A Distributed AI Research Lab

In April, Nat launched AIGrant.org. The idea was simple: fill out an application, get a grant for $5,000 to work on open source AI. No committees to convince. No university to attend. Just a five-minute form.

People seemed to like it: we received nearly 500 applications from more than 50 countries. After a few weeks of screening, we selected our first 10 Fellows. Our projects range from many-body quantum system simulation to hardware-accelerated deep learning in the browser, a GAN to simulate brain activity and more.

So what’s next? I’m happy to announce that I’m joining Nat as co-founder of AI Grant. Together, we’re excited to announce AI Grant 2.0!

Like a good neural net, we’ve back-propagated and improved. AI Grant 2.0 Fellows will receive some new treats, including:

  • $20,000 each in Google Compute Engine credits.
  • Exclusive Q&A with AI experts including Andrej Karpathy (Director of AI at Tesla and previously at OpenAI) and researchers at Google.

This is in addition to the standard perks, which include:

  • $2,500 in cash.
  • Access to the AI Grant network.
  • 250 Tesla K80 GPU hours from FloydHub.
  • $1,000 in ScaleAPI data labeling credits.
  • $5,000 in CrowdFlower data labeling credits.

We’ve learned from the previous cohort that $2,500 will satisfy the needs of most projects. We’d like to thank CRV for their generous contribution, which will fund the next 20 grants.

Our aspiration with AI Grant is to build a distributed AI lab. We’re starting by issuing grants to the smartest people we can find, doing interesting work that might otherwise not happen, and connecting them to mentors and experts, and to each other.

How Do I Apply?

  • August 25th: Applications Due. That’s a month from today. The clock starts now!
  • September 12th: Finalists selected and applicants notified.
  • September 24th: Final winners selected and announced.

Stop reading, and click here to start the application. Don’t wait!

But I have more questions!

Fine. Click here to read our FAQ. But apply today!

The Tech Tree


Elad Gil recently wrote a blog post about a possible “end of a cycle”. I wanted to offer a related view, inspired by many hours of playing Civilization. Instead of a singular “cycle”, I think about the budding technologies we have and the different “phases” they’re in.

Phase #1: A Research Technology

Many facets of cleantech, biology and AI are in this phase right now. We have a lot of promising ideas, but don’t expect any short term deliverables that you might capitalize on.

Phase #2: A “Recently Unlocked” Technology

Technology in this phase is defined by two characteristics:

  1. A recent “tech tree” advance has happened recently. Something that was impossible has just become possible.
  2. Application of the technology is still hard and inaccessible to most engineers, since platforms to simplify it haven’t been built yet.

Past examples include web development in 1994 (Amazon.com was written in C by “brilliant” engineers), 1.8” hard drives in 2001 (Toshiba showed Apple the drives hoping they could “find a use for that” — the iPod) or mobile development in 2009. Current examples might be VR or Deep Learning (ConvNets applied to images).

Since applying the technology requires deep investment, broad use-cases are prioritized. This is the phase that presents an opportunity to build monopolies (Amazon, Uber).

Phase #3: A Commodity Technology

Over time technology is simplified by better abstractions (Python, Weebly, Parse, etc). It becomes accessible to a broader set of engineers. I wouldn’t categorically reject commodity investments, but I’d want to see an alternative moat. Otherwise I get stuck on the “why now” question. Many humans have had the ability to make this product. Why haven’t they?

Sometimes the answer is artistic creativity. Snapchat and musical.ly are like hit TV shows: anyone could make American Idol, but Simon Cowell thought of it first. Sometimes the answer is a downmarket focus (Innovator’s Solution). Gusto and Stripe were optimized for businesses that were too small to garner attention from ADP or PayPal.

Often the answer is less satisfying: specialization. The company is focused on delivering the technology to a long-tail market (“software eats the world”). While valuable, these companies often don’t produce the type of return profile VCs seek as they don’t build a monopoly.

Keeping Tabs On Phase #2

When VCs can’t find anything in phase #2, they often fund #1 and #3 by accident.

Ideally you’d like to set a Google News Alert on tech that has recently become “unlocked”. A good way to train yourself what this feels like is to review past examples:

  1. Oracle. “A Relational Model of Data for Large Shared Data Banks”, the paper that inspired Larry Ellison: https://goo.gl/FNsge8
  2. The Atom Bomb. Einstein’s letter to Roosevelt upon realizing that a bomb is possible: http://www.atomicarchive.com/Docs/Begin/Einstein.shtml
  3. Photo Recognition. In particular, the “search inside the photo” functionality. Much of the recent advancement in this area can be traced back to a paper from 2012: http://image-net.org/challenges/LSVRC/2012/supervision.pdf.
  4. The World Wide Web. Many examples here. One favorite is the article that motivated Jeff Bezos to research Amazon (see growth curve on second page): http://www.quarterman.com/pictures/1991-1994–mn/SCAN0418.html.

Thanks to Elad Gil, Chris Dixon, Chris Howard and Christina Cacioppo for reading drafts of this.