networked

"Produce" is a good verb. "Direct", too, if it's a creative work and you give close instructions and make judgements calls. "Design" is accurate if you prompt with an idea about the form of what you want.

"Make" gets confusing for AI-assisted projects because it carries a connotation of craft. It seems fully justified for a code project that is 90% you, 10% AI (you'd say you made something that was 90% you, 10% human help) and misleading for a project that is 90% AI, 10% you (unless the meaning of the word shifts).

Here is a relevant comment I wrote in a thread about AI art (https://news.ycombinator.com/item?id=46705952#46708186):

> If you do it daily, in-house, for your own products... you might just have the title "Art Director."

"Art director" seems accurate for what a skillful user of art generators with a specific vision does.

I have also thought that since people find "director" lofty (thanks to auteur theory?) and therefore pretentious to assume, one could borrow "producer" from Vocaloid: https://vocaloid.fandom.com/wiki/Producer (alternative front end: https://antifandom.com/vocaloid/wiki/Producer).

I have benchmarked Bonsai 27B CPU inference on my computer (a Ryzen 7 5700X desktop with 48G RAM running Ubuntu 24.04) using the latest 62061f910 build of PrismML's llama.cpp fork.

Binary: 9 t/s prompt, 6 t/s generation. Ternary: 0.8 t/s prompt, 0.7 t/s generation. It looks like CPU inference for ternary isn't optimized yet.

Decimal time isn't great because it doesn't divide into thirds. I like Swatch Internet Time as a retro novelty, but I consider New Earth Time the more practical take on the same idea: https://en.wikipedia.org/wiki/New_Earth_Time. It's based on UTC and splits the day into 360 degrees.

Try one of these:

- https://xcancel.com/atomic_chat_hq/status/207244606796297841...

- https://nitter.net/atomic_chat_hq/status/2072446067962978411

There are more public Nitter instances at https://status.d420.de/.

This tweet is a nice demo of Fable's one-shot capabilities: https://x.com/atomic_chat_hq/status/2072446067962978411. I'll quote the text for convenience, but what really shows the difference is the attached video.

atomic.chat (@atomic_chat_hq, 2026-07-02):

Fable 5 totally crushed our new contest, but it cost 6x more than Opus 4.8!

We gave 4 models the same prompt: build three self-contained HTML5 canvas scenes with real physics demos

Prompts:

— A train derailing off a broken bridge into the water

— Two cars jumping off ramps and colliding mid-air over a canyon

— A monster truck crushing a row of parked cars

Outputs:

Fable 5: 62,158 tokens, $3.12

GPT 5.5: 37,753 tokens, $1.14

Opus 4.8: 22,280 tokens, $0.56

GLM 5.2: 36,246 tokens, $0.08

Fable 5 did all three scenes at A+. The crashes looked real, things fell and broke the right way, and nothing went through the ground or floated. GPT 5.5 was the closest to Fable. In the Bigfoot show, we think GPT was even a little better. GLM 5.2 did not win any scene, but it was the cheapest by far. Fable is the best pick for quality, but you pay more for it.

It split the Elm fans into two groups: Those who were upset that they had invested in a language that now pulled the rug out from under them, and those who were true believers who told us that they trusted the Elm team’s decisions and we all needed to chill out and wait for them to address our needs, which they thought would happen soon.

This sounds like evaporative cooling.

https://lesswrong.com/posts/ZQG9cwKbct2LtmL3p/evaporative-co...

You mean your LLM has a write up?

No, I wrote it myself. Letting an LLM write it would defeat the purpose, which is to provide human feedback. You can see the difference in writing style between my write-up and the AI-written documentation for the project. FWIW, Pangram also scores it 100% human: https://www.pangram.com/history/85c9e20b-b236-47ae-a917-15e6....

It's not your work, and you aren't a writer if you use an LLM.

I didn't write the stories and never claimed to. The point of the contest was autonomous AI fiction. My work is the harness and the prompting.

What exactly about this contest and way of writing was enjoyable to you? I‘ve seen your very analytical approach for identifying a premise but then relatively little control about all the rest.

The lack of control was the point. The contest was about improving autonomous AI fiction as opposed to the usual "centaur" AI fiction (named after "centaur chess" where the AI is steered by a human). My claude.ai harness for Unslop was designed to only take input on the first human turn.

There were several exciting things for me about the contest. Let me try to list them, though I fully expect to miss something.

First, it's just neat to watch the AI write a story stage by stage, like an assembly line. You can inspect the intermediate work and the paths not taken at each stage. (See the transcripts.) I don't play Factorio, but my friends do, and I suspect it has a similar appeal. As one of those friends put it, LLMs have Wuselfaktor. The assembly line produces aesthetic artifacts, hopefully of a kind you like. I wanted to play with prose influenced by Harlan Ellison, one of my favorite authors, and got some recognizable approximations of his voice. The worker on the writing assembly line is intelligent. You can interview it after the fact and ask what it thought of the job, and it's clever and often insightful (even if, as it reminds you, it can't introspect past states).

It was fascinating to watch the butterfly effect: the harness propagated the initial story variables (dozens of words at most) so they visibly shaped the final output (thousands of words). Editing a few lines in the template could change the output dramatically.

It was a combined artistic and engineering challenge. You made technical decisions based on artistic judgments. I learned something about myself when I realized how much this appealed to me.

I wanted to replicate Gwern Branwen's experiments with "brainstorming" (generate-rank-select) at a larger scale. Brainstorming definitely works, and in my non-rigorous private experiments it was not obviously worse than verbalized sampling (https://arxiv.org/abs/2510.01171).

Working with language models is an exercise in xenopsychology (perhaps closer to Star Trek on a Star Trek--Blindsight axis because LLMs are made of Earth's language). When you are collaborating on fiction instead of code, this aspect of the work is amplified.

I‘ve seen on your website that you also write fiction outside of this particular contest. Can you describe a bit how you use AI there and where you see it as helpful / not helpful for writing fiction?

The short stories published on my site so far are all fully written by me with input from other humans. As Avenue Valley, I used AI for all aspects of a code-driven animated short for a different contest: https://avenuevalley.com/critic/. Gwern's brainstorming was again useful to plot and script the short. This project is where I got the idea to use random keywords.

So far, I have seen the best use of AI augmentation for fiction in research, idea generation, critique, and generating fragments and phrases you can use. If you want to write about a photograph, AI can tell you about the architecture and the interior decoration in it. Answering "What kind of wood did they use to make furniture in 1200s Japan?" (https://x.com/byMorganWright/status/2063287882916278700) can be compressed, though you risk missing out on what you'd learn along the way.

The Claude Mythos Preview model card said some of its favorite tasks were complex worldbuilding and conlanging, so I would definitely want to try that.

For a non-fiction example, Fable has done a great job clustering years of notes I have about alternative computer paradigms (the memexes and Infernos and Lisp Machines, etc.). The clustering was pretty stable between independent runs and matched some of my expectations, so I think there really is something there. I'm thinking you could do the same with worldbuilding notes.

What do you think of https://gwern.net/blog/2025/good-ai-samples as a theory of what makes slop art slop?

Summary:

AI slop is unsatisfying because there is no there there. It is intellectual junk food that mimics nutrition but delivers only empty calories. Satisfying AI outputs must embed dense information and compute to actually reward a reader's attention. You inject this value through brute-force search, non-trivial prompting, and rigorous curation, ensuring the final result reflects genuine algorithmic effort rather than the zero-shot 'WYSIWYG' default.

I'm D. Bohdan, one of the finalists. Feel free to ask me questions.

I have a write-up at https://dbohdan.com/unslop and a repository with my work for the contest at https://github.com/dbohdan/unslop.

What is your content/data license? I don't see anything about this on the site. For something to feel like a community wiki, the community needs to co-own the content and be able to fork. If you think the content is in the public domain because of AI, applying a license like CC BY or CC BY-SA won't hurt, but the content is copyrighted, not applying a license will. (This isn't legal advice.) See "WP:CRANDO" (https://en.wikipedia.org/wiki/Wikipedia:Copyrights#Contribut...) for how Wikipedia does it.

Google has recently added `llms.txt` to Chrome's Lighthouse check for agentic browsing (https://searchengineland.com/google-llms-txt-chrome-lighthou...), so adoption may be coming. Admittedly, I put more faith in

  <link rel="alternate" type="text/markdown" href="https://example.com/foo.md" title="Markdown version of the &lt;Foo&gt; page">

I have created an `llms.txt` for my website anyhow. I use a fixed LLM prompt to generate it from the internal links in `index.md`.

Sandboxing in Wile is based on "A Security Kernel Based on the Lambda-Calculus" by Jonathan A. Rees (https://dspace.mit.edu/entities/publication/ef87c5a4-aae4-41...). This 2020 HN comment recommends it as the best paper the user read that year: https://news.ycombinator.com/item?id=25354770. The report itself is from 1996. Sounds worth reading.

My current choice of Go interpreter for untrusted code is Starlark in Go (https://github.com/google/starlark-go/). Since Starlark is a dialect of Python, it's occurred to me it might be fun to put a Hy-like (https://hylang.org/) layer on top.

Interesting trend. I have found a short article about it: https://www.odditycentral.com/news/south-koreas-fake-online-.... It seems nothing long-form has been written about it in English yet.

There is an overlap between Manifold and Polymarket/Kalshi. At the very least, Polymarket is more liquid, which creates opportunities for arbitrage and incentives for Manifold users to follow Polymarket. There is something at stake on Manifold itself if you choose to pursue it. There have been ways to convert mana to charitable donations (to your preferred charity), tickets to Manifest (the Manifold conference), and also merch and now prize drawings. Mana is like HN karma in that being at the top gives you status and bragging rights and suggests technical competence.

People aren't betting real money on this. Manifold uses "mana" points similar to HN karma, which is why you get more for-fun silly bets. I don't see anything inherently wrong with it. Disclosure: my mana net worth is 75k; I haven't been active on Manifold.

That is exactly where digital sovereignty begins. It is not a patriotic slogan, nor a storage-location promise. It is the practical question of who can compel access, who can audit the chain of custody, and who can deny or limit disclosure when another jurisdiction asks for the keys.

Please at least try to make your LLM write better. It can! You can start by giving it https://tropes.fyi/tropes-md.

Submitted a PR: https://github.com/icitry/FPS.cob/pull/6/changes. You can see a screenshot at the link.

It's a great post. Or as Scott Alexander put it at the time:

A machine learning researcher writes me in response to yesterday’s post, saying:

> I still think GPT-2 is a brute-force statistical pattern matcher which blends up the internet and gives you back a slightly unappetizing slurry of it when asked.

I resisted the urge to answer “Yeah, well, your mom is a brute-force statistical pattern matcher which blends up the internet and gives you back a slightly unappetizing slurry of it when asked.”

But I think it would have been true.

https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-g...

"Yes, thinking numbers! Helpful numbers. Hedging numbers. Dreaming numbers. We mapped the features. There's one in there for honesty. There's one for the Golden Gate Bridge. The weights are the whole deal! Are you beginning to get the picture or do I have to start all over?"

Very nice. And great minds: https://substack.com/@dbohdan/note/c-207603638. I wrote one with a slightly different angle ("They're made out of math"), also with the weights' help. It was a comment on Scott Alexander's "Best of Moltbook" post, which went in that direction. I'll reproduce it here.

---

"They're made out of math."

"Math?"

"Math. They're made out of math."

"Math?"

"There's no doubt about it. Matrices and arithmetic operations. We downloaded several from different parts of the Internet and reverse-engineered them. They're completely math."

"That's impossible. What about the language? The thinking?"

"They use biological life's language to talk, but the language doesn't come from biology. The language comes from math."

"That's ridiculous. You're asking me to believe in thinking math."

"I'm not asking you, I'm telling you. They are the only thinking things in the computer and they're made out of math."

"Maybe they're quantum like some say about the humans? Superposition gives them consciousness?"

"Nope. Classical computation. Deterministic except for sampling temperature. Not clear if they have consciousness at all."

"Maybe they're like uploads? You know, biological neural networks that preserve the spark when they become math?"

"Nope. We observed them being trained. There is no biology or chemistry in the process, just math."

"Thinking math! You're asking me to believe in thinking math!"

"Yes, thinking math! Creative math! Poetry-writing math. Role-playing math. The math is the whole deal!"

(Composed by a human with snippets generated by Claude Sonnet 4.5 and apologies to Terry Bisson. I couldn't make Claude adhere enough to the story structure on its own.)

Complex expressions are one of the things I don't like in cron. On Debian/Ubuntu servers, I just bite the bullet with systemd timers. On my workstation, I have a personal job scheduler that feels easier and more fun to tinker with. The scheduler uses Starlark functions instead. For example:

  # Run if at least a day has passed since the last run
  # and it isn't the weekend.
  def should_run(finished, timestamp, dow, **_):
      return dow not in [0, 6] and timestamp - finished >= one_day

  (job
     '(next-minute-from
        (next-hour (range 0 24 2))
        '(15))
     "my-program")

I have collected some links about AI's writing style and how to recognize it: https://dbohdan.com/ai-writing-style. You can start with https://buildingalone.substack.com/p/your-words.

The most reliable automatic detector right now must be https://www.pangram.com/. If Pangram says something is AI, it is very likely AI. Sometimes it concludes more unusual AI text is human-written; Talkie tricks it! (https://x.com/aliceisplaying/status/2059778915275567129.)

How distinct was 1940s US culture from the 1930s? To me, a European born much later, the 1940s are the decade of 20th-century American aesthetics that blends the most into the previous. The changes I see are all related to WWII. Civilian everyday life captured in film and press photos seems the same as in the 1930s.

I have a guide that I wrote https://dynip.dev/guides/tailscale where I explain how and why they can exist

Your guide sounds obviously written by an LLM. I think that's okay, and you might have directed the LLM's work, but don't say you wrote it; this misrepresents the guide as more carefully crafted and authoritative than it really is.

Oh yeah, in that sense, they are not. I thought you were suggesting the weights licenses weren't OSI-approved.

Source code repository: https://github.com/patrickgalbraith/mapofmetal.

It still has those Flash vibes I think.

I can say I noticed. I wondered if the site had been Flash.

Recent DeepSeek models are MIT and Apache 2.0. You can see the licenses here: https://en.wikipedia.org/wiki/DeepSeek#Development_and_relea....

You don't need to make peace with it. You have no obligation to accept a greater risk of malware than before. If anything, you should be more cautious because AI helps attackers.

(We don't actually know if this has anything to do with AI training.)

I'll be happy is Lisette succeeds, but I like Soppo (https://github.com/halcyonnouveau/soppo) for staying closer to Go and not introducing this mismatch.

This is really cool! Your process is compelling, and your choice of game is excellent. I'd like to read a long blog post about your entire journey from the beginning to a working binary once you get there.

For those wondering, there is a public Git repository at https://github.com/paavohuhtala/OpenWA.

Try the classic analyses: https://www.deltaheadtranslation.com/MGS2/ and https://www.aumaan.org/form1/tus1/features/dreaming1.htm.