Behind the video
Tesla FSD v12.3 - Hits and Misses
Overall Thesis
Dave Lee test-drives Tesla's FSD v12.3 around Austin, highlighting dramatically more human-like, confident driving behavior and fewer interventions compared to v11. He frames the rapid year-over-year progress of the end-to-end neural net approach as a positive sign for Tesla's ability to eventually reach robotaxi-level capability, while noting some remaining misses and unresolved uncertainties.
Narratives
TSLATeslaDave Lee is impressed by FSD v12.3's dramatic improvement in natural, human-like driving behavior and reduced interventions versus v11, seeing this rapid progress on the end-to-end neural net approach as encouraging for Tesla's path toward robotaxi capability. He stops short of making any stock price or investment call, focusing purely on product/technology observations.
Key Arguments
- FSD v12.3 handled multiple tricky real-world scenarios well: a confusing temporary two-way parking lot detour from both directions, parallel parking, yielding to pedestrians, and navigating around a bus and a roundabout
- The system feels much more natural and human-like because it's trained end-to-end on video data rather than relying on V11's rigid heuristic-based rules
- Interventions, including minor pedal presses, have gone down significantly in his limited testing compared to V11
- The jump from early end-to-end development (around December 2022) to this level of capability in roughly a year is described as impressive progress
- He believes more compute and more training data will keep improving the system, and highlights the need for 'emergent behaviors' (unexpected good decisions) as a sign of progress toward robotaxi-level driving
Risks acknowledged
- Feeling safer is not the same as being safer; only Tesla's internal intervention data can truly validate safety improvement
- He experienced a clear fail/disengagement when FSD got stuck too close behind a car
- Observed hesitation at a left turn and confusion over a time/day-restricted no-turn-on-red sign
- His assessment is based on only a couple hours of personal driving, not large-scale data
- Highway driving is suspected to still run on the older V11 stack, not the new V12 approach
Hedges & Caveats
- Feeling safer doesn't necessarily mean the car is actually safer; critical intervention data is what matters most and only Tesla has that at scale
- His observations are limited to personal anecdotal driving over a couple of hours, not large-scale data
- He is not certain the highway stack is still the old V11 system, calling it a strong suspicion rather than confirmed fact
- He notes several real misses: hesitation at intersections, a disengagement after FSD got stuck behind a car, and confusion about a no-turn-on-red sign's day-of-week logic
- He flags that FSD still has weak spots relative to human drivers that need to be offset by excelling in most situations before reaching robotaxi-level reliability