If you want something done right...
*** A quick plug: The next Scaling Biotech webinar is tomorrow, this time with Sphinx Bio’s Nicholas Larus-Stone. We’re calling it “Unlocking Biotech Data: How Sphinx uses AI to move biotechs faster and make better use of their data.” It’ll be on Thursday, November 21st at 2pm EST/11am PST and you can sign up here. ****
Last week, I wrote about what I think the limitations and opportunities are for using LLMs to allow (non-computational) biologists to take on some of the work currently done by computational biologists. I’m planning to write about more of these kinds of applications in the coming weeks, but this week I want to suggest a model for thinking about tasks like this that involve creating structured information. This could be data, meta, configuration, instructions, etc.
Here’s my thesis: On the spectrum from keyboard to mouse, I think of LLMs as sitting past the mouse, at the far end from the keyboard.
To be clear, I’m not saying that an LLM is a glorified keyboard… but for these use cases, it kind of is. Hear me out:
Before ChatGPT, there were basically two ways to enter structured information into a computer: 1) Type it into an IDE - code or a config language like JSON or YAML. 2) Click around in a mouse-based UI. In between is the option of using a mouse-based UI with a bunch of keyboard shortcuts so you barely have to touch the mouse.
For folks who are very familiar with the form of the structured information, using just the keyboard tends to be faster. An IDE is ideal, but keyboard shortcuts in a UI are better than just the mouse.
Someone who isn’t familiar with the task - either because they’re new to it or they don’t do it often enough - is going to be a lot slower no matter what. But they will be less slow clicking around with the mouse than trying to figure the out JSON or YAML or code. (And they’re not going to remember keyboard shortcuts.)
In other words, mouse-based interactions are better for less familiar users, while keyboard options are better for experts. Less familiar users who do the task often enough to become experts will often switch from mouse to keyboard. (Usually this means adopting keyboard shortcuts, except in the rare cases where there’s a UI and an API with the same functionality, but that’s a rant for another day…)
We get this dynamic because the learning curve on a UI is much more agreeable than the learning curve on a config language/API. But guess what has an even better learning curve? ChatGPT.
In other words, in cases where a task is so fundamentally complex that even a UI is hard for new and infrequent users, an LLM can do the work of translating a natural language description of what the user wants into structured information. Putting the results in a UI that the user can check and fix is still important. But verifying the config in a UI is much easier for a casual user than generating it from scratch.
On the other hand, I think we’ll still see the same pattern for more expert users: going through an LLM will become more of a pain as users become more expert. In particular, once a user comes to understand the task well enough to go straight to the UI, describing what they want in free text will start to feel like more work than just doing it themselves.
So in the same way that a mouse-based UI often functions as a gateway to the keyboard, LLMs for these tasks will start to function as gateways to the mouse.
Of course, it’s foolish to make predictions about these things. But I couldn’t help it. We’ll see. Next week, it’s back to more applications of LLMs.
Thanks for reading this week’s Scaling Biotech! I really appreciate your continued support, and I read every comment and reply.
As a reminder, I offer several services to help connect biotech teams with tools, practices and expertise to make their organizations more data driven.
The Biotech Reference Stack is a website designed to help biotech data teams identify the tools they need and figure out how to put them together.
For help navigating the Reference Stack, sign up for a free consultation call to clarify a problem you're facing and identify the best options to evaluate.
Or if you’re building software that makes biotech more data driven, find out how to add your app to the Reference Stack.