Trying to replicate human action and cognition is a large but finite opportunity; trying to supplement human behavior (or create entirely new ones) is an infinite and unconstrained opportunity. We're more interested in apps that give users superpowers over the things they inherently want to do and leverage the distinctive capabilities of foundation models to augment human capacity intensely in a narrow domain, or for surprising, unusual, non-obvious, augmentations. A great example is Re:Collect, which allows users to effortlessly connect and retrieve relevant information they’ve seen from across the web for their own creative ideation.

The real unlock with LLMs is that they are large enough to be able to learn semantic structure. As such, products that make deliberate use of semantic reasoning (summarization, synthesis, argumentation, intonation, etc.) are far more interesting to us than ones that purely simulate syntactic understanding. For example, Maestro helps teams synthesize what the teams’ individual and collective goals are and minimize miscommunication errors.

Creativity is not just drafting. It's editing. The conversational and interactive UIs that power these models have a long way to go in terms of using semantic understanding to retain context, personalize and get better at specific tasks. We're interested in companies that are thinking about UIs that focus on the context of interactions over time, like Second Voice, an LLM-based representation of your own inner voice, built to help you process your emotions and feelings via dialogue.

At their core, these models are probabilistic - they cannot create outputs with certainty. As such, they are ideally suited for usage where there isn’t necessarily a right answer. For example, with most search queries people are looking for deterministic answers. Regardless of the model/product’s overall fidelity rate, there will always be misses (false positives and negatives), and the cost of those misses can be catastrophic to users (imagine a car that makes a wrong turn or a medical chatbot that provides a mis-diagnosis). We’re interested in products and activities that are aware of what fidelity their tools need to provide and how to achieve it. A great example here is Wand, which allows artists to create probabilistic versions of some original material - imagine a game designer who can create NPC designs with as much detail as the Main Character.

There are a lot of products that are refining unimodal creation (text, image, audio, video, etc), but human synthesis is always multimodal. Valuable apps will allow us to work with multimodal outputs by chaining different models together, or even using models trained across modalities and types of content. In addition, we think new models will make use of new atomic data structures, graphs, and more. With more than one modality, you’re increasing the surface area of possible use cases, like Fermat, which gives generative toolkits for brainstorming to creative teams that work across modalities, like product designers, storyboard artists, and marketers.

Creativity is collaborative. Startups focusing on the infrastructure required to allow for collaboration, alongside these superpower tools, will be set up for a future at scale, where creators can truly make a living off of these platforms. A great example is Never Before Heard Sounds, a new cloud based audio workstation that allows anybody to make high quality music without any prior experience, while also cleanly attributing the inspiring original artists of any work.

We think value will exist across applications, middleware, and open & closed source foundational models.Every developer has received access to these transformative models, including your competitors. We’re looking for companies that have put immense thought into not only what allows their product to supercede incumbents, but also maintain an edge over future products down the line. In this context, that could mean running/training your own foundational model, fine-tuning an existing model based on contextual inputs, or using RLHF methods to improve your product. We are agnostic to the architecture of your product but have a preference for technical depth.