Embrace the Human

Human assisted ai — it’s a term that’s been misunderstood, overused, and mistreated, but at Elves it’s an ethos.

Imagine this alternate history for a moment — it’s a lazy afternoon in Mountain View circa 2009 and a budding young Travis Kalanick (Uber) sees yet another Google car self-drive itself down University and has an epiphany. “What if I built an app that would let you summon a little self-driving car to your location at the press of a button, you get in, and it’ll take you anywhere you want. Bam — a $65 B idea!” Nope. It would have sucked.

Now don’t get me wrong, it’ll be awesome someday, but with what he had to work with at the time it wouldn’t have performed very well — it would have been what most bots are today.

“Using Elves was the first time I used a bot to do something useful”
- Zak Fassi, Strategic Partnerships at Facebook

Take a look under the hood of your run-of-the-mill human-assisted ai startup. The “human” component will usually comprise one of the following states: 1/ a handful of engineers looking at a minuscule sample of interactions or 2/ the use of ai trainers (think outsourced contractors, Mechanical Turk, or other solutions that are all generally regarded as a necessary evil and kept at arm’s length)

Startups have been at this a while. The folks at X.ai (developers of the inspiring Amy / Andrew scheduling assistant) openly explain their use of ai trainers and their method of Supervised Learning. In their blog posts they explain how people will tackle queues of tens of thousands of data points one at a time — it’s “not necessarily sexy work”.

“So let’s make it clear: today, scores of humans are involved just about everywhere in AI, whether in tiny startups or massive tech companies. In fact, most AI products are very much NOT fully automated, at least not in an end-to-end, 100% bulletproof way. It is probably ok for the general press to get a bit carried away with AI. However, we in the tech industry should probably better understand this reality, and acknowledge it as a necessary step in the process of building a major new wave of technology products.”
- Matt Turck VC at FirstMark

Now, there might be some disagreement over how long it’ll be until Uber’s fleet is fully autonomous but there is no disagreement that it’ll happen. Human assisted ai needs to take the same approach/leap and not shy away from a large, however transient, human component. Apply a Travisian mindset to artificial intelligence / machine learning, really Embrace the Human, and you’ll end up with bots that don’t suck. This could be a key strategy that startups in places other than Silicon Valley could use to get a leg up.

What does that look like in practice? Well, we’ll need a ton of data/traffic so let’s go ahead and drive it.

1. Come up with the most absurdly profound value prop.

Elves will do anything you want; for free

2. Great brand, customer service, cutesy name, and lots of kittens, hearts & teddybears.

3. Execute, execute, execute

“I don’t want to go back to a time before Elves”

At Elves we are human first — the bots serve the Elves. The Elves act as human shields between users and bots. Via chat through our app or on Messenger you can chat with an elf that’ll do or buy anything (legal) you want. We found a sweet spot on the human/machine continuum that allows for delighted customers contributing towards a robust learner.

We have — by definition — attained Human Level Intelligence. We’re smarter than Google — yay! Unfortunately, we’re also radically inefficient — but it’s getting better. Using a human first approach our “bots” never fail to understand intent, utilize creativity and execute on an infinitely wide array of intricate requests. As they build up the repo of use-cases (bot training-data) they get more efficient. As they continually and passively A/B test every sentence that comes out of Elves they get smarter.

Ultimately, success will boil down to:

I. Premise — How big / relevant / cool the problem is. At Elves we are annotating a broad array of human experience and automating the associated execution. Think “Hey Siri — I have a meeting in Dubai tomw; hook it up!” and have her execute on your flight, cab, hotel, dinner reservations, ext vs. “Hey Siri — what’s the weather like outside.”

II. Method — This screenshot is a little dated but shows the logic of our data collection method.

From the vantage point of an Elf using our custom built console it’s a three way conversation between her, the user, and the bot. They teach the machine passively as they work.

  1. A simple counter is involved every time an elf uses a sentence the bot produces; human verified proof that the bot got it right (the green button sends it out to the user and the black flags it for review). If a sentence goes out correctly often enough it can start to go out automatically.
  2. If an elf gets a bad answer from the bot she can write whatever she wants. The new message persists as an alternate response to that intent. The next two times we get that intent the two Elves that handle them will each get served one of the alternates. It’s what we mean by passive A/B testing — and its continual.
  3. If the entire use-case is something we’ve never come across the conversation is recorded, indexed, and served up the next time we come across a similar scenario — Elves never go through the same pain twice.

III. Short / Mid term retention — it’s super important and the most commonly overlooked. All companies will one day have access to autonomous vehicles but Uber will still have retained its dominance. Truly Embracing the Human has allowed us to delight and thrive where other bots have frustrated and failed.

That said, until the Ai is perfected, our kind of services will continue to be nuanced in the US and most of Western Europe where the alternatives are more frictionless. In the meantime they can be profound for users in countries where daily life is just a little more difficult. It’s something we realized early on at Elves. It’s why we decided to focus on MENA while our method built up a large enough repo through the thousands of convo’s they’re now having daily.

  • The top 10% of Elves users use Elves almost every day (well at least more than 30 times a month)
  • 80% of the people that make a purchase on Elves make another within 28 days (Avg ticket size = $260)

Karim Elsahy

Founder / CEO @ elves