Behind the video
Tesla FSD v12.3 - First Impressions!
Overall Thesis
Dave Lee test-drives Tesla's FSD v12.3 across seven Starbucks locations in Austin and finds it dramatically more human-like, confident, and capable than v11, which he attributes to the shift to an end-to-end neural network approach. He is impressed by the pace of improvement but repeatedly cautions that the system is not robotaxi-ready and needs much more data before any such claim is warranted.
Narratives
TSLATeslaDave Lee is impressed that FSD v12.3's end-to-end neural network approach already drives more naturally and makes smarter decisions than v11, despite Tesla still being early in scaling data and compute. He sees this rapid improvement as evidence Tesla can keep advancing the system quickly, though he stops short of calling it robotaxi-ready.
Key Arguments
- v12.3's end-to-end neural net approach produces noticeably more human-like acceleration, braking, and turning than v11's heuristics-based system
- v12.3 made smarter decisions in several situations where v11 historically struggled, including a dangerous recurring lane-merge issue in Austin that v11 never fixed
- The speed of improvement from v11 to a comparable or better system in about a year suggests Tesla can scale the approach quickly with more data and compute
- Tesla's planned buildout of training compute (referenced ~100 exaflops) is framed as key to further rapid gains in FSD capability
Risks acknowledged
- v12.3 made some clear mistakes: failed to complete a U-turn, an unsignaled lane change, and GPS/parking-lot confusion
- Current small-sample personal testing cannot establish true safety statistics or robotaxi readiness
- High-criticality interventions must become extremely rare, which has not yet been demonstrated with real fleet data
- The test conditions (low-traffic Sunday afternoon) were not the most challenging scenario for the system
Hedges & Caveats
- Explicitly states he does not think v12.3 is robotaxi-ready and Tesla isn't claiming it is
- Notes the small sample size of personal testing is not enough to make an objective safety assessment
- Acknowledges several mistakes during the drive (failed U-turn, unsignaled lane change, GPS confusion in parking lots, following too closely)
- Says judgment lapses need to become extremely rare before any robotaxi-level claim is justified
- Notes conditions were relatively easy (Sunday afternoon, low traffic) rather than a worst-case test