The Grand Experiment
JR
There is a grand experiment happening in AI right now, and it’s not just in large language models (LLMs). It’s on our roads. What Tesla’s FSD 12 is showing us tells us everything we need to know about the state of AI and, more importantly, what your company's strategy should be for the next two years.
The Two Eras of Computing: Rules vs. Statistics
For the vast majority of the information age, computer software has run on "expert" or rules-based systems. This is “if this, then that” (IFTTT) programming. It's entirely deterministic: given the same input, you get the exact same output, every single time. This is fantastic for predictability, but it’s brittle. If the world changes, the code must be manually updated to reflect those changes .
We are now in the era of machine learning (ML), which is a completely different paradigm. Instead of hard-coded logic, these are probabilistic systems built on statistics. We feed them large volumes of data, and they "learn" how to respond.
This new approach has benefits and drawbacks :
Benefit: They are adaptive. As the data from the real world changes, the system can automatically retrain itself to react to new scenarios.
Drawback: They are not deterministic. Under the hood, they operate on probabilities. At their core, these systems are designed to perform best in the "middle of the bell curve" of data. They are, by their very nature, terrible at handling edge cases because those cases are statistically rare.
The Tesla Experiment: A Pure ML System in a Safety-Critical World
This brings us to Tesla. For years, self-driving systems (including Tesla's pre-FSD 12 and competitors like Waymo) have used a hybrid model: a mix of ML and old-school, rules-based logic.
With FSD 12, Tesla made the bold—and risky—move to go to an entirely machine-learning system. This is a "grand experiment" in a safety-critical system
Tesla is arguably the only company that could even attempt this. They have more driver data and more sensor-equipped cars on the road than anyone on the planet. If anyone could make a fully ML system work, it would be Tesla.
And yet, we're seeing the problems you'd expect. The system is struggling with the exact thing that probabilistic models always struggle with: edge cases. In driving, a statistically rare event—a deer jumping out, a sudden construction zone, an unusual piece of road debris—is the only thing that matters for safety. The middle of the bell curve is easy. The edges are life and death.
The AI Plateau is Here
This challenge with FSD 12 perfectly mirrors what we are seeing in the generative AI and LLM space. We are starting to see a plateau. The massive, easy gains are behind us. We have exhausted the quality data and the algorithmic innovation that gave us the huge leaps of the past few years. The next large step-function improvement will require more than just incrementally improving what we have now .
For the past two years, the smartest strategic move for many businesses was simple: wait. We had use cases we knew AI could solve, but the models weren't quite good enough. We correctly decided to wait for the next release, which often did fix our problems.
That era is now over.
Your New AI Strategy: Stop Waiting, Start Building
For the next 12-24 months, you must operate under a new assumption: the quality of the base models we have right now is nominally what we will have two years from now.
Waiting for a magic new model to be released by a hyperscaler is no longer a viable strategy. So, what does that mean? It means the incremental value is no longer in the model itself, but in the systems you build around it .
This is where you should be focusing your AI strategy. If the models are great at the "bell curve" but fail at the "edges," then the solution is to build systems that:
- Automate the vast majority of repetitive tasks.
- Intelligently detect when the model is likely to make a mistake (i.e., encounter an edge case).
- Loop in your human experts at exactly the right moment to handle that edge case.
This "human-in-the-loop" approach is how you drive hallucinations down to zero. It’s a system that can’t be overtaken by a simple model update, because it involves your in-house experts and your unique processes.
Where to Find the Highest ROI
The best opportunities aren't always the flashiest. The highest and most immediate ROI we see with our clients is in managing internal processes: policy, procedure, quality control, and compliance.
If you have repetitive tasks where you have good data defining what "right" looks like, but you struggle to get consistent implementation from staff, you have a perfect opportunity. We can help you build a tool into your process flow that provides immediate productivity gains.
My advice is simple: identify those high-ROI, high-value implementations and put them in place now. Don't get overtaken by competitors who are stuck waiting for a "next big thing" that may not be coming.