AI IRL Podcast Episode 43: How Real-World Models Drive Rapid Insights for Everyday Application
The AI industry faces a model problem: typically, we feed our AI invented datasets that don’t represent real life situations and scenarios. These models are preparing AI for an unrealistic environment that doesn’t work in real life.
But James Dean is changing models. He’s using breakthroughs in camera and inexpensive drone technology to capture real geographies at his company, SenSat. They layer these geographies with everyday obstacles — like people, temperature, lighting — to develop models that replicate real life geographies, creating what James calls “digital twins.”
He came on the AI: In Real Life podcast to tell us all about how they’re creating these new models, and how they’re impacting construction, infrastructure, and other industries.
Currently, we feed AI ‘invented’ data
“AI will be stunted if it can’t understand the actual world we live in.” — James Dean
Here’s the current way we create AI models: we invent data, then feed that data into an artificial intelligence. We run simulations and train the AI all from that somewhat imaginary dataset. When the AI meets real life, it’s simply not prepared.
James compared it to sending someone to an online K-12 school — some of the information is available, but none of the nuance and intangible lessons a student learns within a traditional education environment will come through in stripped-down online education.
James Dean’s model solution: create a real world replica (digital twin)
James’s solution is simple in theory, and now, technologically feasible: replicate the real world (what some call a “digital twin).
At SenSat, James creates a digital map of the real world, drawing on LIDAR data, drone data, or satellite data to create a replica 3D geometry. Next, he layers on real world factors, such as temperature, lighting, moving people, etc.
The result? “You can think of it as a SimCity of the real world,” he said.
This ‘SimCity’ IRL is only recently accessible
Modern high-definition cameras and the increase of inexpensive drone technology have made real world replication generally accessible for business, and James said these advancements have only reached this level of accessibility within the last decade.
This sophistication requires an insane amount of data points
The amount of data necessary to create these real-world scenarios is astronomical; currently, SenSat has 110 billion data points. For reference, that’s about 8 times the amount of seconds between now and when Jesus was born.
These models rely on 4 specific elements
Note 4 aspects that SenSat models rely heavily on:
- Large data sets (110 billion data points)
- Granularity of data (temperature, lighting, etc.)
- Additional available data input from any other sources
- SenSat focuses on data that is specific to the problem they’re trying to solve
“If you think about the world, it’s a bit like a jigsaw — no single piece is going to give you the whole picture.” — James Dean
A simple dirty example of a real-life AI model affecting construction
Here at LogMeIn, we’re all about AI in real life. Here’s how SenSat has impacted the construction industry, on a specific 5-mile road project they recently were hired to consult on:
The construction company needed to know how much soil was dug up in point A and moved to point B, which determined how the general contractor would pay their subcontractors. Traditionally, this measurement was taken physically, and you’ve probably seen it being done: construction workers wear high-visibility clothing and use what looks like a pogo stick to measure the volume of dirt moved.
The findings are reported, and the subcontractors are paid based on their work. To measure the soil moved from a 5-mile stretch would usually take about 6 weeks using the traditional method.
Instead, SenSat used drone technology to take pictures and create millions of data points across the road’s landscape. They fed the data into a computer, which turned the data points into a solvable math problem for a computer to calculate — all within a matter of hours. And the accuracy? Quite literally, down to the pebble.
(Humorously, because of the inefficiency in the traditional “pogo stick” method, it was actually commonplace for subcontractors to invoice for the soil removal, then, move the soil back and move it again, allowing them to re-invoice for the same removal.)
And that’s AI, In Real Life.