Why 2017 Isn’t Living Up to the AI Hype

We’re not yet at the AI utopia hyped so feverishly in the media this year. There are still hard questions to ask:

• Has the technology really progressed?
• Why is it so hard to build an AI company?
• When will machines reach human-level intelligence?

Let’s dig into each.

The Clickbait Nature of AI

News outlets are notorious for taking comments out of context, and this is especially true with AI. They’d rather write about doomsday scenarios and post images of the Terminator. And it’s not just mainstream tech columns: Respected international and national publications can’t resist the temptation of a good hook that gets clicks. Albert Einstein once said, “There is not the slightest indication that [nuclear energy] will ever be obtainable.” If Albert Einstein couldn’t predict nuclear energy, we can’t expect reporters to predict the future or AI nor should we be influenced by the media.

What to Read Instead

Turn instead to blogs, newsletters, podcasts, research papers, and books or audio books. Read different and sometimes competing points of view from leaders in their respective industries and make your own judgement from factual evidence. Your educated guess isn’t going to be too far off from the experts, especially given their track record. A few of my favorites are:

Building an AI Company Is Still Hard

While there are lots of open source software libraries and tools to take advantage of when building an AI company, creating one that’s scalable is still very, very hard. Although not as easy as starting an SaaS company, AI ventures can go from idea go from idea to product with relative ease if the employees have a technical background and access to the right tools.

I had the pleasure of chatting with Dennis Mortensen, founder of x.ai, a few weeks ago, and he emphasized that x.ai has been around for three years and they are still working on optimizing a their single-purpose intelligent agent, Amy/Andrew. Three years for a very defined, narrow use case!

How much “AI” is enough?

  1. Understand very early on in the process if you’re building a core/application AI company or industry AI company. H/t to Justin Gage for his blog post describing the difference. Know which you’re building to determine the type of team you need to hire and the kind of AI expertise needed.
  2. Top AI experts have been researching the technology for a long time so have technical expertise that’s unrivaled. Having them on your team is critical if your venture is developing a core/application AI solution, but there’s a limited number of experts to approach; they’re easier to find but harder to recruit to your company. Not needing them can be a competitive advantage if you know what your venture needs to build and the level of technical expertise necessary to get there.
  3. Defining the problem that you’re solving for requires a disciplined engineering and business team that won’t be persuaded to build features just because a single customer asked for them or to chase press that praises you for accomplishing a new task first. Don’t fall into the trap of customizing a solution to every customer. Do so and you’re just a consultancy.

We Need to Leave Accuracy Fetish Behind

Artificial intelligence has been metrics-driven for the last eight years, partly due to the consistent progress in accuracy at ILSVRC (ImageNet Large Scale Visual Recognition Challenge). 2017 was ILSVRC’s last year in part because the work of many researchers is now surpassing human-level accuracy, which was more than 97% in 2016. This first surfaced in 2015 when Olga Russakovsky pointed out the fundamental flaw with such a competition: Its categories are limited ; machines have better accuracy for thousands of classifiers. As a parallel, humans can distinguish 10 million colors alone — and having a singular metric for success does not necessary correlate to being the best possible solution.

Unfortunately, the metrics mentality has permeated to the AI startup world, whether it’s an intelligent agent or an algorithm. In the hundreds of AI applications that we’ve received for the AI NexusLab, the vast majority point to a single metric that defines their company, either percent of human interaction need, human hours saved, sales increased, among others.

Having a great metric isn’t the problem. It’s when you define your business based on a single metric that you run into trouble. Investors, especially now that we’re at what I hope is the tail end of the AI hype cycle, are no longer investing just because you have an AI company and just because you marginally improve a basic human function. While the tech may be cutting edge, investors need returns, and returns mean exits, and exits don’t come from being 10% better than your competitors at completing a task.

Don’t define your company by a single metric

Remember, to an institutional investor you need to have a viable path to a $100M venture. If you are excelling at a single-use case, make sure your product can sell itself, i.e., the value proposition is enough to convert customers, ideal in the same industry but different markets. If you’re in an industry that’s quickly becoming commodified, either scale an already repeatable sales strategy or build out the core tech so that it can be applied to to other use cases. The latter will make a compelling argument for seed financing and help build the growth strategy for Series A funding.

The Promising Road Ahead

There’s still a lot of room for growth for the industry, and there’s an almost limitless amount of opportunities to learn. For example, our Future Labs AI Summit on October 30–31 offers trainings, presentations, and panel discussions — use code FL25 for 25% off tickets — but be careful to sort through the noise. Every day we make progress with the technology and apply it to new challenges in an effort to advance society and humanity — something we can all get behind. As we close out the year and celebrate the tremendous headway AI has made this year, I just hope I don’t have to read about 2018 being the year of AI.

PS : If 2017 was the year of anything, it was the year of AI stock photography.

Thanks to the Future Labs team and Costa from FindMine for feedback on this post.