In the world of Harry Potter, the sorting hat serves as an algorithm that takes data from a student’s behavioral history, preferences and personality and turns that into a decision on which Hogwarts house they should join. If the real world had sorting hats, it would take the form of machine learning (ML) applications that make autonomous decisions based on complex datasets. While software has been “eating the world,” ML is starting to eat software, and it is supercharging trillion-dollar global industries such as healthcare, security and agriculture.
If ML is expected to create significant value, the question becomes: where will this value accrue? I will explore ways that value will be created and captured by three types of companies: traditional companies applying ML, companies building industry-agnostic ML tools and companies building vertically-integrated ML applications.
Machine learning is not just for the tech giants
ML innovation coming out of Facebook, Amazon, Apple, Netflix and Google (FAANG) is well known, from news feeds to recommendation engines, but most people are not as aware of the increasing demand for ML from traditional industries. Global spending on AI systems is projected to reach $98 billion in 2023, over 2.5x the amount spent in 2019, with financial services, retail, and automotive leading the way. Blackrock, an investment management firm with over $7 trillion in AUM, released several ML-powered ETFs in 2018. ML has rapidly gained mindshare in the healthcare industry, and budget for ML-driven solutions spanning medical imaging, diagnostics and drug discovery is expected to reach $10 billion in the next three years.
Across these enterprise customers, three broad customer segments have emerged: software engineers, data scientists and business analysts, sometimes known as “citizen data scientists”. Although business analysts are less technical by training, they comprise a large and growing segment of users who are applying ML to help companies make sense of their multiplying data repositories.
Machine learning tools are embedded across industries
To accommodate these customer segments, companies looking to craft pickaxes for the gold rush have proliferated. “The challenge is not to make ML transparent but rather to make the painful parts like logging, data management, deployment and reproducibility easy, then to make model training efficient and debuggable,” said Stuart Bowers, the former VP of Engineering at Tesla and Snap.
Incumbent vendors, most notably the public clouds, have adopted an “end-to-end platform” approach as part of their strategy to sell more infrastructure services. AWS’s ML platform, Sagemaker, was originally intended for expert developers and data scientists, and it recently launched Sagemaker Studio to expand the audience to less technical users. For tech giants like AWS, selling ML tools is a means to drive additional infrastructure spend from its customers, meaning they can afford to offer these tools at a low cost.
Unicorns have also built value, often in partnership with the cloud providers. Databricks, an ML platform known for its strong data engineering capabilities built on top of Apache Spark, was founded in 2013 and is now valued at $6.2 billion. The partnership between Databricks and Microsoft enables Microsoft to drive more data and compute to Azure while massively scaling its own go-to-market efforts.
However, enterprise practitioners are starting to demand “best of breed” solutions rather than tools designed to nudge them to buy more infrastructure. To address this, the next generation of startups will pursue a more targeted approach. In contrast to the incumbents’ broad-brush platform plays, startups can pick specific problems and develop specialized tools to solve them more effectively. Within the ML tools space, three areas pose significant challenges to users today.
While ML results can be elegant, practitioners spend most of their time on the data cleaning, wrangling and transformation parts of the workflow. Because data is increasingly scattered in different formats across multiple machines and clouds, it is difficult to engineer the data into a consumable format that teams can easily access and use to collaborate.
To solve this, Mike Del Balso, the co-founder and CEO of Tecton, is democratizing the best practices he championed at Uber through his new startup. “Broken data is the most common cause of problems in production ML systems. Modelers spend most of their time selecting and transforming features at training time and then building the pipelines to deliver those features to production models,” he noted. Tecton simplifies complexity in the data layer by building a platform to manage these “features” – intelligent, real-time signals curated from a business’ raw data that are critical to operationalizing ML.
Further upstream, Liquidata is building the open source GitHub equivalent for databases. In my conversation with Tim Sehn, Liquidata’s co-Founder and CEO and the former VP of Engineering at Snap, he emphasized that “we need to collaborate on open data, just like with open source software, at Internet-scale. That is why we created DoltHub, a place on the internet to store, host, and collaborate on open data for free.”
Experiment tracking & version control
Another common problem is the lack of reproducibility across results. The absence of version control for ML models makes it difficult to recreate an experiment.
As Lukas Biewald, co-Founder and CEO of Weights and Biases, shared in our interview, “today, the biggest pain is a lack of basic software and best practices to manage a completely new style of coding. You can’t paint well with a crappy paintbrush, you can’t write code well in a crappy IDE (integrated development environment) and you can’t build and deploy great deep learning models with the tools we have now.” His company launched an experiment tracking solution in 2018, enabling customers like OpenAI to scale insights from a single researcher to the entire team.
Building the infrastructure to scale model deployment and monitor results in production is another critical component in this maturing market.
Anyscale, the startup behind the open source framework Ray, has abstracted away the infrastructure underlying distributed applications and scalable ML. In my conversation with Robert Nishihara, Anyscale’s co-Founder and CEO, he shared that “just as Microsoft’s operating system created an ecosystem for developer tools and applications, we are creating the infrastructure to power a rich ecosystem of applications and libraries, ranging from model training to deployment, that make it easy for developers to scale ML applications.”
Scalability is also rapidly advancing in the field of natural language processing, or NLP. Hugging Face established an open source library to build, train, and share NLP models. “There has been a paradigm shift in the last three years, whereby transfer learning for NLP started to dramatically change the accessibility and accuracy of integrating NLP into business applications,” said Clément Delangue, the company’s co-Founder and CEO. “We are making it possible for companies to apply NLP models from the latest research into production within a week rather than in months.”
Other promising startups include Streamlit, which allows developers to create an ML app with just a few lines of Python and deploy it instantly. OctoML applies an additional intelligence layer to ML, making systems easier to optimize and deploy. Fiddler Labs has built an Explainable AI Platform to continuously interpret and monitor results in production.
To build long-term durable companies in the face of stiff competition from incumbents, startups are asking themselves two questions: To which set of customers am I indispensable? What is the best way to reach these customers?
Many startups pitch the idea of capturing 1% of a large market, but often these big markets are already well-served, if not crowded. Companies focused on winning a core customer segment end up exhibiting strong early traction that translates into long term expansion potential. To reach these customers, most incumbents like Databricks and Datarobot have embraced a top-down, enterprise sales motion. Similar to what we’ve seen in the developer tools space, I expect ML startups will eventually evolve from pure enterprise sales to drive bottoms-up adoption and gain an advantage over today’s enterprise-focused incumbents.
Vertically-integrated machine learning applications are upending the status quo
Some of the most exciting companies in ML are pioneering business models to disrupt entire industries. Auto has been the most obvious example, as $10 billion of funding poured into the industry in 2019 alone. The next generation of verticals where ML will also have a revolutionary impact include healthcare, industrials, security and agriculture.
“ML is most effective when it’s ML plus X,” said Richard Socher, the Chief Scientist at Salesforce. “The best ML companies have a clear vertical focus. They don’t even call themselves an ML company. ” He points to healthcare as a uniquely promising area: Athelas has applied ML to immune monitoring, helping patients optimize drug intake by collecting data on their white blood cell count. Curai leverages ML to augment the efficiency and quality of doctors’ recommendations, allowing them to spend more time treating patients. Zebra and AIdoc empower radiologists by training datasets to identify medical conditions faster.
In the industrials and logistics space, Covariant is a startup that combines reinforcement learning and neural networks that enable robots to manage objects in large warehouse facilities. Agility and Dexterity are similarly building robots that adapt to unpredictable situations in increasingly sophisticated ways. Interos applies ML to evaluate global supply chain networks, helping enterprises make critical decisions around vendor management, business continuity and risk.
Within security and defense, Verkada has reimagined enterprise physical security by intelligently analyzing and learning from real-time footage. Anduril has built an ML backbone that integrates data from sensor towers to augment intelligence in the interest of national security. Shield AI’s software allows unmanned systems to interpret signals and act intelligently in the battlefield.
Agriculture is another vertical that has reaped enormous benefits from ML. John Deere acquired Blue River Technology, a startup that developed intelligent crop spraying equipment. “We are changing the world of agriculture by bringing computer vision techniques to identify individual plants and take action on a plant-by-plant basis,” said Lee Redden, Chief Scientist of the combined company’s Intelligent Solutions Group. Other notable enterprise AgTech companies include Indigo, which applies ML to “precision farming,” harnessing data to produce food more profitably and sustainably.
Where do we go from here?
ML has quietly become part of our daily lives, powering our cars, the operations in our hospitals and the food we eat. Large incumbents have pioneered the state-of-the art so far, but the real promise lies in the next wave of ML applications and tools that will translate the hype around machine intelligence from a Harry Potter-like fantasy into tangible, societal value.
There are many reasons to be optimistic about the value ML can create in the coming years. Traditional companies will train millions of citizen data scientists to reshape broken industries into more productive ones. ML tools will lower the barriers to building intelligent applications, pushing millions of new ideas into production every day. Vertical ML business models will democratize access to healthy food, reliable physical security and affordable healthcare.
That’s where we’ll find the true value of machine learning.