Harder than getting into Berghain
As I am writing this, it is Q2 of 2020 and we have been running a strict policy when it comes to admitting users and prospective customers. And we will for some time. There are two main reasons:
- Because we can
- Because we can't
Let's take them one by one!
Because we can
We are lucky enough to have secured a comfortable amount of funding. The primary expectation attached to it: "Build a product that users love!" After many reaffirmations, we started to understand what this meant: Unlike many other early-stage businesses, our priority would not be on the usual measures like trajectory, revenue growth, managing churn, and so on but to prove a few hypotheses such as these:
- Users onboard themselves
- Users reach the most important product milestones without personal intervention
- Users love the platform so much that they are recommending it to others
To say the least, this is a very comfortable and privileged situation. It does not free us from having a pricing discussion with users at some point but we don't need to push the boundaries and it frees us from the pressure of having the perfect pricing in place from start. And it makes it so much easier to write out this sentence: When we say that we want to give users a great experience, we mean it.
Because we can't
We are not yet at the point where we are giving the software away without any interaction. The core of it works but much of it is not as self-explaining as we want it to be when we open up the "user hose". There are simply too many cases that still need some tweaking and we, therefore, require at least one personal contact with everyone who gets on board. This does not mean that it takes weeks for a user to get a result: Once we let someone in, their case can usually be covered just as our landing pages say. And for every user we take onto our platform, we speak to at least ten others who think they have a case but really don't, be it for the lack of data, business case, commitment or otherwise. In other words: It takes time.
All this could be accelerated by having a dedicated sales team with us (which we don't). There are four reasons why we are not going down that route:
- Despite our funding, we don't have unlimited resources. We are very conscientious when it comes to hiring as this can quickly turn sour if not done in the right way.
- We are aiming for no-touch and this would quickly make this team redundant.
- A salesperson would need to understand a fair bit about the inner workings of our product and while it looks simple from the outside, there is some crazy stuff going on that a company representative needs to know about – so we are essentially hiring another deep learning & cloud expert of which there are not that many.
- We want to get immediate feedback on what people like and what they hate. Not buffered by a CRM or something of that sort. Users are directly speaking to our friendly product team once we let them through the gate.
None of this is meant as an excuse – quite the opposite: It is an invitation to our users to expect a truly great experience from us and our product.
And for anyone who has signed up with us but hasn’t heard anything: This is nothing against you! For sure we have read your description but at the time we didn’t think that it was a real business problem that could be solved through AI. But we might be wrong and if you think we are, please send us a note and we are happy to get back!
Deep learning can be expensive
Despite all optimization, we are doing on the engineering side, training a neural network on cloud infrastructure can get costly – quickly. Very quickly. If we are running at full capacity for one day, this easily costs several hundred if not thousands of Dollars and therefore we want to at least get a chance on a payback. That payback can be twofold: Feedback and/or revenue. And if the chances are slim for at least one of the two, we have to pass.
At the moment, we are running everything on GPU and TPU cloud servers; these are specialized machines for deep learning and blazingly fast – but also disproportionately expensive. The only way to make things cheaper is through training on CPUs. For household applications and Kaggle competitions, this is not an issue but we are not going to sacrifice on speed for early and paying users that want to solve a serious problem of theirs.
Hopefully later this year, we will offer a more affordable option in which we could release a free or less-gated version. Until then, you will need to have a good argument for why you need to train your cat/dog-classifier on our machine!
Customers are scarce
Customers are individual human beings who are just summed up as "customers" by convention but those individuals are the reason why we are building this product in the first place. We truly believe that there are people out there who would like to do fantastic things with machine learning automation but can't for a variety of reasons. For some, we already know that this is true.
We know that customers are not just "things" that you can switch on. They are scarce! And as long as we cannot ensure a joyful and satisfying experience, we don't want to let them in by the dozens each day. We can only handle onboarding, extensive feedback and additional requirements with the capacity we have. If we did let them in by the dozens, we would be wasting their time – time that could be better spent otherwise.
We also know that alternatives are abundant. Not necessarily because there are hundreds of no-code deep learning platforms out there (there aren't) but because people's time can be spent in a variety of productive ways.
If this was not clear enough yet: We treat our customers as an essential part of our business. What I wrote above in a hurry with regards to sufficient funding is true but we are aware that this is just to artificially bridge the time gap between our vision and reality, the reality where people want to open their wallet for what we have built because they are happy using it.