AI as a platform – commodity
For the non-technical investor, the term artificial intelligence (AI) can be either a celebration or kryptonite. A celebration because adding a few AI startups to their portfolio is a great way to justify to their LPs that their deal flow is current. But it can also be kryptonite in the sense that every startup seems to be using the term du jour for a significant valuation increase.
But the truth, like in most things, is somewhat muddy:
It certainly is the dawn of AI startups, and increasingly more and more startups will add AI to its repertoire.
There is a shortage of truly deep technical talent, but that won’t stop the evolution of startups since there are ample open-source libraries in the world.
Data is certainly key for AI startups, but that does not mean that the Googles, Facebooks, and Baidus (GFBs) are the only sources of these data. This means they won’t be the only players in the AI space.
If you’re interested in investing in the AI space (as a VC or an angel investor), there are really three key principles in investing in startups that employ artificial intelligence or machine learning: whether the startup is a VAS or a HAS, how the startup creates value, and their data-product fit.
VAS vs HAS
First, some definitions: vertical AI startups (VAS) are those that apply AI in one focused vertical industry. Horizontal AI startups (HAS), on the other hand, apply AI across a multitude of different industries. You can read about it in more detail in my previous article.
The general pattern is that HAS are comprised of stronger technical teams (PhD’s will do the trick), will need a longer runway to find product/market fit (let alone revenue), and will likely be acquisition targets for the GFBs. On the other hand, VAS can find revenue models faster, need less technical teams, and satisfy a real industry need.
DeepMind is an example of a HAS and satisfies all three criteria. However, the days of HAS are limited, and some would argue that there really just isn’t that many HAS startups at all.
Categorizing the startups you’re eyeing is the first step to understanding the nature of the beast.
Value creation
Value creation in an AI startup can come from two sources: the product itself (short term) or the startup’s data aggregation (long term). This means that startups can either create value through the use of the product itself or through the amount and uses of the data they collect.
AI startups that care about short-term value creation are great, but those that care about long-term value are even better. This is simply because data aggregation, especially of unique and exclusive data, is inherently more valuable.
Take Sero.ai, for example. While farmers can use their tool as a way to track their crops’ health, the startup is actually aggregating crop image data, which is something no one is doing yet—at least not in such a large scale.
Another example is a startup that can aggregate data on elderly falls, which is the sixth leading cause of death among the elderly.
Both these startups have short-term value creation models, but more importantly, long-term ones. It’s critical to check whether startups care about both ends of the value creation and can prioritize one over the other at the right time in their journey.
Data-product fit
Data-product fit (DPF) is the alignment between how the startups think about data and the product they’re creating.
It’s important to ask the following questions when discerning DPF:
Does the data they’re gathering feed into product creation and, more importantly, product improvement?
Does the data they have give them an ability to create a better product?
If they don’t have enough data, can they create a new data gathering process that gives them an advantage in product creation?
How do they think about data and can they make data gathering a repeatable, scalable process?
For example, a startup that has achieved DPF in the HR space may compile CV data that tells their customers and hiring managers whether someone is the right hire for a certain job, free of bias. Over time, the algorithm gets better, eliminating more bias, like grades and pedigree, and surfacing the right candidate for the data job.
Conclusion
Investing in AI startups is the next big opportunity, both from an impact and a returns perspective. But just because a startup is about AI doesn’t mean that it’s exponentially more complex to invest in them. This irrational fear stops investors from catching the next wave. The best way to prevent that fear is to stop sitting on the sidelines and start investing and learning.
Investing in AI
Investing in AI is not an easy job: AI technologies are black boxes and unless you are able to dig into lines of code they may be inscrutable. Simply looking at proof of concepts might not be enough to really understand the underlying stack behind specific applications, and this represents a big barrier for investors to efficiently allocate their capitals.
ii) Data effect: it is common knowledge that neural nets require a lot of data to be trained, and if the startup has a way to create a virtuous data cycle (‘data network effect’) or has access to proprietary data, this is sometimes enough to be deemed as investable;
iii) Team and Patents: the biggest barrier to entry AI/ML is talents and IP. Therefore, if a team is composed of scientists/researchers and has patents(obtained or pending), it would already be a good candidate for an investment even without any revenues. This is driven by top tech companies acquiring smaller startups simply for their ‘brain power’ rather than their actual numbers.
1 – AI is secondary (as is any technology) to the overarching business model.
Bottom line – integrating the latest technology, even AI and machine learning, is not what will get you funded; how you use that technology to produce a concrete and scalable product goal/service of value is the ultimate driver for raising funds.
2 – Go into the pitch knowing your competitive advantage and sell it.
Having a solid track record of sales and a strong management team goes without saying, but there’s got to be more under the hood than just a trusty engine. Investors want to know what sets you apart from every other AI startup in the ring. What do you offer that other companies are lacking? What’s the ingredient in your secret sauce that makes it compelling enough for others to want to try a first batch? Remember for AI companies, a proprietary plume of data is often more important than having the “best” algorithms. B2B AI vendors will be forced to put their answers to those questions into simple language in order to convince early adopters to chose them (over a sea of “safer” alternatives).
Interview Highlights:
The following is a condensed version of the full audio interview, which is available in the above links on TechEmergence’s SoundCloud and iTunes stations.
(3:20) Have you seen the trend recently (that AI is more common in business models), and if so is that an uptick in talent flowing in to the entrepreneurial world or an uptick in the word being used?
Ben Narasin: There’s plenty of businesses who don’t mention it at all, it depends on what you’re focusing on…what’s happened is machine learning, neural networks, and then loosely stitching all that together to say AI (we’re not at general AI yet by any means) has sort of replaced the word algorithm that everyone has some version of…I think everybody believes that having a data plume that they capture and then manipulate on the backend could be called ML, but it’s probably not.
…there is also a well-spring of growing people with expertise in this category…part of what we’re noticing is AI is being open sourced, and as it’s being open sourced you’re having a larger community to contribute to it and also making it available to more people, so very few people are pitching that they have some proprietary ML; what they’re pitching is that they have proprietary data sets which can inform that ML, which is what creates great power.
(5:36) Is it often more compelling to hear about how you’re collecting something at grand scale and of value than it is to know every nut and bolt as to how the deep learning is set up?
BN: If you tell me you’ve got the most compelling deep learning structure in the world, I personally can’t diligence whether that’s true anyway…what I need to understand is when you’re feeding that engine, are you putting high octane fuel in this thing or unrefined oil?…Data is now one, the universal source of truth, and I think that people understand data is also the new oil. Our job as Enigma.io or as any company in this space is to be the new-age refinery of that data, so it’s what do you have as raw materials and what can you do with it.
There’s more on this discussion here. A key consideration, however, is that the open sourcing of technologies by large incumbents (Google, Microsoft, Intel, IBM) and the range of companies productizing technologies for cheap means that technical barriers are eroding fast. What ends up moving the needle are proprietary data access/creation, experienced talent and addictive products.
User-in-the-loop
Human plus machine
Here are some great examples of products that prove that involving the user in the loop improves performance:
- Search: Google uses autocomplete as a way of understanding and disambiguating language/query intent.
- Vision: Google Translate or Mapillary traffic sign detection enable the user to correct results.
- Translation: Unbabel community translators perfect machine transcripts.
- Email Spam Filters: Google, again, to the rescue.
We can even go a step further, I think, by explaining how machine-generated results are obtained. For example, IBM Watson surfaces relevant literature when supporting a patient diagnosis in the oncology clinic. Doing so improves user satisfaction and helps build confidence in the system to encourage longer-term use and investment. Remember, it’s generally hard for us to trust something we don’t truly understand.
Failures – IBM Watson
We must remember that access to technology will, over time, become commoditized. It’s therefore key to understand your use case, your user, the value you bring and how it’s experienced and assessed. This gets to the point of finding a strategy to build a sustainable advantage such that others find it hard to replicate your offering.
Aspects of this strategy may in fact be non-AI and non-technical in nature (e.g., the user experience layer ). As such, there’s renewed focus on core principles: build a solution to an unsolved/poorly served high-value, persistent problem for consumers or businesses.
More and more AI startups are being built and funded. Startups are pursuing to build their product/Service to include an AI component, saturating the market and proving there is demand for such companies at the same time creating a lot of noise.
One of the things worth considering, when a startup claims a certain component of their business has AI component – investigate carefully, as it might not be a competitive advantage as most claim, but a simple sorting mechanism, that can be easily replicated.
“AI” is surely on the rise, but there is a huge amount of BS going on.
The tl;dr: only invest in very focused and niche applications where there is a willingness to pay. There are a lot of areas for that. The bot space is hot so avoid.
Point to note
- No one knows what AI means or is: Data science was ‘advanced analytics’ not long ago. Gartner wrote in a report that they changed their report title as ‘everyone else was doing it’ so needed to align to the new term. AI doesn’t exist, strictly, so no one knows how to define it, but defines it as benefits themselves.
- 99.9% of startups that say they are AI, only use AI: Startups that do ML use a Tensorflow API, or NLP use GATE etc. To ‘do’ AI you need a team of PhDs doing research across a whole bunch of disciplines, which you probably can’t find nor afford.
- Advances will be slow so don’t believe big leaps will happen: A lot of the renaissance happening in the past few years is off the back of papers written in the 1950s… I heard from a friend that many papers (I think Russian ones in particular) can’t even be understood by people today as we got stupid (my paraphrasing, not his). The reality is it is only 1982 that any self-respecting academic would touch ‘AI’ field again, and in recent years that any real processing can be undertaken. I can go on about this, but short version, is we are just scratching the surface so don’t expect a startup to come up with epic stuff.
- Deep Learning is super important but still early: We are starting to see more cool stuff, but it is only in recent years that Hinton et al in 2012 won ImageNet (with a 2-layer DNN). Semi-supervised learning (graph-based learning) can help solve more difficult data sets
- Most of the things startups talk about are pointless without data and startups don’t have access: Machine learning and related techniques need obscene amounts of data. IBM just did a deal with SalesForce for data. Google offers free photo hosting to get their photos for processing! Where the hell is the startup going to get the data to do any kind of proper training?
- 99.9% cant build AI due to talent shortage: Google, Facebook, Baidu etc hoover up every single competent person. There’s one uni in Canada where the faculty have generally refused to leave (listen to the a16z podcast on AI, it’s quite good)
- Founders lie about AI to get valuation – Founders will strap AI into the deck title at any opportunity with the goal of adding a few million to their pre. They know they don’t do anything, but that doesn’t stop them joining the new SoLoMo trend
- Startups can only possibly win in niches: The ‘winners’ will be in distinct niches. Since AI resources are effectively a commodity, industry knowledge is important. My friends in Singapore doing Active AI started as retail banking experts and then added the ‘AI’ part around problems they knew about and had the relationships to access.
- Back startups with a real defined focus and niche and not broad unfocused one
- Founders that tout NLP as core competency are probably lying: NLP is a bit iffy in many ways, or at least incredibly onerous to apply. There are verrry few people in the symbolic camp
- AI implementations will be focused on incremental improvements in the short to mediumish term… but mediumer term there is some transformative potential.