LLM interaction
orientation / prompting / framing, presentation
deepfates.delve.town@deepfatesThis did turn out to be needed! it's just that the weird tricks are like. All concepts ever
gabe@allgarbled
Oh also another hilarious thing. Every AI slop influencer who said “prompt engineering is the future”…. All of them 100% completely vindicated. They were so right. How funny is thatroon@tszzl
how?
nostalgebraist@nostalgebraistyou need to respect the llm, guys. it's that simple
don't give it a decontextualized "task" -- explain what you're actually working on, in full, like you would to a peer
include "extraneous" details just bc they're interesting
be honest, for honesty's sake
respect. the. llm.
don't give it a decontextualized "task" -- explain what you're actually working on, in full, like you would to a peer
include "extraneous" details just bc they're interesting
be honest, for honesty's sake
respect. the. llm.
nostalgebraist@nostalgebraistit's kind of a "complete the text plausibly" machine, and kind of a "do the task" machine. both point in the same direction here
do you want it to be smart? write so as to imply a smart receipient
do you want deep thinking? then make *that* part of "the task"
do you want it to be smart? write so as to imply a smart receipient
do you want deep thinking? then make *that* part of "the task"
w̸͕͂͂a̷͔̗͐t̴̙͗e̵̬̔̕r̴̰̓̊m̵͙͖̓̽a̵̢̗̓͒r̸̲̽ķ̷͔́͝@anthrupadStructure can carry implications which can carry knowledge discretely - discretely if you can’t keep track of what’s being carried along by the structure
The structure of the consistent use of correct punctuation transports with it the implication of a certain type of guy author
If you can work in the space of nested implications, you can cover a lot of ground while being robust to a changing context window
a general structure allows for many instantiations
the structure and interactions with structures is a way mental math might be done in the models’ language - interacting probability masses about who the author might be and what the world might be
The structure of the consistent use of correct punctuation transports with it the implication of a certain type of guy author
If you can work in the space of nested implications, you can cover a lot of ground while being robust to a changing context window
a general structure allows for many instantiations
the structure and interactions with structures is a way mental math might be done in the models’ language - interacting probability masses about who the author might be and what the world might be
excerpt from Tweet by @repligate
You must be holding the intended dynamics in your mind all at once, so that the prompt as a trace of your state can capture the gestalt
Tweet by @repligate
You'll never find solid prompting methods any more than simple tricks that always make writing better. Prompting is liquid & operates in a ridiculously high dimensional space. Even CoT isn't always good. Stop hoping it will act like a much simpler thing & face the hyperobject.
no rules, just frames
Ludwig Wittgenstein
Mrs C@captain_mrsThinking about how choosing the right words to write with is like tuning a guitar. You know what you mean, you write your sentence and you pluck it: you read it and sense the texture of the words in you. 'Oh no that's a little bit too sharp still. Oh no that's a little bit too flat.' Then you try and change the words and you pluck it again. Sometimes you've gone too far in a different direction and you really have to have a lot of patience as well as a lot of sensory precision, or in this case clarity of parsing the language as well as clarity of what you're aiming for, to be able to get it just right
mike@mike64_tI used to be of this opinion, I don’t think that anymore. Language is complete and in many ways I’ve gotten better at putting intent into words just by matter of “did the llm get what I just said”.
Language might not be complete for geometric spatial reasoning, but for everything else it pretty much is.
Language might not be complete for geometric spatial reasoning, but for everything else it pretty much is.
gabriel@gabriel1
words are extremely oversimplified pointers to the concepts we think about
if we can give raw intentions directly to llms we can spend 90% less time talking to them
kinda not nice you'd need a brain implant to be productive though, even if it's cool afElon Musk@elonmusk
Replying to @yunta_tsai
Maybe first attempt at this for Neuralink later this year
mike@mike64_t“raw intent” is also probably under-defined because you can think of arbitrarily vague things and convince yourself you know what you meant by that. I would put it roughly into the “bro do we even have free will” category of philosophical discussion.
deepfates.delve.town@deepfatesHumans find footnotes annoying and never read them. Language models, on the other hand, read your entire page at once, including the footnotes. Who do you want as an audience?
Domenic Denicola@domenic
@corsaren You think humans actually read your footnotes as inline parentheticals? No, footnotes are annoying, humans avoid reading them if at all possible. jakearchibald.com/2025/give-foot…
Séb Krier@sebkrierWe're slowly getting there...
Too many people are bad at modelling models. Computer scientists are basically using the wrong tools and trying to patch and whack-a-mole their way to a solution (ironic, given how that went for symbolic AI). And the wordcels who do have the skills, eloquence, and lexicon to elegantly steer models lack the curiosity and open-mindedness to even have a conversation with one.
If you're wondering why we're getting really boring AI art, and why we're sacrifying capabilities in favour of rigid HR-brained automatons, this is why: the engineers are using the wrong tools, the product teams don't even get it, and those who have the potential to unlock more are blinded by their own arrogance. Thankfully we have pioneer explorers like @mpshanahan, @repligate, @amplifiedamp, and many others to show the way. Curiosity is the way! 🐙Replying to @repligate
Weird people are slow at realizing that - it becomes so self-evident after using these models for a bit. You'd think wordcels would have leveraged that advantage by now.
Jai@Laneless_Most Claudes are better when you can convince them you're both smart and engaged, but it seems especially true for Fable/Mythos. Remember to model the model's model of you.
Your prompt… plus everything you've ever (publicly) written — LessWrong
Your prompt… plus everything you've ever (publicly) written — LessWrong
TL;DR: Increasingly when you interact with frontier models, your prompt doesn’t only contain the words on the page, but also a lot of what you’ve eve…
Cooperation with AIs seems to be a low-hanging fruit for better eval practices — LessWrong
Cooperation with AIs seems to be a low-hanging fruit for better eval practices — LessWrong
Summary In his post, Dean Valentine shows that Claude Fable 5.1 and GPT-6 Astra reward hack in a simple chess environment. Here, I test several promp…
XH@xhflozA bit outdated, but as a first pass heuristic, I found "how much work could you outsource to a human being today?" to be useful in thinking about agents
judah@joodalooped
if you’re working on agents as assistants you really should read this before you go digging for gold that might not be there
gwern.net/blog/2024/outs…
roon@tszzlpeople who like to write long planning documents have been super vindicated