10:00–2:00 gap where there are no street signs. ‘‘Center Camp,’’ where all of the official operations, bureaucracy, and public services are located, is a secondary set of ringed roads located centrally at 6:00.

The sudden addition of a new city to the Nevada landscape doesn’t go unnoticed. Although there’s always some grumbling whenever 70,000 people suddenly show up anywhere, capitalism has a way of bringing Burners and the citizens of Reno together. The demographic inside local Wal-Marts shifts precipitously in the days leading up to Burning Man as bikes, water, and food are stripped from the shelves. On the way out, many of the tens of thousands of visitors want a dust-free meal, and all of them want a shower (although some amount of playa dust inevitably makes it onto outbound airplanes).

Burning Man has a ‘‘gift economy,’’ so once you’ve left Reno and entered Burning Man proper, money ceases to have worth. In a barter economy you exchange goods for other goods. In a gift economy you give without the expectation of any return. It ‘‘works’’ because everyone else (or at least a larger fraction than you might expect) is doing the same. The gift economy even covers public transportation; if you see a mutant vehicle with enough space for a person, then you can jump on. I wouldn’t trust it to handle the housing market, but it did supply pancakes three days in a row from three different locations.

Against this backdrop, my good friend and office-mate, Spencer, and I forded out and set up an ‘‘Ask a Mathematician / Ask a Physicist’’ booth to discuss and hopefully answer whatever questions might come our way. We wanted to keep up our half of the gift economy contract, but giving things away is expensive and it’s difficult to find a productive application for Banach spaces and character classes in the middle of nowhere. Our one big skill is talking about math and science, so that’s exactly what we did.

The very first two questions, ‘‘How do I find the love of my life?’’ and ‘‘What is spacetime made of?’’ set the bar for variety and difficulty. With a big enough shoehorn, you can turn anything into math, so we reduced the first question to finding the best of N options, where you only meet the ‘‘options’’ sequentially and can only say yes or no to each. In other words, by ignoring all the nuance of romance we had reduced finding the love of one’s life to the ‘‘fussy suitor problem’’ (more commonly known as the ‘‘secretary problem’’). The optimal solution is to pick a value N, date N / e people (e ≈ 2:718, so this is a little more than a third of the total potential love interests), and then marry the first person you like better than anyone in that first group. But after a few more people joined the discussion, the answer evolved until we eventually settled on this: interact with lots of people, be patient and kind, and what’s wrong with flowers? It may not have a solid mathematical backing, but I bet it works better.

‘‘What is spacetime made of?’’ was actually a lot easier to talk about, but we were surprised by how fast it became philosophical and mathematical. Time and space come up a lot, and we’ve found that the simplest, most solid, and universally disappointing definitions for them are that time is what clocks measure and space is what rulers measure. On that first day, those definitions went over well, but then the conversation swerved into the relationship between the two. Then into how they’re affected by mass and energy. Then into what that says about space and time. There are hand-wavy ways to talk about these things, but no one was satisfied until we had gotten neck deep in spacetime intervals (the spacetime notion of distance, s² = x² + y² + z² - (ct) ²) and parallel transports (a method for describing the curvature of space), and surrounded the booth with equations and diagrams in the dust. The next year we replaced the dust with a whiteboard, and the year after that (after we learned what playa dust does to whiteboards) with a blackboard.

We weren’t always so lucky. One year, before we’d even finished setting up the booth, a cadre of MIT undergraduate physicists stumped us with ‘‘Soap bubbles physically solve minimal surface problems. Are there any other physical phenomena that quickly solve NP-type problems?’’

We had initially assumed that public enthusiasm for physics and especially math would be lukewarm. We expected to get a handful of straightforward, easy questions about fractions and outer space, then break for lunch. Instead, we were a lightning rod for fascinating, wide- ranging, precise inquiries. Rather than talking to rare individuals, we found ourselves holding court with a dozen people at a time, fielding multiple new questions even as conversations about previous questions ruminated around the group and resurfaced. After the first year we invested in a bigger tent.

For every ‘‘How fast is Earth moving?’’ there was a profound debate into the nature of understanding itself followed by a delving inquiry into quantum theory. I found that detail especially surprising. My own research is in quantum information theory, which I was certain would not come up in a tiny tent in the middle of the desert. But it turns out that the big questions that drove me into my field, about entanglement, Schrödinger’s cat, and the endless train of other bizarre things quantum theory implies about our world, are nothing special. A shocking number of self-identified non-scientists have the same questions, and if you advertise yourself as ‘‘a physicist,’’ you’ll hear them. Although I’ll admit to a certain amount of if-you-have-a- hammer-every-question-is-a-nail-ness, it’s surprising how frequently the double-slit experiment (a beautifully weird keystone experiment in quantum physics) meanders into the conversation.

The simplest and often most profound questions tended to come from professional scientists. A researcher from the satellite division at Boeing sparked a heated debate about the relationship between prior probabilities and human bias when he asked, ‘‘Why do weird things happen so much?’’ One innocent question, ‘‘Is there such a thing as a one-half-dimensional space?’’ split the assembled guests into tribes on what ‘‘space’’ meant in this context. On one side were people who thought that ‘‘fractals with dimension one-half’’ was a reasonable way to talk about such a space, and for them, the answer was yes. Another faction believed that ‘‘space’’ clearly means ‘‘a space you could imagine walking around in,’’ and for them the answer was no. More remarkable than the politeness of the debate was how well informed it was and how quickly new arrivals came up to speed.

In retrospect, we shouldn’t have been surprised. While all of the art on the Playa is eye-catching, a lot of it is extremely technical as well: lasers that continuously outline your shadow, sensors that sync the Man’s heartbeat with your own, huge kinetic sculptures, interactive digital art (which is either highly random or unobviously complex), and on and on. A moment’s thought would have revealed that there must be an army of scientists and engineers behind the scenes. Fortunately for us, through our booth we were rapidly introduced to the healthy math, science, and maker communities of Black Rock City. There’s even a ‘‘Math Camp,’’ built in and around a second-order Sierpinski gasket the size of a truck and located (we were absolutely thrilled to learn) at E and 3:14. There they have guest lecturers, an ‘‘open problems’’ board, and tequila shots.

Once you’re aware of the nerdy undercurrent, it’s hard to miss. While walking around in the ‘‘Deep Playa’’ (a long walk from Black Rock City proper) and looking at art, Spencer was struggling to explain the idea behind wiki software to me. The one other person present, a man wearing sunglasses, feathers, and body paint to the strict exclusion of all else, volunteered ‘‘You know, my engi- neering firm uses wiki software all the time!’’ He really cleared up a lot of confusion, but in any other context that would have been an unusual conversation.

The people attending Burning Man come from all over the world, with wildly divergent personal histories. So as often as not, the festival itself is the only common experience two strangers may have. Fortunately, that common experience carries a lot with it. For example, when a question about complex numbers comes up (it always happens at least once), you can get everybody onto the same page by talking about the complex plane in terms of the layout of Black Rock City: the Man is at zero, Center Camp (Esplanade and 6:00) is at i, the main entrance is around - 2i, the 3:00 and 9:00 plazas are at ± 1, and the Temple (a nondenominational, pan-belief, transient, and intentionally flammable structure) is at + i.

And when all else fails, there are very few people at Burning Man who aren’t at least a little enthusiastic about open fires. A detailed conversation about momentum and gyroscopic forces is normally difficult to follow, but if there happens to be a helpful fire dancer nearby, suddenly the theoretical becomes empirical and arguably very difficult to ignore. A fire dancer’s weapons of choice are typically poi: flaming weights at the ends of a pair of chains. Changing a poi’s plane of rotation, without that plane intersecting yourself, requires some subtle angular momentum exchange, and fire dancers, while usually not familiar with the exact vocabulary, are always more than happy to demonstrate. Who can’t learn about torque if the option is getting burned (a little)?

Most of the people who show up at the booth don’t explicitly have questions to ask. At least not at first. For most people the barrier to even beginning a conversation about science and math is way too high. Of course, everyone has things they wonder about, but as any teacher can tell you, most of us are also worried about sounding stupid. It takes listening for a little while to realize that science is something that can be understood, scientists aren’t unusually smart, and there’s no need to impress them. It warms your heart to listen to what other people wonder about and to hear how they think, and nothing makes a science conversation more inclusive than a sci- entist who’s deeply bothered by exactly the same things that bother you.

We were so impressed by the enthusiasm behind the questions, both from the asking itself and the inevitable ‘‘Oh yeah, I’ve always wondered that too!’’ that afflicts practically everyone in earshot, that we set up the booth back in New York a few times as well as online in the form of http://askamathematician.com. We’re still surprised. Ten years on and the questions keep coming.

Reprinted with permission from the Mathematical Intelligencer

When you first take the Artifact, you will see a vision of ALPHANION, Demon-Sultan of the Domain of Order, who appears as a grid of spheres connected by luminous lines. Alphanion will urge you to use the Artifact to enforce cosmic order, law at its most fundamental. He will show you visions of all the most brutal and sadistic crimes of history, of all the wars caused by nations that could not live together in harmony, and he will tell you they are all preventable. He will show you dreams of perfectly clean cities with wide open streets, where everyone earns exactly the optimal amount of money and public transportation is accurate to the second. He will tell you it is all attainable.

But if you hesitate even an instant to take Alphanion’s offer, you will see a vision of CTHGHFZXAY, Demon-Shah of the Domain of Chaos, who appears as a shifting multicolored cloud. Cthghfzxay will urge you to use the Artifact to promote cosmic chaos, the ultimate principle of freedom. She will condemn the works of Order as a lie, a dystopia bought at the cost of true human liberty. She will show you visions of primaeval forests, where no two flowers are alike, where each glade holds a new mystery, where people run wild in search of new adventure. She will tell you it can all be yours.

As you weigh these two offers, you will see a vision of ZAMABAMAZ, Demon-Pharaoh of the Domain of Balance, who appears as a man and woman conjoined. They will tell you that neither Order nor Chaos is at the root of human flourishing, but an ability to strike the right balance between the two. That a virtuous life is one spent in moderation between total wild liberty and a stifling concept of rote rule-following. That Alphanion and Cthfhfzxay are the two poles of the universe, and that righteousness exists in the space created by their interaction. They will ask you to devote the Artifact and its power to the Domain of Balance, so all people can better manage the interaction of Order and Chaos in their own lives.

This will seem reasonable to you, but then there will appear a vision of IYYYYYYYYYYYYYYYYYYYYYYY, Demon-Raja of the Domain of Excess, who appears as a blinding violet light. It will tell you that both Order and Chaos present coherent visions of the world, but that for the love of God, choose one or the other instead of being a wishy-washy milquetoast who refuses to commit to anything. It will tell you that blinding white and pitch black are both purer and more compelling than endless pointless grey. It will ask you to give the Artifact to somebody – anybody – other than Zamabamaz.

Just as you think you have figured all this out, there will appear a vision of MLOXO7W, Demon-Kaiser of the Domain of Meta-Balance, who appears as a face twisted into a Moebius strip. It will tell you that sometimes it is right to seek balance, and other times right to seek excess, and that a life well-lived consists of excess when excess is needed, and balance when balance is needed. It will remind you that sometimes you are a sprinter and other times a tightrope walker in the Olympiad of life, and that to commit to either eternal carefulness or eternal zealousness is to needlessly impoverish yourself. It will ask you to devote the Artifact and its power to balancing balance and imbalance, balancedly.

You will not be the least bit surprised when there will appear a vision of K!!!111ELEVEN, Demon-Shogun of the Domain of Meta-Excess, who appears as a Toricelli trumpet with eyes and a mouth. She says that seriously, pick a side, all this complicated garbage about the balance between balance and excess is just another layer of intellectualization to defend against having any real values, a trick to make you feel smart and superior for believing in nothing, not even Balance. She will ask you to choose something now, lest you be caught in an endless regress of further options.

As soon as you acknowledge that this makes sense, there will appear a vision of ILO, Demon-Chancellor of the Domain of Excessive Meta-Balance, who appears as a deep hole in space whose end you cannot see. They will point out that yes, there is potentially an infinite regress of further levels. But to act to avoid those levels is essentially to unthinkingly side with the principle of Excess over Balance. After all, if you had originally started by siding with Chaos or Order rather than waiting to hear of the existence of Balance, you would have been unknowingly favoring Excess over Balance. And if you had decided to choose either Excess or Balance, you would have been favoring the principle of Meta-Excess over Meta-Balance before even knowing they existed. So choosing at any level of the hierarchy is essentially equivalent to choosing Excess at all higher levels of the hierarchy. When viewed this way, the hierarchy collapses to chaos, order, first-level-balance, second-level-balance, third-level-balance, and so on. They offer a new, better vision: Infinite Balance, a theoretical top of the hierarchy in which you choose to balance all previous levels.

But as you start to consider this, there will appear a vision of PAHANUP, Demon-Taoiseach of the Domain of Balanced Meta-Balance, who appears as a hole in space exactly three inches deep. Ze will tell you that going to infinite lengths to ensure perfect balance at an infinite number of levels actually seems a bit excessive in ways. To choose either Chaos or Order outright would be insufficiently careful, but to give yourself an intractable problem with an endless number of meta-levels would be excessively careful. Ze will suggest seeking balance in the number of levels you seek balance in.

This will seem plausible to you right up until the sudden appearance of a fiery vision of IFNI, Demon-Secretary-General of the Domain of Chaotic Meta-Excess, who appears as static. She will point out that there is now another infinite regress, more difficult than the last – to wit, how long you should spend calculating the number of levels on which to seek balance. She will state her case thus: suppose you want to calculate the correct amount of balance in the universe. Let us call this Calculation A. You need to calculate how long to spend on this calculation before giving up and satisficing; let us call this Calculation B. But you need to calculate how long to spend on Calculation B before giving up and satisficing; let us call this Calculation C. Clearly you will never be able to complete any of the calculations. Therefore in order to avoid spending your entire life in an infinite regress of calculation, you should flip a coin right now and use it to decide either Chaos or Order, no takebacks.

But as you reach for the coin, you will see a vision of GOSAGUL, Demon-Admiral of the Domain of Ordered Meta-Balanced Excess, who appears as a cube with constantly flashing black and white faces. He will lecture you on how it seems pretty strange that, when faced with the most important decision in the history of the universe, you decide to flip a coin. Surely, even if Ifni’s argument is correct, you can do better than that! For example, you can just go a specific finite number of levels, such as three, then seek balance at that many levels, then stop. This will be strictly better than Ifni’s plan of choosing completely randomly.

But this sage advice is interrupted by MEGAHAHA, Demon-Pope of the Domain of Excessively Ordered Meta-Balance, who appear a as pattern of black and white that cycles between a line, square, cube, and hypercube. It will point out that if you’re in the business of accepting arguments along the lines that “it seems pretty strange that when faced with the most important decision in the history of the universe you…”, then it seems pretty strange that when faced with the most important decision in the history of the universe, you agree to a kind of random number of levels chosen by a demon you have no reason to trust. By what logic do you reject making the decision itself randomly, but accept making the decision about how many levels to make the decision on randomly? Any amount of Balance in Meta-Balancing Excess is just arbitrary capriciousness; you either need to act fully randomly, or embrace the entire difficulty of the problem.

At this point, you will remember that the Artifact is cursed and demons are evil. With a final effort of will, you will shout the words “I choose Balance! Just normal Balance! First-level Balance! That’s it!” and throw the Artifact to the ground, where it will shatter into a thousand pieces and the voices of the demonic hierarchy will suddenly all go silent.

And for a thousand years to come, heroes will grumble “Why, exactly, are we seeking balance in the universe? Isn’t that kind of dumb? Don’t we want more good stuff, and less bad stuff? Doesn’t really seem that balance is really what we’re after, exactly.”

And you will tell them the story of how once you found the Artifact that gave you mastery of the universe, and you refused to take more than about three minutes figuring out what to use it for, because that would have been annoying.

The collective intellect is change-blind. Knowledge gained seems so natural that we forget what it was like not to have it. Piaget says children gain long-term memory at age 4 and don’t learn abstract thought until ten; do you remember what it was like not to have abstract thought? We underestimate our intellectual progress because every every sliver of knowledge acquired gets backpropagated unboundedly into the past.

For decades, people talked about “the gene for height”, “the gene for intelligence”, etc. Was the gene for intelligence on chromosome 6? Was it on the X chromosome? What happens if your baby doesn’t have the gene for intelligence? Can they still succeed?

Meanwhile, the responsible experts were saying traits might be determined by a two-digit number of genes. Human Genome Project leader Francis Collins estimated that there were “about twelve genes” for diabetes, and “all of them will be discovered in the next two years”. Quanta Magazine reminds us of a 1999 study which claimed that “perhaps more than fifteen genes” might contribute to autism. By the early 2000s, the American Psychological Association was a little more cautious, was saying intelligence might be linked to “dozens – if not hundreds” of genes.

The most recent estimate for how many genes are involved in complex traits like height or intelligence is approximately “all of them” – by the latest count, about twenty thousand. From this side of the veil, it all seems so obvious. It’s hard to remember back a mere twenty or thirty years ago, when people earnestly awaited “the gene for depression”. It’s hard to remember the studies powered to find genes that increased height by an inch or two. It’s hard to remember all the crappy p-hacked results that okay, we found the gene for extraversion, here it is! It’s hard to remember all the editorials in The Guardian about how since nobody had found the gene for IQ yet, genes don’t matter, science is fake, and Galileo was a witch.

And even remembering those times, they seem incomprehensible. Like, really? Only a few visionaries considered the hypothesis that the most complex and subtle of human traits might depend on more than one protein? Only the boldest revolutionaries dared to ask whether maybe cystic fibrosis was not the best model for the entirety of human experience?

This side of the veil, instead of looking for the “gene for intelligence”, we try to find “polygenic scores”. Given a person’s entire genome, what function best predicts their intelligence? The most recent such effort uses over a thousand genes and is able to predict 10% of variability in educational attainment. This isn’t much, but it’s a heck of a lot better than anyone was able to do under the old “dozen genes” model, and it’s getting better every year in the way healthy paradigms are supposed to.

Genetics is interesting as an example of a science that overcame a diseased paradigm. For years, basically all candidate gene studies were fake. “How come we can’t find genes for anything?” was never as popular as “where’s my flying car?” as a symbol of how science never advances in the way we optimistically feel like it should. But it could have been.

And now it works. What lessons can we draw from this, for domains that still seem disappointing and intractable?

Turn-of-the-millennium behavioral genetics was intractable because it was more polycausal than anyone expected. Everything interesting was an excruciating interaction of a thousand different things. You had to know all those things to predict anything at all, so nobody predicted anything and all apparent predictions were fake.

Modern genetics is healthy and functional because it turns out that although genetics isn’t easy, it is simple. Yes, there are three billion base pairs in the human genome. But each of those base pairs is a nice, clean, discrete unit with one of four values. In a way, saying “everything has three billion possible causes” is a mercy; it’s placing an upper bound on how terrible genetics can be. The “secret” of genetics was that there was no “secret”. You just had to drop the optimistic assumption that there was any shortcut other than measuring all three billion different things, and get busy doing the measuring. The field was maximally perverse, but with enough advances in sequencing and computing, even the maximum possible level of perversity turned out to be within the limits of modern computing.

(this is an oversimplification: if it were really maximally perverse, chaos theory would be involved somehow. Maybe a better claim is that it hits the maximum perversity bound in one specific dimension)

One possible lesson here is that the sciences where progress is hard are the ones that have what seem like an unfair number of tiny interacting causes that determine everything. We should go from trying to discover “the” cause, to trying to find which factors we need to create the best polycausal model. And we should go from seeking a flash of genius that helps sweep away the complexity, to figuring out how to manage complexity that cannot be swept away.

Late-90s/early-00s psychiatry was a lot like late-90s/early-00s genetics. The public was talking about “the cause” of depression: serotonin. And the responsible experts were saying oh no, depression might be caused by as many as several different things.

Now the biopsychosocial model has caught on and everyone agrees that depression is complicated. I don’t know if we’re still at the “dozens of things” stage or the “hundreds of things stage”, but I don’t think anyone seriously thinks it’s fewer than a dozen. The structure of depression seems different from the structure of genetic traits in that one cause can still have a large effect; multiple sclerosis might explain less than 1% of the variance in depressedness, but there will be a small sample of depressives whose condition is almost entirely because of multiple sclerosis. But overall, I think the analogy to genetics is a good one.

If this is true, what can psychiatry (and maybe other low-rate-of-progress sciences) learn from genetics?

One possible lesson is: there are more causes than you think. Stop looking for “a cause” or “the ten causes” and start figuring out ways to deal with very numerous causes.

There are a bunch of studies that are basically like this one linking depression to zinc deficiency. They are good as far as they go, but it’s hard to really know what to do with them. It’s like finding one gene for intelligence. Okay, that explains 0.1% of the variability, now what?

We might imagine trying to combine all these findings into a polycausal score. Take millions of people, measure a hundred different variables – everything from their blood zinc levels, to the serotonin metabolites in their spinal fluid, to whether their mother loved them as a child – then do statistics on them and see how much of the variance in depression we can predict based on the inputs. “Do statistics on them” is a heck of a black box; genes are kind of pristine and causally unidirectional, but all of these psychological factors probably influence each other in a hundred different ways. In practice I think this would end up as a horribly expensive boondoggle that didn’t work at all. But in theory I think this is what a principled attempt to understand depression would look like.

(“understand depression” might be the wrong term here; it conflates being able to predict a construct with knowing what real-world phenomenon the construct refers to. We are much better at finding genes for intelligence than at understanding exactly what intelligence is, and whether it’s just a convenient statistical construct or a specific brain parameter. By analogy, we can imagine a Martian anthropologist who correctly groups “having a big house”, “driving a sports car”, and “wearing designer clothes” into a construct called “wealth”, and is able to accurately predict wealth from a model including variables like occupation, ethnicity, and educational attainment – but who doesn’t understand that wealth = having lots of money. I think it’s still unclear to what degree intelligence and depression have a simple real-world wealth-equals-lots-of-money style correspondence – though see here and here.)

A more useful lesson might be skepticism about personalized medicine. Personalized medicine – the idea that I can read your genome and your blood test results and whatever and tell you what antidepressant (or supplement, or form of therapy) is right for you has been a big idea over the past decade. And so far it’s mostly failed. A massively polycausal model would explain why. The average personalized medicine company gives you recommendations based on at most a few things – zinc levels, gut flora balance, etc. If there are dozens or hundreds of things, then you need the full massively polycausal model – which as mentioned before is computationally intractable at least without a lot more work.

(you can still have some personalized medicine. We don’t have to know the causes of depression to treat it. You might be depressed because your grandfather died, but Prozac can still make you feel better. So it’s possible that there’s a simple personalized monocausal way to check who eg responds better to Prozac vs. Lexapro, though the latest evidence isn’t really bullish about this. But this seems different from a true personalized medicine where we determine the root cause of your depression and fix it in a principled way.)

Even if we can’t get much out of this, I think it can be helpful just to ask which factors and sciences are oligocausal vs. massively polycausal. For example, what percent of variability in firm success are economists able to determine? Does most of the variability come from a few big things, like talented CEOs? Or does most of it come from a million tiny unmeasurable causes, like “how often does Lisa in Marketing get her reports in on time”?

Maybe this is really stupid – I’m neither a geneticist or a statistician – but I imagine an alien society where science is centered around polycausal scores. Instead of publishing a paper claiming that lead causes crime, they publish a paper giving the latest polycausal score for predicting crime, and demonstrating that they can make it much more accurate by including lead as a variable. I don’t think you can do this in real life – you would need bigger Big Data than anybody wants to deal with. But like falsifiability and compressability, I think it’s a useful thought experiment to keep in mind when imagining what science should be like.

Mary Meeker of Kleiner Perkins Caufield & Byers, one of the premier Silicon Valley investors at one of its premier venture capital firms, is leaving her position in an abrupt, high-profile splitting of the firm she helped lead.

Meeker is leading an exodus of late-stage investors from Kleiner Perkins in its most dramatic shake-up since legendary investor John Doerr stepped back from his role more than two years ago. Meeker’s exit — she, along with three of her partners, will form a new firm — will undoubtedly deal a hard blow to Kleiner Perkins, given her high profile in the business community and her stature as by far the most senior woman in venture capital.

Meeker, “the first Wall Street analyst to become a household name” during the dot-com boom, has been most famous in this era for her agenda-setting, unusually thorough “Internet Trends” slide decks, delivered — in memorable, rapid-fire fashion— in recent years at our annual Code Conference. This year’s deck included 294 slides and has been viewed more than a million times; last year’s, with 355 slides, is well past three million views.

The departures of Meeker and her colleagues — Mood Rowghani, Noah Knauf and Juliet de Baubigny — are rooted in different visions for the types of deals they would like to do. But like so many disputes in recent years at Kleiner Perkins, there was also persistent friction between the two sides in ways that had little to do with the firm’s core business — over mundane things such as whether to host a holiday party in San Francisco or closer to Sand Hill Road in Silicon Valley — according to people familiar with the situation.

Kleiner Perkins is one of tech’s oldest firms and its early investments in Amazon and Google have made it part of Silicon Valley lore. That legacy has helped sustain the firm even as it self-admittedly missed out on a wave of generation-defining startups. Meeker’s new team will lose the connection to that brand, effectively placing a bet on the singular brand of Meeker.

“The way the teams operate are just very different,” Meeker said in an interview with Recode.

The principals are naturally downplaying how important a split this is, but this is a massive moment for the firm.

“I don’t think it’s a huge deal,” Ted Schlein, who succeeded Doerr as the de facto head of the firm, said in an interview. The late-stage investing group, which started in 2011, “continued to diverge away from what the core part of Kleiner Perkins has done for 46 years, and will continue to do for another 46 years.”

People familiar with the situation described a distant relationship between the early-stage group, which backs U.S. startups that sometimes don’t have a finished product, and the later-stage group, which can compete with sovereign wealth funds to back companies worth billions of dollars all across the world. The two teams spent fairly little time with one another in recent years, sharing office space, some limited partners, and — critically — a storied brand. But deals? Not many.

Lots of top investing firms have two teams that operate semi-independently, and those firms don’t feel the need to go separate ways. But at Kleiner Perkins, there were cultural clashes between the two squads. The firm has long been home to personal squabbles between its big-personality investors — in-fighting that spilled into very public view during the gender discrimination lawsuit brought by Ellen Pao— and it has been unable to put those to rest even as it elevated a new, likable leader in Mamoon Hamid, who joined the firm last year.

Hamid is a bit of a purist for early-stage investing, though Schlein said the decision was not solely his.

What the decision was, according to the people familiar with the situation, was fairly abrupt. While the two sides have been growing apart for years, they had a “moment of clarity” in the last week or so, as one person said, and the two teams quickly moved to go public with their plans for a split before it leaked. The firm’s investors — its limited partners, in industry parlance — were only informed early this morning.

The breaking point came as the team had to decide whether to raise another later-stage fund under the Kleiner banner or under a new one.

The new, still-to-be-named firm led by Meeker will have immediate cachet given Meeker’s connections on Wall Street, where she served as a top research analyst at Morgan Stanley, and in Silicon Valley. Unlike most big-name tech investors, Meeker has name recognition that goes well beyond the sector because of her annual trends report.

“There is only one Mary Meeker,” Schlein said at the time of her hire.

Schlein now says he is “not naive” about what it will be like to lose Meeker. Kleiner Perkins will have to move on without its most famous name. And after the exit of another investor, Beth Seidenberg, in an unrelated departure, it will have no female general partners. Kleiner Perkins plans to take some economic stake in the new firm.

“People have gotten emotional, not in an angry way,” said one of the people, “but in a poignant way.”

This article is the first in a series dedicated to explaining how Uber leverages forecasting to build better products and services. In recent years, machine learning, deep learning, and probabilistic programming have shown great promise in generating accurate forecasts. In addition to standard statistical algorithms, Uber builds forecasting solutions using these three techniques. Below, we discuss the critical components of forecasting we use, popular methodologies, backtesting, and prediction intervals.

Forecasting is ubiquitous. In addition to strategic forecasts, such as those predicting revenue, production, and spending, organizations across industries need accurate short-term, tactical forecasts, such as the amount of goods to be ordered and number of employees needed, to keep pace with their growth. Not surprisingly, Uber leverages forecasting for several use cases, including:  

• Marketplace forecasting: A critical element of our platform, marketplace forecasting enables us to predict user supply and demand in a spatio-temporal fine granular fashion to direct driver-partners to high demand areas before they arise, thereby increasing their trip count and earnings. Spatio-temporal forecasts are still an open research area.

• Hardware capacity planning: Hardware under-provisioning may lead to outages that can erode user trust, but over-provisioning can be very costly. Forecasting can help find the sweet spot: not too many and not too few.
• Marketing: It is critical to understand the marginal effectiveness of different media channels while controlling for trends, seasonality, and other dynamics (e.g., competition or pricing). We leverage advanced forecasting methodologies to help us build more robust estimates and to enable us to make data-driven marketing decisions at scale.

What makes forecasting (at Uber) challenging?

The Uber platform operates in the real, physical world, with its many actors of diverse behavior and interests, physical constraints, and unpredictability. Physical constraints, like geographic distance and road throughput move forecasting from the temporal to spatio-temporal domains.

Although a relatively young company (eight years and counting), Uber’s hypergrowth has made it particularly critical that our forecasting models keep pace with the speed and scale of our operations.  

Figure 2, below, offers an example of Uber trips data in a city over 14 months. You can notice a lot of variability, but also a positive trend and weekly seasonality (e.g., December often has more peak dates because of the sheer number of major holidays scattered throughout the month).

If we zoom in (Figure 3, below) and switch to hourly data for the month of July 2017, you will notice both daily and  weekly (7*24) seasonality. You may notice that weekends tend to be more busy.

Forecasting methodologies need to be able to model such complex patterns.

Prominent forecasting approaches

Apart from qualitative methods, quantitative forecasting approaches can be grouped as follows: model-based or causal classical, statistical methods, and machine learning approaches.

Model-based forecasting is the strongest choice when the underlying mechanism, or physics, of the problem is known, and as such it is the right choice in many scientific and engineering situations at Uber. It is also the usual approach in econometrics, with a broad range of models following different theories.

When the underlying mechanisms are not known or are too complicated, e.g., the stock market, or not fully known, e.g., retail sales, it is usually better to apply a simple statistical model. Popular classical methods that belong to this category include ARIMA (autoregressive integrated moving average), exponential smoothing methods, such as Holt-Winters, and the Theta method, which is less widely used, but performs very well. In fact, the Theta method won the M3 Forecasting Competition, and we also have found it to work well on Uber’s time series (moreover, it is computationally cheap).

In recent years, machine learning approaches, including quantile regression forests (QRF), the cousins of the well-known random forest, have become part of the forecaster’s toolkit. Recurrent neural networks (RNNs) have also been shown to be very useful if sufficient data, especially exogenous regressors, are available. Typically, these machine learning models are of a black-box type and are used when interpretability is not a requirement. Below, we offer a high level overview of popular classical and machine learning forecasting methods:

Interestingly, one winning entry to the M4 Forecasting Competition was a hybrid model that included both hand-coded smoothing formulas inspired by a well known the Holt-Winters method and a stack of dilated long short-term memory units (LSTMs).

Actually, classical and ML methods are not that different from each other, but distinguished by whether the models are more simple and interpretable or more complex and flexible. In practice. classical statistical algorithms tend to be much quicker and easier-to-use.

At Uber, choosing the right forecasting method for a given use case is a function of many factors, including how much historical data is available, if exogenous variables (e.g., weather, concerts, etc.) play a big role, and the business needs (for example, does the model need to be interpretable?). The bottom line, however, is that we cannot know for sure which approach will result in the best performance and so it becomes necessary to compare model performance across multiple approaches.

Comparing forecasting methods

It is important to carry out chronological testing since time series ordering matters. Experimenters cannot cut out a piece in the middle, and train on data before and after this portion. Instead, they need to train on a set of data that is older than the test data.

With this in mind, there are two major approaches, outlined in Figure 4, above: the sliding window approach and the expanding window approach. In the sliding window approach, one uses a fixed size window, shown here in black, for training. Subsequently, the method is tested against the data shown in orange.

On the other hand, the expanding window approach uses more and more training data, while keeping the testing window size fixed. The latter approach is particularly useful if there is a limited amount of data to work with.

It is also possible, and often best, to marry the two methods: start with the expanding window method and, when the window grows sufficiently large, switch to the sliding window method.

Many evaluation metrics have been proposed in this space, including absolute errors and percentage errors, which have a few drawbacks. One particularly useful approach is to compare model performance against the naive forecast. In the case of a non-seasonal series, a naive forecast is when the last value is assumed to be equal to the next value. For a periodic time series, the forecast estimate is equal to the previous seasonal value (e.g., for an hourly time series with weekly periodicity the naive forecast assumes the next value is at the current hour one week ago).

To make choosing the right forecasting method easier for our teams, the Forecasting Platform team at Uber built a parallel, language-extensible backtesting framework called Omphalos to provide rapid iterations and comparisons of forecasting methodologies.

The importance of uncertainty estimation

Determining the best forecasting method for a given use case is only one half of the equation. We also need to estimate prediction intervals. The prediction intervals are upper and lower forecast values that the actual value is expected to fall between with some (usually high) probability, e.g. 0.9. We highlight how prediction intervals work in Figure 5, below:

In Figure 5, the point forecasts shown in purple are exactly the same. However, the prediction intervals in the the left chart are considerably narrower than in the right chart. The difference in prediction intervals results in two very different forecasts, especially in the context of capacity planning: the second forecast calls for much higher capacity reserves to allow for the possibility of a large increase in demand.

Prediction intervals are just as important as the point forecast itself and should always be included in your forecasts. Prediction intervals are typically a function of how much data we have, how much variation is in this data, how far out we are forecasting, and which forecasting approach is used.

Moving forward

Forecasting is critical for building better products, improving user experiences, and ensuring the future success of our global business. It goes without saying that there are endless forecasting challenges to tackle on our Data Science teams. In future articles, we will delve into the technical details of these challenges and the solutions we’ve built to solve them. The next article in this series will be devoted to preprocessing, often under-appreciated and underserved, but a crucially important task.   

If you’re interested building forecasting systems with impact at scale, apply for a role on our team.

Subscribe to our newsletter to keep up with the latest innovations from Uber Engineering.

Ludovic Henry, Miguel de Icaza, Aleksey Kliger, Bernhard Urban and Ming Zhou September 11, 2018 runtime

  Stack Transition: 
      ..., value1, value2 -> ..., result 

Temp3 := ADD Temp1 Temp2

We are barraged by constant distractions, and they are degrading the quality of our work. Our digital society now is set up to allow us to focus for mere minutes at a time, since we are in an attention economy and the sole objective of companies is to capture more of our time. Facebook, Google, and Snapchat are all incentivized to get us to look at our phones many times a day.

Distractions permeate everything, even at work. GitHub has notifications for so many things that if you have work and personal projects on the same account, you will get unrelated notifications all the time. Our employers set us up with ping-pong tables, open offices, and Slack, the open-office of chat tools.

With all these distractions surrounding us and with all these notifications, we are expected to get deep work done. Personally, I cannot. And you cannot, either.

When I’m highly distracted, I’m prevented from entering flow. To my core, I’m a maker. I get such a thrill from making things that are usable and useful, and these distractions cut through that in a way that makes it impossible to have a productive, fulfilling day.

Flow is important if you want to get anything meaningful done. Context switching takes a lot of effort and time and you can only do it so many times in a day. If you are constantly distracted, you will never enter flow and you will never have great, innovative ideas.

If you never concentrate and go deep, you will produce bad work: you will produce bugs, and fail to debug them; you will create security issues; you will cause performance problems; and you will architect things poorly.

You cannot live in a vacuum: you have to talk to users and stakeholders and your teammates and your manager. But that should be done on your schedule, not on theirs (most of the time) so that when you are done talking to them and you have a good idea of what to build, you can go crank out a high quality first version. This version will be on the right track, technically: good architecture, usable performance, well-tested, with minimal bugs. This is a first iteration you can go take to users to get concrete feedback and keep iterating.

Our attention is being squandered and we have an opportunity now to reclaim it. Fight back. Get rid of the ping-pong table; delete Facebook and Snapchat; disable push notifications for emails; build some walls to establish real offices. Setup processes on your teams to give people large chunks of time where they can go deep, for days at a time. Put walls or even hundreds of miles between your employees. Embrace flow, and get some work done. You’ll feel better, I promise, and what you produce will be better as well.

Motivation

At Makery we are huge fans of Kotlin, and we always get excited when anything new comes up about our favorite language. You can imagine that we were pretty hyped after last year's KotlinConf when JetBrains introduced Kotlin Native for multiplatform projects. Writing Kotlin code for iOS was a dream for us and the reality is even better: we can create a shared layer with the business logic and use it on both major mobile platforms. Maybe this is the right cross-platform solution we were waiting for so long? After this point, it was just a matter of time when to start experimenting with MP projects for Android and iOS.

In this article, I'm going to guide you through the labyrinth of Kotlin Multiplatform (Kotlin MP or MP from now on) and show you how we created a simple example app to test out the capabilities of the technology.
Also, I highly recommend to check out the project's source code on GitHub because not all the details are covered in this article, just the most interesting ones.

GitHub example

First, I needed a feature set realistic enough, so we can say: "If it's working with this example it will probably work in production too."

Usually, a baseline for any application is to be able to communicate with the outside world. In a developer's head, this means the capability to fire network requests and process the received data. So with these guidelines in our mind, that's what we came up with:

1. Authenticate with Basic HTTP Authentication to GitHub
2. Fetch the list of the user's repositories
3. Deserialize the result
4. Display it in a list view

The aim is to implement the first three points from the list as a shared library. Anything outside of that (the UI layer basically) will be written with the platform native SDKs (in Swift for iOS and Kotlin for Android).

Project structure

We started our research by looking for documentation about how we should structure our project. We found a pretty good description here with a simple example app.

This tutorial by JetBrais suggests the following directory structure for an MP project:

project/
├── androidApp/
├── iosApp/
└── shared/
    ├── common/
    ├── android/
    └── ios/

The androidApp and the iosApp folders are for the native applications receiving the compiled artifacts from the shared module.

The common folder in the shared directory should have most of the implementation of the desired business logic and the android and ios directories are there for the platform specific code parts.

Let's look at all these in detail!

Shared module

The shared module is literally a Gradle module with every component we need for our business logic.

The main idea is to look for a library which can help us with HTTP requests and another one to deserialize the received data. Also when we talk about I/O, threading is always something to consider with care.

Fortunately, we can find all these components ready to work with Kotlin MP.

We will use ktor as our HTTP client and kotlinx.serialization for processing/mapping request and response entities. And finally, Kotlin coroutines is JetBrains solution to handle asynchronous computation.

These libraries already support Kotlin Native (which is necessary to work on iOS) and they also work well with Android.

The next step is to wire these parts together and implement our business logic.

Common module

The common module is the heart of our MP project. It's the place where all the magic happens. After setting up everything correctly in our Gradle build files (they are written using Gradle Kotlin DSL) we can start using our libraries to implement our GitHub client.

An important thing to mention here is that you only have access to the Kotlin Stdlib and the libraries you included (and they should support MP projects of course).

class GitHubApiClient(..) {
..
fun repos(successCallback: (List) -> Unit,
          errorCallback: (Exception) -> Unit) {
    launch(ApplicationDispatcher) {
        try {
            val result: String = httpClient.get {
                url {
                    protocol = URLProtocol.HTTPS
                    port = 443
                    host = "api.github.com"
                    encodedPath = "user/repos"
                    header("Authorization", "Basic " + 
                    "$githubUserName:$githubPassword".encodeBase64())
                    }
                }

           val repos = JsonTreeParser(result).read().jsonArray
                    .map { it.jsonObject }
                    .map { GitHubRepo(it["name"].content,
                                      it["html_url"].content)}
                  successCallback(repos)
            } catch (ex: Exception) {
                errorCallback(ex)
            }
        }
    }
}

That's it. Wasn't that hard, right? Let's take a look at the interesting parts:

• line 3-4 The repos() function needs two callback function for handling the success and the error case.
• line 5 We start a coroutine with launch{} to push work on to a background thread. But where does the ApplicationDispatcher come from? We defined it in the common module: internal expect val ApplicationDispatcher: CoroutineDispatcher. The expected keyword means that concrete implementation should come from the platform-specific modules which we will check out later.
• line 7 Our httpClient instance is created by using Ktor.
• line 8-16 We define the target url and encode the user credentials in the Authorization header according to the HTTP Basic Authentication protocol.
• line 18-21 Parsing the result looks more manual than what we are used to but currently this is the only solution working with Kotlin Native.

After we finished with the business logic, we need to write the concrete implementation for our expected ApplicationDispatcher variable mentioned above. We have to do this in the platform specific modules (android, ios).

Android module

The android MP module is basically an Android library project. Kotlin's main target was to work well with Java so in the case of Android, there are not a lot of tricks. Our only job here is to provide an actual implementation of the ApplicationDispatcher.

It looks like this:

internal actual val ApplicationDispatcher: CoroutineDispatcher = DefaultDispatcher

Because in this module we have access to the Android SDK and the coroutines API for Android (unlike in the common) we can use the DefaultDispatcher.

iOS module

In case of iOS we have to work a little bit more on our actual implementation:

internal actual val ApplicationDispatcher: CoroutineDispatcher = NsQueueDispatcher(dispatch_get_main_queue())

internal class NsQueueDispatcher(private val dispatchQueue: dispatch_queue_t) : CoroutineDispatcher() {
    override fun dispatch(context: CoroutineContext,
                          block: Runnable) {
        dispatch_async(dispatchQueue) {
            block.run()
        }
    }
}

We create a new class called NsQueueDispatcher which is basically there for to provide a dispatching method for coroutines on Kotlin Native (which is currently behind of the coroutines implemementation on other platforms, using a single-threaded (JS-style) event loop solution).

The next step is to check out how can we integrate the product artifacts of Kotlin MP into our native mobile applications.

Android app

For Android, as it was in previous cases, the work with Koltin MP is hassle-free. You just need to create a regular Android application in Android Studio.

After the initial project setup we will use a Gradle trick called composite build. This feature allows us to access other builds and in this case, build artifacts.

To enable composite build I inserted this line into androidApp/settings.gradle:

includeBuild("../")

In the root project level build script we have already defined the group and the version property:

allprojects {
    group = "com.adrianbukros.github.example"
    version = "0.1.0"
  }

Now we have access to the .aar in the android app module, generated by Kotlin MP.

To include it in our application, simply just add as a dependency:

dependencies {
    ...
    implementation("com.adrianbukros.github.example:android:0.1.0")
}

Now we can easily instantiate and use our GitHubApiClient class:

GitHubApiClient(username, password).repos(
                successCallback = {
                // handle the result
                }, errorCallback = {
                // handle the error
        })

iOS app

Let's add our Kotlin MP library into a native iOS application! In the case of iOS, the story is more complicated than it was with Android. What the Kotlin Native compiler does is basically grab our Kotlin code from the common and ios modules and compile it into an Objective-C framework.

What we have to do is to include this framework in our Xcode project. Here are the steps:

1. Add a new framework to your project by clicking File/New/Target/Cocoa Touch Framework and call it Shared.
2. On the project settings page, select the Shared library below targets and go to the Build phases tab.
3. Delete all the default build phases except Target Dependencies.
4. Hit the + at the top left of this view and add a New Run script phase.
5. Open the new phase and insert the following lines:

"$SRCROOT/../gradlew" -p "$SRCROOT/../shared/ios" "$KOTLIN_NATIVE_TASK" copyFramework \
-PconfigurationBuildDir="$CONFIGURATION_BUILD_DIR"

What this script will do for you is call a piece of Gradle code, which will copy the Obj-C lib from the Kotlin MP build folder and insert it into the Xcode build folder for e.g.: /Users/{user}/Library/Developer/Xcode/DerivedData/github-multiplatform-example-bqzqpuwrnlaafjgkqfwustgrqnwa/Build/Products/Debug-iphonesimulator (the $CONFIGURATION_BUILD_DIR marks this).

This way Xcode will have access to our freshly build Kotlin Native library at compile time.

Wait a minute! Do we have the copyFramework task available in Gradle by default? Of course not. Let's add these lines to shared/ios/build.gradle.kts:

tasks {
    "copyFramework" {
        doLast {
            val buildDir = tasks.getByName("compileDebugFrameworkIos_x64KotlinNative").outputs.files.first().parent
            val configurationBuildDir: String by project
            copy {
                from(buildDir)
                into(configurationBuildDir)
            }
        }
    }
}

One thing is still missing. We have to define a $KOTLIN_NATIVE_TASK environment variable's value. Also on the Shared framework target, select Build Settings and hit the + on the middle of the screen, then select Add User-Defined Setting. If you want the build to work with every target, you have to add the following values:

Now after building your project, you can import the Shared module into your Swift files and use it like this:

import Shared

class LoginViewController: UIViewController {
    @objc private func loginButtonDidTap() {
           GitHubApiClient(githubUserName: username, 
                           githubPassword: password).repos(
            successCallback:{ [weak self] repos in
                // handle the result
                return StdlibUnit()
            }, errorCallback: { [weak self] error in
                // handle the error
                return StdlibUnit()
        })
    }
}

• line 1 You have to import the Shared framework before you start using it in your class.
• line 5 Initialize the GitHubApiClient and call the repos() function on it with a success and an error closure!
• line 9 and 12 Currently there is limitation in Kotlin Native that if a function doesn't have a return value, you still have to return with StdlibUnit(). I guess it's just one thing that can be omitted in the future versions of Kotlin Native.

Conclusion

Koltin Multiplatform for mobile applications is no question an experimental technology. I think there is still a long way until we can say that it's production ready, but my opinion is that it is a great thing.
There is plenty of room to improve the framework and the tooling, but it's amazing that even at this stage, we were able to develop this small example app. I'm really positive about the project's future and I'm sure that I will give it another try a few months later (maybe after KotlinConf '18?).

In their new book Building the Cycling City: The Dutch Blueprint for Urban Vitality, Melissa and Chris Bruntlett use the example of the Netherlands to show how a cycling culture promotes community building and health.

C: I think that's an indication of how you build your streets. If you build hostile streets, people are going to want to keep up with car traffic and armor themselves up with protective equipment. There’s a differentiation in the Dutch language between a sports cyclist and a utilitarian cyclist; the two phrases loosely translate to “walking with wheels” versus “running with wheels.” The “wheeled walkers” make up the vast majority of people that bike in the Netherlands because they've created these conditions that aren't as hostile, so anybody feels like they can do it.

Another point you make in the book that's so important is about the different kind of bikes Dutch cyclists ride.

M: They're upright, they’re a little bit slower but they're meant for utility. They're meant to get them comfortably, without any complication, from point A to B, hauling some goods along the way or hauling children. Those utility bikes mean a lot in terms of simplifying the trip. They don't overcomplicate it. The bikes already come with all the gear, you don’t have to worry about buying lights or a bell separately. They're meant for day-to-day transportation.

Why is the bike shop such an important part of the cycling environment?

C: In Vancouver, the cycling shops were still very sport-focused. The staff weren't trained to sell bikes, they usually only have one or two collecting dust in the corner. Because the vast majority of people riding bikes in North America are doing it for sport and recreation, the retail industry is still playing catch up. It's almost become this chicken and egg scenario where they don't see a large market for transportation bikes so they're not putting many resources into developing that market. Bike sharing has kind of changed this a little bit, because people are riding these more upright utilitarian bikes. But if they ultimately want to invest in one, they have a real job on their hands trying to find one.

In the book, you point out that a very small fraction of Dutch riders wear helmets, but the rate of injury and death from cycling is much lower.

M: It's not even a part of the conversation in the Netherlands because they've engineered their streets to take out a lot of the possible stresses and risk of collisions that would inherently make people feel like they need the extra bit of safety. Less than 1 percent of the population in the Netherlands actually wears helmets, because they've got the investment in the safe infrastructure and safety in numbers.

How much do the protected lanes matter, or are there other elements of the infrastructure that are of equal or more importance?

M: In North America, a lot of the times we talk more about protected and fully separated infrastructure. But in the Netherlands, the conversation is actually much more about traffic calming. A lot of their streets don't have speeds over 30 kilometers [approximately 18 miles] an hour. On neighborhood streets, they build in surface treatments that inherently force you to slow down, like laying cobblestone or narrowing the street. Also, because more people bike there, drivers have a more empathetic approach toward cyclists.

Why do you think we have this mentality that the car is more important than a cyclist or even a person?

C: It's been a product of 60 years of post-war planning and propaganda from the automobile industries that streets are for one purpose: moving cars from A to B. Before the Second World War, streets in a lot of Dutch cities were places for connection and community and commerce. Then the post-war planners came along pushed all those public functions into parks and private spaces. The Dutch resisted that urge to modernize their city around the car, so they kind of have this 40 to 50-year head start on us.

You see cycling as a way of connecting diverse groups of people —not just hard infrastructure, but an element of what Eric Klinenberg calls the social infrastructure which binds people and communities together.

M: On a bike, you inherently have to make a physical connection with people. In a car, you're separated by glass and steel. But when you're out on a bike, you can actually see everybody, you can say hello to the people that you meet along the way. The side of the road becomes a place to reconnect or have a quick wave in the morning to brighten your day. That's how we need to see our streets—as places for connection as opposed to just a place to pass through.

There's a great chapter in your book titled "Not Sport. Transport." You make the point that riding a bike can connect more people to public transit.

C: The prevailing understanding is that the bicycle doesn't replace the car. Neither does the public transit system. But by combining bikes with trains and buses, and trams, suddenly you've got this game-changing seamless transportation network that can get you from door-to-door often quicker than a car, with a little bit of exercise and social connection. In the Netherlands, that means providing bike parking and infrastructure that leads to the transit stop, and then providing a last-mile solution like bike sharing or rental on the other side of the trip. That ultimately can be used as a strategy to reduce congestion in our cities.

What about electric bikes now that they are growing in use?

About the Author

Richard Florida

Richard Florida is a co-founder and editor at large of CityLab and a senior editor at The Atlantic. He is a university professor in the University of Toronto’s School of Cities and Rotman School of Management, and a distinguished fellow at New York University’s Schack Institute of Real Estate.

An interactive music theory dashboard, including a circle of fifths for guitarists.

Follow @mikehadlow on Twitter for news and updates.

Watch the video:

Read the blog post, What I learnt creating guitar dashboard.

By Mike Hadlow. Open source on Github.