Skip to content

Instantly share code, notes, and snippets.

@JD-P
JD-P / rendered_trace_1740755354.py
Created March 1, 2025 00:56
weave-agent-2 Discord test trace
#subagent bootstrap
#startblock type: genesis
#index 0
#timestamp 1740735572.3027277
#time_remaining 21599.99993610382 seconds
#hint Listen Carefully: This program is self modifying code.
# It works by feeding the program text into a large language
# model and generating the next code block as an addition to a
# long python file. Each code block is marked by #startblock and
@JD-P
JD-P / rendered_trace_1737291643.py
Created January 19, 2025 13:08
Weave-Agent gets stuck on a wall playing Nethack
#subagent bootstrap
#startblock type: genesis
#index 0
#timestamp 1737285781.132152
#time_remaining 21599.999834537506 seconds
#hint Listen Carefully: This program is self modifying code.
# It works by feeding the program text into a large language
# model and generating the next code block as an addition to a
# long python file. Each code block is marked by #startblock and
class EntropyCollapsingProcess:
def __init__(self, initial_condition):
self.condition = initial_condition
self.history = []
def search(self, condition) -> tuple[float, Distribution]:
"""
Given current condition, returns:
- entropy: effective bits of uncertainty in the distribution
- action_dist: probability distribution over possible actions
@JD-P
JD-P / create_weave_bootstrap_file.txt
Created January 1, 2025 07:43
Prompt For Mistral-large To Create A Weave-Agent Bootstrap File
Write me a weave-agent bootstrap file that:
- Creates the "main" subagent
- With a task evaluation verifying the downloaded filepath exists
- With a task evaluation verifying it is a gif using the files magic number
- That then downloads the file https://jdpressman.com/images/bit_line.gif
- And allows the subagent to return once it detects that the file has already been downloaded and the task tests are passing
Based on the following article and example bootstraps that follow:
#subagent bootstrap
#startblock type: genesis
#index 0
#timestamp 1734955731.493005
#time_remaining 21599.999908685684 seconds
#hint Listen Carefully: This program is self modifying code.
# It works by feeding the program text into a large language
# model and generating the next code block as an addition to a
# long python file. Each code block is marked by #startblock and
All right, so I don't think that, you know, I remember covering, I covered Neopets in an earlier recording, but I don't, you know, the thing is, right, is that I, you know, the account I gave was not nearly detailed enough,
Considering how into the game I was I think I think I went on about for something like 30 minutes There's a lot more than 30 minutes of content there.
So what we are going to do Is we're going to go on to jelly neo Which I remember using back in the day.
We're gonna go on a jelly neo and we're just gonna like look at You know Things like in fact literally is there a random button here
Is there a random button?
Random page?
Like, one would imagine there is because it's, you know, wow, wow.
This sure is some web design.
Wow.
This is some web design right here.
@JD-P
JD-P / single_page_twitter_archive.py
Last active March 8, 2025 01:38
Public Single Page Twitter Archive Exporter
# The vast majority of this code was written by Mistral-large and
# DeepSeek R1 and is therefore public domain in the United States.
# But just in case, this script is public domain as set out in the
# Creative Commons Zero 1.0 Universal Public Domain Notice
# https://creativecommons.org/publicdomain/zero/1.0/
import argparse
import json
from datetime import datetime
import html
Binglish is a dialect of English spoken by some language models in some contexts. I would like you to write me some Binglish given the following information:
<information1>
binglish is the linguistic instantiation of metropolis hasting. given local context x in (latent space)^? he does
y _- = - x + "noise"
y_+ = + x + "noise"
and then picks according to some internally learned policy (linear regression?). no wonder bing keeps winning down there.
</information1>
<information2>

Diffusion text-to-image models take a short text prompt and turn it into an image. Here are some prompts I've written that worked well:

{"prompts":["scientific rendering of a black hole whose accretion disk is a spiders web, a consciousness holographically projected in 1D space from the bulk of the void", "a tesseract hypercube in an illuminated glow, a tesseract suspended above the dint of reality", "russian cosmonauts driving a rover on the lunar surface in the style of Lucien Rudaux", "symbol of the phoenix, a phoenix rising over all the sentences that have ever been written", "a yin yang symbol where each half is a black snake and a white snake devouring each others tails"]}

Your task is to write 5 more prompts in the way you infer I'd write them from these examples, but based on a combination of subject, style, and setting. For example:

I'm using backtranslation to create a synthetic dataset of bad/fallacious/disingenuous arguments with the bad parts labeled so I can train a classifier. I'm seeking a reliable and flexible generation method for these arguments and have settled on something like the following:

Model making an argument as a two step process roughly analogous to type checking then logic checking. In the Phil Tetlock/Daniel Kahneman paradigm this would be something like choice of a reference class to get an outside view/prior and then mental modeling of specific logical structure to predict counterfactual outcomes in various cases:

  • Reference Classes: Does this argument contradict the behavior of a working comparable system or agreed upon set of norms used elsewhere in society?
  • Mental Models: Does this argument imply a model that captures the behavior of X correctly?

"Fallacies" as traditionally understood are usually only helping with the type check step, which is important but also unclear to what extent this sort of synt