The Sequence Chat: Arjun Sethi on Venture Investing in Generative AI

The founder and CIO of an enterprise VC powerhouse shares his thoughts about the generative AI market.

Subscribe to The Sequence below:

Quick bio:

Arjun is a serial entrepreneur, investor, and co-founder and Chairman at Tribe Capital.  Prior to his venture career, he successfully founded and sold two companies (LOLapps and MessageMe) and served on Yahoo’s executive team and a Board Observer on Alibaba.  Arjun has a long history of sourcing, winning and leading early stage deals - he was named to the Midas Brink List and he’s widely recognized as one of the top 100 angels in Silicon Valley.

Since its inception in 2018, Tribe Capital has consistently had top-quartile venture capital performance by using data science and AI to identify and amplify product-market fit. Some of its noteworthy investments are Apollo.io, Carta, Docker, Kraken, Relativity Space, and Shiprocket. Arjun later co-founded Termina in 2023 to help the broader Venture ecosystem to improve investors’ and companies’ ability to diagnose early-stage product market fit. Arjun recently launched Termina to apply AI to the VC due dilligence process.

Please tell us a bit about yourself: your background, current role, and how you got started in artificial intelligence and venture capital.

Looking back on my career, I’ve always chosen to work on problems where it was possible to harness a large amount of data to build and enhance consumer engagement products. From social gaming, to messaging to social networking to data infrastructure - the driving purpose behind all of these problems was to build better consumer experiences.

When people think about AI, what’s really important to remember is that it is built and trained off data sets.  And the more proprietary the data set, the more valuable your interpretation can be. Across most of my career in building products, and now as co-founder and partner at Tribe Capital, we have spent years accumulating non-public data to feed a long-standing passion for deploying quantitative analytics and AI to achieve stronger business outcomes.

Since launching Tribe Capital in 2018, we’ve pioneered the practice of data-based benchmarking of companies, applying the same rigor from an investment and diligence perspective.  The development of our benchmark dataset (now approaching X data points) has required a meticulous blend of AI and market analysis. This data-centric methodology became our unique edge in identifying and nurturing early-stage ventures such as Carta, Shiprocket, Kraken, Docker and Athelas, among others. We’ve recently scaled our research and diligence tools into Termina - a standalone software product built on top of Tribe’s successful investment methodology - which we are now offering to key partners and investors.

The transition to using AI in venture capital was the culmination of experiences, including my time at Yahoo! and my entrepreneurial ventures, such as co-founding MessageMe and leading LOLapps. These prior interactions  paved the way for a data-focused perspective in venture capital.

My role now involves actively applying AI and data analytics to unearth hidden potentials and risks in early-stage investments. It's a rewarding convergence of technology and finance, where each investment decision is backed by robust data analysis and benchmarked against millions of data points.s.

🛠 AI Work  

Tribe Capital has a long track record in enterprise software and has recently ventured into the generative AI space. Could you please elaborate on Tribe's investment thesis and vision for generative AI?

At Tribe Capital we believe that product-market fit (PMF) can be quantitatively assessed early on by examining engagement data alongside traditional venture capital due diligence - see more here in our overview of underwriting product market fit. Over the last five years, we've collected a vast array of labeled transaction level and product engagement data across various business models, geographies, and stages.

The advent of recent AI breakthroughs is expected to significantly reduce the resources needed for products to come to market. This shift broadens the application of our quantitative approach and increases the relevance of our investment strategy, which prioritizes achieving product market-fit - that is, product engagement and the distribution at scale needed to drive that engagement. I’d mentioned earlier that we’d launched Termina as a response to increasing interest from investors) and partners who wanted to adopt our engagement benchmarking techniques for quantitative diligence. Termina is designed to expand the reach and impact of our diligence and investment approach, in anticipation of a surge in demand for insights into early product engagement.

Scaling our approach with a product like Termina includes leveraging large language models (LLMs) to enhance our analysis and benchmarking capabilities. By utilizing our labeled datasets and analysis archives, we are developing fine-tuned models that bolster our ability to distill key insights and analysis, to enhance our ability to evaluate and support emerging companies.

One of the most challenging aspects of investing in generative AI is determining where the long-term value will accrue. How do you differentiate potentially disruptive companies from those that are merely features of other products powered by generative AI?

To answer this question, we would draw a parallel ton how cloud providers affected the development of tech companies, which, while valuable, became somewhat commoditized. The greater value emerged from innovations built upon these platforms. We anticipate a similar trajectory for foundational models.

Success for entrepreneurs hinges not just on what they build,but also on how effectively they create a distribution network for their products. At Tribe Capital, and now with Termina, our investment strategy reflects this understanding. We focus on identifying companies with distinctive product or distribution advantages. Our due diligence process intensifies when we encounter unexpected positive trends in product engagement and benchmarking metrics.

We foresee a future where products reach the market faster and with less capital, generating an abundance of engagement data. Both Termina and Tribe are poised to harness this data for quantitative diligence and benchmarking, using it as a key input for investment decisions. This influx of data is expected to transform the landscape, enabling more informed and strategic investments in emerging technologies.

How do you perceive moats in the generative AI industry? Are these moats fundamentally different from those in other tech markets?

There's somewhat widespread agreement about the moats for foundational model providers, including data, talent, and model training costs. While acknowledging these, it's perhaps more impactful to consider the users of these models. The advent of Large Language Models (LLMs) and recent AI advancements might not alter the core requirements for building a successful business, but they certainly reshape how competitive advantages and value are accrued.

Going back to the example around cloud computing, which diminished the value of managing one's internet infrastructure, prompting companies to shift their focus to market penetration, distribution, and product innovation. A similar transformative impact is likely with the rise of Gen AI.

Consider one of our portfolio companies Shiprocket, an early investment by Tribe, which now serves hundreds of thousands of ecommerce merchants in India. Their success in customer onboarding and service - both areas ripe for AI optimization - doesn’t solely define their competitive edge. Their extensive distribution capabilities provide a more significant advantage, along with their ability to generate unique, proprietary data sets over time.

We have other portfolio companies like Alpaca, whose competitive moat largely derives from a combination of integrated financial ledgers, an API suite, and various brokerage and clearing licenses. The immediate impact of LLMs on Alpaca is relatively muted. Their infrastructure is affected indirectly, as much of their core value proposition lies beyond the direct reach of current AI advancements.

 One of the most intriguing dynamics in the generative AI space is the balance between open-source and closed-source/API-based models. How do you foresee the evolution of these two distribution models over the next few years? Who will ultimately emerge as the winner?

In the generative AI space, the dynamic between open-source and closed-source/API-based models presents a fascinating landscape. Over the next few years, we anticipate a complex evolution of these two distribution models.

In the broadest sense, the best models will depend on where the best data is located and if it’s public or private. This applies to general purpose models, but also tailored, specific use-case models. For example, Bloomberg’s ability to create the greatest financial markets AGI likely exceeds that of the public domain.

The open-source model offers broad accessibility and collaborative improvement, fueling rapid innovation and adoption. However, its reliance on community contributions and public resources can sometimes limit the speed of specialized advancements. On the other hand, closed-source/API-based models, often backed by substantial private funding, can rapidly iterate and specialize, catering to specific enterprise needs with greater efficiency and customization.

As for which model will emerge as the winner, it's likely that neither will exclusively dominate. Instead, we foresee a landscape where both models coexist and cater to different segments of the market. Open-source models will continue to thrive in environments that value collaboration and transparency, while closed-source models will find favor in sectors requiring bespoke solutions and high levels of service.

A hybrid ground is AI model marketplaces, which we expect to play a role in bridging a variety of open and closed source models to adoption. For example HuggingFace or our portfolio company Instabase. In Instabase’s case, there is a potential evolution into a closed-source marketplace for distributing customized models to large enterprises and financial institutions exemplifies this trend. Marketplaces like these will be crucial in applying AI models in practical, value-driven ways, whether those models are open or closed source. They provide a platform for a variety of AI solutions to reach broader audiences, ultimately driving engagement and capturing value in various industries.

Many large investments in generative AI have been in companies developing massive foundational models. How far do you think the scaling laws can be extended in this area? Will we see LLMs that surpass $10 billion in pretraining and fine-tuning costs?

In the context of scaling generative AI, particularly large foundational models, several critical bottlenecks emerge beyond just capital expenditure (capex). The most significant of these include the calendar time and energy costs required for training and, perhaps in the short term, the practical limits on GPU supply shortages and hardware availability.

Regarding the economics of scaling, as long as the anticipated value of a model exceeds its capex, there will likely be investors ready to finance its development. However, if scaling up to the next generation of models necessitates a $10 billion investment in training, its expected value (EV) becomes questionable. Despite this, some entities may still pursue this investment.

This scenario is akin to the dynamics observed in dark fiber networks, where investments made by one entity benefit others. This phenomenon could repeat itself in the generative AI field as well.

A notable challenge lies in the data used for training these models. Past iterations, like various GPT models, have relied on exponentially increasing amounts of text data. However, this strategy is nearing its practical limit, as exemplified by GPT-4, which has nearly exhausted the available orders of magnitude of text data. The next phase of advancement requires models to incorporate some form of ground truth or self-learning mechanism, similar to the evolution from AlphaGo to AlphaZero. AlphaGo was trained on a comprehensive database of human-played Go games, while AlphaZero developed its proficiency through self-play.

The challenge in applying a similar self-play approach for real-world scenarios is that there's no universally accepted method for generating high-quality, rapidly compiled data. While organizations like OpenAI (OAI) or Google (GOOG) might be exploring solutions, there's currently no known answer to this challenge in the wider AI community. This represents a significant frontier in AI research, where breakthroughs could redefine the capabilities and costs of future large language models (LLMs).

SaaS incumbents have rapidly, arguably faster than any other tech trend in history, incorporated conversational capabilities into their platforms. Do you envision a new form of SaaS platforms built from the ground up with generative AI? How might that look?

As more of these companies come to market and accumulate engagement data we will have a better idea, but it could take some time for these business models to iterate before these application cases become more clear.

It all goes back to product-market fit and product engagement. The focal point of our analysis at Termina, and across our portfolio at Tribe, is fundamentally centered on how product engagement influences a company's ability to upsell and accelerate growth. Our approach to tracking this area is ongoing and based on employing a bottoms-up methodology for benchmarking user engagement to diligence early-stage product-market fit. When our benchmark reports identify anomalies or outliers in engagement data, we delve deeper to understand the underlying factors.

Addressing the core question of how new platforms might emerge and evolve, it's clear that advancements in technology, particularly in AI, are poised to reduce the time required to bring products to market. This is especially true for open-ended use cases like conversational platforms. We anticipate that these advancements will catalyze the emergence of new companies and business models. A critical aspect of our work involves closely monitoring and understanding the customer engagement metrics of these nascent entities.

Today's world operates on established computing platforms such as browsers, cloud computing, and mobile phones. However, any sufficiently disruptive tech trend in history has the potential to unlock a new computing platform. Could generative AI power the next generation of computing platforms? Please share an ambitious vision of that future.

We don’t need to overcomplicate here, what you describe has already happened.

💥 Miscellaneous – a set of rapid-fire questions 

Do you think there will be trillion-dollar companies native to generative AI? If so, do you predict that OpenAI will be one of them?

Yes: on a long enough timeline this is obviously true because the potential to augment human labor even without advances beyond GPT-4 is clear. OpenAI has a better chance than most to become one of them but they haven’t won that title yet.

Which major tech incumbent (Apple, Microsoft, Google, Amazon, Meta) is most vulnerable to disruption by generative AI?

The better question is which of the above is poised to pivot their business models to become more valuable by employing GenAI. The winners there will be distribution-driven. We think Google has the greatest global distribution and the most data, and therefore the best positioned in the long term to exploit their distribution and data to capitalize.

What are the most significant mistakes you see entrepreneurs making in the generative AI space? What about the most common mistakes made by investors?

The biggest mistake people make is getting caught up in the hype, and forgetting that at the end of the day, it’s still all about product-market fit.  As much as the world has changed, the fundamental tenets remain the same. 

Who is your favorite mathematician or computer scientist, and why?

My favorite mathematician and computer scientist is Norbert Wiener. Wiener was a pioneering figure in the field of cybernetics, which explores the complex relationships between systems, control, and communication in machines and living beings. His interdisciplinary approach, blending mathematics, engineering, and computing, was groundbreaking and has had a lasting impact on various fields.

What fascinates me most about Wiener is his work on signal processing and the development of the Wiener Filter. This mathematical technique was pivotal in advancing radar technology during World War II and continues to be a fundamental concept in the field of electrical engineering and communications. It's a prime example of how theoretical mathematics can have practical, real-world applications.

Moreover, Wiener's vision of the impact of automation and computer technology on society was remarkably prescient. He foresaw many of the ethical and social challenges associated with technological advancement, emphasizing the importance of considering these aspects alongside technical development. His book 'Cybernetics: Or Control and Communication in the Animal and the Machine' is a testament to his profound understanding of the complex interplay between technology and humanity.

Wiener's ability to traverse and contribute to multiple disciplines, his foresight regarding the societal impacts of technology, and his significant contributions to practical applications of mathematics make him my favorite in the field.