jeremy-ken-anderson

[jeremy-ken-anderson]

The number of people who act like public defenders "get murderers off on a technicality" who do not then seem to give a single shit about prosecutors and cops who get innocent people into prison using coercion, manipulation, and good ol' fashioned lies. The mind boggles.

I have a friend who was falsely accused of a horrific crime. He has spent more than two years on house arrest, paying for an ankle bracelet, while waiting for his court date. He has taken a photo of the process of his plug-in, every night, for two years, when he plugs the device in, because - and this has already happened to him at least once - the device sometimes fails, at which point the company monitoring the bracelets tells the state that he is attempting to flee and he has to prove that the accusation is bullshit or he'll be thrown in jail. During the two weeks he spent in jail between the initial arrest and his bail hearing his cellmate attempted to kill him because the guard told the cellmate what the accusation was and the assumption from the cellmate was that he was guilty because, after all, he was in jail. He's had to deal with all this when they haven't even indicted him.

man the public defender discourse pisses me off so bad. yeah. yeah I do think that every single person deserves representation. yeah that includes people who *have* committed rape and murder and abuse. when I say every single person I mean every single person. if your idea of justice excludes one person it excludes everyone. next question

[jeremy-ken-anderson]

If you're going through a dry spell, you may have been cursed by a sand witch.

[jeremy-ken-anderson]

The Strawberry R Thing

Looking at a broader picture, the folks who love AI or profit from other people loving AI might tell you that people laughing about how the LLMs can't tell you how many Rs are in "Strawberry" are wrong; the AI learned and now will respond with not only an accurate answer but a pedantic one: that there are three lower-case but no capital R.

But of course, that's not what happened under the hood. "Learning" is not what LLMs do.

What happened was that the data set it received gained a larger portion of high-weight answers that align with the truth. And this is learning-adjacent! But there is a problem:

Looking at a still-broader picture than the individual "Strawberry" problem, what happened is that a tiny issue resolved because it went viral. A wide assortment of people posted "no, you idiot, Strawberry has 3 rs in it." Possibly thousands of man-hours, between millions of people typing in their phone, were spent on the individual issue. The other piece - "weight" as I called it - is that someone inside the company likely took the individual answer to the individual question and plugged it in as a high-value reply to that specific set of tokens. Which...That's what Siri was. Fifteen years ago. People came up with individual replies for common problems and, when it couldn't find something in its reply bank, it would perform a web search on your behalf. (Using Google, which is now substantially worse at searches due to the proliferation of AI Slop. Outstanding work, guys)

"Rs in Strawberry" was a sample of a class of mistake. And its resolution as an individual mistake required a circumstance that literally can't be applied to the entire class - and wouldn't fix the whole class if it could be, because tokens comprising the "correct answers" would start overlapping and creating new incorrect answers.

As LLMs process things, the problem at large is completely intractable.

[jeremy-ken-anderson]

I feel like someone saw “sexy banana” on sale as an outfit at Spirit Halloween and said, “I could do better than that.”

this is the type of work they’re gonna make high school art students study in ten years

[jeremy-ken-anderson]

Slop Slops Together

So part of what makes AI Slop into Slop is the lack of nuance in what its model absorbs to inform its decision-making.

To look at this in reverse, you can look at the really blatant theft scenarios, from artists like Sakimichan all the way up to the "Ghibli Filter." The way they get those is by narrowing the input so they're really, really stealing from the works of the individual artist and profiting off of them.

On the other end of that spectrum you have AI Writing, which a lot of people making the products for haven't thought enough about. As such a lot of writing that is book-adjacent is in the pool, such as cliff notes, bookback summaries, and even editorial notes and annotations from the author in special editions (like The Annotated Dragonlance Chronicles by Weis & Hickman).

What I love about this, as a mistake, is the congruence between the idea of physical "slop" - a slurry of messy stuff that's melding together without proper care for what's getting mixed - and the metaphorical "AI Slop."

See, because it includes editors' notes, AI writing will sometimes include a sentence where it describes how the reader should feel about the scene. Now, I could imagine a lampoon or spoof of writing itself, where this could be done carefully and with intent. Occasionally things like this are included in absurdist humor as well for a quick chuckle: "The third judge of today's Iron Chef is Japanese pop star Britney Spears, whose name is a coincidence." But in a self-serious fantasy tale of love and woe, telling the reader what the phrasing suggests about a person's mental state just tells me that the author doesn't trust their own dialogue to express things or, if the character's more taciturn, doesn't feel like providing behavioral cues to fill in the gap.

But yeah. Having writings that are about writing in the pool makes it more likely that this particular kind of bad writing will appear in the outputs of a "writing" LLM.

Anyway if you're a human writer (here, check a box right quick to prove it, okay, now tell me which of these nine boxes has a motorcycle in it), don't do this. And for that matter if you're telling me that a nation is a melting-pot show me first. Let me smell the spices. Have people arguing over cultural differences on a street corner, or praying peaceably alongside each other in completely different ways. I'm not saying never to tell me, but at least try implying it first. A lot of that "showing" will imply other stuff about your world, which can both draw your reader in and give you inspiration about further plot developments!

[jeremy-ken-anderson]

Mono and Mochi is really interesting to me to compare with Garfield because where Garfield strays into a bunch of really surreal territory to have stuff to say, as well as spending a bunch of time on "Jon is Pathetic" as the core joke of a given strip with Garfield entirely absent sometimes...

Mono and Mochi goes the other direction, with almost a quarter of its strips not really attempting to include a joke. Some of the comics are just moments where a cat is being cute.

To be clear I think that's good! I think Mono and Mochi has identified what people who like pet comics are actually there for: Pet Moments.

It's just unexpected, because the idea that 3-5 panels of comic means a joke is forthcoming has been thoroughly ingrained in my head for decades with only a few venerable offerings (Dick Tracy, for instance) holding on to the idea that they could do anything else.

[jeremy-ken-anderson]

The dog burrito!

hang on I’m trying to see something

don’t tell me the name of your pet, just tell me in the tags the name you call them that’s got nothing to do with their actual name

[jeremy-ken-anderson]

Pleased with the Whole Scene

The Kid's Revival scene in my isekai fantasy story delivers on a lot of fronts at once, really well:

The hero Gil is, functionally, a blue mage. But seeing a skill a monster uses just adds it to his list of learnable skills; He doesn't get around having to earn Job Points like everyone else. He's managed to build a bit of a rapport with two other isekai'd folks - Fallen, they're called - Song and Layun, in part by admitting that he died because a friend he trusted betrayed and murdered him. The town guardsman who's more powerful than they are for a little while yet went out to where they saw a fourth Fallen arriving, only to find he was already dead.

The scene starts with an economic argument. Three Fallen have arrived - two have purchased the Language Blessing on credit for four gold, while Song hasn't bothered as she can talk to the other two Fallen and the guardsman because they all have it. But the revival spell is another ten gold. As the rude merchant who'd be a secret corrupt slave-trader or something in most isekai rightly points out, the town is now looking at maybe tripling how much investment they're putting into the Fallen. Another two Language blessings, plus ten for a Revival, and is the public just SOL if this guy - who's literally already gotten himself killed once - gets into trouble again and dies without paying back the first ten gold? Who covers that loss?

The mayor puts the talk on hold by agreeing to loan the ten gold out of his own pocket, not town coffers. It still leaves a bit of questionable ethics around "selling" something that wasn't requested, but it's not too bad as long as it's a bit of garnished wages off of something he was likely to do even if unpaid. And besides, the Fallen have been summoned as the world's last hope against the Demon King.

The priest shows up and

[Mimicry] - 8 JP (Doppelganger)

This is chilling for two reasons: First, Gil knows because he just gained access to a skill that he's looking at a monster who's pretending to be the priest. And second, because he knows doppelgangers are a thing now but he can never catch one this way again. Indeed if this one gets away he can't spot it again. To deal with this one, he has to be able to deal with it now. He tells the 'priest' he reminds him of a friend he had before he was isekai'd, hoping Song and Layun will get the message that the guy's a foe, and asks to talk to the guardsman alone.

He asks the guard captain for info on doppelgangers, and says his insight has revealed the priest as one, but he understands he can't ask the guy to stab the venerable town priest on his say-so. Reading his tooltips, he recognizes that Shriek interrupts concentration, and that Mimicry requires concentration.

When he enters the doppelganger's making excuses for being unable to revive the kid, but on realizing it's in a precarious situation it goes ahead and uses its ability to copy any skill of the person it's duplicating (they have to be in its goop-cocoon to be copied perfectly with memories and skills and all) to cast Revive to throw off suspicion. Gil lets the spell complete, then buys Shriek and uses it.

This unmasks the doppelganger, Song beats it up like a martial arts master, Layun pumps it full of arrows, and then Song and Gil have to duck and cover as the newly-revived kid - who died from casting a fireball point-blank - hucks a fireball at the doppelganger. The explosion still doesn't finish it, but the guardsman is there to behead it and finish the job.

What I like about this scene:

Gil gets an advantage from his power that is pure knowledge

The advantage isn't some spiral-outta-control eternally-repeatable superpower. It's logic applied to his current issue.

He also beats the monster by thwarting its biggest advantage: Uncertainty.

Not for nothing, but Shriek is an ability from Minor Goblins. He turns the tide using a power from the least-powerful monster.

The whole party gets to shine, each interaction in the battle also being related to their character.

He uses the fact that he's allowed to use the skill buy system whenever he wants to go into solving the problem as if every skill costing 3 or less JP was an option.

Reading the power of the monster allowed him to know the weakness in the monster's power as well. This is an angle on Blue Magery that I've never seen explored anywhere ever.

Really the one big drawback is it seems to come out of nowhere, like a Stand Fight in JoJos but without the characters knowing they were against anyone yet.

[jeremy-ken-anderson]

Apropos of nothing, having these two posts back-to-back on my dash really tickled me.

Small experiment with design 🦋

[jeremy-ken-anderson]

I've said several times that my frustration with AI is not tech but hype. Even the negatives of the environmental impact are, as far as I can tell, only occurring because six or seven gigantic businesses are simultaneously trying to outperform each other on promoting worldwide adoption, rather than focusing in on the things the product is good at and just using it for that.

There's a task one could call "protein string guesswork." Essentially, you've got a gigantic pile of potential DNA strands and you want ones that fit a ruleset. LLMs are actually pretty decent at picking through the nigh-infinite set of potential GTCAs and proffering ones that are pretty close to the ones we already know.

This ticks all the boxes that, for me, make AI stuff okay:

Would be hard for a human

Gets double-checked for accuracy / bad results aren't acted on

Doesn't take anyone's job

Doesn't plagiarize anyone's work

Isn't used billions of times per day and therefore doesn't require a community's entire clean drinking water supply just to keep the server farm cool

On the other end I saw someone ask Gemini for examples of nifty tech. It provided an example that did not do what Gemini's answer indicated. As far as the guy could see, the product website didn't claim to do that.

But the part that gave me a chuckle was when the guy was like, "Still, the idea of a product that does that is interesting. Can you name some companies that do that?" The reply was along the lines of "Sure, here are some companies that offer services like [product idea]:

[Another that doesn't exist]

[Another that doesn't exist]

...

...

...

What I had to wonder was whether his request for "something like what I asked for before" got interpreted as "I want you to follow the same process you used to find my previous result." That, in turn, might have functionally worked out to, "Please hallucinate four or five more times for me." Regardless, that's what it did.

[jeremy-ken-anderson]

First Balatro Win

I got a lot farther on my first run of Balatro than with Slay the Spire. On StS I think my very first run I died a few rooms into the second floor. On Balatro I got to the final boss, which amused me relative to StS because it was the Heart. It disabled the linchpin of my silly little strat (which mostly needed the cumulative mult bonus from Ride the Bus) and left me with a 3k hand, and I couldn't make up the difference in the remaining hands and ended up with a frustrating 97k/100k.

My second run was not close to that. I ran a flush deck but the finale was the Driver's License for a mult-mult, and I hit 16 special cards at the second-to-last blind, making Ante 8's Big Blind and Boss Blind actually two of the easiest in the run. I had a foil even joker, for +30 chips and +4 mult per 2/4/6/8/10 played in a hand, a walkie talkie for +10 chips and +4 mult per 10 or 4, the driver's license, the abstract joker (+3 mult per joker), the jolly joker (+10 mult for Flush hands), and the driver's license.

The boss made you delete one joker, and of all of them I figured the Jolly was actually the best. I was pretty consistently getting 3 or more tens and fours, so that was +12 mult (AND 30 chips) for Walkie Talkie and Even, and with one of my five jokers out Abstract would still be +12 too. Despite deleting one of my five jokers my first hand was 75k and my second was 93k. Actually completing Driver's License was huge.

The boss that almost ended the run was the second-to-last, the Eye, which makes you unable to repeat a hand type. I overestimated how much mileage I could get out of the first Flush I laid out, when I should have waited for more tens and fours. I barely squeaked across the line on that one.

[jeremy-ken-anderson]

Mechanics to Narrative Rewrite

The story idea started from a place of mechanics - I knew I wanted my main character to have these powers, and for the challenges he faced to be this nature of challenge at this level of difficulty.

You'll notice that there was nothing in there about the people he'd meet, the political structure, the world, or even anything specific about what my main character wanted.

I just started writing, and then deleted, and then wrote, and deleted, and wrote. Three pages forward, two back.

As I said elsewhere my first draft had him too quickly finding experts. The thing is my guy needs to be a strategic thinker for the challenges I want him to have to play out the right way ("the right way" being "Jojos-esque becleverments") which means it's part of his character to ask questions basically whenever anything happens. The issue with immediately finding experts - especially ones whose job it was to explain stuff to him - was that it was all valuable info but played out as a gigantic info dump. In this case I opted to back up and not give him that advantage until later and expand the testing-things-out phase by making the whole situation new. In other words, there are people with bits of info about their own corner of things, but...it's like The Printout of the Elephant. People who have never seen an elephant are given individual photos of 1' squares of the elephant. Getting even a vague idea of what an elephant looks like requires a lot of people to come together and talk about their very limited experience. Also, everyone's in a race and the first (and only the first) person who has a full picture of an elephant gets a prize. So people might be reticent about sharing their 1' square of [ear and corner of mouth] without you first showing them your [couple toes of the front left foot].

I don't think giving up the info to the character was the problem. In retrospect it would have been okay to have one or two questions, cut to later with the protagonist feeling more knowledgeable, and then have him reveal stuff he learned in that conversation as reminiscences if it became relevant.

Anyway the second draft ended up with the protagonist too quickly making an enemy of the guild that might or might not be trustworthy but is definitely necessary to work with at first regardless. It was an interesting angle - I don't think I've seen someone who just kind of soured a relationship with one of the guild workers - but it cut off the question of whether the guild was trustworthy because the protag wasn't getting along with them anyway, and for reasons that didn't really have to do with trust.

In our current draft the protagonist is himself providing exposition, and then admitting that some of it is guesswork, and he and an ally are headed out to science it up and see whether any of the assumptions hold up to fieldwork. I don't think the starting city burns this evening, but that's less out of pity for the hero and more a) to make sure he's got enough tools to make his escape feel more like a solved puzzle than a lucky die roll and b) so that he's got a little more time for allies to arrive, so that he can escape with them, or make the hard choice to leave them behind, or helplessly watch 'em die, or- You get it. All the regular drama options. More lovable characters means more opportunities for investment. I'm pleased with the foreshadowing I have so far, of things going really wrong soon. Now I just need to remember to keep the foreshadowing if I delete two pages to make way for some other shift in tone...

Anyway I totally get why people go for narrative-driven or character-driven story. Mechanics-driven story eventually needs narrative and/or characters, and then you just have to throw out a ton of work.

[jeremy-ken-anderson]

The closest I could ever stand to get would be a Creative Director for a game. It'd involve a LOT more person-direction than I'd be comfortable with, but getting to live the dream of being The Idea Guy on a project might be worth it, just once.

[jeremy-ken-anderson]

In fairness, my mind is now boggled.

[jeremy-ken-anderson]

I just deleted two pages of text because they meant my protagonist "knew too much too quickly." Can't let a guy get himself out of a situation in the first quarter of the book, that's cheating!

The point of fiction is actually to put that guy in a situation™️, and he might try to tell you the point is to then get him out of the situation, WRONG, second situation

[jeremy-ken-anderson]

He loves his corn pillow

[jeremy-ken-anderson]

This is funnier the longer you've been reading Omniscient Reader, but it's a fun series even if you catch up all the way right now.

hey there *yurifies your yaoi*

based on this textpost by @dip-the-stick