Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boundless Socratic Learning with Language Games

Tom Schaul

TL;DR

It is argued that pure recursive self-improvement, dubbed "Socratic learning", can boost performance vastly beyond what is present in its initial data or knowledge, and is only limited by time, as well as gradual misalignment concerns.

Abstract

An agent trained within a closed system can master any desired capability, as long as the following three conditions hold: (a) it receives sufficiently informative and aligned feedback, (b) its coverage of experience/data is broad enough, and (c) it has sufficient capacity and resource. In this position paper, we justify these conditions, and consider what limitations arise from (a) and (b) in closed systems, when assuming that (c) is not a bottleneck. Considering the special case of agents with matching input and output spaces (namely, language), we argue that such pure recursive self-improvement, dubbed "Socratic learning", can boost performance vastly beyond what is present in its initial data or knowledge, and is only limited by time, as well as gradual misalignment concerns. Furthermore, we propose a constructive framework to implement it, based on the notion of language games.

Boundless Socratic Learning with Language Games

TL;DR

It is argued that pure recursive self-improvement, dubbed "Socratic learning", can boost performance vastly beyond what is present in its initial data or knowledge, and is only limited by time, as well as gradual misalignment concerns.

Abstract

An agent trained within a closed system can master any desired capability, as long as the following three conditions hold: (a) it receives sufficiently informative and aligned feedback, (b) its coverage of experience/data is broad enough, and (c) it has sufficient capacity and resource. In this position paper, we justify these conditions, and consider what limitations arise from (a) and (b) in closed systems, when assuming that (c) is not a bottleneck. Considering the special case of agents with matching input and output spaces (namely, language), we argue that such pure recursive self-improvement, dubbed "Socratic learning", can boost performance vastly beyond what is present in its initial data or knowledge, and is only limited by time, as well as gradual misalignment concerns. Furthermore, we propose a constructive framework to implement it, based on the notion of language games.

Paper Structure

This paper contains 10 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Three Necessary Conditions for Self-improvement
  3. Feedback
  4. Coverage
  5. Scale
  6. Socratic Learning
  7. The Fundamental Limits of Socratic Learning
  8. Language Games Are All You Need …
  9. Higher-level Recursions
  10. Conclusion: Open-ended Socratic Learning is Possible

Figures (1)

  • Figure 1: Cartoon depiction of the key definitions. An observer ( gold) external to a closed system ( black) assesses the performance ( green) of an agent ( red) over time. The process is one of self-improvement if agent outputs affect future agent inputs (i.e., some path like blue exists), and performance improves. Further, self-improvement is recursive if the agent input and output spaces are compatible, and the process is called 'Socratic learning' if that space is language.