Tesla stock drops 13% after neural network replaces code

Tesla's FSD 12.1 drives autonomously but remains Level 2

Does a Tesla drive itself now, hands off the wheel and with no human intervention? That is what testers of the FSD 12.1 beta claim. This version arrived after a radical decision: replacing almost all C++ code with a single neural network, similar to large language models. The change is not cosmetic. Early reports suggest the car completes entire routes without corrections and sometimes drives better than the person sitting beside it.

What changed: from thousands of lines of code to a neural network

Classic autonomous driving was built on rules: if this happens, do that. Tesla broke this pattern and parked much of its C++ to let a trained model decide on its own. Users say the result is no longer robotic: it seems to drive like a human, with more natural reactions. The novelty, they say, is that it no longer loses control frequently.

It was not always so. Two decades ago, neural networks were a mathematical curiosity, recall those who studied them: what was missing then was hardware capable of running them. The system is not geolocated: it works on any road, including sections without loaded maps, which defenders present as an advantage over other manufacturers.

Level 2 or true autonomous driving?

Here the problem begins. The system is marketed as Level 2, meaning assistance requiring permanent human supervision, while other manufacturers already sell Level 3 and 4 systems under limited conditions. A common question is: if the car drives alone, who is liable in an accident?

Some argue that manufacturer certification dictates responsibility. Without certification above Level 2, buyers assume liability and effectively act as testers. Opponents argue that the technical boundary has been crossed and homologation lags behind. This matters for insurers, who need to know if crashes occurred in autonomous or manual mode, and for Europe, where strict regulation is expected.

One forum user recalls Elon Musk's failed attempt to acquire OpenAI as the origin of the autonomous driving bet, while others argue that the future of autopilot does not necessarily lie in electric vehicles.

Country roads, fog, and hardware limits

Enthusiasm clashes with Spanish asphalt. The recurring test is a narrow country road without shoulders, with construction, cuts, and temporary signage. Defenders claim the beta no longer gets stuck in these scenarios as before: where rules fail, the neural network improvises better.

Another doubt concerns sensors. Unlike other manufacturers using LiDAR, Tesla relies solely on cameras. Some argue that fog or rain disables laser sensors while camera vision continues reading the road. Videos recorded in dense fog have become the company's best advertising, but also a source of skepticism: one stretch is different from a whole winter.

The Tesla Semi and numbers that don't add up

Credibility fades when looking at trucks. The company announced two versions of the Semi; the top model carries 900 kWh battery and about 5 tons, with a claimed range of 800 km. Critical users state that real tests with fully loaded trailers drop the figure to 98 km against the promised 220 km with 5-ton towing capacity. Future autonomy promises carry this burden, and ongoing debates about passenger car finishes and materials at current prices do not help.

Stock falls 13%: opportunity or mirage?

In the market, autonomous driving is the main option regarding the company's future. Beta advances did not prevent the correction: shares fell 13% to their lowest level since May 2023. For some, it is time to expand positions; for others, proof that the market competes aggressively, with Chinese manufacturers pressing from below and buying units from rivals themselves.

Two incompatible narratives are on the table. One sees AI driving as the decisive competitive advantage, such that the car would not need to win on finishes. The other notes that the company has struggled to meet demand for years and prices fall because competitors offer cheaper options. Production capacity, in this sense, hides many flaws.

With these credentials—a car navigating a narrow road without shoulders and another reportedly failing to reach 100 km with a load—anyone might think the issue is resolved. Some profit while others wait for the stock to explode.