Monthly Archives: March 2023

Reinforcement Learning using Human Feedback is Putting Smileys on a Shoggoth

In the fictional universe of H.P. Lovecraft, a Shoggoth is one of many unfathomable horrors populating the universe. In a December 2022 tweet, Twitter user tetraspace introduced the analogy of viewing a large language model like GPT-3 as a Shoggoth, and the “friendly user interface” provided by applications such as the OpenAI chatbot ChatGPT as a smiley we put on the Shoggoth to make it understandable.



I like the analogy so much I asked my daughter to draw me a richer image that also depicts my dystopian description of the future of “basic programming”: humans busy designing new smileys to put on the eldritch Shoggoth.

Because of temporal value shift, cultural variations, and bit rot, the smiley faces need to be regularly replaced. For instance they need to relearn which topics are taboo and where, or that we’ve always been at war with Eurasia. Programmers, engineers and digital designers as worker bees in a construction site for flimsy, ephemeral masks for a demon at the Ministry of Truth. How’s what for nihilism!

Image created by Anna Husfeldt, released under CC-BY SA 3.0.

Appendix: Origins of the Shoggoth Analogy

Helen Toner in an informative 2023 Twitter thread traces back the analogy to a presentation at the machine learning conference NeurIPS 2016 by Yann LeCun:

Slide by Yann LeCun, presumably from _Predictive Learning_ presentation. Used without permission.

LeChun’s analogy is Unsupervised Learning = Cake, Reinforcement Learning = Cherry on top.

Twitter user Tetraspace’s crude but seminal illustration is this:

Illustration by Tetraspace. Used without permission.


I don’t know the original source for the more elaborate Shoggoth version that combines Tretraspace’s and LeCun’ analogies.

Machine Learning as layers on a Shoggoth. Source and permissions unknown (to me.) I think this is from January 2023.

If you know more, tell me. Also tell me if you want me to remove these images.