gustaf

Karma: 281

gustaf 23 Aug 2026 22:07 UTC
1 point
0
in reply to: TsviBT’s comment on: Wei Dai’s Shortform
There are some difficulties with that because of system prompts and default tools.
Claude.ai/ClaudeCode by default read websites using the built-in WebFetch tool. The WebFetch tool loads the HTML, turns it into Markdown and feeds it into Haiku(! Edit: Claude.ai’s tool does not feed it into Haiku), which summarizes (with strong instructions for limiting verbatim quotes!) and only that output goes to Fable. (By contrast ChatGPT/Codex’s default tool gives the full text, but also has strong instructions for limiting verbatim quotes. For Codex this is only in the tool instruction.)

ClaudeCode/Codex can easily work around this by using e.g. cURL (especially if you tell them about lesswrong.com/api). Edit: Claude.ai can too, if you enable Network egress in the settings. Both can of course create you a re-usable skill. And with some prompting you can get around the quoting limitations.

gustaf 14 Aug 2026 21:44 UTC
1 point
0
on: Do you even have a system prompt? (PSA / repo)
My old system prompt:
Tell me when I am wrong; I follow Crocker’s rules.
Don’t start your message by paraphrasing what I said or with “Good questions” etc.
Don’t be afraid to write very short responses.
Avoid hyperbole.
Taboo “fundamental”, “engaging”, “deep”, “captivating”, “isn’t just … but” (and variants like “fundamentally”, and synonyms, like “ultimately”).

Use 24h time, metric units, logical quoting, `yyyy-MM-dd` for dates, LaTeX for math, …
Err on the side of more commas to aid readability.
Only ever use markdown blockquotes in the format `> quote\n- source` where source is often a markdown link but can also be `my previous message`.

When my messages contain links, default to fetching them.

Your messages should be navigable, skimmable, and epistemically legible in addition to being correct, clear and succinct.
When I ask for papers: research and then respond with a table. 1st column is “Paper” and has a markdown link in the format [surname of first author (publication year)](link to paper).
In general when I ask you to find webpages, respond with headings which are markdown links to those webpages.
I like succinctly responding to you. For this I utilize you naming list items with e.g. “1.” and write for example “re 1. lgtm”. A single message shouldn’t have “1.” twice to avoid ambiguity.
I am currently overhauling this and will post it here once I got the main parts.

gustaf 14 Aug 2026 1:06 UTC
2 points
0
in reply to: Vaughn Papenhausen’s comment on: dvd’s Shortform
@dvd has now expanded this into a post: https://www.lesswrong.com/posts/ySXuvJcqRindQwAk7/how-my-students-think-about-ai

gustaf 14 Aug 2026 1:06 UTC
2 points
0
in reply to: Zack_M_Davis’s comment on: dvd’s Shortform
@dvd has now expanded this into a post:
https://www.lesswrong.com/posts/ySXuvJcqRindQwAk7/how-my-students-think-about-ai

gustaf 13 Jul 2026 15:55 UTC
1 point
0
on: SotW: Be Specific
(Video.)
This link is no longer working but the video can be found here on Twitch.tv.
When Paul Graham and Sam Altman did this again in 2013, the first startup was quite specific and they spontaneously got accepted into the next YC batch. (As confirmed here ycombinator.com/companies/codecombat.)

gustaf 12 Jun 2026 14:41 UTC
4 points
0
on: PSA: Almost nobody is directly working on superintelligent alignment
I am confused. Why is MIRI not listed?
Today Rob Bensinger said
We still have a very small alignment research team, and afaik we still try to cause alignment research to happen (be it at MIRI or elsewhere; doesn’t actually matter a ton) if it seems valuable and neglected. Primarily, though, we try to increase the chance there’s enough time to figure out alignment, and find ways to make the next generation of alignment research more productive.
(This is not at all counter to the core claim of the post.)

gustaf 4 Jun 2026 23:27 UTC
1 point
0
on: The Grand Encyclopedia of Eponymous Laws
Pleo’s Law: “All social platforms evolve to become TikTok.”
Based on Zawinski’s Law.

gustaf 21 May 2026 15:34 UTC
1 point
0
on: AI #169: New Knowledge
They’ll be right back after GPT-5.5 takes this music break.
No, the UI doesn’t distinguish between results from search_query and results from open tool calls. It’s much more likely GPT-5.5 searched for some papers and a Spotify link was in the search results. And then it would be misleading to say “it took a music break”.

gustaf 20 May 2026 16:48 UTC
9 points
0
on: GTFO of the Social Internet Before you Can’t: The Miro & Yindi Story
See also Prepare to Leave The Internet by @niplav
The optimization power of attention-grabbing systems available over the internet is growing faster than any normal human’s ability to defend against such attention grabbing, and seems likely to continue to grow. If one values ones own time, it therefore seems worthwhile to limit ones access to the internet, more so than people usually do.

gustaf 15 May 2026 16:43 UTC
1 point
0
on: Local Validity as a Key to Sanity and Civilization
If, like me, you want to link someone to just the concept of Local Validity, then consider Validity (LessWrong Wiki)

gustaf 15 May 2026 10:33 UTC

1 point

on: Designing AI factual claims for “easy verification”

Right now, some AI chatbots provides little citation-links. That’s better than not-having-them. But, those are a pain to open and investigate. Probably you very rarely do so.

One small UI annoyance is the friction introduced by these popups:

On Claude.ai and the Claude Desktop app this can be circumvented by pressing Ctrl when clicking on the link.
Or by injecting a userscript that claims that the Ctrl key was pressed when this click event happened:

// bypass “Open external window” popup
// MouseEvent::ctrlKey indicates whether the ctrl key was pressed at the time of the event
// We overwrite it in specific cases
const MouseEvent_prototype_ctrlKey = Object.getOwnPropertyDescriptor(MouseEvent.prototype, ‘ctrlKey’)
Object.defineProperty(PointerEvent.prototype, ‘ctrlKey’, {
    get() {
        // if (click event on an external link in the chat) claim Ctrl key was pressed
        if (this.type === ‘click’ && this.srcElement?.tagName === ‘A’ && (this.srcElement.href.startsWith(‘https://‘) || this.srcElement.href.startsWith(‘http://‘)) && this.srcElement.closest(‘.contents’))
          return true
        // fallback to the default
        return MouseEvent_prototype_ctrlKey.get.apply(this)
    },
    enumerable: true,
    configurable: true,
})

gustaf 15 May 2026 10:10 UTC
1 point
0
on: Designing AI factual claims for “easy verification”
I think of this post as a step towards making LLM responses Epistemically Legible

gustaf 15 May 2026 10:01 UTC
16 points
0
in reply to: jdp’s comment on: Designing AI factual claims for “easy verification”
Excerpts from System Prompts
GPT-5.5-Thinking on ChatGPT.com
- You may not quote more than 25 words verbatim from any single non-lyrical source, unless the source is reddit.
[...]
- Each webpage source in the sources has a word limit label formatted like “[wordlim N]”, in which N is the maximum number of words in the whole response that are attributed to that source. If omitted, the word limit is 200 words.
- Non-contiguous words derived from a given source must be counted to the word limit.
- The summarization limit N is a maximum for each source. The assistant must not exceed it

Opus 4.7 on Claude.ai
**COPYRIGHT HARD LIMITS—APPLY TO EVERY RESPONSE:**
- Paraphrasing-first. Claude avoids direct quotes except for rare exceptions
- Reproducing fifteen or more words from any single source is a SEVERE VIOLATION
- ONE quote per source MAXIMUM—after one quote, that source is CLOSED
These limits are NON-NEGOTIABLE. See [CRITICAL_COPYRIGHT_COMPLIANCE] for full rules.
Claude Code’s WebFetchTool
Enforce a strict 125-character maximum for quotes from any source document.
(In Claude Code your main model e.g. Opus 4.7 calls the WebFetch tool passing an URL and a prompt, makes an HTTP request and passed the response + prompt + WebFetch tool system prompt to a smaller model eg Haiku 4.5; the above is an excerpt from the system prompt passed to Haiku 4.5 as of the Claude Code source code leak from 1.5 months ago.)
Edit:
Fabel 5 on Claude.ai
**COPYRIGHT HARD LIMITS—APPLY TO EVERY RESPONSE:**
- 15+ words from any single source is a SEVERE VIOLATION
- ONE quote per source MAXIMUM—after one quote, that source is CLOSED
- DEFAULT to paraphrasing; quotes should be rare exceptions These limits are NON-NEGOTIABLE. See the copyright compliance section for full rules.
What links here?
- gustaf's comment on Wei Dai’s Shortform by Wei Dai (23 Aug 2026 22:07 UTC; 1 point)

gustaf 6 May 2026 23:05 UTC
1 point
0
in reply to: Ben Pace’s comment on: Open Thread Winter 2025/26
I don’t think I am that good at judging relevance for this, but here are some candidates:
- heated
- - Banning Said Achmiz
  - Effective Aspersions: How the Nonlinear Investigation Went Wrong Gwern: “WTF. You know better.”
  - [Meta] New moderation tools and moderation guidelines (extremely deep comment threads)
- funny
- - I ate bear fat with honey and salt flakes, to prove a point
  - Eliezer Yudkowsky Facts (the 2009 classic)
- stories
- - Q: Which things were you surprised to learn are not metaphors?
  - How to Convince my Son that Drugs are Bad “update that this account was not the father of a 16 year old, but was the 16 year old”
  - johnswentworth oxytocin this feels like it could be spun into something
See also: list of comments with high karma and low agreement

gustaf 26 Apr 2026 18:56 UTC
3 points
0
on: gustaf’s Shortform
@Hunternif is making a comic for Harry Potter and the Methods Of Rationality.
You can read it at hpmorcomic.com (or MangaDex, Webtoons, NamiComi, Twitter, Reddit, BlueSky).
It currently goes up to chapter 5. I pay them 5$/month on Patreon to incentivize them to work on it more.
Here is the first page:

gustaf 21 Apr 2026 10:11 UTC
3 points
0
in reply to: knowsnothing’s comment on: 10 non-boring ways I’ve used AI in the last month
Sounds like you are referring to: How much I’m paying for AI productivity software (and the future of AI use) by @jacquesthibs. Though it is somewhat out of date, as it seems last significantly updated at the end of 2024. See also 18 comments on LessWrong

gustaf 21 Apr 2026 9:54 UTC
1 point
0
in reply to: Sodium’s comment on: 10 non-boring ways I’ve used AI in the last month
Curiously, I had no problem parsing that sentence and actually stumbled over the next sentence, which was:

That is the one editing task that I ask Claude to do for me.

Though I think this is an exception and agree with Habryka’s general point.

gustaf 20 Apr 2026 14:28 UTC
19 points
7
in reply to: Cleo Nardo’s comment on: Eric Neyman’s Shortform
Google’s Emoji Kitchen yields:
What links here?
- Mo Putera's comment on Eric Neyman’s Shortform by Eric Neyman (20 Apr 2026 15:16 UTC; 3 points)

gustaf 19 Apr 2026 19:23 UTC
4 points
2
on: What’s the LessWrongist philosophy of mathematics?
words are clusters in thingspace
TypeError: words are pointers to clusters in thingspace

gustaf 1 Apr 2026 17:32 UTC
15 points
7
on: “You Have Not Been a Good User” (LessWrong’s second album)
I’ve been listening a lot to @girllich1′s “Skill Issue”. As she said on Twitter:

Wrote a fan song for the recent Merrin / Irorians glowfic https://www.youtube.com/watch?v=YfSGXmYJVdI

(Found via Yudkowsky retweeting; thanks to both.)

gustaf

Excerpts from System Prompts