Picture someone breaking into a bakery at 3 a.m., cracking open the safe, and finding forty thousand pounds of flour. Technically, they found the thing that makes the whole operation work. Practically, they found a mess they now have to drive away from.
That is roughly what happened in Fremont, California, late one Wednesday night. Thieves hooked their own cabs up to two trailers belonging to PlusAI, a self-driving trucking company based in Santa Clara, and drove off. The trailers had been associated with Nvidia, which is a name that tends to make certain people’s eyes light up, because Nvidia chips are small, expensive, and easy to resell. What the thieves actually hauled away was 20 tons of sand. About 40,000 pounds of it. The sand exists because you need weight in a trailer to test a self-driving truck properly. Once they broke the trailers open and saw what they had, they abandoned the whole thing. The trailers were later dumped. No arrests have been made.
I write about AI agents for people who do not build them, and I keep coming back to this story because it is one of the funniest and most useful illustrations I have seen of a gap that trips up almost everyone learning about this technology.
AI is heavier than you think
When most people imagine AI, they imagine software. Something that lives in a browser tab, answers questions, writes emails. Weightless. The Fremont story is a reminder that a huge amount of AI work is stubbornly, inconveniently physical.
A self-driving truck is an AI agent in the most literal sense of the term. An agent is a system that takes in information about the world, decides what to do, and then does it without a human pressing the button for each step. A chatbot suggests. An agent acts. And when an agent’s actions involve 40,000 pounds of momentum on a public highway, you cannot validate its decisions with a spreadsheet and good intentions.
So you fill a trailer with sand. Not because sand is special, but because it is cheap, dense, easy to source, and behaves predictably. You need the truck to experience what a loaded truck experiences: longer stopping distances, different weight distribution on turns, more strain going uphill, more push going down. An agent trained and tested only on empty trailers would be an agent that has never met its actual job.
Testing is the unglamorous part nobody photographs
This is the part of AI development that gets almost no attention. The demo is the fun part. The testing is where the real work hides, and it is mostly repetition, edge cases, and deliberately boring conditions repeated until the results stop surprising you.
If you are trying to evaluate any AI agent, whether it drives a truck or processes invoices at your company, the questions that matter look a lot like sand:
- Was it tested under realistic load, or only in ideal conditions?
- Does anyone know how it behaves when the inputs get heavy, messy, or unusual?
- Who checked, how often, and what did they find?
- What happens when it gets something wrong, and how quickly does anyone notice?
A vendor who can answer those questions in detail is a vendor who has been filling trailers with sand. A vendor who only shows you the highlight reel may have skipped that step.
The thieves made a very human mistake
Here is the part I find genuinely instructive. The thieves pattern-matched on a surface signal. A recognizable name, a sealed trailer, an assumption about value. They acted confidently on a shallow read of the situation and were wrong in an expensive, exhausting way.
That is close to the most common failure mode people describe when AI agents go sideways. An agent picks up on a surface pattern that looks like the thing it is supposed to find, commits to an action, and only discovers the mismatch after the work is done. The difference is that the thieves stopped as soon as they saw the sand. An agent without a sensible check at that moment might have kept driving, or filed a report claiming the delivery was a success.
Good agent design includes a version of that moment. Verify before committing. Confirm that what you grabbed is actually what you wanted. Have a way to stop and walk away when the evidence says you misread the situation.
What to take from a pile of sand
Two trailers got stolen and dumped, nobody has been arrested, and a self-driving trucking company is presumably short some ballast. As crime stories go, it is a small one.
Somewhere right now, a trailer full of sand is helping a machine learn how heavy things stop. That is what building trustworthy agents actually looks like. Not a breakthrough demo, just weight, repetition, and someone checking the numbers.
🕒 Published: