Avala: People-First AI
A Primer on Artificial Intelligence
Unless one has been living under a rock in 2023, it is utterly apparent that we are living in a new era of artificial intelligence. OpenAI and its ChatGPT product have dominated headlines, as have prognostications that killer AI is poised to wipe out humanity (or at the very least our jobs). Marc Andreessen recently wrote an essay titled Why AI Will Save the World where he argues to dispel or play down the major concerns around AI. Core to his argument is how he defines what AI is:
“the application of mathematics and software code to teach computers how to understand, synthesize, and generate knowledge in ways similar to how people do it…AI is a computer program…[that] is owned by people and controlled by people, like any other technology.”
AI has the potential to have an enormous positive impact on the world, but our motivations behind its continued development and our systems of teaching the machines will have an outsized effect on its overall positive benefit and how much of humanity receives it.
When many people think of Artificial Intelligence, they think of HAL 9000 or Skynet from the science fiction films 2001: A Space Odyssey and Terminator. These fictional AI machines gain an almost human-like consciousness that predictably turns on their human creators. While these fantasies are more fiction than fact, organizations like OpenAI (valued at nearly $30B!) were formed to make sure this extremely powerful technology is applied to applications that benefit, not harm humans. We live in a world where every company will become an AI company, and understanding and harnessing this technology will be critical to drive the next generation of human progress in safe and ethical ways.
The discipline of AI began in the 1950s like a lot of modern technology: in the wake of a global conflict (in this case the Second World War). AI is a set of sciences, theories and techniques (including mathematical logic, statistics, probabilities, computational neurobiology, computer science) that aims to imitate the cognitive abilities of a human being. The focus has been on creating computing platforms that could perform increasingly complex tasks, which could previously only be delegated to a human.
Today the applications of AI are nearly limitless. Generative AI dominates the current headlines for its ability to use existing content, such as text, audio, or images, to create new plausible content. Perhaps the next sci-fi film will be computer generated. AI is used in drug discovery and development, e-commerce, voice assistants, self-driving cars, fraud prevention, cybersecurity, facial recognition, recommendation engines, search engines, robotics, gaming, agriculture and so much more.
AI is associated with many different terms. At its core, AI is a branch of computer science dealing with the simulation of intelligent behavior in computers. Powering AI is Machine Learning, or ML, which is a field of study that gives computers the ability to learn without being explicitly programmed (but to do this they first have to be taught how). Deep learning is a subset of machine learning, and is a part of a broader family of machine learning methods based on artificial neural networks. Neural networks are a set of algorithms, modeled loosely after the human brain, that are designed to recognize patterns. They interpret sensory data through a kind of machine perception, labeling or clustering raw input. Generative Adversarial Networks (GANs) — technology used to power most “deepfake photos and videos — are a a class of ML systems in which two neural networks contest with each other in a game pitting one against the other. GANs can be taught to create worlds eerily similar to our own in any domain: images, music, speech, prose. They are robot artists in a sense. Natural Language Processing (NLP) is a subfield of computer science, information engineering, and AI concerned with the interactions between computers and human (natural) languages, in particular how to program computers to process and analyze large amounts of natural language data. Finally, computer vision is an interdisciplinary scientific field that deals with how computers can be made to gain high-level understanding from digital images or videos. From the perspective of engineering, it seeks to automate tasks that the human visual system can do.
If one were to create a taxonomy of how these different technologies and systems relate to each other it would roughly look like this:
AI > ML > Deep Learning > Neural Networks > GANs / NLP / Computer Vision
A Better Way to Teach AI
Going back to the subject of teaching AI: underlying and powering everything mentioned above are datasets, or collections of various types of data stored in a digital format. Datasets primarily consist of images, texts, audio, videos, numerical data points, etc., for solving various AI challenges such as image or video classification, object detection, face recognition, emotion classification, speech analytics, sentiment analysis, stock market prediction, and much more. If neural networks are the algorithms that seek to model real-world occurrences, and the outcome of those algorithms are what we call ML or deep learning, then datasets can be considered the “fuel” that makes ML and deep learning, and thus all of AI work as intended.
However, (in most cases) a raw dataset is insufficient to power AI. An untrained computer model on its own does not know anything about what is in that dataset. A computer vision model for instance does not know that something in a dataset is a car, a tree, a stop sign, a lane marker or a pedestrian. All of these items first need to be labeled so that the ML system can have understanding of all the underlying components in a dataset. In ML, data labeling is the process of identifying raw data (images, text files, videos, etc.) and adding one or more meaningful and informative labels to provide context so that a machine learning model can learn from it. To date, most data labeling has been done by humans, mostly in emerging economies with inexpensive labor. Companies like Scale AI (last valued at $7.3 billion) and Labelbox (last valued at $1 billion) are two of dozens of companies that have popped up in the last few years to tackle the large demand in labeled datasets.
Part of the challenge with data labeling for ML models is that much of the data to be labeled is considered “unstructured.” Structured data — typically categorized as quantitative data — is highly organized and easily decipherable by ML algorithms. Developed by IBM in 1974, structured query language (SQL) is the programming language used to manage structured data. By using a relational (SQL) database, business users can quickly input, search and manipulate structured data. Unstructured data, typically categorized as qualitative data, cannot be processed and analyzed via conventional data tools and methods. Examples of unstructured data include photos, videos, sensor data, text, mobile activity, social media posts, Internet of Things (IoT) device data, etc. — essentially much of what goes into modern use-cases of AI.
While machines can be trained to perform data labeling independent of people, until every possible type of unstructured data has been successfully labeled by a human–a task that could truly be never ending–people will be needed to input the foundational data that will train our AI systems. The massive problem to date is that the people doing the labeling, and the systems built to facilitate their work, have been afterthoughts. No data labeling company has been built with a labeler-first perspective. Every scaled data labeling company has been built to focus only on the (mostly enterprise) end user. This would be like if Uber built their app just for the customers needing rides, letting the drivers figure things out for themselves.
Putting insufficient resources (and that’s putting it nicely) behind labeler identification, hiring, onboarding, training, user interfaces and experiences, payments (both amounts and frequency), motivations, management, and scheduling ultimately results in unhappy and unmotivated labelers which leads to a subpar labeling product. And this is what has happened in the industry: underpaid and poorly trained labelers being treated like machines has given way to long labeling lead times, project cost overruns, and datasets riddled with labeling errors and biases.
The incumbent companies in this space remain scaled in part because they have the labelers, and they have the labelers because for a long time they were the only or the biggest game in town. If a single mother in Kenya wanted to make an income as a data labeler, her only likely option would be to work for a company that has a spotty history around how they treat their labelers.
This changes today. We are proud to announce our investment in, and the launch of Avala, a San Francisco-based AI company pioneering a radically different approach to ethical AI deployment. The company is revolutionizing how people can contribute to, develop, and benefit from AI with a collaborative marketplace of datasets, labelers, and models in an ecosystem of products and services that directly address the challenges of AI alignment. Avala delivers market-relevant upskilling and reliable income that can be learned and earned from anywhere, anytime. We are very happy to have invested alongside some best-in-class investors that include Chip Hazard at Flybridge and Dustin Rosen at Wonder Ventures.
Avala is a company born with the equitable treatment of people in its DNA. If there ever was a mission-driven, people-first organization, this is it. Avala was founded in 2020 by Sri-Lankan immigrant Emal Alwis. He has purposely assembled a diverse team made up of immigrants and first-generation Americans from Kenya, Argentina, Taiwan, and Sri Lanka. Emal’s story is an extraordinary one: parents immigrated to Minneapolis when he was young and Emal attended the University of Minnesota where he studied electrical engineering, neuroengineering and pre-med. Upon graduation, he was set to go to medical school when he was contacted by a Tesla recruiter to come work on their autopilot software. Taking a leap of faith, Emal moved to San Francisco to work for Tesla Autopilot in 2014 as the third software engineer on the team reporting directly to Elon Musk. Over time, Emal worked closely with the internal and external data annotation teams at Tesla and saw first-hand the challenge of labeling vast quantities of high quality training data, and an inadequacy in products that were available on the market. Motivated by the idea that if he created a world-class product for the data labelers themselves, a better data labeling platform for the enterprise would naturally follow. This is Avala.
Beyond Data Labeling
Avala has developed a programmatic virtual assembly line for AI Ops workflows that turns raw image data into pixel-perfect annotated datasets. The company steadfastly serves two audiences: their data labelers, who they call their coworkers, and AI teams at the leading edge of software development whose unmet need is high-quality data labeling, dataset management, and model training. Avala’s coworkers are the target of their social mission to provide every human being with access to economic opportunity. These people are the talent that turns raw unstructured data into valuable training data.
The Avala coworker application has been built for mobile in iOS and Android. Unlike other data labeling applications that are mostly PC-based, Avala can be used by anyone with a smartphone, opening up the potential labor pool by several orders of magnitude.
A mobile annotation product by design simplifies the tools and workflow of the annotation process. Avala breaks up the annotation process into distributed microtasks. For example, as seen in the image below, Avala has built out responsive, simple gestures that are used to carry out tasks such as classifying the color of a traffic light or drawing a polygon around an object to segment it. Avala offers its own annotation canvas and tooling for drawing bounding boxes, cuboids, polygons and vectors, with many more labeling use cases on the horizon.
Avala’s ecosystem of products and services directly address the challenges of AI alignment by providing tools to audit the mechanics and trustworthiness of AI models, identify and isolate data issues, and access a global on demand workforce to quickly return the feedback needed to reduce bias and improve performance.
Avala’s Mission Control platform was built to complement customer’s existing AI Ops platforms and support multiple annotation formats, enabling customers to easily integrate with existing third party cloud storage, inference and deployment environments and training frameworks such as PyTorch and TensorFlow.
Benefitting All of Humanity
When technology seeks to ameliorate people’s lives, people must, in turn, play a role in shaping that very technology. AI can be a technology that fundamentally alters the human experience. In the book The Beginning of Infinity: Explanations that Transform the World, author and physicist David Deutsch argues that knowledge creation through creativity and scientific thinking that primarily started during The Enlightenment kicked off an exponential evolution of the human race. AI can be an amplifier of this knowledge creation and creativity. To pull another quote from Marc Andreessen’s essay on AI:
“What AI offers us is the opportunity to profoundly augment human intelligence to make all of these outcomes of intelligence — and many others, from the creation of new medicines to ways to solve climate change to technologies to reach the stars — much, much better from here.”
We backed Avala at MaC because the company puts everyday people at the heart and soul of AI — sharing the transformative opportunities of AI with humanity. Big things start with small beginnings, and providing tools to millions of people around the world to contribute in small but important ways to the advancement of our AI future, and thereby improving their station in life feels like a good foundational step towards aligning AI to benefit all of humanity.
Michael Palank led the Avala seed round for MaC Venture Capital.
About MaC Venture Capital
MaC Venture Capital is a seed-stage venture capital firm based in Los Angeles and Silicon Valley that invests in technology startups leveraging shifts in cultural trends and behaviors. The general partners represent diverse backgrounds in technology, business, politics, entertainment, and finance, allowing them to accelerate entrepreneurs on the verge of their breakthrough moment. The firm provides hands-on support crucial for building and scaling category-leading companies, including operations strategy, brand building, recruiting, sales development, and mission-critical introductions. MaC Venture Capital is the result of a merger between Cross Culture Ventures, co-founded by Marlon Nichols, and M Ventures, co-founded by Adrian Fenty, Michael Palank, and Charles D. King. Find MaC Venture Capital online at https://macventurecapital.com and @MaCVentureCap.
