Quantifying curation Jul 6
Quantifying curation
Measuring optimization from curation of text produced by generative language models with a human in the loop


GPT-3 on Coherent Extrapolated Volition Apr 1
Describing our elusive, recursive volition to (proto-)AGI in words


List sorting does not play well with few-shot Feb 27
How to get GPT-3 to sort a list: make it think it’s running list.sort()


Language models are 0-shot interpreters Feb 10
Mechanisms of meta-learning, beating few-shot benchmark performance with 0-shot prompts, and Bayesian decomposition of prompts


Loom: interface to the multiverse Feb 9
Loom: interface to the multiverse
Preview of loom, a tool for generating, navigating and visualizing natural language multiverses


This Museum Does Not Exist: GPT-3 x CLIP Feb 8
This Museum Does Not Exist: GPT-3 x CLIP
I asked GPT-3 to come up with titles for paintings and had CLIP+BigGAN paint them


Alchemical marriage: GPT-3 x CLIP Feb 8
Alchemical marriage: GPT-3 x CLIP
More figurative fancies from GPT-3’s imagination illustrated by BigSleep (CLIP + BigGAN)


GPT-3 x CLIP worldbuilding Feb 3
GPT-3 x CLIP worldbuilding
A scene written by GPT-3 and illustrated by BigSleep (CLIP + BigGAN)


CLIP hallucinates 1900-2030 Feb 1
CLIP hallucinates 1900-2030
BigSleep (CLIP + BigGAN) generates pictures of “the 1990s”, “the 1910s”, … , “the 2030s”


CLIP art Jan 31
CLIP art
Compilation of images generated with BigSleep from human-written prompts


Language models are multiverse generators Jan 25
Language models are multiverse generators
“Actualities seem to float in a wider sea of possibilities from out of which they were chosen; and somewhere, indeterminism says, such possibilities exist, and form part of the truth.”


The Internet, mirrored by GPT-3 Jan 23
The Internet, mirrored by GPT-3
Combining multipart prompts, logit biases, and counterfactual parsing to make GPT-3 emulate fake Google search results given a query / Wikipedia pages given a title


Methods of prompt programming Jan 12
"Like programming, but more fluid. You’re not programming a computer, you’re writing reality. It’s strange. It’s always different. It’s never the same twice."


Parsing by counterfactual Dec 5
Parsing by counterfactual
Context-aware parsing using counterfactual probabilities predicted by GPT-3


Amplifying GPT-3 on closed-ended questions Oct 30
Improving GPT-3’s performance on SAT analogy questions using task-agnostic prompt pipelines


HITL thought experiment Oct 16
You’re a human with access to a streamlined interface that provides an array of continuations to a given text, as well as continuations spawned from each of those suggestions, recursively ad infinitum…


Pen Oct 16
Imagine that you hold a powerful and versatile pen, whose ink flows forth in branching variations of all possible expressions: every story, every theory, every poem and every lie that humanity has ever told, and the vast interstices of their latent space.