The last few months I have been interacting with some AI startups. I have also been checking out AI startups coming up on a regular basis. The startups can be classified as working on vision, speech or text. But, except for a select few, almost all the startups I have encountered have a fundamentally core problem. I have a very hard time in trying to pin point the core USP of these startups.
(Update: Based on feedback from Twitter, I would like to clarify that I am talking about AI startups that are mining for the gold in the gold rush. Not the ones providing pickaxes)
These startups follow a template which goes something like this:
- Play around with some new open source machine learning libraries or ML architectures.
- Talk to different businesses and find a business which can lend them some data.
- Apply the ML libraries on the data and run the training in cloud GPUs and get some decent results.
- Startup.
But there is a problem with this approach. The ML libraries are open source. It is the customer’s data. And cloud GPUs are available to everyone. So where is the core USP.
I think the fundamental problem with this approach is that most new AI startups are working on solving AI problems as an engineering problem.
For example, I see that most of these startups don’t have a core domain person. Speech startups don’t have speech experts. NLP startups don’t have language experts. And vision startups don’t have image experts. And only a few actually have AI experts.
I know that deep learning has removed some need for these experts. But they are needed in production. When things go wrong, that’s the time you need experts to find the problem and fix it. In production, when people are paying money, we can’t say AI is a black box.
Don’t get me wrong. If a startup is able to make it with this approach, then all the best for them and I will be very happy for them. But if you are going down this path, then it will be harder as you move forward with your startup.
It has never been easier to start an AI startup. It has never been tougher to run an AI startup.
The most important question that’s gonna come up is, if the ML architecture is open source and if data and GPUs are available then how can you beat Google,MSFT, Amazon and FB. They are the ones who provide the GPUs. They have all the data in the world. And their researchers are the ones who open source the ML libraries and architectures.
So if you are an AI startup I would suggest the following:
- Work with domain experts. Get them on board.
- Work on the math. Treat the problem as a math or data problem and not just as an engineering problem.
- Find your USP. Maybe it’s a new algorithm. Deep learning is now showing some cracks. Go after those cracks and fix them.
Following are some common approaches I find people using in their startups.
Speech: DeepSpeech is the default software that people end up using. Being an end to end ASR system, it has caught the mind share of most new speech startups. The idea is that you collect 4000–5000 hours of speech data and you should have a working system. And actually it does give you a pretty good system with 4000 hours of data(though it does cost a lot). But it’s not usable except for very specific cases. And also how do you beat Google with this.
NLP: This is a tough one because the research itself is moving so fast. BERT, Inference, GPT2 Transformers. It’s harder for startups to fix on an approach. But still, many more startups are working in this area as lot more data is available as text. Most of them have chosen a specific domain to concentrate on and provide NLP solutions. But very few actually provide a much better solution than what Dialogflow or Microsoft provide.
Vision: Most startups have picked niche domains and are solving narrow problems. Automated driving is driving most of the enthusiasm. But still we are a long way away. I have encountered fewer startups working on vision than speech or text, maybe because of the domain I am in.
Disclaimer: All this is based on my interactions with a few startups and what I have read of other startups from their websites. There maybe a lot of startups which are going after harder core problems and I may not know them. All the best to them.
