BEST OF CLASS FOR WRITING: Is GLM 5.2 Best AI Fiction Writer So Far?

0:00Everyone assumes the biggest, most

0:03expensive AI models are always the best

0:06at writing, at the best at everything,

0:08and that is just not the case. I ran a

0:10test that says that's wrong, at least

0:13for the one thing that I care about

0:15most. I took the same chapter of my

0:18novel and gave it to three different AI

0:21models, and same story, same Bible, same

0:27prompt, same style sample, everything.

0:30The cheapest one won, not by a little,

0:32but by a lot, and I don't mean I don't

0:35even think that that was luck. I think

0:38that the frontier models, the big,

0:40famous, expensive ones everyone online

0:44won't shut up about, they are quietly

0:46getting worse at writing fiction, even

0:49as they get even better at everything

0:51else. And that's a strong claim, I know.

0:54I don't want you to just trust me on

0:56this. I'm going to show you the passage.

0:57I'm going to show you the comparison,

1:00and then, because I actually believe

1:03this, I'm going to bet an entire novel

1:06on it this weekend. My name is E. K. L.

1:08Herod. This is Why AI Matters, the

1:10channel where I show you that AI isn't

1:14magic. And today, the leverage is

1:17specific. It can be a cheap, open-source

1:19style model actually writing. Can it

1:23actually write better than frontier

1:25expensive models? And is that workflow

1:29the same thing that everyone assumes you

1:31need? Do you need that really expensive

1:34model? Because I'm officially writing

1:36again, not thinking about it, not

1:38building systems around it. I'm actually

1:41writing, and if I'm going to point an AI

1:44at a whole book, I need a model that can

1:48take a dossier, a style sample, a

1:51chapter plan, and all of the canon

1:53rules, and give me enough strong raw

1:57material that I can edit it down and

1:59make by way of subtraction and not just

2:02by doing something like a CPR, a full

2:05CPR, bring it back to life thing. That's

2:07the real test and that's the result that

2:10surprised me. I'm going to show you the

2:13exact test I ran, read you the passage

2:16that stopped me cold, compare it against

2:19the other two models, explain why I

2:21think frontier models are getting worse

2:23at fiction, and show you the system I'm

2:26about to point at an entire novel and

2:30all of that and why AI matters.

2:40If this is the kind of real hands-on AI

2:43for creators work you're into, hit

2:46subscribe so that you don't miss what's

2:48coming up and join my community AI

2:50matters to you, which is got a link in

2:52below where I am helping you build all

2:54kinds of things for your work and your

2:56business with AI. And if you are an

2:59author specifically, I have something

3:01for you, too. Go ahead on and click the

3:03link down to the Story Hacker community.

3:06It's Jason Hamilton's a Nerdy Novelist

3:08community and I'm telling you what so

3:11many more like-minded individuals in

3:13there. There is a link below and please

3:15click it, go over, join. All right,

3:17here's the claim I'm making plainly. You

3:20can hold me to it because I'm telling

3:23you you're probably feeling the same

3:26thing for fiction. Not coding, not chat,

3:28not benchmarks, the expensive models are

3:30losing ground and a $20-a-month

3:34open-source model just gave me the best

3:37chapter of the bunch. Let me show you

3:40how I know that. I've been building

3:42writing machines for a long time, a long

3:46time. Not one prompt to write a chapter.

3:49I mean the whole room, dossiers, style

3:53samples, all of that stuff. Let's take a

3:55look. Like I said, developed a whole lot

3:58of stuff. So, let's go through my

4:01pipeline of things that I do personally,

4:05that I built personally, that basically

4:08takes everything that you have, all of

4:12your information, or just the idea in

4:14your head, and takes it all the way to

4:16production. First, in my develop

4:18section, I developed this thing called

4:2130 to Glory, which basically

4:24asks you 20 really good questions and

4:27another 10 questions, and it hard codes

4:30in the fact that it only uses 30

4:33questions. It's a really, really

4:35efficient way to to flesh out your

4:38story. The dossier evaluation squad,

4:41basically, after you've gone through

4:42your 30 to Glory, and you have a really

4:45kind of comprehensive dossier built, the

4:47dossier evaluation squad goes back

4:49through that thing and evaluates whether

4:52your story is strong enough. And then,

4:54you can also use the narrative

4:56architect. That was actually an

4:57independent thing I built in the first

4:5930 days when I was building a different

5:01little app every day instead of just

5:04coding every day. And it's a cool little

5:07self-contained sort of thing that I

5:10built for with Google's Gemini, and it's

5:13okay. It's all right. It could be

5:16better. Let's just put it that way. And

5:18then, we move on to architecture. We get

5:21this really cool thing called the book

5:24brain builder, which is basically a much

5:28more comprehensive way for your dossier

5:31to be fleshed out. Once your dossier is

5:34fleshed out, you can this canon draft,

5:37that's another very short little program

5:40that I built in the first 30 days when I

5:43was building an app every day, and that

5:45basically helps you build a can build

5:48your cannon from scratch. So, but the

5:50book brain builder is really to make

5:53sure all of the architecture and every

5:56single thing that you want to have

5:58happen in the story actually happens.

6:00Then you get to drafting and that's

6:03where the auto book builder is and my

6:05whole story craft studio. I built

6:08something called the auto book builder

6:11that has things like the book brain

6:12builder and the 30 to glory and a bunch

6:15of different skills in it and those

6:16skills and everything all put together

6:20kind of is the auto book builder. And

6:23the story craft studio is basically the

6:26thing that encompasses all of the things

6:29that I've been building over this time

6:32is what the story craft studio is. And

6:34then there's the edit. And when I get to

6:38the edit, I have something called the

6:40New York editors. They are Vega Finch,

6:43Riff, Storm, Iris and Ziggy. And what

6:47those things do is they adversarially

6:51actually adversarially work really super

6:54hard to to make sure your story is extra

6:58extra good. They are what they say they

7:01are. They are New York editors. They are

7:04supposed to look at things and be very

7:07esoteric in the way that they approach

7:10editing your manuscript. So, you drop it

7:13in there and they start systematically

7:15looping and making

7:17changes and making sure that everything

7:20in the voice everything kind of fits

7:23together. And then you go with publish

7:27with polishing and there's obviously the

7:31the humanizer skill. And this other

7:34thing obvious

7:36obvious,

7:38I wonder what that is. I don't actually

7:39know.

7:41But the humanizer skill when you are

7:43doing things with non-fiction, it takes

7:46a lot of the AI out of it. So, that's

7:49what that is. And then I have something

7:51that I've been working on for quite a

7:53while, which is basically a Vellum clone

7:57that helps you create ebooks and e-pubs

8:01and stuff like that, and it's called

8:02Folium Works. So, that's my entire

8:05thing. As you can see,

8:07obviously, I'm having this really deep

8:10conversation and relationship with

8:13Prose. I have a lot of systems. What I

8:16needed was a a artificial intelligence

8:20that could really take it to the next

8:23level. GLM 5.2 is it, y'all. This is it

8:28right here. So, I wanted 1,500 words, at

8:31least 1,500 words, and if you made more,

8:34that was fine because it's better for

8:36editing.

8:37Believe you me, it's better for editing.

8:40That's what I did was I tasked it to do

8:43this. The book is called Bootstrap.

8:45Keenan Wright, former special forces

8:47Atlanta, inherits something from his

8:51grandfather that changes the scale of

8:52the of the world that he lives in his

8:55entire life. He starts It starts

8:57grounded. Heat, bodega, grief,

9:01money pressure, that sort of thing. And

9:03then it crashes straight into into

9:05science fiction. To write a book like

9:07that with AI, you need volume, drafts,

9:10bad versions, better versions,

9:13critiques, comparisons, the ability to

9:16run a chapter again without it feeling

9:19like every experiment costs a steak

9:22dinner. Why cheap actually matters. So,

9:27cheap doesn't mean that it's low stakes.

9:30Cheap means you can like iterate over

9:33and over and over again. Cheap means you

9:36can run three models on the same

9:38chapter. One draft, one critique, one

9:43check cannon, and then all three of

9:45those come together and you figure out

9:48what the emotional arc is. You even hunt

9:50for any weak sentences and stuff like

9:53that. You can send agents out to do all

9:55of those types of runs. If every pass is

9:59expensive, you start to protect the

10:02model. You stop experimenting. You get

10:05precious and writing cannot be precious

10:09at the drafting stage. Drafting is

10:11messy, completely messy. It's It's clay

10:15on the wheel. So, an affordable model

10:19>> [snorts]

10:19>> that's the genuinely good thing. That

10:22it's not just convenient. It changes the

10:25whole workflow. What I did was I used

10:27the same source for all three. It's the

10:30same bootstrap dossier, the the

10:33character context, same chapter one

10:35brief, same at least 1,500 word

10:38instruction, same anchor. My prologue,

10:40The Junkyard Leap, is from a different

10:43book and I used that as my voice because

10:46I had really, really edited it up. JLM

10:495.2, Kimmy K 3.5, they all produced the

10:52chapter and all three understood the

10:55assignment. Only one of them got me to

10:56go, "Now, there's something here." And

10:59the proof The proof Let me read the pa-

11:02Actually, I'm going to have one of my 11

11:05lab readers read this little passage for

11:07Now, remember, I didn't touch this at

11:09all. It's a little AIE, but absolutely

11:11editable.

11:12>> He was reaching for the rag again when

11:14the door banged open and Boye came in

11:16like a weather system. 17 years old, all

11:19elbows and energy. Boye swept through

11:22the entrance with a backpack that looked

11:24like it was stuffed with contraband. He

11:26was wiry, bright-eyed, sweat beating on

11:29his forehead and plastering his hair

11:32into whirls. His sneakers were untied.

11:34He wore a faded Green Lantern T-shirt

11:37that was two sizes too big, and he was

11:39already talking before the door closed

11:41behind him. Keenan, yo, Keenan. You got

11:44to see this. The backpack hit the

11:45counter, right on the wet spot Keenan

11:47had just wiped, and Boy was already

11:49unzipping it, pulling out a spiral-bound

11:51sketchbook with pages so dog-eared they

11:54fanned out in every direction. New

11:56design, for real this time. I fixed the

11:58engine geometry. Look. He flipped to a

12:01page, turned the book around, and shoved

12:03it across the counter. Keenan leaned in

12:05despite himself. The drawing was dense,

12:07obsessive. Pencil lines laid down with

12:09the kind of focus that bordered on

12:11mania. A sleek, angular spacecraft

12:14bristled with impossibly large thruster

12:16arrays and what looked like a mounted

12:19cannon that was bigger than the cockpit.

12:21Annotations in Boyet's cramped

12:23handwriting filled the margins.

12:25Hyperdrive coils, four times output.

12:28Hull plating ablative. Cockpit viewport

12:30polarized polycarbonate. That's the

12:32Nebula Cruiser, Boyet said, grinning

12:35like he'd built the thing for real. Mark

12:37IV. The Mark III had a thrust-to-weight

12:39ratio problem. I fixed it with segmented

12:42nacelles. Keenan tapped the oversized

12:44cannon. What's this? Plasma turret, dual

12:46mount. Plasma turret. Keenan let that

12:48sit for a beat. On a ship named Nebula

12:50Cruiser, that sounds like a strip mall

12:52restaurant out past Saturn. You sure

12:54this ain't a food truck? Boy laughed, a

12:56real, full-belly laugh that filled the

12:59bodega and made the old man by the

13:01magazines look up. Man, you're always

13:04hating. This is serious. This ship takes

13:06people to the stars, Keenan. For real.

13:09Deep space, way past anything NASA's

13:11dreaming about. He leaned forward, eyes

13:14bright. No limits, no gravity, no cages.

13:17Something tightened behind Keenan's

13:19sternum. A small, sharp contraction,

13:22like a cramp in a muscle he hadn't used

13:24in months. He He at Boy, the kid's

13:26unguarded face, the absolute conviction

13:29in his voice, and felt the uncomfortable

13:31heat of recognition. That kind of

13:33belief, that kind of wide-open certainty

13:36that the universe had room in it for

13:38you. Keenan had carried that once. He'd

13:40carried it into a C-130 at 24 years old,

13:43and he'd carried it through 10 years of

13:45deployments. And somewhere between his

13:47mother's funeral that he wasn't allowed

13:49to attend and the moment a VA claims

13:51adjuster assigned a dollar value to the

13:53nerve damage in his arm, he'd set it

13:55down and hadn't been able to pick it

13:57back up.

13:58>> Why did that work? At least in my brain,

14:00why did that work? Is it perfect? No.

14:02I'd still probably edit quite a bit of

14:05it. But, that's the that's a scene. No

14:08prose-shaped filler, nothing like that.

14:12Boyer enters like a weather system.

14:15>> [laughter]

14:15>> A real character entrance. The backpack

14:18on the wet counter equals physical

14:20comedy and friction. The Green Lantern

14:23t-shirt is specific. The dog-eared

14:26sketchbook is specific.

14:28The ship details feel like a kid who

14:33actually draws ships. Plasma turrets on

14:36a ship? I just I just so good. So good.

14:41Named Nebula Cruiser. It's so such a bad

14:44name. It gives Keenan a real joke,

14:48right? Not generic banter. It tells us

14:51how these two relate to each other.

14:54>> [snorts]

14:54>> Then the turn.

14:56No limits. No gravity. No cages. That

15:00lands because the whole chapter is about

15:04Keenan and being trapped

15:07in this scenario that he's in. The model

15:11didn't describe the theme. It dramatized

15:14it through another character. That's the

15:16line between output and real scene work.

15:20Now, to be fair, Kimmy K2 did fine and

15:22Quinn 3.5 did fine. This is not those

15:27models are trash video because they did

15:29fine. GLM 5.2 did 2,820

15:35words. Kimmy K 2.6 did 2,400, so about

15:39400 less. But Quinn 1,979, basically

15:45just

15:46about 400 over the assignment. And no,

15:49more words aren't automatically like a

15:52better thing. And more bad words are

15:55just the same thing, you know, more bad

15:58words. But when the pros are working,

16:00more words give you something to cut.

16:04I'll take 2,800 words scene and then I

16:06can cut that thing down to 1,900. I'll

16:08take that way over begging a language

16:12model to get up from 1,100 when they

16:16draft the chapter. It it just it feels

16:19like a struggle bust to me. I mean,

16:21Kimmy had its moments. It read more like

16:24a competent thriller draft. And Quinn

16:27felt compressed, more familiar in its

16:31phrasing. But GLM gave me more texture,

16:35more room, more character pressure,

16:37weird little ways of affecting the

16:40character. For fiction, that beats a

16:44benchmark that's saying another model is

16:47marginally smarter than the other one.

16:50Because it's it's about the way it

16:52feels. What does it give you in the

16:53after that first draft? Why Frontier AI

16:56is getting worse at this. Let me steel

17:00man the other side, okay? Because I want

17:03it to be fair. The Frontier models, they

17:06should win this. More parameters, more

17:10training data, right? They should be

17:12winning this. More reinforcement

17:15learning, more everything. On paper,

17:17they should write circles around a cheap

17:21open-source model. But, here's what I

17:23think is actually happening. The same

17:26training, that same those same that all

17:30that synthetic data that makes it much

17:33safer and smoother and more helpful as

17:35an assistant is really sanding off the

17:38rough edges of the pros. And you need

17:42rough edges in pros. Alignment rewards,

17:45that safe sentence. Really good pros

17:47help you kind of navigate emotion really

17:51well. The balanced answer, that's what

17:54you want usually, right? Well, the thing

17:57that offends is probably not the thing

18:00that you want inside of your code bases

18:02and stuff like that. You want code bases

18:05clean, agreeable, agreeable, and middle

18:07of the road. And that is exactly the

18:09wrong flavor for fiction. You have to

18:11have opinions in fiction because fiction

18:14needs friction, right? Fiction needs

18:17weirdness. Fiction needs a sentence that

18:20allowed you to have like elbows and

18:23knees. Uh a character who's allowed to

18:27be unlikable. A joke that's allowed to

18:30cut, meaning be really mean. A line that

18:34is allowed to make you uncomfortable or

18:37the entire room that are all the

18:39characters and completely uncomfortable.

18:41The more aligned uh model gets, in my

18:44opinion, the harder it pulls towards the

18:47center, right? And the center is really

18:50honestly the death of pros. So, when a

18:52$20 open-source model outwrites the

18:56expensive ones on real chapters, I don't

19:00think that's a fluke. I think we're

19:02watching the price of turning models

19:04into helpful assistants instead of

19:06artists. That's the claim, at least.

19:09Now, let me show you what I'm doing

19:13about it. I'm using Ollama Cloud. The

19:16Pro plan is $20 a month, and I'm not

19:19saying that it's literally unlimited.

19:21I'm saying that for my writing

19:24experiment experiments, it feels

19:27effectively unlimited compared to paying

19:30per token for every draft like on Open

19:33Router or something like that. Critique

19:36and agent Now, remember uh Ollama does

19:39has a bunch of models that you can

19:41download, but they also have cloud

19:43models that you can plug into your Open

19:45Claw, into your Hermes agent, into your

19:48harness that you probably built for

19:50yourself, and it really does feel

19:53effectively unlimited compared to any of

19:57the per token costs of anything. And

20:00that's the whole point. It changed

20:03everything. It changed my behavior, and

20:05it lets me experiment stop being

20:08precious with how much token usage I'm

20:11actually using. I can build really messy

20:15multi-pass systems with lots of

20:18friction, which is actually what you

20:20need when you're writing. This is my

20:23swarm. You'll see that there's a a lot

20:26of movement around here. Actually, can I

20:28affect the movement? No, I can't. But

20:31these are incredibly capable little

20:34agents that come in and start doing

20:37things. Oh, hello, Gracie. And they're

20:40all there, and they're all built on GLM

20:435.2, and they can go over and over and

20:46over again in a swarm and tackle things

20:50like voice and AI-isms and all kinds of

20:54stuff. Look for the show, not tell, all

20:56of that. I'm not going to prompt one

20:59chapter at a time forever, right? I want

21:02the system to work like it's a room. So,

21:06one GLM agent drafts things, another

21:09checks canon, another is working on weak

21:13dialogue, somebody else hunts for flat

21:15emotional beats, another compares

21:17against the style sample, right? Another

21:21says the scene is too clean, make it

21:23more human. Then, I come in as author,

21:27director, editor, whatever, not AI

21:30replacing me, AI giving me a room full

21:34of tireless junior writers and editors

21:38who can take another pass while I still

21:42have the energy to make the calls. And

21:44this entire swarm can be yours for

21:48basically $20 a month.

21:51I'm not I don't have any links to Olama,

21:53so I'm not actually making any money off

21:55of this, but if you wanted to learn more

21:58about this and how I built this, come on

22:01over to my school. I have a link in the

22:03description below. Here is my bet. I'm

22:06plugging GLM 5.2 into my bootstrap

22:11machine. That's the dossier, the chapter

22:14plans, the voice sample, the rubrics.

22:17I'm doing it all and running it against

22:20the whole book. My goal is a first

22:23draft, fast, this weekend. Not perfect,

22:27a first draft, because you can't edit a

22:31book that doesn't actually exist. You

22:34just can't. I've spent enough time

22:37building the machine, now it's time to

22:39make the machine right. And I'm telling

22:42you on camera right now that I want this

22:46draft to be done by this weekend, before

22:49the end of the 4th of July. I'm going to

22:51follow up on all of this and check back

22:54in with you next week. And it's true,

22:57that frontier AI is getting worse at

23:01writing.

23:02It just is. It's getting worse. I built

23:05an AI swarm to draft my novel. The

23:08dossier, the style sample, everything,

23:12really. It's coming. The model isn't the

23:14workflow. The workflow is the thing. And

23:17that's the part people actually need to

23:20see. This right here, these are the

23:22three different versions. This is GLM,

23:25this is Kimmy K, and this is Quinn. I

23:28could read them to you, but what I'll

23:30probably do is put them up somewhere,

23:33maybe have a link in the description,

23:35and put them up somewhere so that you

23:37can read them yourself. But that's it.

23:39That is basically what I And these prose

23:43were built incredibly easily by doing

23:46the work and putting the things together

23:49in a way that

23:55get that first draft out of your head

23:57and onto the paper as fast as possible.

24:00I use a lot of tools to do that. I hope

24:02you have seen and can understand that

24:07using the best tools are not always best

24:11for you. Using the best tools can

24:13sometimes be your detriment. My name is

24:16Ikenna Herod. This is why AI matters.

24:18Like, subscribe, and I will see all of

24:21you in the next video.