Examining “Software Eating the World”
I recently read a blog post by Ethan Kaplan which argues that
…the world itself is now possible to represent by polymorphic functions with refactorable logic.
This is what we mean by “software eating the world.” Out of personal experience, technologists can often get excited by the application of logic to all of life’s decisions. So, I’d like to consider the two parameters along which this is occurring:
- We are identifying new, previously un-parameterized areas of life to replace with software.
- We are improving on preexisting functions.
The first is intuitive. 10 years ago, only a human could recommend a song to you. Then Pandora began writing algorithms to do it. In the past decade or so we have begun solving this for finding restaurants, movies, taxis, and even love. There are probably still many untouched fields to explore here (if I could think of them I’d build it!). As an investor, this happens to be the most exciting kind of opportunity to come across.
The second parameter begs an interesting question: What are the upper bounds to quantifying every variable? You can often see them directly in the product experience. Searching for a car online (to continue Kaplan’s central example) takes a list of inputs: make, year, price, horsepower, etc. But what if you get inside the car and the driver’s seat doesn’t go back far enough? Or the steering is surprisingly loose for your taste? Or just that, in person, the car doesn’t look quite as sexy as you’d hoped?
I think some of these variables may be inherently unquantifiable. Others disagree. But either way, some of the functions we use may already be approaching the limit^ of an inherent tradeoff given the current computing form factors we use: UX vs. time. Even if it were feasible for Cars.com to ask me all those questions, it’s much more human to just drive the car for 5 min. So, until the next technological breakthrough, how much more effort will we put into improving these functions?
At this point in history “software is eating the world” for only those parts of life that are most easily quantifiable.
:Commodities are particularly interesting, as they remove all human preference and are thus best served by a function.
Disaggregation of a Bank
There are lots of cool charts thrown around the startup/VC world. One of the best is the “Disaggregation of Craiglist,” originally created by Andrew Parker in 2010 and updated by David Haber in 2012 (shown below):
USV has recently invested in a host of internet-enabled financial services, which prompted the question: what would a “Disaggregation of a Bank” look like?
I’ve started putting together a list of startups providing services that would traditionally be offered by a bank (broadly defined). Please contribute if you have any startups or categories that I’m missing!
I took a first stab at visualizing this in the following deck. Past the Craigslist slide, I tried to diagram out the functions of a bank. Most everyone at USV found this a bit confusing, but, on the next slide, it allowed me to show where companies are innovating. I recommend going full screen to take a look (or feel free to download the PDF):
Any thoughts? Ways to improve? My hope is that this leads to larger questions: Which banking functions benefit from economy of scale and remain defensible vs. which areas are ready for disruption? Should startups be focused on consumer, business or both at once?
Best Google Search Ever
This is an anecdote that got me thinking about how the UI/UX of search will continue to evolve.
I recently was in London for Techstars and ended up eating at a small cafe. I had an excellent eggplant/couscous dish (I’m clearly not a chef). A couple weeks later, a few friends were coming over for dinner and I wanted to recreate that dish.
I couldn’t remember the name of the restaurant, anything resembling a street address, or even what the dish would be called.
But I still had a mental image of how I got to the restaurant. So my search was: - Start Google Street View at the St. James Theatre, where Techstars Demo Day was held. - Walk myself through the streets of London using Google Street View towards the restaurant. It’s kind of amazing to explore a city this way. - Found the restaurant! (I was actually disappointed that I couldn’t Street View my way inside). Now I knew the name to search for. - Foursquare brought it up, and I found a link to their menu. Wasn’t the exact recipe, but I was able to make a passable replica knowing the major ingredients.
The point is, we still use search in mostly two capacities: type in a phrase or look at a bird’s eye view map. Yes, Google Street View has been around for a while, but I’d never used it in this way before. But it was perfect for my brain: I don’t remember names, but I have an excellent visual memory. Everyones memory works differently, and I think search will grow to be able to handle this. Someone else might have remembered the decor in the restaurant, or the color of the sign outside. Or the music that was playing. Or the fact that the bathroom was downstairs. I think all these little things, the oddball remnants in our memory, will help us complete searches like this in the future.
Thoughts on Quantified Self
USV has been interested in the quantified self space for over a year now. As investors in “networks enabled by the internet,” we’ve seen the majority of the innovation so far as being outside our wheelhouse: hardware products and apps in specific verticals. And so while we have not made any investments yet, we’ve met with many entrepreneurs and had enough internal discussion to share some of our thoughts.
Quantified self = personal data collection Wearable = device that captures data (except your smartphone) Quantified applications = apps that either collect (with or without a wearable) quantified self data from a user and/or sit on top of the quantified self data from other sources
The form factor of the mobile phone will determine the wearables market
The past two years has seen an explosion in fitness-based quantified self devices such as the Fitbit, UP and Nike Fuelband. This mass adoption is even more remarkable given that these devices all track the same few metrics around activity: steps taken, calories burned, hours and quality of sleep.1
Yet this data can be captured by a mobile phone utilizing the same array of sensors2 found in a wearable device. Moves and Runkeeper are currently leading the pack in app-only products. Battery life and developers’ utilization of that battery life will continue to improve. And Apple’s recent announcement of the M7 chip, a processor dedicated towards making quantified self data available to app developers, shows that mobile phones are making the current class of wearables redundant. Current mainstream devices will have to innovate to remain relevant.
The quantified self device industry must shift towards data that is impossible for a mobile phone to capture. At the bottom of this post is a partial breakdown of the quantified self data that is and will be captured by devices and apps. It’s unlikely that mobile phones will be the form factor to, say, capture brain waves while you sleep, so there is opportunity for wearable products in that specific vertical.
Data tracking will become more medically relevant
While the current mainstream devices are focused on fitness, next-generation quantified self devices will enable medically relevant data. The list of next generation wearables (at the bottom of this post) shows potential for data in bloodwork. And these companies are mostly in the Kickstarter/IndieGoGo phase of development, potentially only a year or two away from being available. Similar advances in mainstream, big-brand wearables won’t be far behind.
Hopefully, this data can be shared with doctors to improve preventative health.3 It’s possible that this will be a lightweight interaction, such as showing your doctor your accumulated physical activity as part of an annual physical. Or, it could become an integral part of the medical ecosystem, our vitals constantly being monitored and our doctors alerted automatically when necessary.
APIs and data aggregation
The biggest winner in this space will be the API-driven platform that creates a stable ecosystem to aggregate and interface between data sources. Developers will have access to larger pools of data. There may also be powerful social networks built on top of this data. Barriers to entry will be particularly high if this platform player figures out how to be HIPAA compliant and become the point of contact between patients and medical professionals, a particularly difficult barrier for fledgling teams. It’s hard to overestimate the value of medically relevant data coming from user engagement in a network lightweight consumer apps.
Signals in the Quantified Self Industry
The original list is here. It includes links to companies and will evolve over time. Feel free to contribute!
Mainstream (i.e. in Fitbit, Fuelband, UP, Moves, Runkeeper, etc.)
- Calories/Distance: tracking how many steps you take and how active you are
- Sleep: tracks amount and quality of sleep, some products also time alarms to your sleep cycle
- “Active” tracking: requires you to manually input data. A suite of apps have been built for fitness regimes and dietary tracking.
Available, but not mainstream
- Heart rate: measuring your pulse
- Withings Activity Tracker is a little clip-on that fulfills that mainstream categories but also has a built in optical pulse tracker. Requires the user to hold their finger up to the device. They also make a blood pressure monitor.
- Azumio has an app for measuring pulse using a phones native camera and flash.
- LUMOBack monitors lower back posture and pelvic alignment. They’ve built a waistband with inertial sensors.
- Asthmapolis has built a bluetooth sensor to monitor the environment, alerting the user of conditions that could trigger an asthma attack
- Spire has a yet-to-launch clip-on device to monitor breathing, showing people their respiratory patterns and helping monitor stress and emotions.
- Tinke tracks and helps with breathing exercises.
- Advanced Fitness
- Nike Hyperdunk is the next generation of in-shoe sensors. Designed specifically to track and analyze your basketball game.
- Active Mind Technology is soon to launch golf specific sensors and analysis after a successful fundraise on AngelList.
- Ovuline uses machine learning to help couples conceive faster. Works with some of the mainstream wearables technology.
- 23andMe is the leader in affordable DNA sequencing and analysis.
- Microbiome: the ecological community of microogranisms that live inside of you
- uBiome provides a $89 kit to “quantify your gut” and compare it to other user data sets. Dozens of health conditions are correlated with the microbiome — from asthma to diabetes, autism to depression, irritable bowel, Crohn’s, chronic sinusitis, heart disease, and more.
- Sano Intelligence is building the “API for the bloodstream,” made possible by their proprietary transdermal blood chemistry sensor patch
- Sproutling is creating a baby monitor for the quantified self age. Will track temperature, heart beat and sleeping position from an ankle band that wirelessly connects to the parents mobile phone.
- Melon is a headband to measure and improve focus. Completed a Kickstarter fundraise for their initial product.
- Muse is another brain sensing headband, available for preorder now. It picks up brainwaves in real-time and sends them to your phone, tracking this data and even turning them into commands!
- Ballistocardiography: measures mechanical forces of the heart, gives data on heartbeat (cardiac contractions), respiration (chest wall movements) and movement
- Beddit is building a sensor that lies on your bed and picks up on these forces while you sleep. Available for preorder.
- In vivo sensors
- Nutritional analysis (without having to manually input everything you eat)
- Mainstream blood glucose monitoring (as is currently the norm for diabetics)
- Integrated into clothing
- And much, much more…
That there is so much consumer interest in a product with so little power is indicative of how exciting the space will become. ↩
An accelerometer, gyroscope and compass are industry standard in smart phones. ↩
My personal realization came when my grandfather was recently admitted to the hospital for the second stroke in two weeks. Doctors were adjusting his blood pressure medication, causing it to fluctuate. But the only data points they get are when it’s too late, once he’s already had the stroke. It’s insane to me that our current medical system is not using a wearable device to monitor his blood pressure in real-time, monitoring the effects of medication outside of the hospital as well. A simple algorithm could easily have provided early warning signs as his condition worsened. ↩
The best ingredients are fresh. So fresh that they’ve just been picked. How do we get to a world in which most of our food can taste this good? There are two obvious possibilities:
- Grow food everywhere. While there are plenty of small scale efforts to grow crops in urban areas, this is inherently unscalable.
- Speed up the supply chain. Most crops are held by wholesale buyers for long periods of time at warehouses, sacrificing quality for market opportunity as prices fluctuate. Even with vast improvements in the system, it’s not physically possible to transport produce from farm to table on the time scale I’m envisioning.
The answer is a combination of the two, the product of thinking so far out of the box that you end up in one: grow food as it moves through the supply chain.
Freight Farms has figured out how to grow crops inside a shipping container in any external environment. Each 40 foot by 8 foot container is packed with produce growing under energy-efficient LEDs regardless of whether the container is held in a lot, is traveling on a flatbed, or has arrived at its ultimate destination. And while their focus is not on consumer yet, it’s possible to imagine a futuristic Whole Foods made up of a series of these containers. Need carrots? Walk into the appropriate container and pick a few out of the hydroponically nourished soil. Suburban homes could have a container in their backyard. Freight Farms could even periodically airdrop containers onto the roofs of urban apartment buildings.1
I’ve spent the past month doing a deep dive into the agricultural supply chain. There are a couple of talented teams creating online marketplaces to disrupt the middleman-heavy status quo. And while Freight Farms is outside of this model, and not a USV company, it has been one of the most fun to think through the implications of the world that they’re envisioning.
The future is full of awesome helicopters/drones making stuff like this a reality. And most urban rooftops are totally wasted space anyway. ↩
Connecting on LinkedIn is pointless gamification
LinkedIn is an network of interconnected professional personas. I’ve already argued that the current connection model makes for a useless professional address book. This post is about why connections themselves are an arbitrary hindrance on the value (to consumers!) of the LinkedIn graph.
Part of the idea behind being on LinkedIn is the possibility that someone interesting will see my profile and reach out to me. It might be someone I met briefly at a conference, someone who just Googled my name, or, hopefully, a recruiter with my potential dream job that I didn’t even know existed.
And my profile is exactly “who” I want them to find: my resume, education, skills, accomplishment and a picture of me wearing a tie.1 This is my professional persona. So the question is: who do I NOT want to be able to access this information?
The answer is no one. The expectation value of benefit to me increases the more open to the world that profile is.2 Yet for anyone to see my profile, they have to be a “second-degree” connection to me. For me, and most other people I’ve talked to, that interaction only occurs with people I’ve met in real life.
On the flip side, there are plenty of people’s resumes I’d be interested in seeing, but who I either am not personally acquainted with or don’t want to bother with a pointless LinkedIn connection. For example, I’d love to see Reid Hoffman’s career track, to understand how he got to where he is today, without bugging him about it.
Why, then, the connection model? It’s a classic case of a network constraining information to earn money. There are no connections on the backend: so long as you pay LinkedIn enough money you get free rein to troll your way through the entire platform. This is only feasible as a business expense, of course, so just recruiters and companies get to see the whole lay-of-the-talent-land.
Constraining information that does not want to be constrained is an inherently unsustainable business model for an internet company.3 The ease of creating a LinkedIn profile used to be a major technical advantage. But it’s increasingly easy to put up your resume online. There are alternative services specializing in specific verticals. Most members of the tech community just throw up their own website. And Google works just as well, possibly even better, than LinkedIn’s internal search.
As an aside, a byproduct of the connection model is a brilliant reinforcement of LinkedIn’s network effects. Connecting keeps us coming back to LinkedIn, remind us of it’s value and reinforcing the idea that LinkedIn is “the” platform for our professional personas. All while providing us with only a tiny fraction of its value of the platform. So the next time you make a connection on LinkedIn, remember why you even had to do it in the first place.
Citi Bike’s Failure of Information
Citi Bike in NYC is great, when it works. Plenty has been written about the programs engineering struggles. But even more astounding, and ultimately more detrimental to the program, is the shoddy quality of information in the program.
Above is a picture from my phone of the Citi Bike app on Saturday night. I had ridden a Citi Bike back to the East Village from Midtown. All of the Citi Bike docks shown are supposed to have many free spots, represented by being only “partially full” in dark blue. All of the ones I’ve circled in red were, in fact, completely full. It ended up taking me longer to ride around searching for a spot in vain than it would have taken for me to have just walked.
Besides being extremely frustrating for a consumer, it’s much more dangerous to Citi Bike itself. For example, the program is spending significant money and effort trying to dynamically balance the allocation of bikes. How can they do this when all their data is plain wrong?
An open letter to a new VC analyst
The following are my thoughts on sourcing deals after having worked at USV for (gasp!) just over a year now. I will focus on run my own process, separate from the day-to-day collaboration with other members of the investment team. While every firm is different, there is hopefully something in here that will be useful. The underlying lesson is to try and identify the weak spots in your firm’s process and then fill them.
There are two types of sourcing: reactive and proactive. Reactive sourcing is fairly self-evident: troll AngelList, TechCrunch, Hacker News, VCDelta, and any other source that fits your sector; go to events, such as demo days, to meet promising companies; take as many (reasonable sounding) pitches as walk in the door; and, of course, VC is a networking business, so meet as many people as possible.1 Most of the value here is adding bandwidth to your firm.
My lesson from this approach is that building companies takes longer than expected. As an analyst, I started off talking to companies that were not mature enough for an investment from us. While these are all teams that would otherwise not have been in contact with someone else on our investment team, it has taken much longer than I expected to be able to bring these companies to the partnership. And, unlike a partner, I will have a much shorter life at the firm, so I unfortunately can’t be as patient.
Proactive sourcing is much more interesting. These are initiatives you can come up with, an intellectual stance for why should something should exist (and it might not yet!). Come up with an idea or a sector and dive in. If that thesis resonates with the firm, go deeper until you find a company that fits. At least at USV, the partners talk about their best deals coming out of such work. As an example, I’ve been fairly deep into Quantified Self for the past month.
Also try to identify a signal that is currently being missed by your firm. After testing out various analytics projects, I couldn’t find it became clear that tracking pageviews, app downloads, etc. at our stage of investing was not meaningful enough. But since Series A is typically the second round of publicly disclosed funding, we’re in the perfect position to use prior VC investment as a signal for good companies. And on top of this I can layer analytics. For example, given the size of the team and the date and amount raised previously it’s quite easy to estimate when they’ll be raising their next round. Once I get a handle on this signal, I should be able to at least get a quick look at every company that otherwise wouldn’t walk in the door.
My 10% project for the past few months has been to build a Django app that integrates with CrunchBase, AngelList and VCDelta. The code for the entire project will remain private, but I’ll start sharing snippets to help others build their own tools. The first is an update of a Python library for the CrunchBase API.
I highly recommend breakfast meetings every day. Let’s you work through lunch and stay focused. ↩
Music Education Technology
Innovating music education is something I’m passionate about. As both a musician1 and an investor I’ve been following a number of exciting music education startups. I’d like to give an overview of what those companies are doing, then put that in the context of the history of music education and add some of my own thoughts, as well as how I’ve approached building my own (very rudimentary) music education software.
There are two primary technological advancements that many of the following companies are utilizing. The first is the proliferation of mobile tablet devices, the ideal computing form factor for consuming content, whether it is sheet music, audio, or video (and it fits perfectly on a music stand!). The second is pitch detection: a whole suite of apps now use pitch detection to follow exactly what a student is playing on his/her own, non-digital instrument.
The first identifiable group of startups are focused on digitizing sheet music.2Tonara is a pure port of sheet music to the iPad, plus added pitch detection to enable automatic page turning and the ability to scribble notes. In conjunction with Hal Leonard, Tonara includes a digital sheet music store. Chromatik takes this further by adding a social network on top. Their thesis is that learning music should be inherently social and collaborative. Music can be uploaded from the Web, annotated, recorded and, crucially, shared with fellow students, teachers and bandmates. Such a model could fundamentally change the social aspect of learning music.3
To me, the most exciting group of companies are innovating new content and curriculums for teaching music. SoundSlice is building a community around syncing videos and guitar tabs.4 Users upload or link a video of a performance and transcribe the keys and tabs. This is then viewable by all other users who can, if necessary, slow down and/or loop the music/videos to help them learn. Instinct takes a similar approach, albeit pairing scrolling guitar tabs with a instructive animations showing which strings and frets to use. The web app also has great pitch recognition, The team at Instinct has also begun to put together an early curriculum for beginning guitarists. It clearly begs the question, How well can you learn guitar just using technology? A great spin on this idea is the gTar, a digital guitar that requires an iPhone for it’s brains/speakers. This approach allows for some really radical changes to the instrument: because strumming each string creates a digital signal which is processed by the iPhone in real-time, you can begin to learn a song just by playing the correct string and not worrying about the fretboard! The gTar software will still output the appropriate pitch, of course.5
There are also two content companies focused on jazz accompaniment. They are the natural evolution of the famous Aebersold recordings,6 accompaniment tracks used by aspiring jazz musicians since the 1960s. To this, (Tutti)[http://tuttiplayer.com] has added high quality videos and recordings of professional musicians, lead sheets, looping/slowdown and the ability to select only certain instruments to play with. Tutti strongly believes in the power of learning technique by watching professionals play. On the other end of the spectrum, iRealb has built a simple app that generates MIDI accompaniment from lead sheets. The accompaniment itself is unsurprisingly predictable, but also sounds quite good for just a MIDI file. The genius of iRealb is its network of users, who have used the in-app tool to input thousands of standard pieces.
Finally, there is the application of basic marketplace models to the music education vertical. For example, (TakeLessons)[http://takelessons.com] and Opus are marketplaces for local music teachers. And while I can’t find a company focused on this emerging vertical, music teachers are increasingly turning to Skype to provide long-distance music lessons.
Music is sound. An obvious statement, but I think it should be at the heart of anyone considering the possibilities of how technology can change music education. For the first time in history, the average music student is in possession of an electronic device that can provide atomic sound. By this I mean such small, specific amounts of musical notes that the average human ear can start to grasp and learn from them.
Consider a musical prodigy, a Mozart. 100 years ago, she could go to the symphony and pick out exact musical phrases, harmonies, etc. that she could replicate in her own playing. For the past few decades, highly talented individuals could replay records or CDs over and over until they reached the same understanding. Now, we can write computer programs that can teach just a simple phrase, or even just focus on a single note. These programs can be responsive, recognizing your aural progress and adapting the curriculum as needed. They can fill in the blanks, so to speak, providing a beginning musician with enough aural backup to make the experience fun and rewarding.
As an example, a beginning piano student is quickly asked to decipher sheet music, translate this into difficult muscle movements and, on top of it all, infuse some musicality. Imagine, instead, a musical “Simon Says:” the student’s iPad introduces two notes, C and G to play. It then plays a measure using just these two notes, next asking the student to repeat. With an accompaniment track, the student can switch off measures with her digital instructor. This particular intro lesson teaches rhythm and pitch but most importantly it teaches the student to play music that’s in her head. She’s not reliant on muscle memory, or a visual memory of sheet music, or any other construct that she has used to compensate for a lack of aural ability. The greatest musicians are also the best listeners; why shouldn’t we use new technology to help young musicians learn by listening?
There are many markets to be explored in music education. My thoughts above focus on a particular pain point, one that I have personally explored: a classical music education restricts an individual to the particular set of pieces he or she can sightread or has spent countless hours perfecting. There are different problems for jazz, rock or other styles of music. And the largest markets are for the amateur musician, even the beginning adult musician, for which a victory is simply learning their favorite song. But for those interested in the particular set of problems I have outlined, feel free to ping me or check out some very rudimentary code I have written for myself personally.
A brief background: I’ve been a classically trained pianist my whole life, although I didn’t take it seriously until I was 16. I even studied at conservatory in Shanghai for a gap year before college. Now, I’m mainly learning jazz, a whole different animal. Most exciting, I found out yesterday that I’ll be accompanying two brilliant opera singers (friends from college) at Carnegie Hall in November. ↩
Example: as a kid taking piano lessons once a week, there was basically no social component to my musical education until band started in middle school. ↩
While not yet fully launched, this model has the potential for strong network effects, quite similar to what iRealb has built to power their app. Also, YouTab has built a similar product which, while more fully flushed out, suffers from a clunky Flash UI/UX. ↩
The next step in the progression is using the light-up fretboard, as early guitarists don’t know their way around the instrument quite yet. ↩
It’s truly amazing how useful these recordings are, and how only now, 50 years later, are we just starting to improve upon them. The Aebersold recordings are the basis for my current foray into learning jazz. ↩
Navigating Venture Capital
Yesterday I was lucky enough to give a seminar at Princeton’s East Coast Startup Summit. The summit is for a few hundred students interested in entrepreneurship. So, drawing on my own personal experience of having absolutely no idea how VC worked when I was a student, I decided to give a talk on how to effectively navigate the fundraising process. Rather than focus on a history lesson of VC, I went straight into helping them understand basic financial concepts, the angel/accelerator/seed-stage/mid-stage/late-stage equity investing landscape, what VCs are looking for, etc. One of our newest portfolio companies, Science Exchange, was nice enough to let me include some slides from their pitch deck as an example of how to pitch your business. At least in my research, there are surprisingly few presentations like this available online. Hope it’s helpful!