Why is AI like Interstellar Travel?
There’s a curious dilemma in the space sciences known as the Incessant Obsolescence Postulate (IOP). It concerns interstellar travel. Essentially, the IOP says something like: any spacecraft you launch today risks being overtaken by a future one if technology improves fast enough. Hence the dilemma: launch now or wait?
The resolution depends on your expectations for technological progress - which is, of course, uncertain.
There’s a direct analogue to the IOP in the world of AI adoption: deploy today and you risk creating “technical debt” as your system quickly becomes obsolescent; wait and you miss out on the benefits you’d otherwise accumulate in the interim.
You might be thinking, “This isn’t an issue - when new models are released, we simply switch to using them; no technical debt here!”.
Whoa there partner. Maybe; maybe not. Here’s a few ways that tech debt can accumulate:
Your prompts, eval suites (you have those, right?) and guardrails have been carefully designed to work with your deployed models. You do need to revisit them when switching to a new model endpoint.
You may be locked in to a single vendor’s ecosystem via their models’ capabilities and native integrations, making switching to a new best-of-breed frontier model a headache.
What if you have non-engineering processes that constrain your abilities to deploy: model compliance reviews or certification processes, for instance?
What if you fine-tune your models? There could be a non-trivial re-training cost involved.
What about your end users? Are you certain they will prefer the personality and behaviour of a new model over the old? You will likely need to test this before switching.
The fact of the matter is: you probably have work to do when models update. So should you launch now, or wait?
Despite what I just said, in 95% of cases, you launch now. (The 5% of “waits”, since you ask, are those heavily regulated industries where there is a palpable risk to raising your head above the parapet and deploying a technology that regulators are wary of.)
Why?
Setting the current frothiness and hyperbole aside, Generative AI is a game-changing, general purpose technology that will create structural shifts in how we operate businesses once we’re onto the Slope of Enlightenment.
It’s a new technology and we’re all learning how best to use it. How you should be using it in your business is not the same as the way I should be using it in mine. You need to figure that out through a virtuous cycle of trialling, learning and adapting. The process of institutional learning cannot be shortcut, neither will the requirement to learn be eclipsed by the better models of the future.
Take the historical example of electronification: early power grids were expensive and unreliable. The companies that learned how to rewire their factories and redesign their manufacturing workflows saw some gains. Later, electricity got cheaper and more reliable. The early movers could rapidly compound their advantage whilst the late adopters scrambled to start on their own learning journeys.
Future AI models may be more powerful - and continually updating your AI workflows to incorporate them is a headache - but they won’t wire your organization for you. Start learning now; don’t end up fumbling in the dark later.