's Avatar

@isaackarth.bsky.social

71 Followers  |  180 Following  |  35 Posts  |  Joined: 28.08.2023  |  2.4304

Latest posts by isaackarth.bsky.social on Bluesky

Title slide.
We Are Maxwell's Demons: The Thermodynamics of Procedural Generators. Younรจs Rabii. Roguelike Celebration 2025.

Bottom right: pixel art of a tiny guy heating up a teapot.

Title slide. We Are Maxwell's Demons: The Thermodynamics of Procedural Generators. Younรจs Rabii. Roguelike Celebration 2025. Bottom right: pixel art of a tiny guy heating up a teapot.

Tune in to @roguelike.club in 3 hours for my talk about the analogy between physics and procedural generation! We'll see how generators can be solid, liquid or gaseous and why โ€“just like Maxwell's Demonโ€“ designers are underrated little guys. ๐Ÿต๐Ÿซ–

26.10.2025 14:02 โ€” ๐Ÿ‘ 40    ๐Ÿ” 11    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 5

So I think you want Hal Clement, then. Mission of Gravity, etc.

As mentioned, Ander's The City in the Middle of the Night is literally climate fiction on a tidelocked planet, so probably also counts.

Forward's Dragon's Egg is set on a neutron star, but otherwise fits.

25.10.2025 05:41 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Roguelike Celebration 2025 a community-generated weekend of talks, games, and conversations about roguelikes and related topics

This year's Roguelike Celebration steam sale is up! Check out some awesome games, and don't forget to buy tickets for the conference *this weekend*!

(also @shing.bsky.social's art is so good, I'm so glad we got them to do this year's art!!)

store.steampowered.com/curator/4122...

21.10.2025 19:59 โ€” ๐Ÿ‘ 64    ๐Ÿ” 31    ๐Ÿ’ฌ 2    ๐Ÿ“Œ 1
I am a Doctor Now / Bjarke.it Bjarke A. Larsen's Blog

I Am a Doctor Now!
Little blog announcing my dissertation is now published and readable!
(And i have not been posting here so this also serves as the announcement that I am a PhD now lol)
bjarke.it/article/i-am...

14.10.2025 00:44 โ€” ๐Ÿ‘ 21    ๐Ÿ” 2    ๐Ÿ’ฌ 4    ๐Ÿ“Œ 1

For some things it'd be cool to be able to do the tool call completely inline; it's theoretically not a hard technical challenge to wire up a calculator as part of inference, but since it takes both training the model and editing the inference I haven't seen an implementation.

13.10.2025 11:12 โ€” ๐Ÿ‘ 2    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

Healing from my tooth extraction has been rough, please check out my Ko-fi, perhaps leave a buck? Or just share it around

24.09.2025 15:34 โ€” ๐Ÿ‘ 15    ๐Ÿ” 17    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

What grad students? We cut the funding and we're replacing them with AI too.

03.09.2025 23:15 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Join us for:
- a peek at the 2025 Roguelike Celebration social space
- a fireside chat with @perryjon.bsky.social about design & proc gen in UFO 50
- @apepers.bsky.social talking about Designing for Systems Suspense
- @ezra-szanton.itch.io talking about Synergy Webs in Roguelike Deckbuilders

15.08.2025 17:06 โ€” ๐Ÿ‘ 15    ๐Ÿ” 8    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 2

Not at all close. Novel-length plots are very hard. There has been a consistent, odd, overreaction on this.

15.08.2025 00:48 โ€” ๐Ÿ‘ 121    ๐Ÿ” 11    ๐Ÿ’ฌ 17    ๐Ÿ“Œ 4
Post image

This from @tressiemcphd.bsky.social hit me over the head like a mallet of truth. This is the thing. This is what Iโ€™ve been trying to warn people about perfectly crystallized. www.nytimes.com/2025/08/12/o...

12.08.2025 12:02 โ€” ๐Ÿ‘ 1223    ๐Ÿ” 420    ๐Ÿ’ฌ 15    ๐Ÿ“Œ 26

weather: solar eclipse

06.08.2025 03:25 โ€” ๐Ÿ‘ 168    ๐Ÿ” 34    ๐Ÿ’ฌ 13    ๐Ÿ“Œ 0

But using the AI for necromancy is just a complicated way to use false hope to torment the grieving for your own ends.

04.08.2025 22:19 โ€” ๐Ÿ‘ 4    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

Which is why I never trust an AI summary that uses my own phrasing. If it can't paraphrase it, it didn't understand it well enough to generalize it, and therefore cannot be used as a signal for how well it understands the concepts involved.

04.08.2025 22:19 โ€” ๐Ÿ‘ 4    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Humbug puppets are sometimes honest mistakes; the Eliza effect means that it is easy to fool ourselves into thinking that were were talking to something intelligent, only to realize that we were merely speaking to a mirror that is repeating back what we said.

04.08.2025 22:19 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

This shows up in more subtle ways; when a reporter gets a confession from GPT or someone asks a model how good it is at chess: they don't have access to that information, and you explicitly instructed it to lie to you, so why are you surprised when it lied to you? You told it to.

04.08.2025 22:19 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

There are a lot of clever ways to get more information, or better information, or fill in with other information. But there are hard limits. Just like there are a lot of ways to get more energy, but you still can't create a perpetual motion machine.

04.08.2025 22:19 โ€” ๐Ÿ‘ 2    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Lack of information isn't something that can be overcome with better technology; Information Theory describes physical limits that are as immutable as thermodynamics.

04.08.2025 22:19 โ€” ๐Ÿ‘ 6    ๐Ÿ” 1    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Necromancy is impossible because the data just isn't there.

You won't be able to ask Shakespeare about his plays or Socrates about his philosophy. You can converse with puppets wearing their faces as masks, but you ignore the strings at your peril.

04.08.2025 22:19 โ€” ๐Ÿ‘ 5    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

No, to have a Humbug Puppet you have to have it do something that it cannot do, such as bring back the dead.

I've been asked to bring back the dead before. I've always said no. Not only is necromancy a bad policy, but it doesn't work. You cannot recreate a person from the writing they leave behind.

04.08.2025 22:19 โ€” ๐Ÿ‘ 4    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

This goes beyond "surely the AI can't do such-and-such" complaints because there's quite a few things it can do. Sometimes badly, of course, but a language model can translate and summarize and it's silly to pretend otherwise, even if hallucinations mean you might not want to trust it completely.

04.08.2025 22:19 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

A humbug puppet is when the AI is presented as doing something that it cannot do but is treated as if it was doing that thing.

We point at the Great And Terrible Oz and ignore the little man behind the curtain.

04.08.2025 22:19 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Weizenbaum's Eliza/Doctor gave people the illusion of being listened to, merely by being able to reflect back to you what you typed in.

Many humbug puppets are similar: people have mistaken mirrors for meaning.

04.08.2025 22:19 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

The second thing is slightly more forgivable, as many people do it out of ignorance, ascribing the properties to the AI that it does not and cannot possess.

The Wizard of Oz admitted that he was a humbug, puppeting his stand-in, but too many people fall prey to the Eliza effect.

04.08.2025 22:19 โ€” ๐Ÿ‘ 2    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

If the purpose of an artifact is to be a witness for reality, you cannot use generative AI to create it.

04.08.2025 22:19 โ€” ๐Ÿ‘ 5    ๐Ÿ” 1    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Given that a cookbook unmoored from reality has the very real potential to kill someone and at minimum is likely to create a great deal of harm merely from being wrong, I regard it as one of the most dangerous things that someone can do with AI.

04.08.2025 22:19 โ€” ๐Ÿ‘ 4    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

And, in the case of mushroom field guides, deadly.

04.08.2025 22:19 โ€” ๐Ÿ‘ 4    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Cookless cookbooks applies to a lot of different areas. Fashion models wearing generated clothes that have never been made. Mushroom field guides where no one involved ever ate a mushroom.

If the purpose of your thing is to demonstrate something about reality, an imitation is useless.

04.08.2025 22:19 โ€” ๐Ÿ‘ 13    ๐Ÿ” 3    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

My number one criteria for a cookbook is that someone has to have actually made the food. Nothing else matters if that criteria hasn't been met.

04.08.2025 22:19 โ€” ๐Ÿ‘ 20    ๐Ÿ” 2    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

There's two cardinal sins of talking about or working with AI that I think deserve instant excommunication:
- Cookbooks
- Necromancy

04.08.2025 22:19 โ€” ๐Ÿ‘ 53    ๐Ÿ” 13    ๐Ÿ’ฌ 3    ๐Ÿ“Œ 0

Sabina Howard-Khan
Naughty Sailor
she/her
kombuncha enjoyer
memorizes favorite poems

02.08.2025 18:33 โ€” ๐Ÿ‘ 6    ๐Ÿ” 1    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 1

@isaackarth is following 20 prominent accounts