When Will Tesla Solve Self Driving? And when will comma.ai?

Prediction (TeslaFSDTracker Optimist): Based strictly on the current accelerating failure-rate reduction (8x/year), FSD could hit 99.99% reliability on HW4 by Q3 2027.

Prediction (George Hotz / Industry Realist): "The actual timeline is going to be about 8 years" (approx. 2033), with the logic that solving the "long tail" requires multiple orders of magnitude of improvement, not just linear percentage gains.

The Optimistic View: The "March of Nines" (Target: 2027)

Four successive snapshots from teslafsdtracker.com reveal a clear exponential reliability curve. When focusing specifically on the "% Drives with No Critical Disengagement" metric, we see the failure rate collapsing.

The logic: The failure rate has dropped from 8.4% (v12) to 0.4% (v14.1) in under two years. If this 8x annual reduction continues, the system hits 99.99% reliability in Q3 2027.

The Skeptical View: George Hotz (Target: 2033)

In his keynote at Comma Con 2025 (Nov 2025), George Hotz (geohot) offered a more conservative estimate. Referencing the macro data from TeslaFSDTracker, he argued that while the curve is exponential, the distance to the finish line is further than it appears.

His Verdict: "The actual timeline is going to be about 8 years" for Tesla to solve self-driving (approx 2033). He added, "If Tesla will be there in 8, we'll be there in 10."

How did Geohot get to "8 Years"?

Hotz focuses on Miles to Disengagement rather than percentage success rates. This is because "Unsupervised" driving requires matching human safety levels.
 

  The Math of Orders of Magnitude:

• Current State (Nov 2025): As seen in Figure B below, the fleet is currently achieving roughly 4,331 City Miles between critical disengagements on the latest version.  
• The Goal (Human Level): To remove the steering wheel (Unsupervised), the car must be safer than a human. Estimates for human reliability vary, but generally require 500,000 to 1,000,000 miles between critical safety failures.  
• The Gap: To go from ~4,331 miles to ~1,000,000 miles is not a small jump—it is an improvement of roughly 230x (over 2 Orders of Magnitude).  
• The Timeline Calculation: Even if Tesla doubles their reliability (2x) every single year—an incredible feat of engineering—it would still take roughly 8 years to bridge the gap:         
  • Year 0: 4,331 miles
  • Year 1: 8,662 miles
  • Year 2: 17,324 miles
  • ...
  • Year 8: ~1,108,000 miles (Human Level)

This explains why the "Optimistic" view (based on recent rapid spikes) sees 2027, while the "Realist" view (expecting diminishing returns and a logarithmic slog) sees 2033.

Comma.ai's Pivot: Robots & Compute

Hotz admitted that Comma.ai trails Tesla in this timeline ("If they are there in 8, we are there in 10"), but outlined how they are closing the hardware gap.

• Compute: Comma introduced the AMD RX 9060 XT external GPU (~$300), delivering 205 TOPS. This moves Comma from "2 Orders of Magnitude" behind Tesla's compute to "0 Orders of Magnitude" behind.  
• Control: To solve the "Torque Problem" (where legacy car steering motors are too weak for sharp AI turns), Hotz outlined a future where a humanoid robot sits in the driver's seat, physically turning the wheel with human-like force, bypassing vehicle-specific electronic limiters.  

Data Accuracy & Methodology

It is important to note that the charts above from TeslaFSDTracker rely on crowdsourced data. To ensure reliability:

• Automated Collection: Data is ingested via MATT3R K3Y and the EVGuyCanada app, which automatically track GPS distance and allow single-tap disengagement recording to reduce bias.  
• Verified Manual Entry: Manual contributors must provide a screenshot validating their VIN (last 6 digits) and firmware version to prevent data poisoning.  
• Statistical Confidence: Dashboards calculate 95% Confidence Intervals to provide visible error bands, ensuring outliers do not skew the trend lines significantly.