Skip to content
TubeRank.
Get app

Behind the video

Tesla FSD v12 Rollout - FIRST THOUGHTS

The thesis, the calls, and the words behind them. Play a quote to hear it in the original video.
Dave LeeDave LeeJan 22, 2024Watch on YouTube
Overall SentimentBullishStrength: 55%

Overall Thesis

Dave Lee reacts to Tesla's rollout of FSD version 12 to a non-employee, framing it as a major technical milestone in shifting from rule-based code to an end-to-end neural network approach. He is optimistic about the software's trajectory but notes significant remaining challenges before it reaches a customer-ready, unsupervised state.

Narratives

TSLATesla
Bullish

Dave Lee views Tesla's FSD v12 end-to-end neural net rewrite as an unprecedented and highly ambitious technical accomplishment, expressing optimism that compute and data improvements through the year will lead to sizable gains in driving ability. He frames the current limitations as expected growing pains of a necessary, high-risk architectural shift.

Key Arguments

  • Tesla replaced over 300,000 lines of rule-based C code with an end-to-end neural net for planning, a first-of-its-kind shift at this scale
  • Early FSD 12 testing results were promising enough for Tesla to go all-in on the end-to-end approach
  • Tesla increased compute investment specifically to support training the end-to-end neural net
  • The host expects sizable improvements in FSD 12's abilities over the next several months as larger compute clusters come online
  • Giving a non-employee access to FSD 12 signals Tesla's internal confidence in the current iteration

Risks acknowledged

  • FSD 12 still stops with 'red hands' and gives up control when confidence drops below a threshold
  • The jump from rule-based to purely neural-net planning is described as 'almost an impossible task'
  • It remains unclear how long it will take to reach unsupervised FSD and robotaxi capability
  • Removing safety guardrail rules could introduce new, previously unguarded dangerous situations

Hedges & Caveats

  • The first version rolled out is 'far from perfect'
  • FSD 12 still 'gives up' and hands control back to the driver in low-confidence situations
  • Uncertain how long it will take to reach unsupervised FSD and robotaxi capability
  • Speculation about Tesla's internal timeline and compute needs is explicitly presented as a guess
Analyzed with claude-api-scheduled | Extraction manual_v1 | Cost: —