From Self-Driving Cars to Self-Driving Knowledge Workers Jesse Hu

A roadmap from deep learning to AGI, and where we are in the hype cycle.

From image classification to language generation to reasoning and planning, we’ve quietly learned how to build a repeatable engine: define a problem, codify it into data and benchmarks, and then hill-climb using models, logs, and evals.

This talk explores how that engine (I'll call this the “Imitation Engine”) is a a very repeatable (but boring!) way to get from here to AGI. It lays out a practical blueprint for AGI as the aggregation of millions of discrete task automations, played out by hundreds and thousands of agent developers.

We’ll cover:
- How we moved from classification to multi-modal to actions and autonomy
- What some of the "secrets" are to self-driving, such as simulation and self-play
- Why the key to AGI is task definition and benchmark design
- What PMs and AI engineers can learn from researchers and roboticists (and vice versa)
- Where current methods fall short—and what’s needed next

Видео From Self-Driving Cars to Self-Driving Knowledge Workers Jesse Hu канала AI Engineer