Earlier, I wrote about new primitives available through LLMs to build new types of applications.
Across all the AI companies we are meeting, we’re seeing a common set of features built on top of these primitives.
Instant content generation is the first one we’re seeing pop up all over the place. All the writing assistants have this baked in eg Jasper and Copy.ai. This is the obvious one (in hindsight).
Hyper-personalization – models trained on personal data (with consent, hopefully) provide personalized experiences. We’ve seen products ingesting all of someone’s email or digital notes and then train a model to provide very personalized responses to prompts. This personalized model could then travel to your search box, your email, your slack to suggest answers or even answer on your behalf. Rewind is this on steroids.
The end of specialized knowledge – everything becomes a natural language query – applies to software development, technical manuals, language translation, and more. $10/month for an AI-based software development bot. Natural language interfaces to operate heavy machinery.
Chaining of prompts/actions – output from one prompt is used as input to the next. We saw a demo of a ChatGPT prompt which asked for the best recipe for minestrone soup, then sent the ingredients in a name-value pair format to the Instacart plug-in for immediate ordering. While this is basic, think of how this can be expanded to compress multiple otherwise pretty basic steps in a normal workflow into one uber-step. Such “light intelligence” in the LLMs will remove friction from use cases that vary slightly from user to user or situation to situation.
Workflows for specific use cases – the defensability in enterprise/consumer apps will come partly from domain-specific playbooks (aka how we work, live and collaborate) being codified into workflows. In other words, the real world is still very tough to model and there will be a gap between what a gen AI model can tell us about a CT-scan and what a human can tell us and then workflow on top to get approvals, initiate the right set of other lab tests and treatments and get insurance payments (I know, this can never be approved by regulatory authorities).
Closed loop learning – systems will automatically gather input, create output, then plug the results of that output back into their input. The result is a continously improving system, chasing its tail. Lots of agency workflows get either automated or amped up with this… more on this later.
Authors
