← Blog

When is a problem truly solved?

· 5 min read

For the past couple of weeks, working on Vulture Labs, I’ve been thinking pretty heavily about this topic. Namely, when do you truly solve a problem, and how do you figure out if your solution is truly optimal.

What we do is context extraction from Video, we’re very good at understanding WHAT, WHO, WHERE, and WHY in video or live video streams. Now with adjacent context, such as context about the environment, personnel on-site, or other company infos, you can tie in even more information to the gathered data.

For example: Interpreting an asset’s interactions in a form of timeline, tying in the whole chain of custody. Cross-referencing with the people that actually touched the thing, and exposing it as a continuously running service. Gathering a bunch of assets of a certain type(e.g. a Server Cart) and recreating timelines and interaction graphs for all of them.

Sounds cool, but if you look at this from other angles, is it truly an “Optimal” solution. Solving this problem in this way?

Continuing on the server cart example: Datacenters around the world constantly receive equipment that should not be touched under any circumstance by anyone until reaching some conditions. Meta has robots(AGVs) in their datacenters to move equipment around.

Now, Information Theory says that if the concept is encoded in the data, there should be some way to retrieve it. Tracking stuff across cameras and interpreting it all works fine. But is this all necessary?

Couldn’t they just put something like a space blanket on the thing, and track capacitance(similar to touchscreen)? It sounds right, and it should work for those purposes.

Now, my dilemma is. If you come up with a suboptimal solution, that can solve multiple problems. Is that more valuable than developing 1,000 companies that fix 1,000 problems, but using an optimal solution? Is there a function to markets that limits the amount of solutions that can coexist? Does generalization always lose to specialization in the long run, or is there some equilibrium point?

Think about it this way: Video context extraction isn’t the most efficient way to track chain of custody for server equipment. You could do RFID tags, capacitive sensing, weight sensors, or a dozen other purpose-built solutions that would be cheaper, more reliable, and easier to deploy for that specific use case. Each one would be the “optimal” solution for its narrow problem domain.

But here’s where it gets interesting. What happens when you need to know not just IF someone touched the cart, but HOW they interacted with it? Whether they seemed rushed or methodical? If they followed protocol or improvised? Whether the person on camera matches the person who was supposed to be there? Suddenly your optimal single-purpose sensor becomes useless, and you’re back to video anyway.

So maybe the question isn’t about optimality at all. Maybe it’s about solution density. How many problems can you solve per unit of complexity? A capacitive sensor solves exactly one problem with minimal complexity. Video context extraction solves maybe fifty problems with substantially more complexity. The ratio might actually favor the general solution, even if each individual problem could be solved more elegantly in isolation.

But then you hit the deployment problem. The datacenter already has cameras. They’re already recording. The infrastructure is already there, paid for, maintained, integrated into security systems. Adding your context extraction layer is marginal cost on existing infrastructure. Meanwhile, wrapping every piece of equipment in capacitive blankets requires new hardware, new protocols, new failure modes, new maintenance schedules. The “optimal” solution has to overcome the activation energy of changing the entire system.

And this is where I keep getting stuck. Because from a pure engineering standpoint, purpose-built solutions should win. They’re more efficient, more reliable, more predictable. But from a market dynamics standpoint, general solutions have this weird property where they can absorb more use cases faster than specialized solutions can be deployed.

There’s something Clayton Christensen-esque happening here. The “good enough” general solution gets deployed first because it’s easier to integrate, even though it’s technically inferior. Then it gets better. Then it gets entrenched. Then the optimal solutions never get their chance because the switching cost is too high.

Or maybe I’m just rationalizing. Maybe we’re building something suboptimal because it’s what we know how to build, and I’m trying to convince myself that general solutions have some inherent value beyond their technical merit. Maybe the market is actually efficient and specialized solutions do win, just slowly, and we’ll get disrupted by someone with capacitive blankets and RFID tags in five years.

But then I think about AWS. Amazon didn’t build the optimal solution for any specific computing problem. They built a general platform that was suboptimal for almost everything but good enough for most things. And that “good enough for most things” turned out to be worth more than a thousand optimal point solutions.

So maybe the question isn’t “is this optimal?” Maybe it’s “is this deployable, and can it capture enough use cases that the generality becomes the moat?” Because if you can solve 60% of someone’s problems with one system, even if each solution is only 80% as good as a purpose-built alternative, you’ve created something that’s really hard to displace.

- Teo