INNOVATIONOngoing practice88% confidence

Behavioral-Cloning Shortcut to AGI

You don't have to re-run evolution's compute to reach AGI — you can behaviorally clone everything humans already know.

Problem it solves

How to reach general intelligence without paying the compute bill of redoing evolution.

Best for

AI researchers and founders choosing between brute-force RL-from-scratch and bootstrapping from human-generated data.

Not ideal for

Domains with no large corpus of human demonstrations to clone, where the behavior must be discovered from scratch.

Overview

Why this framework exists

Against the biological-anchors view (a Berkeley professor's claim that AGI requires reproducing all the FLOPs that went into evolution), Luan argues there is a giant shortcut: behaviorally clone everything humans already know. LLMs already solved this for written knowledge; multimodal models extend it to the visual world, heading toward a "universal byte model" where high-signal tokens come in and any combination — text, voice, image, video — comes out. The next phase recombines this cloner with the lessons of the RL era.

Core principles

3 total
  1. Cloning existing human knowledge is vastly cheaper than rediscovering it via de-novo RL.
  2. Self-supervised pre-training is the data-efficient engine of the clone.
  3. A universal byte model learns any high-signal mapping once you have cloned the modality.

Origin story

How this framework came to be

Luan's long-running debate with a Berkeley professor over what it actually takes to build AGI, revisited after the LLM/Dota-RL era at OpenAI.

Source

Traced to primary
Source · PODCAST
Why Google failed to make GPT-3 + why Multimodal Agents are the path to AGI — with David Luan of Adept
Latent Space (swyx & Alessio) · 2024
Open source →

Related frameworks

Browse all Innovation →