Why Technologists Keep Guessing Wrong On the Outcome of Technologies

An article in the Wall Street Journal yesterday highlighted that Uber and Lyft haven’t lived up to their promises.  Ride sharing was supposed to make traffic better but study after study has shown that it has made traffic worse.  The thing no one is talking about is, this is a bigger trend in tech – believing something will be good and it turns out bad.

Remember how the Internet was supposed to save democracy by making every citizen well educated and encouraging civil discourse?  I don’t feel like we live in that future.  Remember how social media and smartphones were going to connect us and help ease loneliness?  The data says people are lonlier than ever.  Remember how blogging was going to free the Internet from control of the big media companies and highlight the truth?  Instead, we are in a a world of spam and clickbait and are now begging traditional media to fact check everything for us.  And remember Wikinomics and Wisdom of Crowds?  The trends they reference changed things, just not necessarily in the ways we thought.

How did tech get everything so wrong, and not just wrong but almost entirely opposite of how it really turned out?  And why does it keep happening?  

The simple answer is – human beings continually miss second order effects.  The truth about ride sharing is that “all else being equal” ride sharing probably would indeed reduce congestion.  But all else isn’t equal.  When ride sharing started, it changed behavior patterns for public transit, walking, biking, etc.  Humans need to be trained in complex systems to understand that the world is dynamic and adaptive and nonlinear.  I joined the Board of the New England Complex Systems Institute because I am very passionate about this problem.  It’s one few people understand.

When it comes to AI, we are making many predictions, as a society, about the impact it will have on many areas of life.  The only thing I know for certain is that we are likely to be very very wrong, particularly if we don’t consider second order effects.  So how do you think about this appropriately?

When thinking about “will AI put us out of work?” It isn’t enough to ask if AI will take jobs – it will.  You have to ask how people will respond when AI takes jobs.  Humans will continue to do what is in their best interest, so, the way the job loss happens matters.  It’s path dependent.  Losing blue collar jobs vs white collar jobs vs American jobs vs foreign jobs first may all have different impacts. 

There are lots of predictions about AI, and it isn’t helpful to go through them one by one, but, the broader issue is that whatever we are predicting, we are most likely to be wrong if we don’t think about possible second order effects.  Tech has a bad history of this, and it is to our benefit not to repeat it.

Beyond AI: Why I Joined The Board of NECSI

Several years ago I read the book “Logic of Failure.” The book is about how poorly humans deal with complex systems. To give you a simple example, when I was in college I had a roommate who didn’t understand how a thermostat worked. There is a time delay between changing the thermostat and changing the temperature of the apartment, and he didn’t really understand that, as a result, you need to use small increments, not large movements, in changing the thermostat settings. So, he spent a lot of time moving it up 5 degrees, getting too warm and moving it down 7 degrees, etc.

I’ve been fascinated with complex systems and the problems they cause since those days. Humans are mostly terrible at dealing with complex systems because we are trained to look at the entities in a system more than the interactions between the entities. The whole field of engineering is about breaking down a large problem into smaller problems, until they get small enough you can solve one, then building it all back up.

Complex systems are very common in the world, yet we humans continue to make mistakes when dealing with them. So I took a course a few years ago taught by Yaneer Bar-Yam at the New England Complex Systems Institute. It was so interesting that I started getting more involved with NECSI, and became friends with Yaneer.

As an entrepreneur and investor focused on AI, I’ve spent a lot of time thinking about what is next. One of the things that is next is the complex systems mathematics that NECSI uses to solve problems. I had the chance to see first hand how complex systems mathematics beat existing AI techniques to solve supply chain problems, financial problems, and more, and so when NECSI decided to start spinning off more research for commercial purposes, I was very interested.

After a bit of discussion, I agreed to join the board of NECSI and help advise on some of the technology commercialization. It fed both my personal interest of complex systems, and my professional interest in learning about what the future may hold beyond artificial intelligence.

If you run a company dealing with complex systems (most likely your company does), you should think about getting some formal training for your key executives. Learning to think about second order consequences, and other key complex systems skills, is really important in today’s economy. And if you have problems that current AI methodologies can’t always solve, I hope you will reach out and see if NECSI can help.

Startup Marketing In 2020

I’ve done two enterprise software startups, and the biggest change between them was the shift in the customer acquisition landscape. In the early days of Backupify (2009 – 2011), if you were a company doing SEO, lots of content marketing, re-targeting, and social media marketing, you were cutting edge and not everyone did those things. You could still arbitrage keywords or some marketing channels. But by the time Talla started marketing to customers (late 2016), everyone was good at all of those things. Marketing platforms make it all so easy now that it was hard to hack marketing to acquire customers on the cheap.

The other problem is that no new customer acquisition channels have emerged in the last 10 years. The last real “innovations” (if you can call them that) are Account Based Marketing and lead capture with chatbots. Those might make your marketing linearly more efficient, but they aren’t game changers.

On the consumer side, it isn’t quite as bad because at least the Instagram influencer channel is relatively young and some startups of the last few years have been able to take advantage of it, but, customer acquisition costs have gone up for consumer companies too.

I was reminded of this today when I read Fred Wilson’s post on the topic. I wanted to write about it because most of my investing theses are around robotics and machine intelligence but I have one that isn’t – that B2B marketing is broken and needs to be fixed. I’m going to take a different approach than Alex Iskold did in his post. The customer networks he references are definitely rising as a thing. (This is part of why I invested in Capiche) But I think the fundamental problem is deeper. I think how companies buy and sell software needs to change. This process of pounding people with calls and emails from SDRs is terrible for everyone, but, it’s exactly what we all do because it’s still the best strategy.

For a clue to what the new world will look like, I want to reference the MidMarket CIO Forum. We used to go to these at Backupify. The conference flies in a bunch of CIOs for free, throws a party, has some keynotes and sessions from influential peers, and then makes them sit in small groups through 6-8 vendor pitches over 2 days. When I asked the CIOs why they came to a thing that forced them to listen to pitches, they all gave roughly the same answer. “Well, I do need to buy stuff, and I do want to know what is out there and what is new, I just don’t want 30 phone calls and 60 spam emails everyday – I want a more controlled process.” I think the secret to solving the customer acquisition problem is to flip the buying process on it’s head and somehow give more control to the buyer to manage the process rather than the seller to force them through it, and I’m excited by new emerging ideas like Vetd, which has a real shot at solving it.

But we still have some time, probably 3-4 years, before a new model of buying takes hold. If you are a startup that hasn’t hit major scale yet, what do you do in the meantime? Here is my advice.

  1. Focus on one channel, whichever one is best, and make it ultra efficient. You are trying to find one channel where you can eke out some gains others can’t, and one way to do this may be hyperfocus.
  2. Try partnerships. For the last decade, I’ve discouraged startups from looking for customer acquisition related partnerships early because getting those deals in place is slow and difficult for startups. But now that all customer acquisition is slow and difficult partnerships earlier in the life of a startup sometimes make more sense, particularly when big companies are looking for innovative ideas.
  3. Consider channel sales. While channel sales can also take a long time to get off the ground, and require just as much, or more work than direct sales, in a world where customer acquisition has become so expensive, it makes sense to consider it.

If you are starting a new company, customer acquisition is a bigger challenge than ever before. I’m already seeing more companies with Business Development roles earlier in the company lifecycle, and more partnerships signed in the first year, but the whole customer acquisition ecosystem (particularly for B2B) still needs to change. If you are an entrepreneur looking for a problem to work on, this is a good one to consider.

Defining The Services-As-Software Business Model For AI

My angel investment in Botkeeper has been one of the most influential in my thinking on how AI strategy is evolving. When new high impact technologies come along, they often shake up status quo business models and because no one understands what business models might emerge on the other side, it’s a wonderful time to make a few bets on some startups. As an investor, these initial bets help me learn how the space is evolving, which means when the real wave of startups comes that are embracing this new tech, I’m much more educated than most people who sat out the initial round. And on top of that, sometimes you get lucky on the early bets too.

An example that I give some times during talks is how the web changed graphic design. Pre-Internet, most companies had an established graphic design process to crank out brochures and flyers and such. When the Internet first arrived, they used that same process. The designer created a graphic of a web page, which was then turned over to a developer to turn into HTML. But the Internet isn’t paper, it has hyperlinks and dynamic text and so you can’t just tack on Internet to a paper-centric graphic design process and get a good outcome. Yet companies do that today with AI. They take their existing workflow and try to tack some AI on at one step and thing that’s how it works.

Botkeeper is one of the first companies to really incorporate machine learning in a more economically viable business model. The CEO is a phenomenal entrepreneur who realized before almost everyone else that there were many work tasks you could automate if you had an example data set to do it. At inception, it had a bunch of accountants who did all the tasks, and most investors I spoke to who looked at the company passed because, in their words “it’s just services.”

What they missed was that, by watching what the accountants did, Botkeeper automated away more and more of the tasks. Gross margins moved from services gross margins to software gross margins, and they grew revenue without adding operational headcount because of increased automation efficiency from machine learning.

The services-as-software business model has a lot of interesting characteristics and some of them may surprise you. I expect this to be a type of business model that grows quite a bit in this decade, as machine learning systems make it possible to tackle types of task automation that weren’t possible before.

The defining characteristic of a services-as-software business model is that you take a task previously done by humans and keep the interface the same, but strip out everything behind it and make it an algorithm (if the workflow is digital) or robotic (if the workflow is physical). Most robotic companies that are going to follow this model today are actually software and model training companies using off the shelf robotic parts so, if the business use robots but resembles a software business economically, I still consider it to fit into this model.

When people first hear about the services-as-software business model, they miss a lot of nuance. They typically say “oh so it’s about cost savings via automation.” No. What makes it such a powerful business model is that historically, companies had to choose between quality and price when building their business models. If you want better performance at some step, you have to spend more on that step. Services-as-software business models have the rare ability to improve quality while lowering costs.

Here are some properties of the services-as-software business model:

Consistency – Right now, your experience of a haircut, for example, varies place to place because it is done by humans. Having a robot do it means that it could be done exactly the same way in NYC and in Tokyo if the robots/algorithms are connected to each other. Consistency is enabled by the network that software can build that humans can’t. Contrast that with something like CafeX, where you can get the exact same cup of coffee made the exact same way even if you are on the other side of the world at a CafeX machine.

Accuracy – For many tasks currently done by humans, the human can only absorb so many examples and anomalies in their work lifetime. A machine can absorb many times more. So Botkeeper, for example, seeing the expense classification of so many more companies than a human bookkeeper ever could, has a 2% error rate versus an 11% error rate for humans.

Personalization – A machine providing a service can know and learn more about you than a human can, in most cases, meaning a much more personalized service.

TAM Expansion and Abundance – By lowering the cost of a service using machines, we make it available to more people, thus improving quality of life and expanding the market opportunity simultaneously.

These characteristics imply that services-as-software businesses drive both cost savings and performance improvements simultaneously, which is unusual.

Here are a few possibly non-obvious points about services-as-software.

  1. Companies miss the opportunity because the change in workflow usually filters up and down the value chain more than they anticipate, and requires them to change other steps in the process, which they don’t want to do. It isn’t as simple as “insert software at step 3 in a 7 step process.” It usually requires you to revamp all steps so that step 3, a machine learning step, has the biggest impact. It requires employees at other steps to label data, evaluate model outputs, and correct the system.
  2. In general, I believe you want to leave the initial interface to the consumer the same, and strip out everything behind that for automation, but there are exceptions. CafeX, for example, is different than having a Starbucks that is mostly robots but 1 cashier and 1 person filling the bins the robots use to make and serve coffee. The choice of whether or not to keep the existing services interface in tact will have other business model implications. For example, CafeXs can be put in more places, at smaller price points, than automated Starbucks can if the rest of the Starbucks store remains the same.
  3. Customer acquisition in services industry is often based on geo, and could lead to expensive switching costs. It may be more cost effective to buy a business than acquire a customer. Going back to our haircut example… say a robot does a haircut for $20, but acquiring a consumer who has a regular hair stylist costs $60. You could probably buy a hair salon for not much more than 1x revenue, possibly less, an acquire the customers that way, instead of 3x revenue it takes to steal them away.
  4. Service-as-software businesses may require full stack business models. For example, Squirrel AI, the educational software startup in China, opened learning centers instead of selling their software into existing learning centers. It’s probably because of point #1 above – the existing learning centers wouldn’t adapt their other workflow steps to make best use of AI automation.

The trick here for an entrepreneur is to figure out which spaces are ripe for services-as-software and which aren’t. My guess is haircuts are not a good one, but accounting is. I think you have to look for situations where the value provided is not at the interface level of the service, and where there is maximum complexity behind the interface that could benefit from automation. You also have to maximize the opportunities for personalization and consistency of service because relying solely on cost cutting will be of limited market value compared to true services-as-software models.

I’ve been thinking about this services-as-software model for a few years now, and have invested in several companies pursuing it, some of which I can’t yet mention. I finally had enough clarity to write about it, but it is still being proven out by the market so, feel free to challenge these ideas or iterate on them – I’m sure there are many key insights I am still missing. And of course if you are building a services-as-software company and are looking for investment, I hope you will reach out.

Why I Became a VC and Joined PJC

Today I’m excited to announce that I’ve officially joined PJC as a General Partner. I’ve thought about becoming a VC for a long time. In fact, back in 2011 I was talking to my head of Sales at Backupify and we were discussing “doing what you love.” I talked about how what I really loved was business models. I loved the complex systems aspect of business – how go-to-market and product and customer segment and team all have to line up in the right way, and all the different levers you can pull to influence the business and the market. “That sounds like you want to be a VC” he said. I didn’t realize it at the time, but he was right.

When Backupify got acquired I was 38, and so was a good profile to become a VC. I like to write and network. I had founder/CEO experience, and had a good exit. And I was young enough to do it for a long time. Several firms approached me but it didn’t make sense at the time. So I decided to start Talla and do some angel investing with my friend Todd Earwood, through an entity we formed called Half Court Ventures.

Talla took much longer to get to product/market fit than Backupify, and in the process, I learned that I don’t really like the CEO job. At Backupify, I liked learning the CEO job, I didn’t actually like the things you do in that role. It’s a lot of culture building, a lot of people-stuff, and I’m a mediocre people manager.

On top of that, to be a good CEO, you have to have to like that feeling of walking into the office and being THE guy/gal that everyone wants to talk to. You go to an event and everyone wants your attention. You have a press release and everyone wants to talk to you. While I am not shy, I am a heavy introvert, and I get energized from reading, writing, thinking, etc, so, the social/political burden of being CEO, the people stuff, it weighed on me. I’m happy to be moving to a role where I can play supporting cast instead.

But what I learned making 74 angel investments with Todd over the past 5 years is that I really really loved the investing side. I often have unique insights about markets and as an entrepreneur I could act on one of those insights. As a VC I can act on many of them. And being a natural contrarian, very comfortable going against the crowd, VC is a better fit. As a CEO, trying contrarian tactics is a difficult thing to get a team to swallow. As a VC, betting on a company you believe in that everyone else doesn’t can get you a big win.

This summer, I decided it was the right time to hire a CEO for Talla, and was able to find an excellent candidate (stay tuned for the announcement). I talked to a few firms but, PJC was the right fit for a few reasons. First of all, PJC was a small firm that wants to build. It piqued my entrepreneurial nature to grow something, and meant I could go in as a legitimate partner and do deals, not be a junior partner who had to spend years proving myself.

Secondly, PJC is one of those Boston based firms that does deals everywhere. A lot of firms are focused just on Boston, or just the East Coast. I like the flexibility of finding the best deals anywhere. About half of my angel deals are west coast, so I am excited to maintain an investing presence there.

Third, PJC is a very flexible firm that will do many types of deals. Some VC firms have very narrow ownership targets and deal criteria. While we do have those at PJC, the range is a lot broader than most funds.

VCs get a bad rap, for a lot of reasons. And while I too have had some bad experiences with VCs, I’ve also had some fantastic ones, and I hope I can be one of those helpful VCs, rather than one that is draining to work with.

I’m excited to start this new chapter, and so if you are building early stage AI or B2B companies, please reach out.

I also want to thank all the people that helped me get here. There are way too many to mention but a few who stand out are my partner at Half Court, Todd Earwood, who was always free to discuss and debate business ideas (he is joining PJC too as a operating partner). Jamie Goldstein from Pillar taught me a lot about investing by letting me be a venture partner there for 2 years. David Orfao, Rich Levandov, Brady Bohrmann, and Rudina Seseri were all mentors of sorts who taught me pieces about investing. Dharmesh Shah and Jason Calacanis were instrumental and supportive of my early angel investing and gave me great advice about how to go about it. Jeff Bussgang was one of my biggest advocates years ago when I moved here from KY in 2010, and helped me get plugged in to the Boston ecosystem. Kent Bennett and Eric Paley were helpful as I thought through the transition from Operator to Investor. Phil Beauregard always challenged me, pushed me, debated me, and showed me some good deals along the way as I was angel investing. Vivjan Myrto started investing in AI even earlier than I did so meeting him convinced me I wasn’t crazy to focus on AI back in 2015. And of course I want to thank David Martirano and Matt Hayes for having me join them in their partnership.

I think I’ve also knocked heads with everyone on that list above at some point, and so, have been very lucky to have the kinds of supportive people who tolerate strongly opinionated people around them. So thank you all for your help getting me to this point, despite that fact that I have aggravated the hell out of all of you at times.

I’ll be writing more about the specific theses I have and where I am investing at PJC, in addition to a final advice post for entrepreneurs that I want to write before I am gone too long from that side of the table. Stay tuned.

Mistakes That VCs Cause Founders To Make

VCs get a bad rap, but I’ve always enjoyed talking with them. They see a lot of stuff, so, they are great sources of market information. It’s true that they sometimes screw up companies, but I don’t think it is in the way most people assume. It’s rarely that they are greedy, forcing you to take more money than you need or selling a company that shouldn’t be sold. I think the mistakes are of a different category – advice that lacks business context. Founders often take this advice too seriously, and often have their own biases that are stupid, just like the VCs, but let’s look at some examples.

Channel Partnerships

Founders, particularly first time founders, often have this same idea that one massive channel partnership will make their business suddenly jump to hypergrowth. I’ve seen it across my investments too. “Company X wants to sell our product, and even though we are a 5 person company, they think it could $20M in revenue next year for them.” I’m sure somewhere, at some point, this has worked, but usually it’s death for the startup. Why? Because the startup runs out of money before Company X ever makes a sale. A deal like this might take 6-9 months to finalize even if Company X really is into it. Investors see this over and over, and every founder thinks their situation is unique and so, as an investor it gets old explaining why this is a bad idea.

That said, there is a flip side to this. Investors often overly favor the last go to market model of their best portfolio exit, which is stupid because GTM really needs to match your market and buyer and product. But I believe investors overindex on direct inside sales. They say things like “own your own destiny” which is entirely true unless your buyers like to buy through an existing channel. I could write an entire post on what to look for to figure this out but suffice it to say VCs often recommend you sell direct and build a brand even when it’s a bad idea.

Luckily, back in my Backupify days, we had a great Board member who gave us good advice on channel sales. In our first year of rolling out a channel program, we signed up 65 channel partners and they were 3% of our revenue. I told the Board at the end of the year we were shutting down the program. He said “wait, has any partner sold anything?” Three partners had. He said “fire the other 62 and focus on those 3 for a year.”

It seemed stupid to me. If 65 partners can only sell 3% of our deals, how will 3 sell anything? But those 3 sold 9% of our deals the next year. We were too distracted with too many partners, and so we learned, with these 3, how to really do a channel deal. Then we started to expand slowly and when we sold the company, the partner channel was 30% of our revenue.

Buyer Personas

Warren Buffett always says “never mistake precision for accuracy.” I think this happens with personas. Investors always want to hear that “we target B2B companies 100 – 500 employees that use MySQL and sell to the head of I.T.” It’s so crisp and clear. And sometimes, you can do that. But often in new markets, you don’t know where the adoption is going to come from, or sometimes the market will support a very horizontal product, or the buyer profile is psychographic instead of a demographic, but investors like to push you to a clear demographic profile.

At Backupify, we had 9000 customers when we sold, and our largest customer paid us $360K per year, and we had hundreds that bought one seat and paid us $30/year. What mattered wasn’t size or industry, it was whether or not you thought backup was important. We sold to “the broccoli eaters” as we called them. There are people who always wear their seatbelts, floss their teeth, file their taxes in March, and eat their broccoli. Those people also love to backup their data. That was our market. Investors often push entrepreneurs to “focus” when it’s too early to actually do so.

Hire The Stud

This is probably the worst mistake investors make – pushing entrepreneurs to hire some marketing or sales or engineering stud who led XYZ product at ABC big company. It’s almost always a bad idea. That person is almost never the right fit, but every first time entrepreneur feels pressure to do it. It’s compounded by the fact that as a first time CEO, you feel constantly incompetent because of everything you don’t know, and these people seem to have a lot of answers, but, they almost never work out. They can’t deal with your stage and size. Investors though, feel a lot better if they are around.

My point in writing this is to give some examples of specific areas VCs influence founders in bad ways. I also want to show that it isn’t that VCs are morons (although some are), but that they are giving reasonable advice that is just often out of context. It might be great advice at a different stage or size.

Synthetic Social Media Via AI

One of the favorite parts of modern video games, for many people, is designing their avatars.  I’ve been thinking about this because I’m looking for AI to create an area I call “synthetic social media.”  But so far I haven’t seen any companies in this space.  Since it doesn’t exist, I don’t know exactly what shape it will take but let me try to explain what I mean.

In Japan, there is a synthetic pop star.  In many immersive video games, like Second Life, people use them as a way to escape and be someone else in a virtual world.  I think AI will allow us to combine these two ideas and create “friends” who act the way we want and do the things we wish they would do.  

Imagine a platform that allows anyone to create a virtual pop star.  You design your avatar, give it a name, and then AI can write the music.  You could say you want it to be 40% Rolling Stones, 10% Taylor Swift, 20% Beastie Boys and 30% random other mix, and the AI can write a song that sounds like that blend.  Now the avatar can perform concerts or shows for your friends, singing the songs it wrote.  The most popular songs and concerts (and thus avatars) develop followings.  And it isn’t just in music.  This could work for art or video content – anything digitally created by humans that AI could step in and perform reasonably well at.

Now it gives you the chance to be friends with, or be the manager for, this synthetic AI start on a social media platform that will be mostly synthetic identities.  If you aren’t talented enough to sign and play music, your avatar can.  You can live vicariously through it, and many people will.  Longer term, you could see the idea expanding into sports, and even into the workplace to some extent through synthetic ai workers.  It’s out there but, it feels like a logical extension of where AI is going, and the time is right to set the seeds in motion for phase 1 of such an industry.  

If you are building something along these lines, please reach out as I’d love to invest. 

Does Customer Development Really Matter For Startups?

I was chatting with someone the other day about customer development and I asked a simple question. Now that customer development has become so common among venture backed tech startups, has it moved the needle at all on outcomes? I don’t have any hard data but, it seems to me that just as many startups fail now as did before. And, anecdotally, in my own angel investing (74 companies) it seems that the more obsessed a founder is with customer development, the less likely they are to build a big successful company.

I want to explore a hypothesis today, one that some of you will hate, but will make others of you nod your head and agree. I hypothesize that customer development hasn’t really worked out all that well, and has led to many small outcomes and companies that won’t die, and thus doesn’t fit the venture model. This, in turn, has led to the worship of the visionary entrepreneur and is why we sometimes get bad situations like WeWork was a few weeks ago.

When people say “customer development” they mean a few different things. I mean the more formal version of it as laid out in Steve Blank’s book Four Steps To Epiphany. I think Steve was actually on to something, but like most ideas, it was summarized and bastardized until it became generic and mostly wrong.

The problem is, talking to customers has some value, but it’s not a panacea. Customers often don’t know what they want, or they want a problem solved in a way that is stupid, wrong, not scalable, or not repeatable. Listening too much to customers often gets you stuck in a rut where you do a bunch of services work. Or more often, it gets you stuck in a local minimum with a handful of customers that have your specific problem and desire your specific solution.

Blank’s book actually gets you out of this – he has 4 major steps, and one of them is “customer creation” but, it seems like 95% of what is written about customer development is just about making sure you talk to customers before you build anything.

Here is my hypothesis.

  1. Customer development was bastardized from a really nice smart process into “interview customers before you build anything.”
  2. As entrepreneurship became “cool” circa 2011 or so, and everyone wanted to start a company, there were too many interviews with people who wanted niche solutions, but these interviews (if a few people agreed) were proof that it should be built.
  3. People without vision started using metrics and “data” to iterate their way to success via customer feedback.
  4. But it never worked and didn’t move the needle on outcomes overall. All we got was a bunch of walking dead startups that were really lifestyle businesses.
  5. Smart investors realized this tired of it and thus and started being drawn to big visions and the entrepreneurs who swung for the fences. (For the record, I personally think it’s better to swing big and miss than to swing small.)

What should have happened, and what I think Blank intended, is that companies that found product-market fit but then realized they:

a) couldn’t reach customers cost effectively

b) were in market with a small TAM

c) couldn’t find a scalable repeatable business model

These companies would re-trench and start with a new product for their market. But that didn’t happen. They adopted the YC “don’t die” philosophy and, well, now you have a bazillion 3 person companies clinging on for life support. And you have a bunch of VCs who grew up in the customer development age and think that a lean company clinging on for life support is what they should be looking for.

All I know is that, in my own experience running companies and investing in them, sometimes they just pop. It’s often hard to explain why. The timing matters, and the it’s a function of all the pieces lining up at the right time. But when they do, you know it. Until they do, you just keep trying stuff. I don’t think it’s a good idea to just stay small because it doesn’t fit the venture model, and if you want to raise venture, you have an obligation to try to go big. (Note, there are many paths other than venture, and those are all good too, but, if you take VC, it’s about going big).

I’d love to see some numbers if they exist, but, my guess is the customer development wave hasn’t improved returns for VCs or success rates of startups. It’s hard, and the factors for success are complex. But feel free to tell me in the comments why I’m wrong.

The Coming Rise of Synthetic Data in AI

More and more often, I am hearing startups talk about “synthetic data.”  I’ve seen my existing startup investments start to use it, and I have seen entire companies formed around it.  So, what is it?

Put simply, it is data created by a machine.  Now, why would we do that?  Imagine that you want to train a machine vision model to identify a Tesla.  Now imagine you only have 10 pictures of Tesla’s to train on, so you need a bigger data set to train a better model.  One way to get a bigger data set is to go get thousands of more Tesla pictures.  Or, you could consider doing some simple manipulation to the pictures you have to create new pictures instead.  

For example, maybe you don’t have a picture of a red Tesla.  You could photoshop one of your other pictures to make the Telsa red, and add that red Tesla to your data set so you model performs better at classifying Teslas.  What most people use synthetic data for is to test under different conditions.  They take an image and change the lighting, shadows, etc to simulate different conditions so the machine learning model learns what an object looks like from different angles.

A common use of synthetic data now is to build data sets for autonomous vehicles.  You could create an entire machine generated city, drive around that city obeying traffic laws, and feed that data into the autonmous vehicle model.  This allows you to simulate things that may be harder to capture in real life (e.g. a car running a stop sign).

Now, synthetic data isn’t always good for a model.  In NLP applications, one of the criticisms is that synthetic data sets generated for training are often very simple (because our language generation techniques are still weak compared to other types of AI).  So training a model on all of this language data fails to capture the nuance and vagaries of messy real human language.  But in other situations, like machine vision, synthetic data tends to work really well.

From a business perspective, there are a few use ways to think about synthetic data.  First of all, can you use it to generate new variations of things in ways that are valid for training.  Secondly, can you use it to label data about things that humans no longer need to label?  And finally, should you create the synthetic data yourself, or not?  

My current hypothesis is that synthetic data will mostly be done by a few third party platforms in a market that develops into an oligopoly.  I think the way that software debugging works today:  report a bug -> code a fix -> test on a staging environment -> deploy to production and verify, will be the way a synthetic data workflow evolves.  It would look like this:  report a model failure (e.g., model doesn’t detect things well at night) -> use a synthetic data platform to generate new items for a data set that increase the data for that problem (what things look like at night) -> rebuild model -> test model -> deploy new model to production.  Someday it will be push button easy.

This means if I am right, synthetic data business opportunities will come in 2 flavors.  The first is synthetic data for common objects, where there is lots of data.  These platforms will win by being the easiest to use, connecting to the most workflows, and having the most common options for data generation.  The second is cases where generating the synthetic data is hard because of the nature of the problem space and the lack of existing data sets to start with.  This will lead to specialized providers who can master specific domains.

If you are working on AI, soon you will need a synthetic data strategy.  And if you are company in the space, please reach out if you are looking for investment.

What Ethan Zuckerman Could Teach Googlers About Debate Culture

When I read recently about Google’s attempts to corral internal debates, I thought the way Ethan Zuckerman handled his leaving the MIT Media Lab was a very good model for how to handle debate in general. I’ve met Ethan twice while sitting on panels together and have always been impressed with how he handled himself. On one of the panels he even criticized many of my personal views about AI and the news media, but did so in a way that was very factual and respectful.

Here is the key text I like from Zuckerman’s response:

That’s okay. I feel good about my decision, and I’m hoping my decision can open a conversation about what it’s appropriate for people to do when they discover the institution they’ve been part of has made terrible errors. My guess is that the decision is different for everyone involved.

The decision is “different for everyone involved.”

As someone who has spent a lot of time in debates over the years, one of the things I’ve noticed is that the reasoning behind the debates, in general, has gotten worse. People rarely take Zuckerman’s view that rational people can come to different decisions or conclusions about the same thing that happened, or a stance on a major topic. I suspect if most college students today had to write a resignation letter from the MIT Media Lab, they would demand that everyone else resign too, and the lab shut down, or something like that.

What struck me about Zuckerman’s response is that he doesn’t demand anyone do anything else. He is mostly concerned with his own behavior and feelings, and he acknowledges that others may feel differently and that’s ok. It is an extremely mature perspective in today’s world, and I wanted to highlight it because I feel like it should be the norm, not the exception.

I’ve had a lot of friends leave Google over the past couple of years, as people say “it’s not the same.” I know some of their consternation is due to the attitudes of Googlers, the way the debates happen, and the my-way-or-the-highway debate types that often dominate these discussions. Maybe Googlers could learn something from Zuckerman about respect, perspective, and maturity.