Knut Behrends knbknb

Beast Mode

Beast Mode is a custom chat mode for VS Code agent that adds an opinionated workflow to the agent, including use of a todo list, extensive internet research capabilities, planning, tool usage instructions and more. Designed to be used with 4.1, although it will work with any model.

Below you will find the Beast Mode prompt in various versions - starting with the most recent - 3.1

Installation Instructions

Go to the "agent" dropdown in VS Code chat sidebar and select "Configure Modes".
Select "Create new custom chat mode file"

Okay, I need to write a story using only the vowel 'o'. Let me start by brainstorming possible words. Words with 'o' as the only vowel. Hmm. Let's see. First, the characters. Maybe a protagonist named Tom? Wait, no, 'Tom' has 'o' as the vowel. That works. Or maybe 'Bob'? Yes, Bob is good. Let's go with Bob. Now, setting. Maybe a small town? 'Town' is okay. Or a forest? 'Woods' has 'o' and 'o', so that's good.

Plot: Bob needs to solve a problem. Maybe he's looking for something. A lost dog? 'Dog' works. So Bob's dog, maybe named Spot? No, 'Spot' has 'o' as the vowel. Wait, 'Spot' has an 'o' and a 'o'? No, 'S' 'p' 'o' 't'—only 'o' as vowel. So Spot is okay. So Bob's dog Spot is lost. He goes to look for him. Maybe he goes through the woods, meets a fox? 'Fox' is good. The fox helps him. Then they find Spot. But how to structure sentences without other vowels.

Let me list some words. Nouns: Bob, dog, Spot, fox, log, rock, pond, frog, box, sock, clock, shop, top, fog, moss, lot, job, cot, lock, drop, plot, trot

Reinforcement Learning for Language Models

Yoav Goldberg, April 2023.

Why RL?

With the release of the ChatGPT model and followup large language models (LLMs), there was a lot of discussion of the importance of "RLHF training", that is, "reinforcement learning from human feedback". I was puzzled for a while as to why RL (Reinforcement Learning) is better than learning from demonstrations (a.k.a supervised learning) for training language models. Shouldn't learning from demonstrations (or, in language model terminology "instruction fine tuning", learning to immitate human written answers) be sufficient? I came up with a theoretical argument that was somewhat convincing. But I came to realize there is an additional argumment which not only supports the case of RL training, but also requires it, in particular for models like ChatGPT. This additional argument is spelled out in (the first half of) a talk by John Schulman from OpenAI. This post pretty much

	PROMPT FOR LLM (copy+paste below, then append your plan)

	You are converting a generic plan of action into a production-ready research/engineering `todo.md` that is directly executable by a capable coding agent. Your output must be a single Markdown document with a fenced XML `<workflows>` block. Do not include any extra commentary, explanations, or chat—output only the document.

	### Transformation requirements

	* Be specific and operational. Replace vague goals with concrete steps, commands, checklists, acceptance gates, and explicit assumptions.
	* Prefer compact, high-signal prose. No filler. Use short paragraphs and terse bullets.
	* If information is missing, make minimal, clearly labeled assumptions (e.g., “Assumption: …”). Do not ask questions.
	* Compile the spec into executable oracles: pre/postconditions, invariants, consumer/provider API contracts, and metamorphic properties. Generate both tests and runtime guards.

	Draziw.Button.Mines
	ag.video_solutions.wedotv
	ahf.dummynation
	ai.socialapps.speakmaster
	air.com.beachbumgammon
	air.com.freshplanet.games.SongPop2
	air.com.gamesys.mobile.slots.jpj
	air.com.goodgamestudios.empirefourkingdoms
	air.com.kitchenscramble.goo
	air.com.lalaplay.rummy45

	#lang racket

	(require metapict)
	(require metapict/pict)

	(struct M-edge (turn) #:transparent)

	(struct X+ M-edge () #:transparent)
	(struct X- M-edge () #:transparent)
	(struct A+ M-edge () #:transparent)

	library(tidyverse)
	library(jsonlite)
	library(shiny)

	# get_json_data(date)
	# score_letters(letters)
	# read_keywords() on mac
	# split_char(word) into words
	# limit_words(words, str_length) to specific length and lower case
	# top_words(word, words_in, top_n = 50) - combo of split_char, limit_words

	library(tidyverse)
	library(stringr)
	library(jsonlite)

	# create a function to solve the Keyword game
	# this game uses a 6 letter horizontal word at the
	# intersection of 6 other words, where a missing letter from each of the vertical words
	# accounts for one letter of the mystery 6 letter length horizontal word

	# how to play

	#!/usr/bin/env bash
	set -Eeu

	trap 'STATUS=${?}; echo "${0}: Error on line ${LINENO}: ${BASH_COMMAND}"; exit ${STATUS}' ERR
	trap 'rm -rf ${tempDir}' EXIT

	readonly supportedUbuntuVersion="22.04"
	readonly tempDir="/tmp/setup"
	readonly devDir="${HOME}/dev"
	readonly scriptsDir="${devDir}/scripts"

	Rcpp::sourceCpp("mandeldrop.cpp")

	xlims=c(-0.65,0.65)
	ylims=c(-0.6,2.2)
	m <- mandeldrop(xlims[[1]], xlims[[2]], ylims[[1]], ylims[[2]], 1560, 2160, 512)

	# Colour palette from https://stackoverflow.com/q/48069990
	rainbow=c(rgb(0.47,0.11,0.53),rgb(0.27,0.18,0.73),rgb(0.25,0.39,0.81),rgb(0.30,0.57,0.75),rgb(0.39,0.67,0.60),rgb(0.51,0.73,0.44),rgb(0.67,0.74,0.32),rgb(0.81,0.71,0.26),rgb(0.89,0.60,0.22),rgb(0.89,0.39,0.18),rgb(0.86,0.13,0.13))
	cols=c(colorRampPalette(rainbow)(100),rev(colorRampPalette(rainbow)(100)),"black")