"Tool for the job" vs "Good enough"
Before we had smart phones, we had handheld devices like this:
It didn’t take long before smart phones largely supplanted these handheld devices.
The iPhone offered games and a built-in calculator:
Games like Doodle Jump and the iPhone calculator were solid options, but they don’t really hold a candle to Legend of Zelda or chunky graphing calculators.
A yearn for more utility (along with some nostalgia) prompted me to download emulators for both TI-83 (GraphNCalc83) and Game Boy (RetroArch) so I can enjoy them on my phone:
TI and Nintendo didn’t give up completely with handheld devices, they continue innovating to this day:
But at the same time, the smart phone and tablet gaming and calculator ecosystem have innovated too, and the net result is iPhone and Android are good enough to not warrant buying separate gaming and calculator devices for many people.
Healthcare devices
My capstone in college involved looking at heart data for both consumer wearables and medical devices. We compared the data quality and predictive utility of a FitBit and a Holter monitor.
The FitBit wasn’t bad, but the Holter monitor was measurable superior.
When it comes to safety-critical use cases, “good enough” doesn’t cut it.
The only time when “good enough” should be used is if the alternative is “nothing at all” — like in field medicine.
Unbundling
When something gets used like a Swiss Army Knife to solve a variety of diverse problems, people start trying to unbundle it. They believe there has to a better way to do this. And they are correct — there likely is a better way. But when something is in place and works “good enough”, you’re fighting against inertia.
There’s been a lot of analysis on the millions (if not billions) of VC dollars spent trying to unbundle Excel and Craigslist.
Despite many attempts to unbundle them, Excel and Craigslist are still going strong (with some minor turbulence along the way).
LLM’s
People are trying to apply LLM-based technology to a bunch of applications right now.
Instead of understanding peoples’ real needs then designing a solution that happens to use AI, many of these attempts are a sprint to try to apply AI to as many areas as possible in the absence of a valid problem statement.
This is analogous to how companies feel like they need to put touchscreens everywhere:
The truth is LLM’s are really good at a wide variety of tasks. They can clean unstructured data, they can convert prompts with well thought-out logic to workable code, they can de-noise search results, and more.
I think we’ll see LLM-related applications wipe out a huge swath of existing software applications. The results may be better software in some cases, but I suspect we’ll see a lot of “good enough.”
And like any new tech, it’ll take more time and effort to be applied to safety critical use cases.
Machine learning helps human learning
Systems like ChatGPT and NotebookLM are changing the way humans learn and interact with existing knowledge bases.
In a talk about AI and leaning, Sal Khan, the founder of Khan Academy, expressed his optimism about the future of learning:
If you look at most of human history, we’ve always known what the gold standard in learning is… If you were privileged enough to get an education, you usually got a tutor. If you needed to speed up, they’d speed up. If you needed to slow down, they’d slow down… [AI can] increasingly approximate that type of one-on-one personalized experience that the best educations have always been throughout history, but to do it at scale.
Effective learning is an important input to innovation in any field.
Even if companies bolt LLM’s onto everything just because it’s trendy, I am very optimistic for the ways that AI can be used to help bring knowledge and personalized learning to curious minds around the globe1.
Hallucinations are still an open issue, and LLM’s heavily bias the English language, which obviously undercuts my “around the globe” statement. I think these 2 areas deserve a lot of attention if we’re to fully reap the education benefits of AI.