Examples of AI Improving AI

Author: Thomas Woodside, Center for AI Safety
Contributors: Herbie Bradley, James Campbell, Jun Shern Chan, Aidan O'Gara, Dan Hendrycks, Esben Kran, Nathaniel Li, Mantas Mazeika, Aaron Scher, Zach Stein-Perlman, Fred Zhang, Oliver Zhang, Andy Zou.
Last Updated: October 2, 2023

As machine learning algorithms become more capable of outperforming humans on some narrow tasks, they are increasingly being used to make improvements to themselves or other machine learning systems, or inputs to those systems such as hardware. In some cases, human feedback used to improve models has been replaced with AI feedback; in other cases, GPU circuits that were once designed by humans are being designed by AI systems. Some have warned that this "recursive self-improvement," if scaled up, could lead to AI spiraling beyond human control [1][2][3].

The table below collects some current examples of AI systems being used to improve AI systems. It should not be taken as an exhaustive list, since these applications can occur in many subsets of AI and we have not been able to review all recent AI papers. The "author" and "author affiliation" columns refer to the authors of the paper; the "submitter" column refers to the person who originally brought the paper to my attention. If you know of an example not mentioned here, you may submit more here.

[1] Nick Bostrom, Superintelligence
[2] Joseph Carlsmith, Is Power-seeking AI An Existential Risk?
[3] Dan Hendrycks, Natural Selection Favors AIs Over Humans

ID
Description
Source
Date Published
Authors
Author Affiliations
Submitter
38
LLMs used for genetic algorithm to generate LLM prompts.
9/28/2023
Fernando et al.
DeepMind
Zach Stein-Perlman
37
LLM used to help refine prompts for vision-language models.
9/12/2023
Liu et al.
CMU
Aidan O'Gara
36
Language model used to generate prompts that could correspond to training data, filter them, with the data used to train a stronger language model.
8/3/2023
Li et al.
Meta
39
Fine tuned decision transformer used to generate data that is then used to retrain a new version of the base model in a loop.
6/20/2023
Bousmalis et al.
DeepMind
Thomas Woodside
32
Uses LLMs to help automate ML experiments.
5/4/2023
Zhang et al.
UT Austin
33
Uses LLMs feedback for LLM prompt engineering.
5/4/2023
Pryzant et al.
Microsoft
Zach Stein-Perlman
30
GPT-4 can be used to perform neural architecture search.
4/21/2023
Zheng et al.
Various
Fred Zhang
34
Uses LLMs to debug LLM code outputs without access to unit tests.
4/11/2023
Chen et al.
Google & UC Berkeley
Zach Stein-Perlman
26
Language model gives itself feedback to improve its own generations.
3/30/2023
Madaan et al.
Various
Andy Zou
28
Uses a language model to pick the best refinement of an LLM response based on language feedback.
3/28/2023
Scheurer et al.
Various
Thomas Woodside

1–10 of 39