Abstract: AI Startup Ideas

[Disclaimer ‘These Views Are My Own’]

3. my new understanding of AI startup ideas, 08/2024

After watching two YC videos (Tarpit Ideas: The Sequel, Consumer AI means some tarpit ideas are no longer tarpits.) discussing AI’s impact on conventional startup ideas and spending several months exploring concepts, I’ve shifted my perspective on popular business approaches. I now believe it’s acceptable to work in general, well-established fields. However, the key is to develop a niche idea within that space: you need to propose a specific solution that addresses distinct customer pain points.

2. Types of AI startup ideas, 05/2024

Now that I have been discussing startup ideas with friends, people on YC’s cofounder matchmaking, and members of On Deck and South Park Commons, I’ve had the opportunity to explore a wide range of startup ideas. As a result, I feel I have a better understanding of the AI startup ideas that people are pursuing. I am now developing my own criteria to judge whether an idea is good and whether it is a good fit for me to work on. I have some learnings to share here.

In my opinion, in the age of AI, some of the most popular B2B ideas—with at least five teams I know working on them—are AI marketing, AI recruiting, and AI sales. For B2C, popular ideas include AI companions, AI mental health, AI productivity, and AI meditation.

As always, there are some tarpit ideas in the age of AI, such as AI travel planners, which we should definitely avoid. But what about popular non-tarpit ideas?

Working on popular non-tarpit ideas doesn’t mean you can’t build a successful startup. You can succeed if: (1) you are the top team pitching the same general idea to VCs, who recognize its popularity and choose the best team to execute it; or (2) you have unique insights and advantages that allow you to develop a distinctive solution or present it in an innovative format. However, even then, there are often many venture-backed teams executing the same idea well, making it difficult for one company to dominate the market. In the long run, many companies might survive by focusing on a subfield, with no real winners dominating the market.

On the other hand, if no one else is working on the same problem, that could also indicate something is wrong. (As discussed with Sophia on SPC Slack)

2.2 B2B developer tools

I also changed my perspective on B2B developer tools. Initially, I was very interested in this area because I am an MLE and have benefited greatly from popular frameworks like HuggingFace Transformers, LangChain, Chroma, etc. However, I quickly realized that finding product-market fit (PMF) for these developer tools is very challenging. While you can get a lot of stars on your GitHub repo and many PyPI downloads because developers find it useful for free, it’s a different game when you want to grow your revenue. Additionally, working on developer tools in a rapidly evolving field like RAG or AI agents can be difficult, as you need to keep up with the latest research; otherwise, developers may stop using your framework.

2.3 Is your AI expertise a real moat?

I have an AIML background and have not only deployed models into production like other MLEs but also published original research and applied science papers at top AI conferences (< 600 citations). When I entered the startup world, I believed my AIML background was my greatest asset. I’ve noticed that some startups are founded by researchers with impressive backgrounds, such as Arthur Mensch (Mistral AI, DeepMind, ~9k citations), Curtis Northcutt (Cleanlab, MIT, ~2.5k citations), Ari Morcos (DatologyAI, Meta AI, ~7k citations), and Chenlin Meng (Pika, Stanford, ~6.8k citations). This made me wonder: to what extent is an AIML research background a real asset? Does having more papers, more citations, or a prestigious former employer or PhD school make a significant difference?

For non-deep tech startups, I believe Perplexity AI’s success is not solely due to Aravind Srinivas being a UC Berkeley CS PhD graduate with thousands of citations, but rather because it is a good product that found product-market fit (PMF). If you’re working on an AIML startup, it’s advantageous to be able to do technical proof-of-concepts by yourself and come up with various technical solutions. However, you don’t need over 5k citations to prove your capability, and this alone is not enough for an AI startup entrepreneur. You also need to have a good product sense and be a proficient coder. The latter is not difficult in the age of ChatGPT, but the former takes time and is similar to writing grants, albeit with different strategies.

For deep tech startups, such as those working on foundation models, quantum computing, or on-device model compression, it is true that the more citations you have, the better your vision, and the more likely you are to succeed. These types of startups are not rapid-fire; you will need venture capital to get started, as you will at least need significant computing resources. Thus, you need your reputation to convince VCs to invest in you.

2.4 Is training a model always more advanced?

Many MLEs and researchers believe that training models is always more advanced and that any wrapper over the OpenAI API is replaceable in the long run. I don’t agree. There are numerous papers that use ChatGPT to collect training data for specific tasks and then fine-tune a smaller model with this data, either through full parameter tuning or parameter-efficient fine-tuning. Is this meaningful research? I don’t think so. It is straightforward, given the abundance of training frameworks and platforms (including ChatGPT’s fine-tuning of GPT-3.5) that require minimal effort to train a model. Researchers need to explore the full potential of ChatGPT and develop an intuition on when to use closed-source models and when to train small models. This is especially important for AI agent research.

1. My understanding of different types of AI startup, 03/2024

Since March 2023, with the rise of ChatGPT, I have participated in discussions about AI startups with some of my friends. I had the opportunity to offer some opinions of my own. Unfortunately, from September 2023 to January 2024, I didn’t have the bandwidth to work on this, as I was too busy with my work to take on any side projects.

Beginning this February, I restarted discussing startup ideas with others as a way to refresh my mind. After talking with many people who shared their ideas and experiences of starting a startup, I gained more insights into what an AI startup is like, the different startups out there, and their interests.

My understanding of the different types of startups is as follows: (1) Foundation model-related, (2) Ecosystem-related, (3) Infrastructure-related, (4) Application-focused.

(1) Regarding foundation models, startups such as OpenAI, Anthropic, Cohere, Together AI, Eleven Labs, Mistral, and Voyage AI are examples of companies in this category. They are building foundational pre-trained models, including (multimodal) LLM architecture, embedding models, and reranking models. Additionally, they provide or will offer customers the flexibility to fine-tune their models using customer-specific data. Competing in this space requires significant domain expertise and funding since training pre-trained models demands extensive GPU resources.

(2) Companies like Hugging Face, Langchain, and LambdaIndex represent open-source community ecosystem startups. They are creating a paradigm for developers to develop their applications more efficiently. If we consider LLMs as operating systems and AI-powered applications as software, these ecosystem companies could position themselves as programming languages like Python or programming frameworks like Django.

(3) As for infrastructure-related companies, there are numerous firms that build platforms for vector database, efficient training, model deployment, and data labeling. Examples include Pinecone, Chroma, Databricks, Anyscale, Scale.ai, Snorkel, and many other fine-tuning platforms and ML infrastructure/MLOps automation companies.

(4) Speaking of application-focused startups, there is a vast array of AI startups in this space, many of which are consumer products. For instance, many are developing chatbots, voice assistants, and various types of AI copilots and agents. These AI agents are used to enhance user productivity, automate selected workflows by users or the agents themselves, and assist users in complex workflows. (I am working on a post dedicated to AI agents and autonomous systems, so stay tuned.)

A personal AI agent is one of the most straightforward ideas when thinking about AI agents. However, I feel this space will likely be dominated by operating system providers like Apple (iPhone, iPad, MacBook), Microsoft, and Google (Android). They can integrate AI agents into the operating system, leveraging direct access to all data. Apple has a unique position here; if they continue optimizing inference on M3 chips or optimizing the M3 chips themselves, they could potentially develop the first such AI agent integrated seamlessly on phones or laptops.

Many startups in this category aim to build specific AI agent for a specific domain. Entrepreneurs in this space are more diverse than in others. For example, many attorneys want to build AI attorney-facing customers and attorney copilots/chatbots to assist attorneys. They believe they have unique positions in this space due to their domain expertise and the apparent barriers that prevent software engineers without industry experience from building such tools. However, they typically define the workflows and such domain-specific AI agents themselves, using LLM for function calls or to implement a series of function calls. At the same time, I would say there are still many more startups focusing on building AI agents as developer tools, since it’s straightforward for developers to think about those ideas to facilitate their day-to-day work. For example, automatic code debuggers, automatic code reviewers, coding copilots (there are already many out there like Github Copilot and Codewhisper), and automatic cluster management, etc.

[To Be Continued]