Experts say Tesla's imitation learning-based approach to building fully autonomous driving systems needs AI breakthroughs that may not be achieved for some time
Tesla is pushing to build true self-driving vehicles by 2026. The way its CEO wants to do it seems risky.
This assessment extends a longer-running dispute over whether progress from driver-assistance systems can translate into reliable autonomous operation. Earlier expert warnings likewise held that fully autonomous driving could take longer than expected.
First-order effects
The expert critique raises the execution bar for Tesla’s 2026 autonomy target: imitation learning alone may not deliver the general-purpose driving capability Tesla is pursuing without further AI advances.
Tesla must reconcile its autonomy narrative with a technical approach that outside specialists characterize as risky, putting greater weight on evidence of safe, reliable real-world performance.
Second-order effects
A longer path to true autonomy would make Tesla’s promised self-driving economics more dependent on continued investment in AI development and validation rather than a near-term software transition.
The gap between assisted driving and unsupervised autonomy becomes a sharper competitive distinction, favoring deployments that can demonstrate reliable operation rather than relying principally on future-model claims.
Third-order effects
If the technical hurdle persists, autonomous driving is likely to remain an AI capability and safety-validation problem, not simply a feature rollout problem—lengthening commercialization cycles across the sector.
The recurring mismatch between autonomy timelines and expert assessments may push the industry toward more cautious definitions, staged deployments, and closer scrutiny of claims about driverless capability.
The trend: Autonomous-vehicle development is increasingly being tested by the difficulty of turning AI progress into dependable, scalable operation in open-ended driving environments.
Here's the thing about Tesla, the source of much of Musk's wealth (for now) — he's bet the entire company on something that's very unlikely to work. — And even if it does, eventually, then certainly not on the timetable he's promised. — Or even, for that matter, any timetable…
Chris's excellent column is here: https://www.wsj.com/... My deep dive on Tesla and the limitations of imitation learning is here: https://www.understandingai.org/ ...
@_justinbruen @mims I think Waymo's professional safety drivers are likely to provide much better data for reinforcement learning, since they let the software drive and carefully annotate the reason for each intervention.
Imitation learning suffers from a problem of compounding errors: mistakes move it away from its training distribution, which makes subsequent mistakes more likely, as CMU researchers demonstrated 15 years ago. [image]
One of Tesla's core bets seems to be that they can get to full autonomy mainly through imitation learning. I told @mims I'm skeptical that this will work. [image]
@binarybits @mims I agree data isn't a silver bullet. Most people have more data than they need. We train on only a fraction of our data, more doesn't seem useful yet. Though not sure what that has to do with RL? If you do RL in the real world, you need tons of real disengagement…
@binarybits @mims You have a source from Waymo that says they are using RL and how? AFAIK nobody has done pure RL successfully for self-driving. Did they ship this? https://waymo.com/... Most people in ML agree you need RL in the longterm, everyone will use RL if RL algorithms im…
So much unrealistic talk about $TSLA stock on both sides of the aisle over the past few days. On the overly bullish side, many investors are treating as gospel $TSLA management's expectation that they will have robotaxi operations up and running in TX and CA by late 2025. That [i…
@___Harald___ @mims My claim isn't that Tesla doesn't use any RL. Just that if RL is ultimately required (which seems likely) then Tesla's massive dataset of human driving is likely to be of limited value because you can only use it for imitation learning.