Y Combinator takes machine intelligence startups to school and learns a thing or two


Machine intelligence startups are the black sheep of the startup world. The new kids on the block are challenging investors to do their technical homework and differentiate themselves in intentional ways. joined a growing list of investors offering exclusive services to these companies for its latest S17 batch of startups.

In the competitive world of investing, Y Combinator has to work to convince top startups to apply to the program. Today, many startups that fit the bill are working to solve challenging AI problems. And with the amount of money sloshing around for AI startups, the sense of urgency isn’t always there for prominent researchers who have their choice of financial partners.

and the brains behind the AI track, explained to me that his aim was to offer founders desirable data sets, compute resources and technical mentors, among other things. With experience founding a company and solving machine learning problems for Apple, Gross’ relatable technical background helps to emphasize the legitimacy of the storied accelerator.

Of course, YC also recognizes that much of the current machine intelligence space is hype. In an effort to make sense of the madness, Gross prioritized startups that were working on problems of perception, autonomy and machine learning services.

That last bucket includes startups like , building a speech-to-text API, and , building a natural language processing API. These were perhaps the most controversial of the S17 batch. Many VCs on Sand Hill have on the grounds that tech giants like Google and Amazon are more likely to swallow the market than any singular startup.

But Gross asserts that a -esque possibility remains for companies that are capable of building services that are easier to use. In true contrarian form, Gross argues that the actual machine learning prowess of each of these teams comes secondary to their ability to craft a product that developers actually like and would use by choice.

.text .crunchreport h3 {color:#fff}Latest Crunch Report

Moving past APIs and other developer services, perception and autonomy were easily the most populated spaces for startups within the YC AI track. The perception subcategory includes startups like, automating store checkout, , facility management, , classifying skin conditions, , robotic farming, and, obfuscating faces for security. With a bias towards deep learning, these companies are exploiting unique data sets and readily available compute to automate tasks that were previously inconceivable.

Meanwhile, on the autonomy front, startups like , creating autonomous vehicles, and , building its own radar, areas automakers and suppliers look to stand their ground in the rapidly evolving transportation space.

“There’s a common theme here,” explained Gross in an interview. “A breakthrough algorithm creates a temporary moat that allows you to get to another moat.”

The reality is that algorithmic advances become outdated on a nearly weekly basis in the world of AI. Things tend to go open source faster than they can even reach full deployment. This means that a startup initially using off the shelf AI tools might actually have a speed advantage to collect critical data over its competitors. This ability to forge gold from iron combined with meaningful domain expertise is the difference maker between successful and unsuccessful machine intelligence startups.

With and emerging from the woodwork, the question remains as to what resources actually move the needle for highly technical startups. I tend to believe that most AI startups fail because they are unable to effectively productize. Those whose biggest problem is tuning hyper-parameters are probably in the minority.

Gross mostly agreed with me, adding that a big challenge is helping customers who purchase services from machine intelligence startups understand what it means to rely on a stochastic product. Outcomes aren’t always predictable and often they’re not even explainable.


This is where I think the distinction between concrete and soft problems comes in handy. Concrete problems tend to be easily automatible. They are typically quantitative and highly repetitive in nature. Humans are very good at them but they are labor intensive. Try to think of standard classification problems like grouping images or extracting numbers from a document.

Meanwhile, soft problems tend to be things that humans are not particularly good at. Often qualitative in nature, soft problems require a lot of domain expertise to solve. Point being, I would trust an AI today to look at my photo library and organize it but I wouldn’t trust an AI to look at my photo library and use the knowledge within it to write a letter to my mom.

Applying this heuristic to YCs batch would seem to favor a business like , classifying images of skin conditions, over a startup like , evaluating potential teachers, or , ranking job applicants by culture fit. There really isn’t a great reproducible methodology for predicting teacher performance or culture fit.

But Gross insists that it’s important to remember the cost of making a mistake. The cost of an inaccurate dermatology diagnosis could be very serious but the cost of accidentally rejecting a potentially fantastic job applicant is relatively low. And I’d mostly agree that regardless of the performance of a startup like Nimble or Headstart, anything is better than the status quo of crappy platforms employing keyword search.

“Algorithms can actually be useful at soft skills,” noted Gross. “These are areas where an AI can make more fair decisions where a human would be irrational.”

Time will tell if YC’s framework for investing in AI startups is accurate. In many ways this inaugural batch encapsulates broader trends in commercializing AI. For as long as the chance remains that one of these startups could be the next Dropbox or Airbnb, YC has nothing to lose from investing in verticalizing its storied accelerator.

Featured Image: Bryce Durbin


About Author

Leave A Reply