Self-Programming Artificial Intelligence

me@ryanorban.com (Ryan Orban) — Sun, 02 Oct 2022 00:00:00 +0000

Summary

Self-programming AI refers to systems that can modify or generate their own code, algorithms, or learning procedures — a longstanding aspiration in artificial intelligence research that sits at the intersection of meta-learning, program synthesis, and automated machine learning. The central challenge is the bootstrap problem: a system modifying itself needs to reason about how its modifications will affect its future behavior, which requires a kind of self-model that is computationally and representationally difficult to maintain accurately.

GPT-3 Complete-Program Synthesis (Riley Goodside)

me@ryanorban.com (Ryan Orban) — Sat, 20 Aug 2022 00:00:00 +0000

Summary

Riley Goodside describes a “novel, powerful method of complete-program synthesis using GPT-3” that combines instruction prompts with what he calls “contextually informative templates — a generalization of the format trick” shown to him by Boris Power. The approach generates complete programs in a single call at temperature 0 (deterministic), not just code snippets.

Program-Synthesis on Ryan Orban

Self-Programming Artificial Intelligence

Summary

GPT-3 Complete-Program Synthesis (Riley Goodside)

Summary