Three companies I didn’t start

Below find three ideas for companies that I spent some time thinking about over the last few months. For various reasons, I didn’t end up starting any of them myself. But I’d love to hear from you if you think I’m making a huge mistake.

1. Mechanical Turk for the Masses. Mechanical Turk is ten years old. It feels like one of those things that should have enabled all kinds of new businesses, but never seemed to reach its full potential. I had two approaches to open up Mechanical Turk:

  • The Digital Ocean approach: by making the API and documentation better, could you reduce the friction of getting started and unlock new usage? I suspect there are many projects where the overhead made the initial Mechnical Turk integration seemed not worth it, but where the result ended up very valuable.
  • The Heroku approach: you don’t care about virtual machines, you care about your application. Let’s change the abstraction layer of Mechnical Turk from people doing tasks to the job you want to get done. For example, maybe I want to clean up a 100,000 rows in Excel. Humans could do the first 60,000 from which I could then train a machine learning model to both go backwards and verify quality as well as go forwards and complete the last 40,000 rows. The user shouldn’t have care at all. Maybe the interface is a plugin in Excel, where the user pays $10 and the job is done.

You’d start by building on top of Mechanical Turk until you had enough volume to switch over to your own workers. Note: this is a individual-first approach to Mechanical Turk compared to Crowdflower’s enterprise-first approach.

2. A online bank for small businesses. It’s interesting that the checking account is the first thing that all small businesses need and the only thing they all need. This allows you put the bank in the center of a small business’s world, and allow the other services they’ll need to orbit around you. You get to focus on customer experience and customer acquisition. Here’s how I was thinking about it:

  • There’s been an explosion of third party financial services for loans, money transfer, etc. Thus it’s now possible to create a “faux stack” (fake full stack) startup, where you offer the full set of financial services to a customer without having to build it yourself. This includes a whole range of financially related products like payroll, accounting etc that businesses need.
  • The largest challenge for those third party services is acquiring customers. You can offer a “company in a box” guide to small businesses by curating a marketplace financially-related services (if you’re this kind of business, we recommend…). Each vendor in your marketplace would pay you a recurring referral fee, allowing you to offer your core checking account product at or below cost. And then you can aggregate those referral fees to outspend any of the individual vendors for customer acquisition, while you own the customer relationship. Thus what you’ve created is a defensible marketplace with a great “single player” mode in order to avoid the chicken/egg marketplace problem.
  • Use hyper-verticalized product marketing, where you develop a unique product offering for each of your different segments. Existing banks aren’t doing this nearly enough. Four ideas for possible SMB bank customer segments: startups (probably the worst one), sellers on eBay/Shopify/Etsy/Amazon, small professional offices of doctors/lawyers, and marijuana dispensaries.
  • Ultimately, you’ll have a full financial history of every small business on your platform which allows you to become the best investment bank of all time– one click sale of your small business to anyone in the world who is interested.

3. Tamagotchi for your smartphone. We need human contact to be happy due to inherited biological constraints, but can we fake it? Can a computer be a better friend to a human than a human? Here are a few advantages computers have over humans as a friend and partner: infinite attention, complete customization, a different set of sensors, and ability to cross-polinate insights from large sample sizes.

  • The Quantified Self movement largely failed because it turned out that looking at charts of past activity weren’t that effective at changing future behavior. Could you be more effective by first trying to create a social bond with your software?
  • It seems like applied coaching might be one of the best areas to start with. For example, began by filling in the gaps between appointments with human psychologists for folks coping with eating disorders, addictions, etc.
  • It’s quite likely that software could do better than the median psychologist at a radically reduced cost. If you can demonstrate improved patient outcome and decreased rates of recidivism, there’s a big play for insurance dollars to carry you over as you develop a more generalizable AI.

What could you do with a friend on your phone who spent 100% of their time trying to make you better?