Can Novel Foundation Models Play a Role in National Security?
The purpose of this piece is to discuss the national security implications of foundation models for policymakers and defense and intelligence (D&I) partners.
As the CEO of a company building no-code AI enablement tools, when the recent ChatGPT posts started cropping up everywhere across my social feeds, I had to answer the question for myself, my team, and my VCs: “What does the future hold?”
The short answer is that the future will be different than we thought, but it's still ours to claim.
The Future in Context
My prediction is that large AI research organizations, like OpenAI, will continue to build great foundation models that are data-hungry, driven by openly available data, and computationally intensive (expensive) to train. ChatGPT and other similar offerings will continue to pique human interest and generate clickbait media buzz about potential controversies and ethical dilemmas. And, of course, the onslaught of hot take LinkedIn posts predicting doom for X industry Y company, and Z tool” will continue.
From a personal perspective, we’ve witnessed large groups of savvy professionals and resourceful students experimenting with generating art and AI essays. But from a business perspective, OpenAI and others will focus on expanding their business by building easy-to-use API integrations in AI’s steady march toward democratization. However, at the intersection of both these realities, is still a race between big and small organizations, and a competitive niche to fill. For example, when Stability AI released its model parameters, it forced OpenAI to then significantly lower the cost of their API calls.
But the historical problem is that early-stage AI startups often don’t have the resources to keep building foundational models. Generically-trained models for natural language processing (NLP), vision AI, etc., will increasingly get better and faster than most startups with limited resources can reckon with. As this new generation of “startups” continue to go to market with purpose-built models, skepticism from VCs and customers alike will center on two major questions:
So, the deeper question at hand is: “If you’re a startup, and you aren’t trying to be reliant on the large scale models as your core business, how do you ultimately build a billion-dollar company and scale to IPO?”
Forecasting the Future
I believe the real revenue opportunity lies in understanding that there will no longer be a sustainable economy around AI startups, but instead software startups that use AI to enable automation for businesses. This is where creativity and truly understanding your customer comes into play.
In the short term, companies building specific narrow models deployed on the edge will be able to succeed because foundation models are too large and too expensive to run on the edge. Many large factories (including many of our customers at Fortune 500 companies) aren't cloud native, let alone digitally mature. Here is where generalization vs specialization comes into play: a model that was trained on relevant domain data will outperform a generalized one that lacks experience in that domain.
In the long term, revenue growth will be determined by software that enables a full end-to-end pipeline of hardware integration, data pipeline maintenance, and connectivity with new or existing user interfaces—ultimately automating one or more manual processes.
When it comes to scaling an AI-powered business, horizontal applications no longer achieve those goals. Software will be purpose built for scaling across specific processes.
For instance, when thinking about counting and tracking maritime vessels, a successful startup will be focusing on a holistic solution stack by:
Startups will need to focus on building repeatable software that has the ability to adapt and scale to answering multiple automation process questions via a modular end-to-end platform, enabled by AI.
The key will always still be to solve practical business problems from start to finish, but using foundation models as one of the tools in the toolkit is the new name of the game. Foundation models are not the death of AI startups, but they have, quite frankly, placed a magnifying glass on the need for better product-market fit. The market for AI enablement is gigantic and will not slow down until most processes across the enterprise are automated.
I am excited by what the future holds. There is a lot of work to be done, but with continued interest in the AI field, entrepreneurs will keep adapting. Stay strong, the future is still ours to claim!
Note: Huge thank you to Michael Kanaan, Cliff Massey and Dr. Jean-Baptiste Boin for edits and technical feedback.