The Data Arms Race, "Background AI", and Defensibility
Lessons from YC W24 and beyond
Welcome to the 27th edition of Black Box. This is the third of a series on some of my recent thoughts on generative AI that I’m calling Thesis StAItments; read the other parts here and here.
I have met dozens of AI startups since I started scouting for Punch Capital in the new year. What has struck me is that in the eight months since I last sourced as an investor at DCM Ventures, the “wild west” of AI has started to close. Founders seem to have a better sense of where the opportunities are, how to build with AI, and what makes a product defensible. Here are some of my learnings:
Data arms race
Investors often ask “Why now?” One category of AI startups whose answer I really like are those that use LLMs to analyze or synthesize data at a scale that were previously inaccessible due to sheer volume or lack of structure. For example, I previously observed that many software engineers publicly their ability on code repositories like GitHub and Hugging Face. In theory, a recruiting startup could build a masked contrastive model that compares an engineer’s code with functionally equivalent code generated by an LLM to evaluate their skill. After I published that article, a reader reached out to tell me about Prog.AI, which is building essentially what I described.
What is interesting about this “Why now?” is that there is another category of startups working on the exact opposite. They want to prevent companies from leveraging data in new ways. I previously highlighted Nightshade and Glaze in my deep dive on adversarial machine learning. Another startup on this side is Ceartas, which removes copyrighted content and content that is AI generated without permission under DMCA. With a 98% takedown rate, Ceartas is the official DMCA partner of OnlyFans and the youngest provider to have a formal relationship with the platform — whose content, I imagine, must be among the most pirated and AI modified in the world.
Neither category will back down nor win, so I think this is the beginning of an arms race. But fortunately, investors do not have to pick a side. In fact, a VC who invests in both not only hedges the outcome but increases their upside if the stalemate holds. Not many investment strategies can say that!
“Background AI”
I have long believed that the best AI products are not about the AI. In fact, I always ask AI founders to explain what problem they are solving and how it works without using the word “AI”. This helps me see through the hype and understand the product better since AI is a great way to handwave past a lot of difficult questions. And of course, it is very telling if the founders cannot answer!
Some of the startups that I have met, especially those that are building B2C applications, are starting to de-emphasize AI in their messaging in favor of “jobs to be done”. This is a shift in the right direction, but these startups are often clones of existing software with built-in AI layers to automate certain workflows. It seems these companies could make the shift only because AI is a feature and not a core part of the product. Was it possible for a product to rely on AI without its users realizing — for the AI to be in the background, as a means to an end?
One of the startups that I met from the most recent YC batch answered yes. I thought that Speck was simply a no-code RPA browser extension until the founder explained that the product worked by recording the workflow that the user wanted to automate and translating the steps into a prompt for its LLM agents. This makes Speck not only much easier to use than traditional RPA but also less brittle to changes and small errors since the agent focuses on the intent of the workflow (e.g., click on this button, wherever it may be) instead of literal rules (e.g., click on this specific pixel). But had I not asked, I would have had no idea that Speck used AI at all!
Defensibility
VCs have long questioned the defensibility of AI startups since most use the same models. How do you build a moat when the core of your technology is available as an API to anyone with a credit card? I had some ideas but it was not until recently that I became optimistic that it was possible, thanks again to a couple of YC W24 startups.
One path to defensibility is input curation, which is a key part of how Hona helps physicians quickly understand medical records. While LLMs can now take in this huge mess of multimodal data due to longer context windows, it is not cost or time efficient to do so. Furthermore, LLMs have a tendency to “forget” details when given a large amount of information — which could be fatal in medical settings! To avoid these issues, Hona pre-processes the data to pull out relevant information. This is defensible because what is relevant depends on the patient and what they are coming for, which are not part of the underlying model.
Another path is using LLMs as reasoning engines. Sonia realized that other AI therapy startups had trained their models to sound like a therapist rather than think like a therapist, which meant they did little to guide users toward the goals that they needed to reach in each stage of therapy. As a result, the team trained their model on best practices from literature and case studies from therapists to create a reasoning engine that understands a user in the context of a stage of therapy and its goals. For example, Sonia recognizes if a topic should be discussed further in a future session or when to introduce a tool that might help.
The idea of augmenting standard models with domain-specific information is not a new approach to defensibility. What Hona and Sonia demonstrate is just fine-tuning on that knowledge is insufficient. Defensibility comes from how it is implemented — namely, as an important and non-obvious part of a larger workflow. I am excited to see what other ways startups come up with to integrate particular expertise into AI-driven processes. ∎
Thanks to Lucas at Speck, Adam at Hona, and Chris at Sonia for informing my views on these topics. Let me know your thoughts @jwang_18 or over LinkedIn!