In the last 6 months, the world of generative AI has seen an explosion of interest and hype. There is no denying that GenAI is changing the landscape of software development. GenAI is disrupting the industries, software and the way we work, and it is happening at an incredibly fast pace. Every knowledge worker is playing the game of catching up with so many AI models and sexy tools getting launched every week.

SaaS as an industry is undergoing a complete disruption; nearly all the startups we have evaluated in the last 3 months were using GenAI in some form. The GenAI tech stack has been rapidly evolving, with the emergence of several new models and tools.

We broadly look at the existing stack in five layers-

Infrastructure layer includes hardware and cloud platforms which provide compute hardware and GPU. This includes companies such as Nvidia, AWS, GCP, Azure.

Foundation Models are GenAI models on top of which the entire stack is being built. They are AI neural networks trained on massive unlabeled datasets, enabling them to perform a diverse range of tasks, including test/ image/ audio/ code generation, text translation, and summarization. GPT-3, PaLM2, LLaMa are the well-known LLMs. Some open-source models trained on much smaller parameters are also getting interest for specific use cases.

Companies are using a combination of different models to increase accuracy and performance. A layer of domain/ vertical-specific models (health, legal, ecommerce, finance etc.) over foundation models has become a well adopted practice.

Next is a layer of tools that enables the use of these models, which we call Enablers. These tools are critical for the rapid adoption of foundation models by developers/ businesses, facilitating the shift where every company large or small is working towards adding GenAI capabilities to their products. Building production-ready AI apps is a challenging task that involves various steps. Starting with infrastructure setup, model hosting, database configuration, data collection and preparation tools, model selection, training or fine-tuning with your data, orchestrating different models based on use cases, integrating with your systems, deployment and maintenance of models, monitoring performance and cost of these models – all are part of this complex process. Hence, the emergence of enablers. Enablers offer capabilities such as orchestration, model management, observability, compliance, security, and more. Pinecone and Langchain are the most popular ones in this layer. Pinecone is a managed, cloud-native vector database with a simple API and no infrastructure hassles. Langchain is a framework for developing applications powered by language models.

Next is the Application layer, where customer-facing applications are built on top of GenAI models. It includes tools like Jasper, Glean, Copy.ai, Rephrase.ai. With GenAI, companies are providing better personalized customer experience, the conversation interface is taking over the clunky software UI and reducing their time to value.

Beyond all the excitement, it’s important to note that building and scaling an AI native tool is not as easy as it looks from all the Twitter chatter. We have spoken to many GenAI founders and tried to understand the challenges they are facing in customer adoption and product building. There are few interesting insights –

1. Most of the GenAI tools from India are very early. They are seeing a lot of excitement and adoption from the market. Users are trying out different tools, but there is high churn. Prosumers and SMBs are the initial adopters of these products, but founders have to crack the retention and path to monetisation.
2. Generational and non-critical use cases are seeing more traction. Generational use cases are some kind content generation- text, image, blog, code, audio etc. Non-critical use cases are business use cases where the cost of failure is not high and can work with not-so-high accuracy – sales, marketing, hiring etc. In high value use cases like cybersecurity, there is a cautious adoption of GenAI.
3. GTM motion for enterprise SaaS remains the same. There is a sense of curiosity among enterprises regarding GenAI and acceptance to use it in their workflow but they are cautious. Their concerns about data privacy, security, and compliance persist, and there is limited trust in public LLMs. We are seeing a trend of building on the client’s VPC, which can increase the time to adoption.
4. ChatGPT has changed the customers’ expectations. Customers are conversing with tools in natural language, they are not only asking for information or insights from the data, they are expecting end-to-end tasks to be done. The expectation is AGI (Artificial General Intelligence). We spoke with an AI native marketing product from India where customers are spending more than 30 min every session and asking in one single command to look into the data, identify cart dropout and design and send a campaign to them.
5. There is a lot of adoption of picks and shovels- libraries, tools helping developers build apps for different use cases. Autonomous agents have become very popular. SuperAGI is one of our portfolio companies, which has gained more than 7k GitHub stars in two weeks. It is one of the trending repositories on GitHub. It is a dev-first open-source autonomous AI agent framework, enabling developers to build, manage and run useful autonomous agents quickly and reliably.

Based on our learnings over the last few months, we are more bullish on a few spaces in GenAI where we believe large outcomes from India can be generated:

1. Picks and Shovels for developers: GenAI tech stack has become fairly complex and it is changing rapidly, only LLMs can’t help with production level use cases. As discussed in the above section, it is a long iterative process and these tools help developers build, experiment, train, and compare quickly. This is a new need that has emerged while developing in GenAI. Vector database, framework, model orchestration layer, autonomous agents are some examples of the types of tools getting created. Pinecone and Langchain fall under this category.
2. No code enablers for businesses: We have seen copilot (GenAI) in a box model, where the entire AI stack is taken care of by the tool, you just have to integrate with your existing software. This new layer is emerging in the tech stack and it can be an opportunity for new players.
3. ModelOps/LLMOps:MLOps used to be a not-so-attractive industry until 6 months ago. With GenAI, more models are going into production and none of the existing MLOps were made for that. Existing MLOps companies are adding capabilities and a new set of ModelOps tools are also emerging. ModelOps/ LLMOps tools include model lifecycle management, observability, data security and privacy, and model monitoring.
4. Autonomous Agents:As discussed above, customer expectation is for AGI which can be provided by autonomous agents. They have the ability to analyze intricate problems, solve them iteratively, and take actions. AutoGPT is the most popular one, it has more than 139K stars on GitHub. In today’s form, it is not easy to use AutoGPT for business use cases. AGI is still in the early phase and will take some time to reach the customers’ expectation levels.
5. AI native companies: On the application layer, we are keen on these companies. In many use cases, incumbents have the right to win because of distribution advantage. That’s why we believe all the niche use case products will get commoditised as incumbents will launch it as a feature. However, if you are building an AI native software with some vertical focus and data flywheel, then there is an opportunity to deliver better accuracy compared to incumbents.
6. GenAI-powered Vertical SaaS: Vertical LLMs are trained on curated high-quality data from a specific industry. This allows Vertical LLMs to generate more accurate and relevant results. Legal, health, and finance are among the industries where a lot of knowledge resides in the massive historic data, which is the play of Vertical LLMs. GenAI-powered Vertical SaaS companies are getting popular, such as Hippocratic in healthcare and Evenup and Harvey.ai in legal.

A lot is happening in the GenAI world and we have been continuously learning about the latest developments. Our thesis will keep evolving. We will release a series of articles on GenAI to keep you updated on our learnings. If you are a founder building in GenAI or an enthusiast, we would like to have a discussion. Please feel free to reach out to veenu@kae-capital.com or sarthak@kae-capital.com.