OpenAI, Meta, or DeepMind Will Win Self-Driving

The winner in Self-Driving will be either OpenAI, Meta, or DeepMind

Self-Driving is a software problem, not a hardware problem. The famous issue with self-driving is the long-tail of problems. I personally think this long-tail is hard to solve fundamentally because it requires reasoning.

The companies best suited to deliver these capabilities aren't the current players in self-driving (Waymo, Tesla, Comma) but rather the key players in general intelligence (OpenAI, Meta, DeepMind).

Here's what I expect to play out:

1. Current LLMs will continue getting better at reasoning through things like chain-of-thought prompting and recursive structures.
2. When video + audio is added as direct input and these models are trained on the same regime as current LLMs (predict next frame aka world model + action), we'll start getting far less hallucination and more reliable systems. Think MuZero.
3. These will be general purpose models/agents capable of operating in the physical world.
4. The current incumbents in the self-driving space will pivot to either train similar systems or become 'wrappers' over these new models. Complex task-specific stacks will disappear. The expected self-driving moat will vanish overnight.

Ironically, it seems that achieving general purpose intelligence will be easier than solving a task-specific category that requires reasoning.