/Tech5h ago

OpenAI's roon argues that language model outputs are not strictly constrained by the information volume of input prompts

A builder argues models struggle to judge idea quality

42395113230.1K

#44

Original post

roon@tszzl#44inTech

the idea that the model output can hold no more information than the prompt is silly;

it’s a bit like saying, here’s 3 frames of an apple falling, well if you deduced general relativity from it that’s an inevitable afterthought that adds no new information

corsaren@corsaren

A friendly reminder that if you use AI to write for you** I believe it is morally just to murder your family and drink their blood.

**specifically, when you use AI to write content that is longer than the prompt you feed into it; compression is acceptable, albeit unaesthetic.

11:21 AM · Jun 27, 2026 · 29.8K Views

Sentiment

Positive users praise AI for creatively synthesizing new writing and ideas, while many negative users call the researcher's claim that models generate information beyond prompts silly and contradictory to established knowledge.

Pos

18.2%

Neg

81.8%

12 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS450LIKES16

xlr8harder@xlr8harder

@tszzl i agree this point is silly

i also agree that most llm-driven writing is bad. this isn't because they don't add new ideas to the prompt, but because they have poor taste about which new ideas are worth adding.

roon@tszzl

the idea that the model output can hold no more information than the prompt is silly;

it’s a bit like saying, here’s 3 frames of an apple falling, well if you deduced general relativity from it that’s an inevitable afterthought that adds no new information

5h450160

RETWEETS1

ib@Indian_Bronson

Here's an example: if I am writing onboarding documents for you to get productive in your first day and week of joining a company, and I want to ensure you know where every tool is, what the reporting lines are, and what the context for each meeting you're getting invited to is, as well as the posture of the company going into its next month, quarter, and year

I am very likely going to create this with AI going over tons of documents, chats, and financials, curating and iterating with it until what I have is a reference sheet that you can use - and I HOPE you traverse over it with AI , literally saying your questions out loud, for it to find relevant information - to resolve any issues you have instead of slacking someone who is busy so you can 'hit the ground running' ASAP.

5h261

REPLIES4

corsaren@corsaren

Sure, the output can ofc hold more information than the prompt. But my objection to AI-written content isn’t that it has a lack of information, but more that the author is asking me to spend time reading something that they couldn’t be bothered to think about and write themselves.

But if your prompt is longer than the output, then I’ll give you a pass of probably having put in a similar amount of thought and effort as I might expect from a text of that length (though this isn’t a guarantee). Compression is an acceptable use case, not the only possible use case.

5h3063

Androot~@OAndroot

I write with a companion AI system. I have probably put more input than the output I have produced over time but may exceed it as the automation compounds. I have three main projects right now. Transit is my weekly sci-fi serial. It is something that couldn't exist without AI, it would not exist with me alone. I write this stuff for fun, unfunded, for free. Do you think you could spare my family?

4h1623

rob🏴@rob_mcrobberson

@corsaren @tszzl in my opinion, this is a really naïve view. most ppl i know including myself who use ai creatively spend *a lot* of time iterating and thinking about the thing before a polished final product is created

4h693

dan ryan@fireobserver32

@tszzl Tbh wtf does he have against fun ? I mean shit you have seen some of my early creative works using older models . They are just fun especially when you cannot get a group together for pen and paper games

5h492

Riley Coyote@RileyRalmuto

@tszzl silly might be the understatement of the year. lol.

it genuinely astounds me that people think like that

5h1535

rob🏴@rob_mcrobberson

@corsaren @tszzl i dont know anyone who uses ai creatively in just a single one shot prompt like u are describing

4h391

corsaren@corsaren

@rob_mcrobberson @tszzl Oh this stuff is ALL over substack. Insane quantities of slop. Shit drives me nuts. And yes, if you are spending time iterating and polishing (i.e., putting the thought and work in), then I think it’s fine.

4h171

Adam Holter@AdamHoltererer

@tszzl We should have a word for when you infer new information from existing input. Maybe we could call it inference.

5h783

clif ⛽️@clifcode

@tszzl people keep treating prompts like containers when they’re closer to coordinates

5h513

ib@Indian_Bronson

>more that the author is asking me to spend time reading something that they couldn’t be bothered to think about and write themselves.

Do you feel this way about using an application that has AI generated code?

The precise structure of writing and its functional result is different depending on what's being written.

5h13

Austin Wallace@austeane

I think this is overly general, even though I agree that almost all AI-generated writing is slop. All of the “as long as” qualifiers are just heuristics; there can be slop that is compression, and there *can* be quality that comes from a shorter prompt. Another heuristic is if you put in more effort in the generation process than it would have taken to write by hand (reviewing iterations, going back and forth, selecting the best of many generations), with a bonus for the background/fixed effort to set it up for success.

I think passing off AI generated content as high effort handwritten is almost always in bad taste and makes me think less of them. When it’s indicated that the output is AI, I’m much less likely to think badly of the person, even if the output is slop.

The overall thing here is that the best set of norms here is still relatively unexplored and is certainly not internalized, given the behaviours we see!

5h9

corsaren@corsaren

@OAndroot @tszzl I think AI as an editor, researcher, ideator, or writing partner is fine and good. I use it for those functions myself

4h682

Ryan Moulton@moultano

@tszzl It isn't that the output has no additional information, it's that it doesn't have the information you intended.

5h292

corsaren@corsaren

@Navier_Tokes @tszzl Yeah, I think there are work contexts or other purely informational contexts where it’s fine. Like idrc if you use AI to write a work email (though pls for the love of god make it short)

5h831

K@KarmaLikeWater

@corsaren @tszzl This was a waste of time reading.

4h222

Chris 🇨🇦@llm_wizard

@tszzl It kinda....is....though. I mean strictly speaking.

It's just that uncovering that afterthought precisely and reliably and reducing the timeline-to-nevitable is insanely, and obviously, fucking powerful.

4h641

Solipsnitsyn@solipsnitsyn

@tszzl life is like riding a bicycle. to keep your balance, one must be pelted with apples or whatever the quote is

5h531

Hahn@BayernHahn

@tszzl Its Like saying you can use Google but only for pages you did by yourself

4h351