The startup scene is buzzing with excitement around AI and GPT4. It’s common to see founders boasting about their AI-powered businesses, but is coolness enough to ensure success? History has shown that it’s not. A similar pattern is emerging with generative AI companies where investors are funding technology in search of a problem rather than the other way around.

This article aims to provide a comprehensive guide for founders and investors in the generative AI space. The goal is to help them navigate the growing landscape of AI startups and identify the ones that are truly unique, defensible, and sustainable.

The Business Value Proposition

When evaluating a generative AI startup, it’s crucial to ask whether the business can stand on its own without the AI component. A successful business must have a core value proposition that is not solely dependent on the technology. This is because the technology may change, become outdated, or be replicated by competitors, but the core business value should remain relevant and valuable to customers.

One example of a successful generative AI startup that follows this principle is Grammarly, which provides a writing assistant that uses AI to help users improve their grammar, spelling, and writing style. While AI is a key component of Grammarly’s product, the company’s success is based on its ability to provide a valuable service that users are willing to pay for. Grammarly’s value proposition is based on its ability to help users communicate effectively and professionally, which is not dependent solely on AI.

Another example is Zest AI, a company that uses AI to help financial institutions make more accurate credit decisions. Zest AI’s technology is used to analyze customer data and predict credit risk, but the company’s value proposition is based on its ability to help financial institutions reduce risk and increase profitability. Zest AI’s product is defensible because it provides a valuable service that competitors cannot easily replicate.

The Correctness Spectrum

AI models are probabilistic solutions to problems, which means they can get things wrong. For example, in most LLMs, the model guesses the next most probable word, which means it can hallucinate facts, draw wrong conclusions, or lie. While this issue is improving with smarter models and larger attention windows, it will always be a problem.

Founders need to consider how often their product can fail at a task and still be valuable. If a chatbot is giving medical advice directly to a patient, the correctness bar needs to be essentially perfect. However, if the chatbot is giving medical advice to a doctor who uses it as one of many tools, the bar is lower.

It’s important to focus on the right industry and user persona to turn this problem into an advantage. In some creative fields, where there isn’t necessarily a wrong answer, the model’s hallucinations can be valuable expressions.

The automated car industry has been working on the correctness problem and developing frameworks to solve it for complex models and systems. This is called the operational design domain (ODD). By bounding the ODD of an AI, it becomes much easier to test and validate the correctness of the pipeline. For self-driving cars, this means restricting the vehicles to specific road conditions, maps, and scenarios and simulating millions of driving miles.

For simpler models, an LLM with a verified ODD of nothing can simply respond that it doesn’t know how to answer a question. “I don’t know” may not be helpful, but it’s never wrong.

Optimizing Your AI Pipeline for Success

As the world becomes more reliant on AI, it’s crucial for startups to have a strong foundation in AI research. However, building and maintaining this foundation can be costly, with billions of dollars needed to stay on the cutting edge.

To make the most of your AI investment, it’s essential to optimize your pipeline with the right tools and strategies. One way to achieve this is by leveraging NFX’s 5 Layer Generative Tech Stack. This stack can help you streamline your AI workflow and build a solid foundation for future growth.

But foundational models are just the beginning. To truly differentiate your AI pipeline, you need to focus on the areas around the foundational model inference. Here are some actionable tips to optimize your AI infrastructure and drive innovation:

• Choose the right foundational models for the specific tasks you want to achieve. For example, using GPTNeox to summarize information before feeding it into GPT3’s context window can improve performance and reduce infrastructure costs.
• Invest in prompt engineering and pre-processing data with custom embeddings to enhance the performance of your generation tasks.
• Implement post-processing techniques to ensure your results align with your ODD’s bounds.
• Fine-tune your foundational models to fit your data set and achieve unique performance.
• Build a robust infrastructure to support end-to-end validation pipelines, allowing for quick testing, iteration, and continuous improvement.

Assessing Teams at AI Startups: A New Perspective

One of the biggest challenges for AI startups is building a strong team. Many founders get carried away with the hype surrounding the latest AI research and believe they need to hire PhDs from top universities to build a successful product. However, this may not always be the case.

Instead of focusing solely on research-focused ML engineers, founders should also consider hiring experienced infrastructure engineers. These engineers can help build and maintain a cutting-edge product and a sustainable business.

While having in-house experts who can read the latest papers and incorporate the latest technology into the product is important, incorporating those changes effectively often comes down to having good infrastructure. Key infrastructure hires can make a significant difference in a startup’s ability to build and maintain a sustainable business.

In the generative AI era, startups may not need as many people as they think. In fact, they might be surprised by the types of people they do need. It’s all about striking the right balance between being flashy and being sustainable.

Investors and founders alike should think of good infrastructure engineers as some of the best production ML engineers. Those who have experience working on productionized systems likely spend a significant amount of time working on infrastructure, and this experience can be invaluable to a startup.

The key takeaway here is that a successful AI startup needs a well-rounded team that includes both research-focused ML engineers and experienced infrastructure engineers. By finding the right balance, startups can build cutting-edge products and sustainable businesses that thrive in the long term.

Essential Questions to Consider for Your AI Startup

As you build your AI startup, it’s crucial to ask yourself the right questions to ensure your success. Investors will also want to know that you have thought deeply about these issues. Here are some key questions to consider:

Are You Prepared for the Future of Foundational Models?

Foundational models are rapidly advancing, and it’s essential to keep up with these changes. Ask yourself: What are you doing to stay on top of these developments? How can you leverage these models to improve your product? What AI pipeline features can you patent to differentiate your product from others in the market?

What Sets Your Product Apart from Competitors?

If someone else had access to the same models and ideas as you, what makes your product unique and hard to replicate? Consider whether your differentiation comes from pre-processing, post-processing, testing pipelines, or elsewhere. Think about how you can use your user experience, business model, or problem focus to create a product that is distinct from others.

How Important Is Accuracy for Your Product?

Consider how often your users need the correct answer to make your product viable. Build guardrails around the known unknowns and unknown unknowns, and measure correctness. Define your ODD (Operational Design Domain) to ensure your product performs consistently and reliably.

Can Your Product Be Easily Replicated Using AI Language Models?

Ask yourself how long it would take for a user to complete the same task using an AI language model like ChatGPT+. Is it a matter of finding the right prompt to add in, or does it require significant setup? Consider whether your product is easily replicated and think about ways to add value that cannot be easily replicated.

How Do You Validate Changes in Your AI Pipeline?

When making changes to your AI pipeline, you need to validate that these changes improve your product. Consider what metrics you are using to measure these improvements and any inherent biases they may have. Is your process manual or automated, and if automated, does your validation pipeline have semantic meaning? Ensure that your metrics are clear and easily understandable.

Why Is Your Team the Right Fit for This Project?

Investors will want to know why your team is the best fit for this project. Consider how well you know your customer persona, your experience in standing up and optimizing production ML infrastructure, and your ability to evaluate and test the strengths and weaknesses of AI models. Be confident in your team’s abilities and communicate this effectively to investors.

By asking yourself these questions, you can ensure that your AI startup is set up for success. Keep up with the latest developments in the field, stay focused on adding value to your users, and be confident in your team’s abilities.