Phillip Liu is a pioneer in cloud infrastructure; he built his career at early Saas leaders like Marimba and Opsware, helped Facebook scale in its early hypergrowth years, and leveraged his expertise to co-found SignalFx, where he built one of the most successful platforms to ensure reliable and performant cloud infrastructure.

Zetta couldn’t be more thrilled to partner with Phil’s next pursuit: building the intelligent trust platform for technology partners.

At first glance, you might wonder what “compliance” has to do with Phil’s background in cloud infrastructure, or Zetta’s focus on ML. In fact, they’re deeply interconnected.

Throughout his career, Phil saw that a key challenge of cloud infrastructure is the interdependency of all our systems. Pre-cloud, every part of the stack, top to bottom, was self-contained. The rise of cloud means that for better and worse, our systems adopt dependencies on one another. E-commerce sites rely on payment APIs which rely on Saas APM tools which rely on cloud infrastructure providers. We no longer have a well-defined and self-contained stack — we have a complex web of technology and infrastructure dependencies.

Every technology service depends on third parties; when a service makes commitments around reliability, efficiency, performance, and security, it is committing on behalf of the third parties as well. While this architecture has immense advantages, it exposes every partner in the chain to a degree of third-party risk that didn’t exist in the pre-cloud world; and the more mature and ubiquitous cloud infrastructure becomes, the more complex the web of dependencies.

Compliance has historically been seen as a discipline focused on financial transactions. But the mission-critical nature of our reliance on third-party technology systems, and our customers’ reliance on our own systems — has made technology a core target as well. In the financial world, compliance is all about ensuring adherence to financial best practices and obedience to tax and other legal regulations. On the technology side, compliance is increasingly focused on R&D and DevOps best practices, as well as data privacy and security laws — and customer commitments! So now compliance is the CTO’s focus, not just the CFO’s.

The SOC 2 has become the de facto standard in Saas technology best practices. It’s a priority even for pre-product-market-fit tech startups because they can’t onboard enterprise pilots and design partners without it. That makes it the compliance entry point for most new tech companies, and a vibrant ecosystem of technology and services firms has been growing quickly to serve the growing demand.

The first generation of approaches are all about the compliance workflow: libraries of content and checklists, all aimed at making a mysterious process simple and easy for humans to execute with as few mistakes as possible, and help the engineering team to assemble the necessary policies and documents needed by audit firms to verify a company’s compliance.

Phillip Liu’s approach with Trustero is the next-generation approach: bringing modern AI techniques to bear on the problem — automating information gathering and more importantly, validation; not just accumulating content, but understanding, assessing and validating it, enabling human auditors to scale to more customers at greater velocity.

The SOC 2 is an urgent need for every tech startup with a B2B business model, but it’s just a starting point. Once a year is frequent enough for a box-checking exercise. But the interdependency of our technology SLAs is 24x7 and year-round; Trustero’s ambition is to build the platform for continuous assurance, an intelligent foundation that enables technology partners to go faster and farther because they can rely on one another. We couldn’t be more proud to support the team on their journey!

Trustero: A Foundation of Trust was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

There are two things that currently motivate me, a strong dislike of terrible software and a passion for creating generative organizational cultures. Over the years I have had the privilege and luck to work on some amazing software in some organizations with remarkable cultures. In this current chapter of my life I’m interested in connecting with teams that are looking to improve the software they deliver while actively creating generative cultures.

Better Software

I realize that ‘better’ is subject to interpretation. So, allow me to provide a brief definition and rubric that I use to assess software and another for culture. To build better software I believe that it starts with measuring the value to the user. Ideally, software enables a user to accomplish a task more quickly and with greater joy than if they were to attempt the same task without the software. While this may seem like a low expectation, I assure you accomplishing this consistently is hard.

One of the greatest challenges is the requirement that we put on software to continually improve. This need for better performance, more features, and greater joy for the users is relentless. Interestingly, the organizations that have learned to do this well perform well on a number of core metrics. We know this thanks to the incredible work of Dr. Nicole Forsgren, Jez Humble, Gene Kim, and the rest of the team at DevOps Research and Assessment (DORA). The rigor they brought to the ‘State of DevOps report’ in 2014 changed the way folks thought about software development metrics . At this point there is general consensus amongst practitioners that the core metrics identified in this report are remarkable predictors of a software delivery organization’s ability to deliver more performant and stable software. For those that are not familiar with DORA the core metrics are:

• Deployment Frequency — How often an organization successfully releases to production
• Lead Time for Changes — The amount of time it takes a commit to get into production
• Change Failure Rate — The percentage of deployments causing a failure in production
• Time to Restore Service — How long it takes an organization to recover from a failure in production

As for user joy, net promoter score (or NPS) continues to be the best leading indicator. NPS can also be evaluated alongside net dollar retention (NDR) for B2B companies and daily/monthly active user (D/MAU) for B2C companies to gain a more complete understanding.

Progressive Improvements

Learning from the teams that have used these metrics to recognize success, I’ve spent the past 5 years helping teams adopt Progressive Delivery as a new software development model. This model aligns the ability to deliver more performant software with the joy and value perceived by the customer. For those familiar with models for the software development life cycle (SDLC), Progressive Delivery is the natural evolution of Continuous Delivery driven by an increased standardization towards a cloud-based Software as a Service (SaaS) model. The two key enhancements with Progressive Delivery are:

Release progression — progressively increasing the number of users that are able to see (or are impacted by) new features.

Delegation — progressively delegating the control of the feature to the owner that is most closely responsible for the outcome.

I’m currently working with James Governor and Kimberly Harrison on a book covering this model.

People are the hard part

On the culture side, there are two individuals that I regularly credit with my passion for culture. The first is Dr. Ron Westrum, for his work and definitions of the types of cultures in organizations. The second is Jocelyn Goldfein, for a wonderful blog post on culture, which I reference regularly. The work from Dr. Westrum is deeply incorporated into the DORA metrics when looking at organizational culture. I had the opportunity to meet Dr. Westrum through my work with Gene Kim and the DevOps Enterprise community. I also had the pleasure of working with Jocelyn early in our careers at VMware.

In both cases, the value that I have gotten from these individuals lies in the eloquence they have distilled from their experience. And in the case of Dr. Westrum, over 30 years of academic research. Dr. Westrum has been able to reduce the types of culture down to three types:

With these behaviors in mind, we can start to understand how the culture of a community or organization formed. Jocelyn makes this very point with a quote she shares in her blog post:

“That’s your culture. Your culture is the behaviors you reward and punish.”

- Charles O’Reilly, Stanford GSB

I like to think of this as expressed as an equation:

Simple, right? Turns out the most complicated part about building software is not the technology or the process, it’s the people. And I find people fascinating.

So, what’s next?

Over the past 10 years I served as an Advisor to over 15 companies including Shape Security (acquired by F5), Lightstep (acquired by ServiceNow), Turbine Labs (acquired by Slack), Buddybuild (acquired by Apple), Weaveworks, Kong, Sourcegraph, CODE2040, /dev/color, and others. The engagements always fall into these two areas of interest, building better software that enables the success of customers and building better leaders that enables the success of their teams.

The consulting names for these activities are typically GTM (or go-to-market) and executive coaching. GTM spans across Product Management, Marketing, Sales, and Customer Success. All of these functions and roles contribute to enabling the success of the customer. It also happened that I have held positions in each of these job functions. The executive coaching engagements are more focused on enabling better leadership practices. Too often in our industry, folks in management roles have never been trained as leaders. This often leads to problematic behaviors being rewarded and praised and behaviors that support psychological safety are discouraged or shunned. I believe that the job of a leader is to enable the success of others. By working with executives to view leadership through this lens I’ve been able to help teams adopt practices leading to more positive cultures.

This past year I’ve begun to expand the reach of my advising by working with early-stage VC firms that share my passion for building both great products and great cultures. Last fall I was excited to start work with Zetta Venture Partners as an Advisor-in-Residence and Vertex Ventures as a Vertex Fellow. These roles allow me to meet new start-ups to advise in building better software and remarkable cultures. It is also a tremendous opportunity for me to engage in my own personal growth to learn more about the current trends in applied AI and ML applications. I look forward to working with members of the portfolios and welcome folks that are interested in better software and culture to reachout on twitter or linkedin.

A passion for software and culture was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

I’m thrilled to share that I’ve joined Zetta Venture Partners to do an Executive in Residence (EIR) stint.

I have been privileged to spend the better part of the last year thinking deeply about what professional endeavors I want to pour my energy into next. Admittedly, venture investing is one of the last places I thought I would find myself, but after much reflection, I’ve decided it’s a priority for me to return to working directly with startup teams that are building AI-driven science and technology products–and this EIR position is a great way to do it. Ultimately, I’m hoping that this tour will help me inform whether I want to commit full-time to founding something myself, join another team doing promising work, or something else entirely.

Zetta is a small firm focused on early stage investments in AI-native companies. They care about mission-driven founders, and the firm’s leadership –Mark Gorenberg and Jocelyn Goldfein–are two veteran investor-operators who I have come to know well and deeply respect for their intelligence, integrity, and care for the teams that they support. I couldn’t be more excited to work alongside them.

Beyond the people, I was also drawn to Zetta for their focus on the path from the lab to the market for leading-edge, applied research in AI and ML; their force of interest in the intersection between innovation, public policy, and responsible development and deployment imperatives; and their commitment to new markets for data-driven technologies that can bring about world-changing applications in arenas such as biotechnology and healthcare, materials science, and climate science–all of which I’ll be focused on in my new role.

I have spent my career building bridges across sectors to enable conscientious innovation because I care about maximizing technology’s ability to advance society and to make a positive impact on the world. I’ve had the privilege to work on related issues from the public sector, from academia, and from civil society. And I’ve invested most of my time in early-stage organizations, helping to build new ventures or support the growth of new teams working on boundary-pushing projects crossing public and corporate policy, product development, and research. My work with Zetta is an extension of that journey, which will provide me an opportunity to understand the power of influence of venture investment, and support and affect how early organizations are built in the private sector.

During my time helping build and lead the Partnership on AI in the last several years, we worked hard to identify levers for change to impact more effective, safe, responsible, equitable AI in collaboration with researchers, product developers and designers, and policymakers. The public interest technology community and civil society are key constituencies in these efforts. So are the individuals building products every day which go to market and impact the world in very real ways–and those funding them. I found these levers often overlooked in my previous work. Important decisions about values, market fit, and other key aspects of product and organizational design are ground-floor conversations in early stage companies, and I’m so excited to be working with founders who are thinking about these questions from the outset. Now more than ever, it’s important to support the next generation of companies to grow responsibly to compete with tech incumbents, and to advance innovation in high-impact sectors that have an opportunity to be progressed by capabilities supported by AI.

I’m looking forward to sharing more of my learnings with all of you along the way. Until then, feel free to get in touch at terah@zettavp.com.

Terah Lyons Joins Zetta Venture Partners was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

As artificial intelligence moves out of the lab and into production, the infrastructure, tooling, and workflows to deploy and manage machine learning systems — a category often referred to as MLOps — has become one of the most exciting frontiers of innovation. MLOps exists because the iterative and data-intensive nature of the machine learning process demands an alternative to the traditional DevOps tools built in the SaaS era. Machine learning practitioners need a modern, deliberate, and industry-hardened workflow.

Many world-class teams have emerged to form this new, MLOps category. Some, like Zetta portfolio company Domino Data Lab, are building the end-to-end workbench for data scientists. Others are building tools that address a specific pain-point along the machine learning pipeline, like DVC, in data versioning; Weights & Biases, in experiment tracking; Kubeflow, in model orchestration; and BentoML, in model serving.

Of all the problems to solve in MLOps, operationalizing machine learning features is one of the most acute. Practically speaking, features are the data streams that enable a predictive model, and a feature store is the software that is used to manage, transform, and serve them to that model. Data, of course, is the primitive element in AI, and the feature store layer is such a crucial foundation that large technology companies like Linkedin, Uber, and Airbnb have each devoted meaningful resources to building and maintaining in-house feature stores for their sprawling data science teams. These features stores are almost always built in the context of a larger, end-to-end ML platform, like Uber’s Michelangelo.

Machine learning is rapidly diffusing outside of hyper-scale consumer technology companies, both into other types of technology companies, and across global economic sectors like financial services, healthcare, and real estate. The demand for feature stores has grown in lock-step, yet there are few workable solutions for organizations without near-unlimited data science and machine learning resources. Designing a feature store for this large and under-served market is a major challenge in MLOps that has yet to be solved.

Enter Simba Khadder. We first met Simba as the CEO of Triton, an analytics company building recommender systems for subscription businesses, including some of the nation’s largest publishers like The Wall Street Journal. Like other applied AI companies, Simba and his team at Triton found that their machine-learned content recommendations improved dramatically when they incorporated a richer set of features into their models.

At Triton’s peak, Simba and his team processed the behavioral data of over 100 Million Monthly Active Users (MAUs) through their system. To perform at this scale, and to operationalize their ever-expanding feature set, Triton was forced to build several foundational MLOps technologies that simply weren’t available in the marketplace at the time, including a production-grade feature store.

Simba quickly realized the potential impact his new feature store product could have, and started a company to amplify its reach. We see the same opportunity that he and his team do, and are privileged to support Simba’s new company, Featureform, through its next phase of growth.

Simba’s real-world experiences building ML systems for a diverse set of companies led Featureform to the unique product experience that it boasts today: a virtual feature store. Featureform won’t ask you to rewrite thousands of existing features already in production, or retool your existing data infrastructure. Featureform sits above the infrastructure layer and provides data scientists a standardized way to define, manage, and share features. Adhering to its set of user-centric design principles has allowed Featureform to begin serving the needs of the fastest-growing enterprise customers in the market.

Simba and his team have established themselves as an important and rising voice within the global MLOps community, and they’ve assembled a small group of seasoned practitioners who built and scaled some of the most important technology companies of the last decade.

We’re thrilled to share that Featureform has launched from stealth and announced the open source release of their first product: the Featureform Embedding Store. We couldn’t be more excited to support Simba, Shab, and the rest of the team on their mission to standardize and accelerate the machine learning process. If you’re a little intrigued, check out the embedding store and give us some feedback. If you’re a lot intrigued, Featureform is hiring.

Featureform: The ML Feature Store was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

The last three decades have seen office work transformed by successive major technology platform shifts: the desktop computer, the internet, the cloud, and mobile. At Zetta, we believe AI is the next generational shift, with the potential to power new tools and automation in almost every industry. That’s one reason we were excited to meet Skan, a company that seeks to better understand the nature of work itself with the help of computer vision and machine learning.

“Information work” — task-based work done by humans with computers, like data processing — has been frequently outsourced (the Business Process Outsourcing market was over $200B in 2019.) This kind of work is also a favorite target of automation with RPA or special purpose machine learning tools. But offshoring and automation are not goals in and of themselves — they are a means to the end of improving efficiency, reliability, and speed. But not all processes are well suited to automation, and stories of failed automation are all too common. If we can’t clearly describe the work, how can we possibly expect to automate it — or even to know if automation is the right prescription?

Human workers are creative and resourceful, which makes them good at performing complex tasks and adapting to exceptions and changing conditions. That’s great news for delivering results, but it means that the real work being done by office workers can diverge sharply from the documented process, even for supposedly simple tasks like processing an insurance claim or sending an invoice. When Skan’s founders worked with big enterprises on digital transformation projects, they found that businesses didn’t really have a handle on what their team members had to do to accomplish their work. Even enterprises with sophisticated process mining technology didn’t have a clear understanding. Examining the logs in your system of record shows you what a human entered into the system, but doesn’t show you the ingenuity and tribal knowledge that went into achieving that result.

Skan provides (privacy assuring) instrumentation that helps customers map out what people are actually doing when they are doing their jobs. With that insight, enterprises are finally equipped to make better decisions. Let me show you what I mean.

Skan’s telemetry provides a graphical view of the work being done, the time spent, the interdependencies between tasks, and more. Now we can see bottlenecks. We can see what’s happening with “outliers” — instances of a task that take much more time than average. We can also see if shortcuts are being taken that might risk violating SLAs or compliance with regulatory rules. We can see redundancy or opportunity for parallelization.

This kind of granular view into the work itself enables enterprises to make informed decisions about process design, organizational design, adoption of tools, and yes, outsourcing and automation.

Skan achieves this with computer vision and machine learning — systems that watch pixels on a screen and abstract that raw input into a higher level understanding of work tasks situated in a context.

Without a product like Skan, enterprises make their best effort to figure this out, typically by hiring consultants to sit and watch employees do their jobs. It’s manual, slow, error prone, and based on much too sparse a data set. And the results are sub-optimal. One CIO told me, “I can’t imagine doing a digital transformation project without this. It would be like driving full speed down the highway with snow all over my windshield.”

Too many AI partisans advocate automation across the board. Skan and Zetta think this is exactly backwards. You have to start with a deep understanding of work, and only then does it make sense to ask where automation has a role to play, and how to implement it.

Skan’s founders are Avinash Misra and Manish Garg, two industry experts who led large scale enterprise digital transformation projects at Genpact. They predict that the future of work will not be a rigid separation of work done by humans and work done by robots, but rather a mosaic of humans, tools, and automation, working in tandem. We’re inspired by this vision and are excited to be supporting them on their journey!

Skan: Visualizing the Future (of Work) was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

At Zetta, we’ve been following the rise of enterprise machine learning since our founding in 2013. As the first focused fund in the area, we’ve been fortunate to partner with a number of the pioneering AI-first enterprise startups, which has given us a front row seat to the first wave of enterprise AI adoption. Over the past seven years, we’ve seen machine learning go from obscurity to curiosity to delivering real business value. But machine learning has a long way to go before it can be widely deployed and fully utilized by businesses: production-grade systems are expensive to build, hard to operationalize and vulnerable to a range of threats we are only just starting to understand.

In her famous book on technological revolutions, Carlota Perez described the tendency of new technologies to follow similar s-shaped development cycles, split into two distinct periods. The first, which she calls the installment phase, is marked by rapid technological development, heavy investment and hype which leads to bubbles and recoveries but ultimately paves the way for a second period which she calls the deployment phase in which the technology is more widely adopted and real value is created. While we haven’t seen a full collapse and recovery cycle in machine learning, we seem to have arrived at a similar inflection point in which both the promise and current perils of the technology are becoming clear. If the last decade was about getting AI to work, this one will have to be about getting AI to work for people and businesses.

To get to the deployment phase, machine learning will have to overcome some key technical limitations and tackle the operational and strategic obstacles standing in the way of broader enterprise adoption. At Zetta, we believe these challenges are among the biggest opportunities that startups can take on. As we kick off the new decade, here are some of the areas we’re most excited about.

Making AI Cheaper and Easier to Build

The first wave of real-world machine learning shows how hard it is to build intelligent systems at enterprise scale. The cost of AI projects have soared with some estimates suggesting that up to 87% of projects never even make it to production. That explains why nearly two-thirds of all AI spending is being funneled to consultants and cloud vendors. Companies looking to build internal capabilities are facing intense competition for data scientists who are commanding higher salaries and churning more often than ever. The imbalance between supply and demand is a hole that won’t be plugged by talent alone, but will require taking aim at the drivers of cost and complexity that plague machine learning today.

Data Quality

Outside of talent, data is the first and most visible cost center for AI projects. While many businesses sit on troves of proprietary data, it’s rarely in a form that is conducive to AI. As a result a multi-billion dollar data preparation market emerged to help businesses unify, enrich and annotate their data for the purpose of training machine learning models. Today, this work is highly manual, dominated by a mix high-end consultants wrangling data and low-cost labor doing annotation. The high costs and explosive demand for these services, especially in markets like autonomous vehicles, has led to some eye-popping valuations in the private markets while underscoring how unsustainable these kinds of services are likely to be in the long term.

Making machine learning accessible to businesses will require new, more efficient ways of improving data quality. Better, more automated tools for data cleaning and enrichment will be important but the cost of data annotation or labeling is by far the biggest obstacle to overcome. A range of new tools are making annotation more efficient by reducing the amount of labeled data a model needs to train. Techniques like active learning highlight the most impactful datum, allowing labelers to focus on the highest yield examples. Similarly transfer learning allows data scientists to train their models on the learnings of another. For smaller or more sensitive datasets, augmenting training sets with synthetic data has been shown to boost performance while improving privacy.

However, the biggest hope for less laborious labeling is in semi and self-supervised learning; techniques that use the structure of raw data to infer labels or the output of one model to train the inputs of another. Some of the leading AI researchers, like Facebook’s Yan Le Cunn, believe these techniques will cover a lot more ground than today’s supervised methods; pointing to the fact that most human learning is gained through experience and reasoning rather than explicit instruction. Compared to supervised learning, these techniques are in their early days but open projects like Snorkel out of Stanford are bringing them closer to real-world applications.

Developer Tools

The tools available to data scientists are in their infancy and so even with the right data and team in place, building production-grade machine learning is needlessly complex. Today, data scientists spend far too much of their time building and maintaining infrastructure, stitching together disjointed tools and building piecemeal pipelines to move models from development into production. Many of today’s most widely used tools are showing their age and failing to meet the challenge posed by new kinds of data and system architectures as machine learning has become more performant and real-time. This has led some of the big tech companies to develop their own internal tools, leaving others to scale back or shelve their more ambitious projects.

To bring machine learning into the enterprise era, we’ll need a new suite of tools for data scientists that are more tightly integrated, production oriented and capable of managing real-time data and distributed systems. Collaborative, cloud-native notebooks are an important starting point to bring more of the data science workflow online and make it easier to connect the different pieces of the pipeline. Better tools for model training, tuning and testing are sorely needed as well. Lastly, we need smarter tools for deploying models across distributed architectures and managing complex infrastructure at scale, like Project Ray out of Berkeley.

While tools like Ray are making data scientists more productive, another class of tools like Google’s AutoML are trying to take AI development out of the hands of data scientists all together, allowing non-engineers and experts to build intelligent models. A range of tools from Google, Oracle and startups like Akkio, Runway and Lob are trying to make machine learning more accessible through standalone tools and integrations into existing products like Google Sheets and Photoshop. While these tools are early, they’re being buoyed by the rising popularity of no-code and low-code tools which are opening up enterprise app development to more non engineers.

Making AI More Robust and Safe to Deploy

The last decade of AI research has produced some truly impressive demonstrations but reproducing those results in the real-world has proven more challenging than expected. Under normal real-world circumstances, models can behave in unexpected and sometimes dangerous ways while also being vulnerable to intentional attacks. The opaque nature of deep learning makes it difficult to predict machine behavior and understand why specific predictions were made. Together, these factors have made it difficult for businesses to use machine learning in mission-critical applications and in regulated industries where they could do a lot of good. To bring machine learning into the deployment phase we’ll need new approaches to building more robust and secure models that can be safely deployed.

Model Robustness

Since the early days of ImageNet, the brittleness of deep learning has been on display: from harmless classifiers mistaking chihuahuas for blueberry muffins to more disturbing cases involving facial recognition and race. Beyond unintentional errors, researchers have found ways to deliberately manipulate models: designing objects to fool security cameras and making small changes to signs and road markers to steer autonomous vehicles into traffic. It’s unclear if these kinds of attacks have been waged on live systems, but for a lot of real-world applications it’s not worth the risk.

Building robust machine learning is one of the core technical challenges of the field which has become a major focus of research over the past few years. Techniques like adversarial training, make models more resilient by feeding them intentionally mislabeled data or examples with small amounts of noise while other approaches have used generative models, like deepfake producing GANs, to synthesize clean examples from malicious ones. The ability to test model robustness under different circumstances is an important piece of the puzzle around which researchers are starting to build a new class of model testing tools. Long term, researchers hope that more reasoning-based approaches to machine learning will be immune to the brightness of today’s deep learning but if an until then better tools for robust training and testing are sorely needed.

Data Security

In an age of frequent data breaches and growing concerns around users privacy, machine learning introduces a range of new concerns. Training models often involves moving large pools of user data out of secure databases and into local and unprotected environments where it is vulnerable to theft or exposure. But even without the underlying data, researchers have shown it’s possible to infer sensitive information from the model itself; a risk that actually grows with more explainable models. As machine learning has grown in popularity so has the range of different attacks and scope of vulnerabilities, which has made it hard for regulated industries — like healthcare and consumer finance — to adopt machine learning for sensitive applications.

Fortunately, there are a number of new techniques strengthening data privacy in machine learning which we expect to be commercialized over the next few years. The most high-profile, federated learning, allows models to train on local devices without pooling user data in the cloud. Google, who developed the technique, has been using it to power auto-suggestions on the Android keyboard since 2017 to great avail. Other approaches are using cryptography — like secure multi-party computation and homomorphic encryption — to allow models to train and make inferences on encrypted data. An encrypted version of Google’s TensorFlow framework has even been gaining popularity in security circles. It’s not yet clear how well these solutions will scale or the kinds of effects they will have on performance, but it’s obvious that privacy-preserving tools will be a critical piece of machine learning’s deployment phase.

Explainability

For certain critical applications, robust and privacy-preserving machine learning won’t be enough. In areas like diagnosing disease and assessing creditworthiness, understanding how decisions are made has important ethical, business and regulatory implications. The ‘black box’ nature of many machine learning models makes it difficult to interpret algorithmic inference, making it hard to adopt AI in certain industries. These concerns are likely to increase as more countries adopt privacy laws like GDPR which includes an explicit “right to explanation”.

Different approaches to explainability have emerged over the past few years, focusing on different domains and application types. In computer vision, researchers at MIT built a tool that dissects the layers of a neural network and lets users identify the individual nodes in a network responsible for identifying certain features in a scene. More recent work builds on this approach by using generative models to offer plain-language, domain specific explanations accessible to business users. Other approaches, like functional transparency, aim to combine deep learning with statistical and causal modeling by training deep neural networks to uncover and follow causal relationships in the data they are being fed.

These are just some of the areas we’re excited about at Zetta and just a few thoughts on the many things that will have to come together for businesses to truly leverage AI. We’d love to hear what you think and especially what you’re working on!

Machine Learning in the Deployment Age was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

Can the Manufacturing Supply Chain Learn Resiliency From Food Logistics? (Spoiler Alert: Yes, and No.)

Sign up for our newsletter to be alerted when new pieces are live.

The reality of the COVID-19 pandemic hit the U.S. early in the spring of 2020, prompting a wave of panic buying. For a few weeks, grocery shoppers might find empty coolers or bare shelves where they were used to seeing arrays of yogurts, or pastas, or whatever.

The thing is, our supermarkets were restocked nearly as quickly as they’d been emptied. The food supply chain proved far more resilient than the manufacturing supply chain — which is still disrupted with no real end in sight. America’s recent obsession with supply chains encouraged Zetta to reconnect with Elliott Wolf, who was the subject of our first Zetta Bytes Live talk back in July, 2020. We asked for his recent observations in light of that 2020 conversation.

Wolf has a lot of insight into how our food supply overcame those early disruptions. He also has a BS in Math from Duke and an MS in Statistics from Stanford, but his insight into food logistics was acquired on the ground, as Vice President and Chief Data Scientist at Lineage Logistics since 2013.

Lineage is a privately held company that happens to be the largest temperature-controlled warehouse owner and operator in the world. Every year, about over 90 billion pounds of food transit Lineage’s 375+ warehouses. Lineage touches over one-third of all the refrigerated or frozen food in the U.S.

As impressive as those numbers are, a business of warehouses and forklifts and tractor-trailers coming and going might not seem like a natural hotbed of AI and data science. Yet, as Wolf pointed out: “It’s got all of the logistics complexities of an Amazon, plus you have to keep everything cold, which is an exercise in thermophysics. It’s a super science-heavy business.”

Wolf was the first data scientist at Lineage Logistics, which was the first company in its industry to start up such a team. These days, Lineage’s Data Science Team has grown to about two dozen people; has an org chart packed with Stanford, Berkeley, and Harvard grads; and Lineage was #23 overall and #1 in Data Science on Fast Company’s 2019 “Most Innovative Companies” list.

Most people see trucks and warehouses. A few see gradient descents, graph theory problems, Monte Carlo experiments, and — oh, yes — bin-packing problems galore

Most logistics companies are just getting up to speed in data science and AI. Lineage’s Data Science Team was a first in the industry and it’s the largest.

In our experience at Zetta, the problems that established companies have when it comes to adopting AI are not technology problems, per se; it’s not lack of expertise or resources, either. Rather it’s building the cultural bent towards innovation. Lineage Logistics had that leaning early on, even though the company didn’t quite know what to expect; eight years ago, when the leadership recruited Wolf they said, “We hear there’s math in logistics. Will you come take a look?” (We presume that many such conversations are happening right now throughout the manufacturing supply chain.)

One of the first projects for Lineage’s nascent Data Science Team was, as computer science majors would call it, a bin-packing exercise. The goal was to cram as many different-sized pallets into a given number of racks as possible.

“What can math do for you?” he recalled. “It’ll build warehouse space.” What he meant was, we’ll create new warehouse space with math instead of construction equipment. Once built, a refrigerated warehouse is very hard to modify. But the steel racks that hold pallets of foods are relatively easy to tweak. Think of a pallet of Sabra Hummus or a 55-gallon drum of Chick-fil-A honey barbecue sauce as paying tenants, whereas empty air above them or beside them is a vacancy.

“If you can just change the pegs on your Ikea bookshelf according to what your mathematician said and achieve the same effect [as expanding a warehouse] that’s a pretty fundamental economic driver in the industry considering [that the only other way to accomplish that is] to buy land, pay $300 a square foot on refrigerator construction, and then spend millions of dollars a year keeping it at zero Fahrenheit.”

“That’s probably a good illustration of how we operate,” he told us. “The objective is crystal clear and no one can disagree with it, even if the actual mechanics of it are opaque to someone who hasn’t gone to graduate school in math or statistics.”

“The job of Lineage Logistics, which is privately held, is to deploy capital to turn it into more capital; not unlike a venture fund or any other private equity fund. Take warehouse rack design: We throw a quarter of a million dollars at rejiggering the rack heights and it makes us $10 million worth of warehouse. So on a valuation basis, we’ve turned $250,000 into $10 million, which is a trade that anyone at any fund would make all day long.”

When Lineage CEO Greg Lehmkuhl spoke to REIT Magazine last year, here’s how he defined the company’s strategy. “We buy these properties and optimize them by increasing occupancy and density,” Lehmkuhl said. “We go in and attack every aspect of revenue and cost to get more out of them.”

One of several ways Lineage increases yield is via Data Science. Once the team had proven its value-add, other math problems presented themselves. Optimizing shipments between hundreds of warehouses and thousands of local and regional distribution centers was classic graph theory; modeling the performance of new blast freezers called for computational fluid dynamics.

“You could call [it] AI, you could call [it] data science, it’s actually applied physics but whatever — it’s a giant high dimensional gradient descent,” he said of the freezer model. “If you know the mathematical methods to do AI/ML, then you also know how to do stuff like this.”

Another thermophysics project involved load balancing power demand. There’s tremendous thermal mass in a 200,000 square-foot warehouse filled with 50,000,000 pounds of frozen food. So Lineage schedules power consumption to avoid peak rates. “It’s as if you put a behind-the-meter battery system on your cold storage warehouse. Now you’re actually intentionally thermally cycling the food supply,” Wolf told us. “We added two zeros on the sensor counts so that we knew exactly what was going on, not because we were concerned about losing control in a particular area or that we were going to thaw something accidentally, but because if you’re going to do that, you need to have a precise idea of whether what you think is going to happen is actually going to happen. How close are your models to your sensor data?”

Lineage Logistics was forced to learn fast when Covid hit

“The food supply chain has existed for millennia,” he reminded us. “The entire history of humanity has been geared towards figuring out how to feed ourselves. And the modern incarnation of that is a whole bunch of pooled infrastructure whereby almost no one who manufacturers food is responsible for its handling and distribution all the way down to the consumer.”

In the spring of 2020, a combination of panic buying and disruptions at packing plants and warehouses resulted in some bare shelves at supermarkets. Lineage quickly pivoted to a different distribution strategy during the brief period when the food logistics system was disrupted.

“We intentionally simplified the orders,” Wolf explained — describing a quick reaction by thousands of people spread across hundreds of warehouses and tens of support teams. “The objective of the supply chain under normal circumstances is to guarantee the availability of a cornucopia of yogurts. You can stand in the grocery store and gawk, thinking ‘I want gluten-free boysenberry,’ or whatever. Lineage decided, ain’t nobody got time for that **** and, in conjunction with our customers, we said, ‘OK, we used to ship five items; now we’re shipping two. You used to want half a pallet; now you’re getting a full pallet.’ It worked out well because they had the demand to use that up downstream. But the critical thing is that when you have a crisis like COVID, the work that you do before it matters more than the work that you do during it.”

One example of work done before the pandemic was a model Lineage built to predict turn rates on pallets entering the warehouse. By putting pallets expected to leave the warehouse quickly in the most accessible spots, Lineage effectively increased labor productivity. It also monitored operational performance to determine how to triage facilities as different areas of the country saw surges in COVID and/or panic-buying at different times.

But labor shortages continue to plague supply chains, even as the direct impact of COVID appears to be winding down.

“The new normal for the supply chain is one that’s less dependent on humans. So that’s a robotics thesis; I want to build this robotic warehouse. But what does it need to do? That’s a statistical question. Okay. I have this candidate layout. Is it capable of doing that?

“It’s an extra exercise in scientific computing because we don’t know what the future supply chain is going to look like. We get to build this building once and modifying it is going to be prohibitively expensive. So now you’re going into this stochastic exercise of, what will the future look like and how future-proof am I? That involves doing a whole bunch of Monte Carlo simulations, to break your warehouse before you build it.”

The U.S. food supply chain has recovered its equilibrium, but the manufacturing supply chain hasn’t. In a recent email, Wolf pointed out some of the differences.

• The US is the largest agricultural producer and the largest agricultural exporter in the world. Common products that do come in from abroad (e.g. cheese) have domestic substitutes. Very little is manufactured wholly within the US with zero foreign dependencies, but agricultural commodities are. The plurality of the food that the US does import comes from Mexico, and thus does not need to transit a container port.
• The feedstocks to any food product are necessarily simpler than something like a car. There are fewer individual components and fewer different suppliers of those components. If one has 10 necessary components vs. 10,000, the system with 10 will prove more resilient.
• Manufacturing operations get to choose when they produce. They often make the choice to minimize inventory or working capital, maximize plant utilization, or otherwise achieve some financial objective. We in agriculture don’t get to choose. We produce at the times and in the amounts that Mother Nature allows.
• Since Mother Nature doesn’t care about working capital, we have to stockpile for times when we don’t produce, leading to much higher inventories. Even if we wanted to produce and ship just-in-time, we can’t. We instead do what humans have done since the dawn of time — stockpile for the (literal or proverbial) winter.

The most international part of the U.S. food supply chain is seafood. Unlike the rest of the food we eat, most of our seafood is harvested outside U.S. waters, and even fish like salmon that are farmed or harvested here are often sent to offshore plants for processing. Harvests are notoriously unpredictable; the supply chain pivots to alternative suppliers in dizzying fashion, but few consumers even notice.

“Shrimp is a very heavily imported commodity into the United States. The Gulf is the only provider of shrimp in the U.S. and gulf shrimp is primarily distributed regionally. Most of the shrimp that we consume comes from Asia; historically Thailand was our largest supplier.

“But the rankings have been wildly gyrating for several years. Thailand’s shrimp were hit with an animal pandemic; something called yellow head virus — it’s a distant cousin of the coronavirus. Suddenly their exports to the U.S. are down 80%. Meanwhile Ecuador has jumped from number seven to number two.

“We’ve had huge global warming-induced squid shortages. Global warming has had catastrophic effects on certain fisheries; we see it in our data. So in 2013 you had 118,000 tons of calamari come into California and in 2019 we had less than 10% of that.

“These are things that happen when you are dependent on mother nature for all of your production; El Nino is going to take out a bunch of fruit, or some fishery is going to fail, or some pork virus is going to take out 60% of the Chinese pork population. There’s a working assumption in our industry that you’re just going to get kicked in the head.”

Advice for startups seeking their first clients

Lineage Logistics is one of the companies that’s willing to engage startups. Since Zetta works with founders, we were curious about Wolf’s insights vis-à-vis engaging startups. He told us that the ones that get contracts are the ones who have taken the time to get to know Lineage’s business.

“People use the word ‘empathy’. A lot of startups who make sales pitches at us are not terribly empathetic… To them [their startup] is the center of the universe. And so they’re interested in this one particular application, and good for them.

“The startup’s application is hyper-focused. It could be computer vision for X; it could be this particular type of inventory management; a freight execution platform, or whatever.

“But imagine in the context of COVID and our problem set in general; we’re fighting fires in all different directions. It’s usually not that successful to come at us directly like, ‘Here’s my product and you need to buy it because I think it’s important to your business.’

“That may be true but Lineage has problems with six zeros on them; we have problems with seven zeros on them; problems with eight zeros on them; sometimes we’ve got problems with nine zeros on them. So, which one is it?

“We might have a two-year-long sales cycle where they come to understand our business and we understand their capabilities. And then suddenly we’ve got a nine-digit problem and we can make their day.”

So what about that troubled manufacturing supply chain? Can it learn from food logistics, or not?

In spite of the many differences between the two supply chains, it’s still natural to wonder whether the manufactured goods chain could learn some useful lessons in resilience from food logistics.

In a recent email exchange, we asked Wolf for any last thoughts on the ongoing challenges facing the manufacturing supply chain. He referred us to an online talk that he and Dr. Danel Wintz, a Principal Data Scientist at Lineage, had participated in.

The talk was sponsored by Duke University’s John Hope Franklin Center; it took place just 6 weeks after that brief period of disruption in the food supply chain. In that conversation, Wolf pointed out that business has embraced the notion of just-in-time inventories to reduce the amount of capital tied up in inventory and improve cash flow.

“That’s what your activist hedge fund wants you to do,” Wolf said. “They want you to free up cash in inventory, so you can return it to shareholders.”

“The statistics of rare events is difficult,” Dr. Wintz interjected. “It’s clear that resiliency is a way to deal with these rare events, and a lot of companies haven’t done that expected value calculation for their supply chain.” By ‘resiliency’, Dr. Wintz meant some of the things that are baked into food logistics: Keeping critical feedstock production onshore; keeping enough inventory on hand to weather a disruption.

Wolf picked up that thread. “I’m proud of the fact that of all the supply chains implicated in this, the only one that really didn’t fall down was perishable food. I wish I could claim some kind of moral high ground in that, but in reality we’ve been trained; the system has trained us not to rely on agricultural production; history has trained us. The history of humanity is, in large measure, the history of agricultural disruptions.”

“And so the [food supply] system regularly deals with things that the rest of the economy is shocked to learn are a factor,” he concluded. “That’s gotta’ change.”

Sign up for our newsletter to be alerted when new pieces are live.

Can the Manufacturing Supply Chain Learn Resiliency From Food Logistics? was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

In 2021, there are over 20 billion connected devices in the world, and experts predict that by 2025, IoT devices will be generating almost 80 zettabytes of data. Connected devices are in our homes, factories, power plants, airplanes, stores, and offices. Hardware from TVs to turbines are intelligent, connected, and generating massive amounts of data.

A revolution like this requires reliable infrastructure, which is why we are so excited to announce our investment in Golioth, a cloud backend service for IoT devices.

When we speak to hardware makers, from startups to enterprises, we hear the same things: reinventing the wheel on the cloud stack; not enough software and devops talent; projects bottlenecked; and most of all — time and resources going to work that isn’t what they love to do. Fundamentally, hardware engineers want to design hardware, not kubernetes clusters and data warehouses.

Founder Jonathan Beri knows these problems intimately. In his early career, he was a product leader at Nest and then Google (where he spearheaded the OpenThread OSS project). He went on to drive product at Particle, building hardware platforms for hardware makers, and then to WeWork where he was responsible for managing the device fleet for physical security across WeWork campuses. Golioth CTO Vit Prajzler was an IoT pioneer who helped build the LoRaWAN protocol at IBM Research. Jonathan and Vit have experienced the pain firsthand and have evolved the solution they wished they’d had.

Golioth is building the systems every IoT developer needs: secure connectivity to the cloud, firmware updates, data collection, device management and health telemetry. Ready built infrastructure will give hardware makers months of acceleration in time to market and dramatic improvements in cost and reliability. But most importantly, Golioth is building these services in a cloud and platform agnostic way: they will give hardware engineers the freedom to select the chipset, protocol, RTOS, and cloud provider (public or private) that make sense for their device, unconstrained by their choice of cloud backend.

We couldn’t be more excited to support Golioth on its journey, and we are thrilled to announce the launch of their private beta today, June 2nd, with support for Zephyr and CoAP. Learn more and sign-up here.

Golioth: Revolutionary IoT Requires Reliable Infrastructure was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

Venkat Chary is currently the CEO of Zenon, a company that delivers AI-enabled automation to financial service, healthcare, and tech-forward companies in both the B2B and B2C categories. He was previously Chief Data Officer at American Express where he began applying artificial intelligence and machine learning to real-world business problems decades ago — in spite of the challenges posed by the hardware of the period!

In a recent Zetta Bytes Live chat with Jocelyn Goldfein, Venkat told listeners that his first exposure to AI and ML came when watching and reading science fiction. “For me it was Star Trek and Isaac Asimov,” he recalled. “That was the future. I wondered if that was ever going to be real, or if we’d see it in our lifetimes.”

Over the course of our conversation, Venkat emphasized the need for founders and entrepreneurs to find ways to work within legacy systems and cultures, acknowledging and smoothing sources of friction rather than insisting on some idealized all-or-nothing technical revolution.

But in some ways his vision is true to the AI in SciFi, in which humans and machines formed seamless teams, without an intervening layer of technologists. “It’s fantastic that we are able to see the beginnings of the transformation; the beginnings of a value-add to society,” he said. “We’re seeing more human-machine augmentation, which is really going to make a difference to the future.”

What follows is an edited summary of Venkat’s chat, which covered Amex’s early adoption of AI and ML and some very pragmatic advice to listeners and then circled back almost to where it began. Star Trek, after all, was fundamentally optimistic; so is Venkat. He even made a few ethical observations that recalled Asimov’s “First Law of Robotics”.

But to match science fiction, we first must fight friction…

“When I joined American Express in 1998, we were on tapes,” Venkat recalled. “That first request would be to load the tapes onto the main frame before I could even extract data or run a model.”

Most people probably think the FAANG companies were the first enterprises to put machine learning into production use. But in spite of the technical challenges, Amex already used ML for real-time fraud detection and real-time credit assessments.

As Venkat explains, the banks didn’t necessarily want to innovate. “If you’re making two, two-and-a-half percent on a transaction, and you’re losing that much on credit and fraud… you’re forced to innovate,” he says. However, those early models were necessarily built on samples of data. “If you tried to process the entire model or build a model on the full sample of data, um, even as recently as 10 years ago, SAS on Teradata, it would just collapse.”

“I remember in 2000, implementing some of the first rules to approve transactions on credit cards that were over the limit or transactions which were delinquent, but using a risk-based methodology,” he told us. At the time, they were logistic or regression models and they weren’t built on a hundred percent of the data. But within 10 years Amex had replaced the rules with models.

Amex’s first use of Big Data was prosaic: personalizing merchant offers for customers. “When you have millions of merchants and hundreds and hundreds of offers and millions of card members, it’s not a problem you can solve with logistic regression,” Venkat recalled. “It’s a problem that can only be solved with collaborative filtering or matrix factorization. Netflix was using it; Google was using it. The question was: How do we bring this to American Express to solve this problem because there’s no other way to solve it?”

Once Venkat’s team had built the infrastructure required to solve Amex’s offer problem, it was then scaled to solve things like fraud and credit risk, and to model customer response and lifetime value. “You’re using a hundred percent of the data now,” he recalls. “And when you go from using a 5% sample for fraud models to a 100% sample — and fraud is not a common occurrence; credit risk is much more common — it significantly improves the accuracy of the model. You’re not only catching more fraudulent transactions but you’re not stopping good spending, which is equally important.”

That notion of not having your card declined — that an algorithm worked invisibly in the background of your life — is, when you think about it, pretty sci-fi in a good way. But as someone who once acquired AI solutions and now provides them, Venkat is uniquely positioned to explain sources of friction that bedevil today’s class of founders and entrepreneurs, many of whom are struggling to close their first sales.

It’s easier to work within legacy cultures than it is to overthrow them

“Often in large enterprises you’re dealing with legacy technology and processes that were built up over very long periods of time,” he said. “There are legacy people and culture, and there’s a strong motivation to deliver results on a quarterly basis — which forces you to build things quickly.”

All those things were factors when Amex struggled to get a merchant recommendation program off the ground. “There are a lot of priorities at a company like Amex; there’s a lot of siloed teams; there’s a lot of decision-makers and there’s not a lot of incentives that a startup would have,” he said by way of explanation, noting that, “It’s beyond just technology. It’s really about culture and people.”

Those challenges create opportunities for start-ups. In the last few years Amex has acquired Resy and Mezi– two recommender platforms that are consumer-facing and provide dining and travel recommendations and reservations. “Now all they have to do is feed their data into those existing platforms and make it work versus trying to build those platforms themselves,” Venkat says.

Mentoring founders and advising start-ups was a big part of Venkat’s job at Amex and it has occupied even more of his time since leaving the bank in 2018. One key question that start-ups need to be able to answer is, why hasn’t the customer you’re targeting built a solution for themselves?

In Venkat’s experience, it usually isn’t because the technology is too hard or that the enterprise doesn’t have the software engineers or the data scientists. It’s usually because the customer lacks one big advantage that the start-up has in spades: the start-up can focus on one thing. But even understanding that advantage isn’t enough unless you can eliminate most or all of the natural friction involved in relationships between start-ups and legacy enterprises.

Large enterprises are focused on their priorities. They trickle down their goals and priorities for every calendar year; every person has their priorities. And then it ladders back up as budgets and planning happen in Q3 for the next year. So trying to get something done in the middle of the year–if it’s not on the priority list, it’s not gonna fly.

People in different departments have different priorities; some are focused on growth, others on customer experience, so what are the real priorities and who is the real decision-maker? Start-ups have to really understand how customer processes work and how things get implemented.

“Convincing those people to open up is very important,” he notes — especially when they’re opening up about the ways they may be constrained by legacy systems or operations.

“Frankly, the implementation issues and buy-in issues are real,” he warns. “People need to syndicate stuff across multiple teams and organizations. And they may not even be able to make the decision on their own. So really be honest with yourself about how much friction there is when it comes to getting your pilot or your product running — whether that’s organizational friction or technological friction, how do you eliminate it?”

Don’t wait for 100% data readiness. Find customers who have enough data ready.

Data readiness — or rather, the lack of it — is a common problem. Despite being a data scientist himself, Venkat’s frustrated with data scientists who constantly push to have 100% of the data before running the models.

“What I would say about data readiness is, you don’t need it,” he says. “It’s not about, do they have all their data ready? It’s about, do they have the data that your startup or your product needs access to in an easy way” You have to figure out which companies have easy access to the data readiness that serves your product.”

Another limiting factor is the current shortage of data scientists. He’s honest about the first hurdle: compensation. People will at most take a small pay cut to work on an interesting project. Beyond that, his secret is not a secret: Network, meet people, build relationships, and stay in touch.

I’m not asking you to fake interest in a person. I think you should have a real interest in these people because you want them to be part of something you’re passionate about. You can’t suddenly email somebody out of the blue or text them or LinkedIn message them out of blue saying, ‘Hey, remember I met you four years ago? What are you up to?’

“But the next stage of evolution is, don’t hire more data engineers and data scientists,” he adds, asking, “Where are your no-code or minimal-code platforms to help you do things faster? Where is the automation?”

Venkat sees AI automation as having a big advantage over the traditional disruption model as defined by Clayton Christensen. “In the past there was always a downside,” he says. “You know, it’s going to make it faster but the decision is actually going to be worse. The one thing with AI and automation is that the decisions are better and the cycle time is lower; it’s not a trade off the way it used to be.”

It’s not as expensive as it used to be either. Enterprises used to ask themselves: Buy, build, or partner? Those decisions are often based on whether or not something’s a core priority or a distraction. “Frankly, if you look at a lot of large enterprises — where they’re investing people and time and effort — it’s not core to their business,” he notes. “So startups and entrepreneurs can help alleviate those problems. [Enterprises] are into that.”

“The other comment I would make is that technology and machine learning are tools,” he said, adding a caution. “Saying you have AI or a data-driven product is very different than transforming them into a product. People care about the business outcome you’re producing, and the value you’re creating for them.”

Value creation is a more enduring pitch, too. “Thirty years ago it was logistic regression on a mainframe; today it’s machine learning in the cloud; tomorrow it’s going to be something else, but we’re still going to be sitting here solving the same business problems. Like, how do I acquire more customers or how do I create more efficiency or how do I service them better?”

An obsession with delivering value may have shaped Venkat’s opinion on the notion of regulating monopolistic juggernauts like Google or Facebook.

If it’s so addictive, you can regulate the hell out of it, but it’s not going to make a difference. When it comes to something like Facebook or Instagram, people are very addicted. They want to be on those platforms. [But] if giving up your privacy gives you enough value, I think you’re okay. If you’re using Google maps and you want to know how to get to wherever you’re going, you are going to give them your location.

He’s unimpressed with some of the opt-in features that seem to exist primarily to make someone, somewhere, feel better. “Who doesn’t click ‘Yes’ when you see that little button that says ‘Accept Cookies’?” he asks.

Not that he’s totally into a free market in which users unconsciously trade personal information for convenience or a momentary dopamine hit. “We are trading off privacy for value. So now the question is, are we valuing some things too much? Or are we undervaluing our privacy until it matters? In the case of social media, maybe it’s mental health; maybe there’s other considerations, which aren’t quite clear. And particularly for individuals who are not capable of making decisions yet if they’re younger, regulation is important. There’s a place for these things; there’s a place for government. But at American Express we used to say, ‘We’re not doing this because the law says we have to do it; we’re not doing this because regulators say we have to; we are holding ourselves to the standard of what our customers expect from us, which should be frankly, a higher standard.’”

Another hot topic on which he has a contrarian opinion is algorithmic bias. “Everybody spends a lot of time on Twitter and other places talking about algorithmic bias and how something is biased or not biased. I’m not saying I’m okay with algorithmic bias, but it can actually be monitored. It can be regulated; it can be improved. How could you ever fix the behavior of a single human loan officer in a small town that was rejecting mortgage applications for underserved people?”

He has a point. We’ve been trying to eradicate bias in humans for eons, without success.

Having read this far, you won’t be surprised to learn that Venkat’s work at Zenon has reaffirmed both his short-term pragmatism and long-term optimism. “We’re building solutions,” he says. “Manual solutions are a waste of time, but in the large enterprise space, there’s a lot of problems that can’t be fully automated yet — maybe in five years or 10 years, there will be product solutions; maybe one of you on the call will be offering them. But right now we’re helping them build this mix, which is human-augmented automation.”

“The last thought I’d leave you with regarding AI is a definition of AI based on it doing what humans can do,” he says. “For me, it’s not about AI doing what humans can do. It’s about going beyond that. We’ve brute-forced our way into a semblance of AI by pushing more data, pushing more compute. If it’s a game of chess, the first way was figuring out all the permutations and combinations; now it’s adversarial again, fine. But I think we’re really reaching the phase of AI where elegance will really drive superiority. And the brilliance of the solution will be in design thinking. And so one of the things that we are focused on as a company is to bring more design thinking. That’s Zenon very quickly, but I think our premise applies generally; it’s not about one company, it’s about, what is the future for all of us?”

Signup for our newsletter here

Science Fiction and Fighting Friction: a Conversation with Venkat Chary was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.

Sign up for our newsletter to be alerted when new pieces are live.

Telm.ai: Great Tools Make Great Data

It’s a real pleasure to welcome new startups to the Zetta family, and Telm.ai is no exception. Zetta invests exclusively in AI-centric startups whose products are powered by data and machine learning, or who are building the infrastructure and platforms to enable those products. Telm.ai happens to be doing both.

We work on a daily basis with startups deploying ML in production for enterprise customers. While code and algorithms tend to capture the imagination when outsiders think of the tech industry, our passionate belief is that data is at least as fundamental. We hear again and again, from entrepreneurs and enterprises, that ML projects suffer from “shiny object syndrome” and that the toughest challenge in getting value from AI projects comes from the unsexy problems: getting clean, reliable and usable data. It doesn’t sound so difficult, but it’s where most of the industry is getting stuck today.

Telm.ai founders Mona Rakibe and Max Lukichev lived it first hand as early product and engineering leaders at Reltio, helping customers find value in their data — if only they could first establish clean and reliable data feeds. Max’s background at SignalFX and Splunk also led him to the insight that observability techniques pioneered in cloud infrastructure management could be applied towards the quality and reliability of the data itself. Mona and Max were inspired to start Telm.ai, both from seeing customer pain up close and personal, and from seeing the potential to solve those problems with machine learning.

Since we made our investment, Mona and Max have built a team, joined and graduated from YC, and accelerated quickly through product development and early customer discovery. Today we’re thrilled to join them in announcing their funding alongside one of our favorite co-investors at .406 Ventures. More importantly, we are thrilled to share the news of their publicly available product and their first customer wins: Dun & Bradstreet and Myers-Holum.

If you follow Zetta, chances are you work with data. Give Telmai a try for yourself and see what a difference great tooling makes to data quality.

Sign up for our newsletter here

Telm.ai: Great Tools Make Great Data was originally published in Zetta Venture Partners on Medium, where people are continuing the conversation by highlighting and responding to this story.