Driving rules are easy to list. Stop at red lights. Follow lane markings. Yield to bikers. “Ped Xing.” It took me a while to understand that last one when it first appeared in Luneta.
Those are not the difficult parts. Computers can memorize those rules without effort.
Driving becomes difficult when the rules stop explaining what to do. A car abruptly stops or turns without signaling. A black cat darts across the road (let’s talk about superstition next time).
Drivers make a quick decision based on experience.
This is the gap autonomous driving systems struggle with, and it is what NVIDIA is trying to address with a new set of open-source AI tools for self-driving vehicles.
The tools are meant to help systems handle situations that are uncommon or unpredictable. Engineers refer to these as edge cases.
One part of the release is Alpamayo 1, described by NVIDIA as a 10-billion-parameter Vision-Language-Action model.
That sounds intimidating, but it is simply a system that tries to do three things together. It looks at what is happening, understands context or instructions, and then decides what to do next.
Remember the early days of large language models like ChatGPT or Llama, before “thinking” or “reasoning” modes existed?
You would ask a question and get a direct answer, with no explanation of how the model came up with it.
That was before chain-of-thought reasoning became standard. A simple test was asking how many “r”s are in the word “strawberry.” The model often got it wrong.
I enjoyed running that test, and many others, on every new model release while it lasted.
Chain-of-thought (CoT) reasoning fixes this by forcing the model to break problems into steps instead of guessing based on patterns.
In the “strawberry” example, the model spells out the word and counts each letter instead of assuming.
Alpamayo 1 applies the same idea to driving decisions. It does not just choose an action or an answer. It can explain why it chose it.
Instead of “brake now” with no context, the system can explain something like seeing construction ahead, noticing the vehicle in front slowing down, recognizing a narrowing lane, then deciding to reduce speed and keep distance.
This “show your work or solution” approach helps developers test decisions and catch bad logic before it becomes a real-world problem.
People do not become better drivers by memorizing more rules. They improve by encountering awkward situations and adjusting over time. The tools try to recreate that process in a controlled environment.
The main challenge in autonomous driving today is not computing power or sensors anymore. The more difficult problem is deciding what to do when the situation is neither clear nor familiar.
Drivers rely on small cues and some Pinoy habits you won’t find anywhere else.
These cues are difficult to encode directly.
This does not mean the technology is unreliable. It means it depends heavily on the conditions it was trained for. Or in AI terms, overfitting.
This is why NVIDIA’s tools use volume and variety. They do not remove confusion. Instead, they expose the system to it early.
Making the tools open-source creates exposure to different environments and therefore creates more situations to learn from, when rules alone are not enough.
Training AI to drive shows how much driving depends on human behavior. Teaching machines to mimic those takes time and repeated exposure to unfamiliar situations.
I used to drive my son to school and pick him up every day when he was still a kid. If you have driven there long enough, you know there is a special traffic algorithm that activates during drop-off and dismissal.
At these times of the day, every driver employs a different driving playbook. It works somehow, but only because everyone is respecting everyone else.
If AI can figure that out, it’s ready for San Juan.