In the big wave of mobile lean startups many companies focus on doing one thing good without vertical integration – usually an app. That has worked out quite well as Apple, Samsung, HTC, Huawei, Xiaomi, Oppo, etc shipped billions of smartphones that could run those apps.
The same approach already hit a wall in the drone war, as the vertically-integrated DJI thrives with $1.6B revenue in 2016 while all the other "lean" drone startups struggle to find large-volume drones to deploy their superb OS or apps.
I wouldn't be surprised if this happens to AI eventually as well.
The main problem for most pure-software AI/ML/DL startups is that: they do not have the input data, otherwise known as the training sets in AI/ML/DL jargons.
Without training sets you can't train your "machine" to "learn". The main reason most AI researches and competitions have focused on image recognition is because Stanford open-sourced its image database ImageNet. AI researchers could therefore sit back and focus only on algorithms. Better for the geeks: they and their competitors all work on the same input (ImageNet) so it's easy to just benchmark against each other on error rates and speeds. May the best brain win.
However, while most Machine Learning applications will work on input data that could be "visualized" and therefore applying the same image-recognition practices, there are no other available large training datasets for other AI/ML/DL applications that pure software startups are trying to build.
If you're using AI/ML/DL to automate and optimize commercial real estates, there is no existing smart building open-source database comprised of temperature, humidity, energy, etc.
If you're using AI/ML/DL to build self-driving car, there is no existing public road video database comprised of street videos at all driving speeds.
If you're using AI/ML/DL to build auto-piloting drone, there is no existing drone flying video database featuring all kind of fast-approaching trees, rooftops, dams, electricity towers, etc.
Most of the time the training sets you need are not available.
Now you understand why all internet giants are building their own hardware products, right?
Amazon's Echo (Alexa) continue to expand rapidly and accumulate more and more voice data. Google rolled out Google Home hoping for the same and launched a completely own-design PIXEL Android smartphone. Facebook bought Oculus VR and continued to pump money into it. Microsoft built Hololens.
Crap – even the super lean Snapchat built a pair sunglasses called Spectacle!
Artificial Intelligence, after all, is meant to mimic and replace intelligent beings. All the 5 senses that we care about, AI should also care about. To build a system that could react to inputs from all 5 senses in a way that human beings would react is but the first major milestone of AI:
- Sight: be able to quickly identify a cat from a tiger
- Hearing: be able to understand a conversation directed at oneself at a noisy party
- Smell: be able to differentiate between a fire and a cigarette
- Touch: be able to discern a caress from a tickle
- Taste: be able to tell sweet from bitter-sweet
Machine Learning could be employed to create a model that reacts as close to human beings would as possible, but without the input dataset to start with, even a mathematician/data scientist with a 240 IQ could not do anything.
This is why full-deck AI startups make more sense, IMHO, and where we're betting our money.
Without full decks, one is at the mercy of internet giants such as Google or Facebook in terms of access to training sets. And we all know what that means.