The vertical software data gold mine 

Automation opportunities from vertical software’s data gold mine mean it has never been a better time to create purpose-built operational tools for overlooked industries. Plus, our market map of AI-enabled vertical software new entrants. 

Vertical software founders build applications across many industries but with a common mission: to empower expert operators and owners to streamline and grow their businesses with software purpose-built for their own industry. 

Over the past decade, Pear has supported founders building powerful applications for industries ranging from supply chain (Expedock, Beyond Trucks), construction (Gryps, Miter, Doxel), energy (Aurora, Pearl Street) and insurance (Federato) to home services (Conduit), travel (JetInsight, Skipper), agriculture (FarmRaise, Lasso), real estate (Hazel) and live events (Chainpass) – and more. These companies and plenty of others we admire outside of our portfolio have created hundreds of billions of dollars in enterprise value for themselves and for their customers. 

Today, with artificial intelligence capabilities easier than ever to deploy in B2B products, we think it has never been a better time to build vertical software that automates core aspects of customer operations. 

Any new vertical software opportunity — any sector that is not well-served by purpose-built software for industry-specific workflows — now looks even more promising. And, many existing vertical software tools have a rare opportunity to ramp up ACVs, increase stickiness, and build a powerful moat. 

To understand why we might be on the verge of a golden age of intelligently-automated vertical software, we’ll take a look at two problems: First, the ACV problem that has limited vertical software opportunities. Then, the defensibility problem afflicting new AI application entrants. 

The ACV Problem 

Vertical software companies often sell into fragmented industries with many small-to-medium businesses. These businesses operate on thin margins and have limited ability to pay for new operational software. Contract values – and therefore market size – for many verticals have historically been constrained, which means that many sectors remain underserved by purpose-built, cloud-based software.

Historically, vertical software companies have sought to increase the value of their customer relationships by adding bolt-on monetization features like payments or banking. Other vertical software companies have initially sold to the fragmented base of a sector and steadily added operational features to move upmarket to larger customers within their industry. 

In either case, the fundamental margin structure of the end customer remains unchanged, and few vertical software companies can reliably claim to impact their customers’ profitability in a major way. A vertical software customer might love their software’s intuitive UI, centralized system of record, navigable scheduling tools, and modern payment processing system. But these benefits rarely reduce the customer’s overall operating cost in a significant way. 

The Defensibility Problem

Despite the excitement over applications of large language models in late 2022 and early 2023, many initial products were dismissed as “thin wrappers” over an off-the-shelf model. Critics argued that these initial applications lacked product differentiation and long-term defensibility.

At Pear, we believe that proprietary data is one of the keys to a defensible AI application. Proprietary data behind B2B AI applications comes in three forms: 

Many early tools built over groundbreaking LLMs offered no form of proprietary data. At best, some products built over lightly-adapted models offered bronze or silver-level data-based defensibility. But we’ve been on the hunt for game-changing applications that build an advantage in proprietary model-training data from the start.  

The Vertical Software Data Gold Mine 

Traditional vertical software products generate enormous amounts of gold-level data, capture substantial silver-level information, and often themselves aggregate bronze-level data within their products.

Customer business logic flows through vertical software features. From operational decisions and administrative record-keeping to sales and product performance, customers of vertical software deposit reams of data daily into a rich system of record. 

The most impactful B2B applications in the next decade will rigorously structure, mine, and harness gold-level product-generated data to enable workflow automation for their end customers.

Any business process with decisions and steps encoded in a vertical software feature set will be a candidate for automation. We’re most excited about automation that enables faster information processing tied to sales growth or cost-saving opportunities for an end customer: instant diagnosis and repair commissioning for field technicians, predictive inventory capabilities embedded in B2B marketplaces, copilot-style knowledge bases that help small business owners understand the impact of every decision on their bottom line. 

Automation across many business functions will mean that vertical software companies can finally impact their customers’ margin structure – and as a result, help these customers break any linear scaling trap they face when they otherwise expand their business.

Early entrants: A preliminary market map 

We have seen a proliferation of promising AI-enabled vertical software products in a handful of sectors, and we are proud to be the earliest supporters of teams like Expedock, Pearl Street, Gryps, Hazel, and Federato.  

Many initial intelligently automated vertical software products target the largest sectors of the economy (we’ve looked separately at the ecosystem of AI in healthcare companies here). We’ll update this market map over time, and we’re eager to include companies unlocking automation potential in industries that have seen fewer capable purpose-built tools in the past.

What we are looking for 

We hope to support many more founders delivering on a new and bigger promise of vertical software. Standout teams that we currently support – or just simply admire – typically excel on a few dimensions: 

  1. They have an unfair data advantage.
  2. They have identified substantial automation potential. 
  3. They can communicate their value without invoking AI. 

We want to hear from you 

If you share our conviction that we’re entering a new golden age of vertical software and you’re exploring startup ideas that help expert operators streamline and grow their businesses through intelligent automation, we would love to hear from you. Reach out at if you’re working on something impactful.

Perspectives in AI: From LLMs to Reasoning with Edward Hu, Inventor of LoRA and μTransfer

I recently hosted a fireside chat with AI researcher Edward Hu. Our conversation covered various aspects of AI technology, with a focus on two key inventions Edward Hu pioneered: Low Rank Adaptation (LoRA) and μTransfer, which have had wide ranging impact on the efficiency and adoption of Large Language Models. For those who couldn’t attend in person, here is a recap (edited and summarized for length).

Aparna:  Welcome, everyone to the next edition of the ‘Perspectives on AI’ fireside chat series at Pear VC. I’m Aparna Sinha, a partner at Pear VC focusing on AI, developer tooling and cloud infrastructure investments. I’m very pleased to welcome Edward Hu today. 

Edward is an AI researcher currently at Mila in Montreal, Canada. He is pursuing his PhD under Yoshua Bengio, who is a Turing award winner. Edward has a number of inventions to his name that have impacted the AI technology that you and I use every day. He is the inventor of Low Rank Adaptation (LoRA) as well as μTransfer, and he is working on the next generation of AI reasoning systems. Edward, you’ve had such an amazing impact on the field.  Can you tell us a little bit about yourself and how you got started working in this field? 

Edward: Hello, everyone. Super happy to be here. Growing up I was really interested in computers and communication. I decided to study both computer science and linguistics in college. I got an opportunity to do research at Johns Hopkins on empirical NLP, building systems that would understand documents, for example. The approach in 2017, was mostly building pipelines. So you have your name entity recognition module, that feeds into maybe a retrieval system, and then the whole thing in the end, gives you a summarization through a separate summarization module. This was before large language models. 

I remember the day GPT-2 came out. We had a lab meeting and everybody was talking about how it was the same approach as GPT, but scaled to a larger data set and a larger model. Even though it was less technically interesting, the model was performing much better. I realized there is a limit to the gain we have from engineering traditional NLP pipelines. In just a few years we saw a transition from these pipelines to a big model, trained on general domain data and fine tuned on specific data. So when I was admitted as an AI resident at Microsoft Research, I pivoted to work on deep learning. I was blessed with many mentors while I was there, including Greg Yang, who recently started xAI. We worked on the science and practice of training huge models and that led to LoRA and μTransfer.

More recently, I’m back to discovering the next principles for intelligence. I believe we can gain much capability by organizing computation in our models. Is our model really thinking the way we think? This motivated my current research at Mila on robust reasoning.

Aparna: That’s amazing. So what is low rank adaptation in simple terms and what is it being used for? 

Edward: Low Rank Adaptation (often referred to as LoRA) is a method used to adapt large, pre-trained models to specific tasks or domains without significant retraining. The concept is to have a smaller module that contains enough domain-specific information, which can be appended to the larger model. This allows for quick adaptability without altering the large model’s architecture or the need for extensive retraining. It performs as if you have fine tuned a large model on a downstream task.

For instance, in the context of diffusion models, LoRA enables the quick adaptation of a model to particular characters or styles of art. This smaller module can be quickly swapped out, changing the style of art without major adjustments to the diffusion model itself.

Similarly, in language processing, a LoRA module can contain domain-specific information in the range of tens to hundreds of megabytes, but when added to a large language model of tens of gigabytes or even terabytes, it enables the model to work with specialized knowledge. LoRA’s implementation allows for the injection of domain-specific knowledge into a larger model, granting it the ability to understand and process information within a specific field without significant alteration to the core model.

Aparna: Low rank adaptation seems like a compelling solution to the challenges of scalability and domain specificity in artificial intelligence. What is the underlying principle that enables its efficacy, and what led you to develop LoRA?

Edward: We came up with LoRA two years ago, and it has gained attention more recently due to its growing applications. Essentially, LoRA uses the concept of low rank approximation in linear algebra to create a smaller, adaptable module.This module can be integrated into larger models to customize them towards a particular task.

I would like to delve into the genesis of LoRA. During my time at Microsoft, when GPT-3 was released and the OpenAI-Microsoft partnership began, we had the opportunity to work with the 175-billion-parameter model, an unprecedented scale at that time. Running this model on production infrastructure was indeed painful.

Firstly, without fine-tuning, the model wasn’t up to our standards. Fine-tuning, is essential to adapt our models to specific tasks, and it became apparent that few-shot learning didn’t provide the desired performance for a product. Although once fine-tuned, the performance was amazing, the process itself was extremely expensive.

To elucidate, it required at least 96 Nvidia V100s, which was cutting-edge technology at the time and very hard to come by, to start the training process with a small batch size, which was far from optimal. Furthermore, every checkpoint saved was a terabyte in size, which meant that the storage cost was non-negligible, even compared to the GPUs’ cost. The challenges did not end there. Deploying the model into a product presented additional hurdles. If you wanted to customize per user, you had to switch models, a process that took about a minute with such large checkpoints. The process was super network-intensive, super I/O-intensive, and simply too slow to be practical.

Under this pressure, we sought ways to make the model suitable for our production environment. We experimented with many existing approaches from academia, such as adapters and prefix tuning. However, they all had shortcomings. With adapters, the added extra layers led to significant latency, a nontrivial concern given the scale of 175 billion parameters. For prefix tuning and other methods, the issue was performance, as they were not on par with full fine-tuning. This led us to think creatively about other solutions, and ultimately to the development of LoRA.

Aparna: That sounds like a big scaling problem, one that must have prevented LLMs from becoming real products for millions of users. 

Edward: Yes, I’ll proceed to elaborate on how we solved these challenges, and I will discuss some of the core functionalities and innovations behind LoRA.

Our exploration with LoRA led to impressive efficiencies. We successfully devised a setup that could handle a 175 billion parameter model. By fine-tuning and adapting it, we managed to cut the resource usage down to just 24 V100s. This was a significant milestone for our team, given the size of the model. This newfound efficiency enabled us to work with multiple models concurrently, test numerous hyperparameter combinations, and conduct extensive model trimming.

What further enhanced our production capabilities was the reduction in checkpoint sizes, from 1 TB to just 200 megabytes. This size reduction opened the door to innovative engineering approaches such as caching in VRAM or RAM and swapping them on demand, something that would have been impossible with 1 TB checkpoints. The ability to switch models swiftly improved user experience considerably.

LoRA’s primary benefits in a production environment lie in the zero inference latency, acceleration of training, and lowering the barrier to entry by decreasing the number of GPUs required. The base model remains the same, but the adaptive part is faster and smaller, making it quicker to switch. Another crucial advantage is the reduction in storage costs, which we estimated to be a reduction by a factor of 1000 to 5000, a significant saving for our team.

Aparna: That’s a substantial achievement, Edward, paving the way for many new use cases.

Edward: Indeed. Now, let’s delve into how LoRA works, particularly for those new to the concept.  LoRA starts with fine-tuning and generalizes in two directions. The first direction concerns which parameters of the neural network – made up of numerous layers of weights and biases – we should adapt. This could involve updating every other layer, every third layer, or specific types of layers such as the attention layers or the MLP layers for a transformer.

The second direction involves the expressiveness of these adaptations or updates. Using linear algebra, we know that matrices, which most of the weights are, have something called rank. The lower the rank, the less expressive it is, providing a sort of tuning knob for these updates’ expressiveness. Of course, there’s a trade-off here – the more expressive the update, the more expensive it is, and vice versa.

Considering these two directions, we essentially have a 2D plane to help navigate our model adaptations. The y-axis represents the parameters we’re updating – from all parameters to none, which would retain the original model. The parameters of our model exist on a plane where the x-axis signifies whether we perform full rank updates or low rank updates. A zero rank update would equate to no updating at all. The original model can be seen as the origin, and fine tuning as the upper right corner, indicating that we update all parameters, and these updates are full rank.

The introduction of LoRA allows for a model to move freely across this plane. Although it doesn’t make sense to move outside this box, any location inside represents a LoRA configuration. A surprising finding from our research showed that a point close to the origin, where only a small subset of parameters are updated using very low rank, can perform almost as well as full fine tuning in large models like GPT-3. This has significantly reduced costs while maintaining performance.

Aparna: This breakthrough is not only significant for the field as a whole, but particularly for OpenAI and Microsoft. It has greatly expanded the effectiveness and efficiency of large language models.

Edward: Absolutely, it is a significant leap for the field. However, it’s also built on a wealth of preceding research. Concepts like Adapters, Prefix Tuning, and the like have been proposed years before LoRA. Each new development stands on the shoulders of prior ones. We’ve built on these works, and in turn, future researchers will build upon LoRA. We will certainly have better methods in the future.

Aparna: From my understanding, LoRA is already widely used. While initially conceived for text-based models, it’s been applied to diffusion models, among other things.

Edward: Indeed, the beauty of this approach is its general applicability. Whether deciding which layers to adapt or how expressive the updates should be, these considerations apply to virtually any model that incorporates multiple layers and matrices, which is characteristic of modern deep learning. By asking these two questions, you can identify the ideal location within this ‘box’ for your model. While a worst case scenario would have you close to the upper right, thereby not saving as much, many models have proven to perform well even when situated close to the lower left corner. LoRA is also supported in HuggingFace nowadays, so it’s relatively easy to use. 

Aparna: Do you foresee any potential challenges or limitations in its implementation? Are there any other domains or innovative applications where you envision LoRA making a significant impact in the near future?

Edward: While LoRA presents exciting opportunities, it also comes with certain challenges. Implementing low rank adaptation requires precision in crafting the smaller module, ensuring it aligns with the larger model’s structure and objectives. An imprecise implementation could lead to inefficiencies or suboptimal performance. Furthermore, adapting to rapidly changing domains or highly specialized fields may pose additional complexities.

As for innovative applications, I envision LoRA being utilized in areas beyond visual arts and language. It could be applied in personalized healthcare, where specific patient data can be integrated into broader medical models. Additionally, it might find applications in real-time adaptation for robotics or enhancing virtual reality experiences through customizable modules.

In conclusion, while LoRA promises significant advancements in the field of AI, it also invites careful consideration of its limitations and potentials. Its success will depend on continued research, collaboration, and innovative thinking.

Aparna: For many of our founders, the ability to efficiently fine tune models and customize them according to their company’s unique personality or data is fundamental to constructing a moat. What your work has done is optimize this process through tools like Lora and μTransfer. Would you tell us now about μTransfer, the project you embarked upon post your collaboration with Greg Yang on the theory of infinity with neural networks.

Edward: The inception of μTransfer emerged from a theoretical proposition. The community has observed that the performance of a neural network seemed to improve with its size. This naturally kindled the theoretical question, “What happens when the neural network is infinitely large?” If one extrapolates the notion that larger networks perform better, it stands to reason that an infinitely large network would exhibit exceptional performance. This, however, is not a vacuous question.

When one postulates an infinite size, or more specifically, infinite width for a neural network, it becomes a theoretical object open to analysis. The intuition being, when you are summing over infinitely many things, mathematical tools such as convergence of random variables come into play. They can assist in reasoning about the behavior of the network. It is from this line of thought that μTransfer was conceived. In essence, it not only has practical applications but is also a satisfying instance of theory and empirical applications intersecting, where theory can meaningfully influence our practical approaches.

I’d like to touch upon the topic of hyperparameter training. Training large AI models often involves significant investments in terms of money and compute resources. For instance, the resources required to train a model the size of GPT-3 or GPT-4 are substantial. However, a frequently overlooked aspect due to its uncertainty is hyperparameter tuning. Hyperparameters are akin to knobs or magic numbers that need to be optimized for the model to train efficiently and yield acceptable results. They include factors like learning rate, optimizer hyperparameters, and several others. While a portion of the optimal settings for these has been determined by the community through trial and error, they remain highly sensitive. When training on a new dataset or with a novel model architecture, this tuning becomes essential yet again, often involving considerable guesswork. It turns out to be a significant hidden cost and a source of uncertainty.

To further expound on this, when investing tens of millions of dollars to train the next larger model, there’s an inherent risk of the process failing midway due to suboptimal hyperparameters, leading to a need to restart, which can be prohibitively expensive. To mitigate this, in our work with μTransfer, we adopt an alternative approach. Instead of experimenting with different hyperparameter combinations on a 100 billion parameter model, we employ our method to reduce the size of the model, making it more manageable.

In the past, determining the correct hyperparameters and setup was akin to building proprietary knowledge, as companies would invest significant time experimenting with different combinations. When you publish a research paper, you typically disclose your experimental results, but rarely do you share the precise recipe for training those models. The working hyperparameters were a part of the secret. However, with tools like μTransfer, the cost of hyperparameter tuning is vastly reduced, and more people can build a recipe to train a large model.

We’ve discovered a way to describe a neural network that allows for the maximal update of all parameters, thus enabling feature learning in the infinite-width limit. This in turn gives us the ability to transfer hyperparameters, a concept that might need some elucidation. Essentially, we make the optimal hyperparameters the same for the large model and the small model, making the transfer process rather straightforward – it’s as simple as a ‘copy and paste’.

When you parameterize a neural network using the standard method in PyTorch, as a practitioner, you’d observe that the optimal learning rate changes and requires adaptation. However, with our method of maximal update parameterization, we achieve a natural alignment. This negates the need to tune your large model because it will have the same optimal hyperparameters as a small model, a principle we’ve dubbed ‘mu transfer’. Indeed, “μ” in “μTransfer” stands for “maximal update,” which is derived from a parameterization we’ve dubbed “maximal update parameterization”.

To address potential prerequisites for this transfer process, for the most part, if you’re dealing with a large model, like a large transformer, and you are shrinking it down to a smaller size, there aren’t many restrictions. There are a few technical caveats; for instance, we don’t transfer regularization hyperparameters because they are more of an artifact encountered when we don’t have enough data, which is usually not an issue when pretraining a large model on the Internet.

Nonetheless, this transfer needs to occur between two models of the same architecture. For example, if we have GPT3 175 B for which we want to find the hyperparameters, we would shrink it down to GPT3 10 mil or 100 mil to facilitate the transfer of hyperparameters from the small model to the large model. It doesn’t apply to transferring hyperparameters between different types of models, like from a diffusion model to GPT.

Aparna: A trend in recent research indicates that the cost of training foundational models is consistently decreasing. For instance, training and optimizing a model at a smaller scale and then transferring these adjustments to a larger scale significantly reduces time and cost. Consequently, these models become more accessible, enabling entrepreneurs to utilize them and fine-tune them for various applications. Edward, do you see this continuing? 

Edward: Techniques like μTransfer, which significantly lower the barrier to entry for training large models, will play a pivotal role in democratizing access to these large models. For example, I find it particularly gratifying to see our work being used in the scaling of large language models, such as the open-source Cerebras-GPT, which comprises around 13 billion parameters or more. 

In our experiments, we found that using μTransfer led to superior hyperparameters compared to those discovered through heuristics in the GPT-3 paper. The improved hyperparameters allowed a 6.7 billion parameter model to roughly match the performance of a 13 billion parameter model, effectively doubling the value of the original model with only a 7% increase in the pre-training cost.

Aparna:   It appears that the direction of this technology is moving towards a world where numerous AI models exist, no longer monopolized by one or two companies. How do you envision the utilization of these models evolving in the next one or two years?

Edward: It’s crucial to comprehend the diverse ways in which computational resources are utilized in training AI models. To begin with, one could train a large-scale model on general domain data, such as the Pile or a proprietary combination of internet data. Despite being costly, this is typically a one-time investment, except for occasional updates when new data emerges or a significant breakthrough changes the model architecture.

Secondly, we have domain-specific training, where a general-purpose model is fine-tuned to suit a particular field like law or finance. This form of training doesn’t require massive amounts of data and, with parameter-efficient fine-tuning methods like LoRA, the associated costs are dropping significantly.

Finally, there’s the constant use of hardware and compute in inference, which, unlike the first two, is an ongoing cost. This cost may end up dominating if the model or domain isn’t changed frequently.

Aparna: Thank you for the comprehensive explanation. Shifting gears a bit, I want to delve into your academic pursuits. Despite your significant contributions that have been commercialized, you remain an academic at heart, now back at Mila focusing on your research. I’m curious about your perspectives on academia, the aspects of research that excite you, and what you perceive to be the emerging horizons in this space.

Edward: This question resonates deeply with me. Even when I was at Microsoft, amidst exciting projects and the training of large models, I would often contemplate the next significant advancements in the principles and fundamentals underpinning the training of these models. There are myriad problems yet to be solved.

Data consumption and computational requirements present unique challenges to current AI models like GPT-4. As these models are trained on increasingly larger data sets, we might reach a point where we exhaust high-quality internet content. Moreover, despite their vast data processing, these models fail at executing relatively simple tasks, such as summing a long string of numbers, which illustrates the gap between our current AI and achieving Artificial General Intelligence (AGI). AGI should be able to accomplish simple arithmetic effortlessly. This gap is part of what motivates my research into better ways to structure computation and enhance reasoning capabilities within AI.

Shifting back to the topic of reasoning, it’s an exciting direction since it complements the scaling process and is even enabled by it. The fundamental question driving our research is, “How can we convert computations, or flops, into intelligence?” In the past, AI was not particularly efficient at transforming compute into intelligence, primarily due to limited computational resources and ineffective methods. Although we’re doing a better job now, there’s still room for improvement.

The key to turning flops into intelligence lies in the ability to perform effective search processes. Intelligence, at its core, represents the capability to search for reasons, explanations, and sequences of actions. For instance, when devising a move in chess, one examines multiple possible outcomes and consequences—a form of search. This concept is not exclusive to games like chess but applies to any context requiring logical reasoning.

Traditional AI—often referred to in research communities as “good old fashioned AI” or “GOFAI”—performed these search processes directly in the solution space. It’s analogous to playing chess by examining each possible move directly. However, the efficiency of these processes was often lacking, which leads us to the development of modern methods.

The fundamental challenge we face in computational problem-solving, such as in a game of chess, is that directly searching the solution space for our next move can be prohibitively expensive, even when we try to exhaustively simulate all possibilities. This issue escalates when we extend it to complex domains like language processing, planning, or autonomous driving.

Today, deep learning has provided us with an effective alternative. Although deep learning is still a form of search, we are now exploring in the space of neural network weights, rather than directly in the solution space. Training a neural network essentially involves moving within a vast space of billions of parameters and attempting to locate an optimal combination. While this might seem like trading one immense search space for another, the introduction of optimization techniques such as gradient descent has made this search more purposeful and guided.

However, when humans think, we are not merely searching in the weight space. We are also probing what we might call the ‘concept space.’ This space consists of explanations and abstract representations; we formulate narratives around the entities involved and their relationships. Therefore, the next frontier of AI research, which we are currently exploring at Mila with Yoshua, involves constructing models capable of searching this ‘concept space.’

Building on the foundations of large-scale, deep learning neural networks, we aim to create models that can autonomously discover concepts and their relationships. This approach harkens back to the era of ‘good old fashioned AI’ where researchers would manually construct knowledge graphs and scene graphs. However, the major difference lies in the model’s ability to learn these representations organically, without explicit instruction.

We believe that this new dimension of search will lead to better ‘sample complexity,’ meaning that the models would require less training data. Moreover, because these models have a more structured, lower-dimensional concept space, they are expected to generalize much better to unseen data. Essentially, after seeing a few examples, these models would ideally know how to answer the same type of question on unseen examples.

Aparna: Thank you, Edward. Your insights have been both practical, pertaining to present technologies that our founders can utilize, as well as forward-looking, providing a glimpse into the ongoing research that is shaping the future of artificial intelligence. Thank you so much for taking us through your inventions and making this information so accessible to our audience.

Join me for the next Perspectives in AI fireside, hosted monthly at Pear for up to date technical deep dives on emerging areas in Artificial Intelligence. You can find an archive of previous talks here.

Generative AI Tech Stack

We recently launched a dedicated AI track to the PearX program and have received a great response. Founders often ask us for guidance on how to build a moat for their AI startup. There are many aspects to this question but to kick things off, we are sharing a presentation I gave at SF Tech Week that covers background on the emergence of Generative AI, the highest priority areas of application particularly in enterprises, and what we believe enables a ‘moat’ for AI startups.

Generative AI Tech Stack Presentation at SF Tech Week

Generative AI is a game changing technology for humanity. A quote from one of my heroes, Professor Fei Fei Li at Stanford, and also was head of AI at Google Cloud for a while captures the excitement well:

“Endowing machines with generative capabilities, has been a dream for many generations of AI scientists” 

Seminal technologies which have led to the recent Generative AI breakthroughs include Cloud computing and within that advancements in GPUs, Kubernetes and open source frameworks like PyTorch provide an efficient and widely accessible substrate for model training and inference.

Research breakthroughs on the transformer neural network, its use on internet scale datasets and recent advancements in AI alignment are at the heart of most of the Generative AI capabilities today.

By no means are we at a peak yet, as research continues to improve efficiency at the hardware, software and services layers. Most interestingly to increase context lengths and optimize AI application architectures for accuracy, latency, and reliability. We cover some of these topics in depth in our Perspectives on AI fireside series.

It is clear that Generative AI techniques apply to multiple modalities. There has been a steady stream of models, both open source and proprietary in the major areas of NLP, Image, Video, Voice and also physical synthesis of Proteins.

Applications to both consumer and enterprise software abound and are already starting to change the shape of what software can do. We highlight some of the opportunities to build vertical and horizontal applications as well as tooling and infrastructure.

Of course there is hype when it comes to Generative AI, and in some sense it is almost too easy to create new functionality by building a thin layer over a foundation model that somebody else has built. While there are some businesses to be built in that way, for a venture scale business, we posit that a deeper moat is required to build. A large business that benefits rather than crumbles from rapid evolution of technology at the lower layers of the AI stack requires several moats.

Our thesis is that Applications will be composed of ensembles of specialized models, not just foundation models, but specialized models that are customized via fine tuning or in-context learning or a range of other techniques to complete part of a use case or workflow. These specialized models should utilize proprietary data specific to a domain and help to personalize the output of the application as well as ensure accuracy. A by product may also be lower cost to serve. Overall such an architecture will be a way to build lasting value and be more immune to disruption.

Tooling and infrastructure supporting the development of new applications of this kind is second part of our investment thesis. In particular, data and tooling companies to evaluate and ensure safety, accuracy, and privacy of these applications will be in demand. Lastly a few new infrastructure companies and capabilities will advance the development of these applications. We see emerging companies at every layer of the AI stack (slide 9). With that thesis in mind,  building a moat is fundamentally not that different in AI than in any other emerging space (slide 10).

Enterprise readiness for adoption of AI is arguably higher than it has ever been with the widespread acceptance of cloud computing, API integrations, and existing investments in data analytics teams and software. The hurdles to enterprise adoption are also not new, these are the same requirements that any cloud service has to meet, with perhaps a stronger need for ease of use and simplicity given the lack of existing AI/ML talent.

We conclude by quoting what many others have already said, that this is a great time to start a company!

Unleashing the power of Generative AI: an invitation to ambitious founders

We kicked off this week with the announcement of our 4th seed stage fund, one of the largest of its kind, raising $432M to seed the next generation of startups. Today we are thrilled to announce a dedicated Artificial Intelligence startup package, PearX for AI, which offers each founding team $250K in cloud credits, access to beta APIs, expanded check sizes up to $5M, 1:1 expert technical advice, customer introductions, AI talent matching and a curated community of AI practitioners who connect and learn from each other. Applications are open now through June 10th.

PearX for AI, is dedicated to the most ambitious technical founders, interested in building groundbreaking applications, tooling and infrastructure to power the Artificial Intelligence driven revolution. This program provides the resources, expertise, and customer connections you need to build the future. 

Who should apply?

PearX for AI will be a small tight-knit community with up to 10 startups selected into our inaugural summer batch. Pear specializes in working deeply with our portfolio companies to solve the most challenging problems startups face during the 0 to 1 journey from idea to product market fit. No idea is too early or too controversial. The program requires technical depth with a focus in AI, and the application process is tailored towards CS / Engineering graduates, PhDs, researchers and other technical professionals who have built AI driven applications in the past. We look for a combination of market knowledge, technical strategy and coding skills. The program will further build upon these skills and help round out your team’s capabilities in any areas that may need support. It will also connect you with Pear’s community of AI entrepreneurs.

Artificial Intelligence is a horizontal technology with the potential to impact many industries. We believe generative AI is a game-changer for consumer, social and enterprise applications. Particularly Healthcare, Legal, Financial Services, Retail, Logistics, Fashion, Design, Media, Gaming, Manufacturing, Energy, Industrial and Biotechnology to name a few verticals ripe for AI innovations. Pear’s AI team is especially skilled in enterprise AI adoption and will work with exceptional founders to craft solutions for the most pressing enterprise needs. 

We are deep technologists ourselves and value founders working on next generation Natural Language Understanding, Image generation, Computer Vision, Protein and Molecular synthesis and Robotics and Simulation technology. We have experts focused on applications of AI to developer tooling, open source software, sales, support, HR, R&D, design and education. We are bullish on infrastructure software, data services, and tools that reimagine the tech-stack for optimal AI performance.

What benefits will I receive?

Pear’s AI track comes with $250K in cloud credits, early access to new APIs and models, technical support from practitioners, mentorship from specialist AI experts, enterprise customer introductions, access to a talented and like-minded community, and an extended virtual cohort of AI founders. Pear will also extend larger check sizes for AI startups that require additional upfront investment, and provide introductions to strategic angel investors. Finally Pear will prepare all founders in the program for Series A fundraising. Our programs have an 87% success rate for series A and beyond investment by top tier funds.

Why now?

We’re living in an era of unprecedented technological advancement. AI is re-shaping our present and enabling some of the most significant breakthroughs of our lifetimes. The pace of this technological shift is breathtaking. The last time we witnessed a transformation of this nature was in 2000 with web technologies. There was a bubble then, just as there is significant hype around Generative AI now, but from it emerged breakout companies like Google, VMware, Salesforce and more. Breakout companies will be built now in this similar environment.

What does this mean for founders? Opportunity! There has never been a better time to start a company, but navigating the hype and shifting landscape that surrounds technical breakthroughs that are still in progress, requires expertise, judgment, and partners who will tell you hard truths and stand steady through difficult times, helping you secure the resources required – both financial and human.

Why Pear?

We have been investing in AI startups for several years and have a portfolio of AI powered companies that span verticals including consumer social, gaming, retail, healthcare, fintech, insurance, property technology, infrastructure, databases, deep tech, and more. 

With PearX for AI, we have pulled together the resources, community, and expertise to help founders discern signal from noise and succeed in building industry defining companies using the latest breakthroughs in Generative AI. The first track of PearX for AI is set to start in July, with a small cohort of founders who have a proven track record in applying AI to real-world problems. If you’re ready to take on this challenge and shape the future with AI, we invite you to apply now here. Let’s build the future together!

AR/VR/XR/PEAR: our call for mixed reality builders 

In the past few months, we have seen the beginnings of rising interest in building AR, VR, mixed reality, and extended reality infrastructure and applications (we’ll just call it “XR” for short): XR applications to PearX are up 5x this year, dedicated XR hacker groups are proliferating at top engineering schools like Harvard and Stanford, and our tracking shows hundreds of founders have left venture-backed and major tech companies to build in the XR space.  

We think this is because XR has the potential to represent one of the most consequential new platforms of the coming decade, and there is substantial alpha to be had for builders who study this burgeoning space and seize early opportunities. 

We expect interest in building in XR only to increase dramatically – particularly following Apple’s upcoming Reality Pro headset announcement. We see builders with a measured sense of urgency having an advantage, and we’ve put together a high-level guide for exploring ideas in XR. What follows is merely one way of cataloging opportunities; we would love to meet and speak with founders who are building early, quickly, and curiously in the broader XR space. 

XR Builder’s Guide to Exploring Ideas

A Familiar “Infrastructure and Applications” Approach

With any new technology, there are opportunities in foundational infrastructure (making the technology easier to deploy or adding capabilities to what can be built for / with it) and novel applications (tools built with the new technology to help users achieve something they could not previously do).

This approach often starts by asking what new infrastructure developers are building, and then asking what applications can be built atop it. 

In XR, substantial existing infrastructure will be first-party specific to headset makers. So, it is worth considering what initial applications may be built on foundations purpose-built for available devices – and which use cases may find breakout success among users of those devices. 

XR applications


Gaming appears to be the first breaking wave of consumer XR, and will likely lead the way for the foreseeable future. Unity’s 2023 report showed that more than half of all console developers are now building for VR platforms, too. It’s been said that “Every game developer wants to be a VR game developer – they just need to find a way to get paid for it.” This may not be a problem soon enough.

According to The Economist and Omdia, global gaming spending will eclipse $185B this year, with half of consumer spend going to mobile games. As AAA titles and boredom-busting mobile games alike are rendered in XR, it stands to reason that anyone with access to an XR device will prefer gameplay in a more immersive form – meaning that a sizable share of the gaming market may shift to XR in the next decade.

Gamified consumer training is already proving its effectiveness in athletics: thousands of baseball players, amateur and professional alike – including the reigning MVP – use WIN Reality’s Quest app to perfect their swings, with measurable improvements on their performance. 

We are also excited about consumer applications in more passive streaming entertainment, social content sharing, education, and commerce. Many of the hours of daily screen time spent watching asynchronous content – social media feeds or professional productions – or browsing e-commerce sites may feel more vivid in an immersive environment. 


Hardware fragility and cost may prevent near-term widespread enterprise adoption of B2B XR across business types. Meanwhile, few people – including us – are excited about open-ended co-working in an immersive XR environment. 

But, there are impactful vertical and horizontal applications alike that may soon make enterprise XR familiar and essential in many use cases, especially at large companies. Horizontal enterprise tools may include general-purpose training and demo environments: collaborative spaces built to allow anyone to host a classroom or sales presentation. Early deployments of immersive training tools have shown efficacy in use cases as diverse as nursing, UPS driver safety, and law enforcement.    

For more specialized B2B applications, initial verticals may include architecture and interior design, product design and mechanical engineering, repair and maintenance diagnostics and installation, and healthcare diagnostics and treatment simulation – among many other sectors.

Key questions for XR application builders 

With any application, we encourage prospective founders to consider first-party risk: What core applications are platform creators likely to want to own? FaceTimeXR may be a more likely winner than a new developer’s video chat app for Apple RealityPro. But, a social app to flip through your photo library immersively, in real-time and alongside the friends in those images, may be less likely in Apple’s first-party domain.  

We also encourage XR builders to have an airtight “why XR” case for any application: what vital benefit does an immersive, co-present environment offer to your user over alternative interfaces? 

XR Infrastructure

Developer tools 

Wise founders will study the initial wave of XR adoption and ask which high-use applications are breaking, underwhelming, or borderline impossible on existing first-party or open infrastructure. Many of the most most compelling opportunities will be in bolstering core experiences: interactivity and copresence, audio, messaging, 3D asset generation, 2D/3D web interoperability, streaming reliability and observability. 

Monetization enablement 

An entire ecosystem of companies will support familiar business models, applied to XR use cases. While many elements of these business models may be unchanged for a new interface, there will undoubtedly be novel components. E-commerce checkout flows will feature transaction-splitting for social shopping. A wave of analytics and marketing tools will help businesses identify lucrative, impression-rich actions users take in XR applications, and usage-based billing providers will emerge to track and monetize novel ways of product use. 

Key questions for XR infrastructure builders 

Although removed from the end XR application consumer, XR infrastructure builders should start with this headset-adorned or AR-phone-app wielding user profile in mind. In a nascent space, an infrastructure builder needs to be sure there are enough end users who will in turn adopt what your own customers build with your tool or on your platform. Even if a developer can build something novel and powerful with your tools, your success relies on that developer’s distribution.   


These are merely a few of the possible areas to explore. The promise of XR lies in the experiences we cannot yet clearly see, and the infrastructure built to enable them. If you’re building now — send us a note to And if you’ll be in LA for LA Tech Week, join us on 6/7 at our XR breakfast — there will be plenty to discuss the week of Apple WWDC!