🎙Albert Azout/Level Ventures on the state of AI market and the areas to pay attention to
No subscription is needed
Since Money in AI is one of our most popular sections in the Sunday Scope, we decided to add an investor’s perspective to our interview series. Albert Azout has a unique knowledge and experience: started as a software engineer, becoming an ML-practitioner, than an ML founder with a few successful exits, now he observes and explores the AI space as an active venture capitalist. Currently, Albert is the founder and managing partner at Level Ventures, as well as venture partner at Cota Capital. You can ask Albert a question in the comment section below. Please share these interview if you find it insightful.
👤 Quick bio / Albert Azout
Tell us a bit about yourself. Your background, current role and how did you get started in machine learning (ML)?
Albert Azout (AA): I fell in love with computers and coding when I was 12 years old. After college in Boston, I applied to be a software engineer at Lehman Brothers on Wall Street. I was accepted and, incidentally, my first day of work was on September 11, 2001 (and my office was on the 40th floor of the World Trade Center).
I spent 2.5 years at the firm, coding large-scale enterprise Java systems. Early on, I was focused on reporting and analytics, and towards the end I began working on a trading engine, which aggregated price and macro data for trade signaling. At that point, I became fascinated with quantitative finance, which inevitably led to time-series analysis, and then to statistical learning. I decided that I would rather be building companies, and in 2005 I left the firm to start one on my own.
In 2007, I co-founded a company called StyleCaster, which became one of the largest vertical ad networks (now owned by Penske Media). That was my first attempt at raising venture capital. The company’s core thesis was to drive user engagement and CPMs by recommending and personalizing content at large scale. Around that time, I remember reading a book called Social Network Analysis by Stanley Wasserman. It had a tremendous impact on me. As part of our work on this large-scale recommendation system, I was exposed to collaborative filtering, unsupervised learning and clustering, and large-scale matrix factorization problems. We decided to spin out the core technology into a company called Velos. At Velos, we developed a proprietary set of algorithms which enabled clustering and collaborative filtering at scale. By formulating the problem as a message-passing algorithm and forcing the data workloads into a MapReduce setting, we were able to cluster billions of data points very quickly. Finding that selling recommendation systems was a hard pitch, we eventually generalized our thinking by building data pipelining for “predictive analytics” workloads. This eventually led to building self-service ETL and feature engineering orchestration on Spark (version 1.0), and consequently going very deep on (shallow) machine learning and large-scale inference.
We sold the company to Verizon/AOL in 2015 and I moved from New York City to Palo Alto, where I worked as a Senior Director of Product for Verizon/AOL for about 1.5 years.
In 2016, I joined Cota Capital as a partner. Cota is focused on enterprise tech investing multi-stage (public and private). I spent 4 years there, helping build the firm, which today has over $1B in AUM. While there, I made many investments in application-level and deep technologies (robotics, advanced computing, life sciences, AI, etc).
During COVID, I decided to move back to the east coast and started Level Ventures, a next-generation, tech-enabled fund of funds. By ingesting large-scale venture capital (and other) data, we are building the largest knowledge graph in venture capital, which allows us to evaluate firms and companies using a proprietary algorithmic approach. Our vision is to democratize access to the rapidly changing VC seed ecosystem.
🛠 ML Work
You have been both an ML-startup founder and an active venture capitalist in the space. How different is to build an ML startup today than 10 years ago?
AA: Things have certainly changed. When I started, enterprises were mostly focused on data infrastructure. They were pushing data from data warehouses to HDFS to be used for MapReduce analytical workloads. The notion of a data engineer did not exist. It were IT engineers who were developing these ETL pipelines. The community around downstream predictive analytics was nascent. Every pipeline an enterprise created was mostly unique, without much tooling or open-source libraries to rely on.
Today is very different. As a data engineer or even a data scientist, you can get from data to training easily. The pipelining is there, at the data/storage level, at the distributed computing and systems level, and at the machine learning training and inference level. This democratization of ML forces value creation to happen at the application level, which I think, generally, is good for the world.
I think ML is one of the trends that have defied the principles of the Innovators Dilemma given that a lot of innovation is coming from big AI labs, which have a lot of data and talent resources. How can ML startups compete with big technology incumbents in these circumstances?
AA: I think it is close to impossible to build a horizontal AI infrastructure in today’s market. A company cannot compete with the resources of large incumbents and cloud service providers. At the application level, every company will become an AI company, as software evolves from being coded by humans to being coded by data. I think application-level AI companies compete on orthogonal vectors (i.e. go-to-market, proprietary data moats, network effects, etc) and implementing the best-in-class ML infrastructure to ensure fast iteration, non-brittle feature engineering pipelines, model stability, seamless model deployment, etc. AI is just software fed by data. I also believe the next generalization of SaaS companies will be workflows embedded within marketplaces. Applying AI as the intermediary in those settings (and also in consumer settings) will be differentiating. The last thing I would add is that AI is more than just pattern recognition and function approximation. There are areas that need to be worked on, including common-sense reasoning, program induction, etc.
A common debate in the venture circles is whether to invest in ML infrastructure or applications. If we can fast-forward five years (not 2, not 10), what area do you think will produce more value: ML infrastructure technologies like MLOps or ML-applications that leverage existing infrastructure?
AA: Again, at the moment I don’t believe investing in horizontal ML infrastructure is a durable investment strategy. But in reality, timing is everything. The market oscillates between infrastructure-level innovation and application-level innovation. The surface area of enterprise workflow have not completely been automated and digitalized, but that is inevitable. The next wave will incorporate AI and automation more aggressively: to reduce reliance on repetitive human tasks, optimize business objectives, streamline collaboration, etc. These use cases will in turn require improvements to horizontal AI infrastructure. I would be looking for those second-order opportunities (which of course takes some vision).
What are new areas of deep learning that you are excited about as a venture investor?
AA: The area I am most excited about is AI at the intersection of science and technology, and specifically deep technologies and life sciences. AI applications are fed by data, and new data sets are emerging in life sciences, given the continually dropping cost of gene sequencing, maturation of gene editing, microfluidic technologies, and new sensors. These areas will have tremendous impact on our healthcare systems but will also positively impact our environment. Quick stat: as much as 60% of physical inputs to the global economy can be, in principle, produced biologically (i.e. synthetic biology). This is a huge opportunity for AI techniques to discover compounds in silico and automate high-throughput manufacturing (via robotics). In general, I look closely to the areas where new data sets are being created.
ML startups have seen an explosion of venture investing in 2020-2021 and, yet, the space is very nascent and not many startups have achieved relevant traction. Is the space overhyped and are we headed for a Series B-C crunch in the ML space, when many startups would not be able to complete new rounds of fundraising and would be acquihired for low valuations?
AA: Generally, venture returns are power-law distributed, where few companies attract the most of the returns. In addition, venture capital sub-sectors are influenced by innovation and liquidity cycles. Having a portfolio-centric approach and timing investments properly is of utmost importance, generally. The worst investors are reactive and thematic, chasing hype. AI is in one of those hype cycles. It is likely that the great AI infrastructure companies have already been mostly created. The flood of capital and mark-ups that we see coming to this over-heated market are false positives. I believe there will be both a macro crunch as well as a sub-sector crunch.
TheSequence is the #1 AI Newsletter on Substack. Subscribe to our Premium newsletter and become more knowledgeable about ML and AI every week.
💥 Miscellaneous – a set of rapid-fire questions
Is the Turing test still relevant? Is there a better alternative?
AA: Having an agnostic third-party decide whether someone is human or not-human based seems to be generally irrelevant to me. Our consciousness is really a generative thought model—a black box. We can never predict our next thought, though we can choose to attend to some thoughts and not others. Our thoughts are tightly coupled with emotions, and inevitably with our actions. What makes something conscious at the lowest level is its ability to sense (and possibly actuate). I think we can create machines that sense and can communicate with language, and thus would be able to pass the Turing test. The harder question is whether we can create systems which evolve thought from more simplistic primitives.
Favorite math paradox?
AA: I love the Liar’s Paradox. But I am lying.
Any book you would recommend to aspiring data scientists?
AA: I think Elements of Statistical Learning is an amazing book. I also like Theory of Complex Systems, which discusses systems and complexity.
Is P equal NP?
AA: I have always felt that logical verification is an artifact of an elegant universe. A world in which deep mysteries can be subjected to verification likely does not require those same mysteries to be solvable by the same machinery. So I would vote for P not equal to NP.