Machine learning deployment

In 2012 or so, if you’d asked most people in tech about ’neural networks’, if they had any answer at all they might well have said that it was an obscure idea from the 1980s that had never really worked - rather like VR. Then, in 2013, Imagenet gave us an explosive realisation that this could work now - again, rather like VR in 2013. Since then, the tech industry has been remaking itself around machine learning. There’s a naive view that ‘Google will have all the data’ or China will have all the AI’ or ‘Data is the new oil’, but it’s more interesting to look at how many different kinds of deployment are now happening. 

The first phase was the creation of companies building platforms (or ‘primitives’ or ‘substrates’) for specific low-level ML applications - image recognition, voice recognition, sentiment analysis etc. Some of these have thrived but most of this space has, I think, been subsumed into larger platforms that offer many other primitives as well, such as storage, database or computing - AWS, Google Cloud or Microsoft’s Azure. 

Second, pretty much every other tech company has picked this up and wandered around their products to see where it could be applied. This often reminds me of the old joke about the man with a hammer who thinks every problem is a nail - the tech industry has been hitting everything with a hammer to see if it’s an AI problem, or can be made into one. Generally, it is. Hence, Everlaw (an a16z portfolio company) might use machine learning to do sentiment analysis or find similar documents, as part of a much broader product offering. A security company uses ML to look for weird transactions. At the extreme, if you take a photo on your smartphone the photo app will now extract the text - that’s ‘AI’, but it’s also just text. And, of course, smartphone cameras themselves are now as much machine learning as they are hardware image sensors. That is, machine learning is being absorbed by existing companies and products, often invisibly. 

Third, there is a wave of companies that are created to use machine learning to solve complete problems in new ways. People.AI uses NLP to analyse text going through Salesforce and identify parts of the pipeline that might be going wrong. Descript uses speech recognition to let you edit audio of people talking with a text editor instead of snipping up wave forms. These aren’t really ’AI’ companies anyone - they’re companies with an actual product that solves a specific problem. Descript doesn’t go to Tony Hall, head of the BBC, and say ‘let us tell you how [deep voice] AI (!) can transform your business!’. Instead, it goes to radio producers and says ‘this tool can can save you some pain every day’.  There are lots of really cool new companies like this, generally solving a very specific problem you didn’t know existed in an industry you know almost nothing about, and very often that doesn’t look like an ‘AI problem’. Quite often, they’re also using vision to solve something that doesn’t look like a vision problem, or audio to solve something that doesn’t look like an audio problem. 

It’s also important that these companies don’t have to build the whole thing themselves. The fact that voice recognition was subsumed into AWS, Gooogle Cloud and Azure means that Descript can leverage that to build an actual product on top, just as the fact that storage was subsumed into AWS means that startups don’t need to spend millions of dollars on storage arrays before they can launch. 

That is, part of the consequence of the growth of giant tech platforms is that they’re platforms, and they enable a wave of company creation, and that now applies to AI as to anything else. We stand on the shoulders of giants. While it might not be a good idea to create a generic low-level voice recognition platform in competition with AWS, AWS is never going to build Descript. 

Fourth, these technologies are now diffusing far beyond the tech industry. Part of the a16z model is that we spend a lot of time hosting big industrial companies to meet our portfolio, and they tell us what they’re working on, and most of them now have lots of ‘AI’ projects. My favourite example recently was the manufacturing company that wants to check its product for a specific defect as it goes down the line - they could never automate that, but now they have a neural network on a DSP with a smartphone image sensor on a stick over the production line. It looks at each unit and says ‘defect/no defect’. Google doesn’t have that data, and neither does China. Moreover, this was built for them by one of the big outsourcing consultancies. This isn’t rocket science that only Google can do. It’s just software. 

Then, of course, people pick these up and unbundle them. Drishti takes that camera on a stick and builds telemetry, metrics and analytics as a complete solution. But is that ‘AI’, or computer vision? Or, say, an industrial process optimisation company? (And no, Google will definitely never build this). 

This is always the life-cycle of the deployment of new technology. I often compare machine learning to relational databases - 40 years after they were an exciting new idea, every big company today has dozens or hundreds of databases from dozens of different suppliers. Some of these are highly specific and you might build them yourself, some are specific to your industry and also used by your competitors, and some of them are completely generic. 

This is probably also a useful way to think about what’s happening more generally. One of the recurring conversations in Silicon Valley is to wonder what the next ’S Curve’ is - we had PCs, then the internet, and then smartphones, but smartphones are boring and well-understood now, so what’s the next rocket ship? We often talk about machine learning and crypto here. But, I’d suggest we should actually look at another time series - over the past few decades we moved through databases, ‘productivity’, client-server, open-source, SaaS and Cloud. In parallel with new client platforms, we had new waves of architecture or development model, and that’s really a better way to look at machine learning - ML is the new SQL (and maybe crypto is in part the new open source). And so if you want to know ‘what’s our AI strategy?’ or ‘how do we choose an AI vendor?’, the answer is, well, how did you choose a cloud vendor or a SaaS Vendor, and how did you identify opportunities for databases?