AI’s eigenquestions
What should companies building on top of foundation models be asking themselves?
Every paradigm shift in technology raises new eigenquestions – the irreducible set of questions founders must answer to build category-defining companies. Some past examples:
Birth of the internet: How do you make digital content easy to create and access?
Shift from desktop to mobile: What can people do now that everyone has a computer with them all the time?
Migration from on prem to cloud: What new collaborative workflows are possible when you can read / write information to / from any device with an internet connection?
Asking the correct eigenquestions is often as important as identifying the best applications. Get the basic shape of a great idea right but the eigenquestions wrong and you’ll still likely have trouble building a breakout product. For example, Seamless’s mobile app helped unlock new demand from people ordering food from their phones. But DoorDash grasped that mobile didn’t just mean new customers, it also created a new supply network of dashers. Whereas Seamless focused on answering “how do we port our desktop UX to mobile?”, DoorDash realized that the right eigenquestion was “now that people carry a computer in their pocket, what new types of coordination problems can we solve?”
Foundation models have the potential to be another such paradigm shift. They are already radically improving product experiences – try taking away Github Copilot from your engineering team. For start-ups building products on top of foundation models, coming out on top means asking a new set of eigenquestions.
I’ve outlined 7 candidate eigenquestions below. There are several ways this paradigm could play out (for instance, it’s unclear how the tension between open source and closed source models will resolve), so I’ll frame some of the trade-offs each entails.
How can foundation models enable a quantum leap in UX?
Users may come to a new platform for a mindblowing AI feature, but they will only stay if the platform offers a radically better UX.
Foundation models make it possible for people to use software by simply communicating their goals instead of worrying about the logical steps required to get there. Adept, RunwayML, and Perplexity.ai have leaned into this philosophy aggressively with their emphasis on declarative, text-based interfaces. Why spend hours on painstaking rotoscoping when you can tell Runway to remove a video background for you?
RunwayML rotoscoping via text interface
Similarly, there is an opportunity to create new UXes that dramatically streamline the effort, time, and number of software platforms required to complete digital workflows. Think populating Salesforce CRM entries by having an AI listen to a customer call, for example.
Iconic AI companies will achieve this by deeply integrating foundation models with user workflows. In the Runway example above, the model just removes a background from a video while the experience of describing what you want to remove and seeing that outcome instantly is what makes the product feel magical.
Can you build a new system of record by using AI to interface with legacy platforms?
During the 2010s, the difficulty of migrating data from legacy systems of record (e.g., Netsuite, Salesforce, Oracle) slowed innovation cycles. Silicon Valley is littered with the corpses of start-ups that tried to make better CRMs and ERPs. SaaS companies that succeeded in building breakout products often did so by sitting on top of these clunky databases instead of replacing them.
Foundation models could break that pattern. Migrations are hard because they involve transforming data schemas, re-indexing data, and managing dependencies. Foundation models require none of this. Since they accept data of any form and are inherently searchable, model application companies can become new systems of record. These models are able to frictionlessly port data from existing datastores and intake data without requiring users to learn a new UI. Imagine being able to seamlessly export your Salesforce data to a sleek, intuitive CRM in a matter of days. No army of consultants required.
Building a new system of record is never easy. Deploying foundation models to enterprises entails handling messy concerns around proprietary data, privacy controls, and idiosyncratic workflows. But if model application companies figure out how to remove the largest roadblock to new product adoption (the legacy datastore), they can enjoy faster time-to-value, grow more rapidly, and capture larger market share. They could end up being more disruptive than their SaaS predecessors.
Will AI moats come from the model, data, or workflow?
SaaS moats can be distilled into a few core primitives:
Owning the system of record: Becoming the datastore customers rely on for critical business information and logic creates a gravitational field around your product.
Acting as the system of engagement: Building an abstraction layer above the system of record allows your customers to do their jobs more easily.
Network effects: The only true flywheel in software.
Most foundation model application companies will be a special breed of SaaS. Their power comes from using these models’ to become a discontinuously delightful system of engagement and, ideally, new systems of record. Doing so requires developing an edge in at least one of: (1) the underlying model, (2) the data the model is fine-tuned on, (3) how the model is adapted to the problem you are solving, and (4) the workflow that the model enables.
Model: With researchers releasing new models every few weeks, AI product companies that seek to build and maintain a proprietary model face a steep uphill battle. Not only do they have to create software that users love but also do so while staying a step ahead of big tech, AI labs, and open source researchers. Models don’t stay private for long.
Text foundation model democratization timeline
Data: Models are only ever as good as the data powering them. Unique datasets can lead to unique product capabilities.
Model adaptation: Off-the-shelf foundation models don’t yet perform phenomenally well on many domain-specific tasks (e.g. medical, legal) out of the box. Coaxing them into outputting relevant, reliable results is the difference between magical and glitchy products.
Workflow: The core insight behind some of the most successful SaaS companies has been a UX that simplifies or solves a painful problem. AI companies will be no different.
Is your data strategy optimized to improve your UX?
In an AI-native world, the best UX involves a UI optimized for data capture. TikTok overtook Twitter, YouTube, and Instagram because it carefully designed each screen to pick up on user reactions and used that information to rerank its video queue in real-time.
TikTok dominates consumer social time spent
As foundation model capabilities grow, maximizing performance will increasingly depend on getting your pretrained model to generalize well to your problem domain. Broadly speaking, there are three ways you can leverage data to achieve this goal:
Fine tuning the model with pre-existing data: If the task you want your model to perform is substantially different than the type of data it was trained on, you will need to fine tune it on proprietary data. GPT probably can’t write drug prescriptions without learning from a new corpus of doctor notes.
Reinforcement learning from human feedback (RLHF): This is most similar to the TikTok example. RLFH is how model application companies will develop a data flywheel: user interactions provide feedback on model outputs, teaching the model how to produce better results in response to similar prompts. Crucially, companies can only acquire this data by putting a product in the hands of active users.
Prompt engineering: While foundation models may soon develop a deep understanding of user intent that renders prompt engineering obsolete, for now, companies must carefully select model prompts to get intelligible results. Translating user inputs into semantically equivalent prompts that produce the best output (so that users are shielded from the need to prompt engineer themselves) will drive a meaningfully better UX.
How much research DNA should your team have?
Foundation models were born in corporate research labs. Their lightning-fast pace of advancement gives an edge to people who can see where the field is headed. Will start-up teams without research DNA fall behind?
Early pioneers have taken a range of approaches. On one end of the spectrum, Adept’s founding team consists of some of the most distinguished researchers in the field. Striking a middle ground, RunwayML’s C-suite combines product, design, and hardcore AI research backgrounds. On the other extreme, Jasper.ai hired their first machine learning engineer just a few months ago only after scaling to nearly $80M ARR.
If history is any guide, scrappiness and product ambition will matter more than research credentials for company success. But, early on, a research mindset could prove important when wrangling with foundation models. Creativity around prompt engineering, fine tuning, and error handling will likely be one of the hallmarks of breakout innovators.
Underpinning this choice is a decision about what parts of the foundation model stack to build vs buy …
Is it better to outsource or insource model operationalization?
Progress in the foundation space has been blazingly fast. In the short-term, training a model from scratch means setting yourself up to be rug pulled unless you are building around a novel data modality (e.g. Adept’s action model). Yet, AI product companies must still decide whether to use a third party model API (for instance, OpenAI’s GPT) or to fine tune, host, and deploy a pretrained open source model.
For now, these options represent a tradeoff between simplicity, on one hand, and control and margin on the other. Building with an API promises faster iteration cycles and requires less machine learning expertise. However, this approach leaves companies with less ability to customize model behavior and subjects them to vendor pricing, terms-of-service, data policies, and latency. This final attribute is particularly critical considering how much it impacts the UX.
Conversely, an open source approach gives you the freedom to optimize model output tradeoffs (e.g., between latency, cost, and accuracy) based on your target application. And bypassing third party APIs can juice unit economics. But figuring out how to distill, host, serve, and update a foundation model yourself involves much more R&D – and doesn’t necessarily make your beer taste better.
Should you be using a foundation model at all?
Admittedly, this is a funny question to end on. But in many cases the answer is still no! While foundation models are a powerful tool, they are not a panacea. Foundation models excel at tasks where:
The goal is translating high level intent to digital output (content generation / manipulation, information retrieval, action completion).
Input data is unstructured.
You have only a few examples to train / fine-tune on. For instance, classification problems like detecting credit card fraud.
For jobs that require extreme accuracy, lots of structured data, minimal latency, or low cost, smaller models are your best bet for now.
What did I miss?
The companies that navigate these eigenquestions correctly have an opportunity to reinvent the internet. Just as the cloud paradigm shift made human-to-human multiplayer UXes possible, the foundation model paradigm shift should usher in an era of human-to-machine (Copilot-like) interfaces.
If you think I missed any big eigenquestions, shoot me a line at philip [at] thrivecap [dot] com.
RLHF? ;)
I think this is a worthwhile exercise, but how do you distill your 7 to ones as fundamental as what you listed for past era's.
If you look at generative models, one would be: How should you design tools knowing that one can create infinite filler or started content in any medium?