Sean Lees
A dark word cloud of familiar AI-writing tics surrounding the phrase ‘load-bearing’.

Frontier Models Don't Write So Good

Share on X

GPT-3.5 ar­rived on the scene during the final year of my law degree. I used it as a study and re­search part­ner: it could reason across com­plex legal texts, ex­plain them to me, and help me cram more ef­fec­tively than ever before. Back then, fron­tier models weren't yet 'agen­tic'. Sure, they had a few tools in their har­nesses (e.g. web search), but pri­mar­ily they were text-based as­sis­tants:

user inputs text model reasons model outputs text

And they were damn good at that!

Beyond the Chatbot

It was clear that LLMs had po­ten­tial far beyond the Chat in­ter­face. The race was on to make models more 'agen­tic': able to do more in the real world. Tool call­ing needed to become more re­li­able, and models needed to be able to reason longer and try harder before giving up. This was a post-train­ing arms race, with each major model re­lease herald­ing a step-change in agen­tic ca­pa­bil­i­ties. In 2025, the models started get­ting good. Opus 4, in the nascent Claude Code, could sud­denly work across and main­tain an entire code­base. By late 2025, Opus 4.5 was good enough that pro­fes­sional soft­ware en­gi­neers were trust­ing it with real work. The GPT-5.6 models are re­lent­lessly agen­tic: they can work for days, call com­plex tools, and write great soft­ware. METR es­ti­mates that the du­ra­tion of tasks models can com­plete with a 50% suc­cess rate has dou­bled roughly every seven months since 2019.

Making models more agen­tic was half of the goal. The other half was making them more in­tel­li­gent. Nobody wants an in­de­fati­ga­ble as­sis­tant if it's stupid. Likewise, an eru­dite chat­bot isn't very useful if it's im­po­tent. Frontier labs' in­ex­orable drive to scale their mod­els' raw in­tel­li­gence has been in­spir­ing. They're com­mit­ting tril­lions in capex and hiring the world's smartest minds to work on my gen­er­a­tion's Manhattan Project. The gains kept coming through 2025, but I really started to feel the ac­cel­er­a­tion in 2026: GPT is solv­ing Erdős prob­lems; an Anthropic math­e­mati­cian has just used Fable 5 to dis­prove the Jacobian Conjecture. And this in­tel­li­gence isn't just ac­ces­si­ble to math­letes: even I used Fable to solve a re­search prob­lem I'd been tin­ker­ing with for eight years, let­ting me im­ple­ment a novel state-of-the-art loss­less audio codec (but that's a sep­a­rate ar­ti­cle!).

Claude Fable

Unfortunately, Fable also pushed me to in­ves­ti­gate a sus­pi­cion that had crys­tal­lized over 18 months of work­ing daily with fron­tier models. The first thing I did with Fable was point it at my notes. I main­tain an Obsidian vault filled with my life's work: ideas, re­search, brain dumps, jour­nal en­tries, and even song­writ­ing (I used to be a record pro­ducer!). I use my own hybrid search tool, talon, so agents can search and reason over this knowl­edge base with­out load­ing the whole thing into con­text. I'm ob­vi­ously in­ti­mately fa­mil­iar with these data: taken to­gether, they're my life's story! I'm also used to what agents do with it. So when I asked Fable:

"Analyze Sean's character and the state of his life based on this vault. Identify his weaknesses and blind spots. Write him a plan to help him realize his potential"

I al­ready knew with some ex­ac­ti­tude what a good answer looked like. I'd asked pre­vi­ous fron­tier models the same ques­tion, and I ru­mi­nate over it in my jour­nal every day.

Fable blew me away. It found con­nec­tions be­tween my dis­parate ideas that I could never make myself, though I can't dis­close more de­tails with­out leak­ing half my "Dear Diary" en­tries to the in­ter­net! Although I had yet to ask it to do any­thing new, the leap in in­tel­li­gence was un­de­ni­able: Fable had that sweet 'big model smell'. But all I had in front of me was the text it pro­duced, and that text fi­nally con­firmed the shift I'd been sens­ing for 18 months:

AI Slop

The writ­ing was shit. The prose was un­nat­u­ral, the struc­ture pre­dictable, and the vo­cab­u­lary full of in­fe­lic­i­ties. It was rid­dled with 'tics': phrases that rarely occur in human writ­ing but appear con­stantly in model output:

load-bearing; it's not just X, it's Y; you're absolutely right; I hear you; hand-waving; smoke-test; goblins; the em dash; and many more!

Distinctive, iden­ti­fi­able slop is only half the prob­lem. The other half is harder to put your finger on. The best de­scrip­tion I have is men­tally taxing. As a de­vel­oper, I am ac­costed by pas­sages like this thou­sands of times a day:

"Independent review is still running. One thing I've already challenged explicitly: a log line that says only 'commitTransition failed' is not enough if the nested cause says 'generation rotation requires settled durable work'. I'm holding the logging branch to actionable-cause output, not merely non-JSON prettiness." - GPT 5.6 Sol

I for­give you if your eyes glazed over while read­ing this. Mine did too. This is my code­base, so I know ex­actly what the pas­sage is talk­ing about and what every word means. I enjoy chal­leng­ing lit­er­a­ture and con­sider myself a good reader. And yet the pas­sage is a schlepp to read. I have to return to the be­gin­ning sev­eral times and exert myself to make sense of it. The action is buried be­neath in­vented com­pound words, stretched metaphors, opaque con­trasts, and double neg­a­tives.

Anthropic has no­ticed this too. Its Mythos Preview system card de­scribes almost her­metic prose: the model 'writes densely, and as­sumes the reader shares its con­text'; some testers found its state­ments 'dif­fi­cult to un­pack'. My hy­poth­e­sis is that as models get better, their com­mu­ni­ca­tion gets worse.

The Problems

As model in­tel­li­gence climbs, the user's IQ is in­creas­ingly the bot­tle­neck. Likewise, as the models become more pow­er­ful, it be­comes more im­por­tant for the user and the model to remain on the same page. For the time being, we're still in the driver's seat: very few agent sys­tems are com­pletely au­tonomous. If a mis­un­der­stand­ing of intent arises be­tween the agent and the user, it can mean hours of wasted work or costly, dan­ger­ous mis­takes. Intelligence has raised the stakes.

Direct model <> user mis­com­mu­ni­ca­tion wor­ries me less than the degra­da­tion of human <> human com­mu­ni­ca­tion that LLMs have pre­cip­i­tated. AI use is ex­plod­ing, and as people rely on it more heav­ily in daily life, they nat­u­rally want to au­to­mate their writ­ing: emails, mar­ket­ing copy, social media replies, Hinge mes­sages I re­ceive (don't ask), and ev­ery­thing in be­tween!

When we use a mod­el's poor output to com­mu­ni­cate with an­other person, we foist the burden of com­pre­hen­sion upon the re­cip­i­ent. The time we save draft­ing is time they spend de­cod­ing slop. If we're aware of this, we might in­stead spend much of that saved time man­u­ally clean­ing up the slop before it goes out.

Because anyone fa­mil­iar with AI prose can spot it, AI-as­sisted writ­ing has become dé­classé. When read­ers sus­pect AI in­volve­ment, their opin­ion of the ma­te­rial plum­mets. They assume less effort and hold it in lower regard. Their judge­ment often does­n't stop at the con­tent: the author falls in their es­ti­ma­tion too. Even if the point is great! So there's a cog­ni­tive tax and a trust tax.

I'm a per­fect ex­am­ple of this: I use AI every day and I'm a huge fan (if you didn't know al­ready!). I also know that, in most ways, fron­tier models are leagues smarter than me. Yet when I read an ar­ti­cle and get even an inkling that AI wrote it, I have a vis­ceral re­ac­tion and im­me­di­ately think less of both the ar­ti­cle and its author. I see this as a model fail­ure:

A model's job is to translate the user's intent into their desired outcome.

If the de­sired out­come of your star­tup's land­ing-page copy is to win cus­tomers, and they balk at the AI copy, then that's a fail­ure. And if Fable 5 writes worse copy than Haiku 3, then that's a model re­gres­sion!

The Obvious Intervention

Frontier models are great at fol­low­ing in­struc­tions now. Surely, if we can easily iden­tify AI slop, we can simply tell the model to rewrite its prose until it's no longer shit, right? That's the premise behind 'writ­ing skills'. Many exist, and indeed I have my own. Each is a com­pendium of an­tipat­terns, ref­er­ences and rules that a model must follow. Mine is almost like a linter for the English lan­guage: if the pas­sage does­n't pass CI, it does­n't get pushed!

But I would­n't be writ­ing this ar­ti­cle (by hand lol) if 'writ­ing skills' worked. Writing skills are in­her­ently sub­trac­tive. They forbid overt model tells. I do find that prose run through my writ­ing skill is gen­er­ally cleaner, more con­cise and slightly less rec­og­niz­ably AI. But they are about rewrit­ing and edit­ing. They cannot create some­thing that was never there.

In fact, I find that some­thing funny hap­pens: by the time a pas­sage passes my writ­ing skill, I can no longer put my finger on why I don't like it. It no longer has ob­vi­ous AI tells. I can no longer write rules to ex­plain what to do dif­fer­ently. I've run out of ways to ex­press why I don't like it, but I don't!

Poor model com­mu­ni­ca­tion is a huge prob­lem. It af­fects our abil­ity to align models with our intent and push them to their po­ten­tial. It af­fects our abil­ity to over­see them and make sure they're be­hav­ing. And as AI me­di­ates more of our con­ver­sa­tions, it rends the fabric of our social con­tract: in busi­ness and in per­sonal re­la­tion­ships. Something must be done, and we're gonna need more than a SKILL.md!

Why are LLMs getting worse?

Transformers are like a sand­wich: the bread is human lan­guage, but the fill­ing is math. On one side, the input text is split into tokens and em­bed­ded as vec­tors. Inside the trans­former, the text is gone: the model sees only the em­bed­dings that rep­re­sent it. This is why Mythos can't simply count how many Rs are in 'straw­ber­ry': those em­bed­dings may encode the word's mean­ing, and even knowl­edge of its spelling, but there are no lit­eral let­ters to count with­out first re­con­struct­ing them or call­ing a tool. The model trans­forms those vec­tors through layer after layer to pro­duce a prob­a­bil­ity dis­tri­bu­tion over the next token. It sam­ples from that dis­tri­bu­tion, de­codes the chosen token back into text and re­peats: the other side of the sand­wich.

When we're train­ing models, we keep op­ti­miz­ing the fill­ing. The loss may be cal­cu­lated from lan­guage tokens, but gra­di­ent de­scent op­er­ates on a high-di­men­sional pa­ram­e­ter space of ma­tri­ces and vec­tors.

Post-train­ing often con­cen­trates prob­a­bil­ity mass on re­sponses that score well against ca­pa­bil­ity and align­ment ob­jec­tives. This makes models more ca­pa­ble, but not nec­es­sar­ily better com­mu­ni­ca­tors. In one com­par­i­son of two model gen­er­a­tions, newer in­struc­tion-tuned models showed less syn­tac­tic and, es­pe­cially, lex­i­cal di­ver­sity than older non-in­struc­tion-tuned models. Another study found the trade-off di­rectly: RLHF im­proved OOD gen­er­al­iza­tion over SFT while sig­nif­i­cantly re­duc­ing output di­ver­sity.

Output di­ver­sity is not in­trin­si­cally good: losing di­ver­sity can mean re­mov­ing errors, pri­or­i­tiz­ing good an­swers and in­creas­ing model align­ment. But nar­row­ing also makes a mod­el's ca­dence and struc­ture more pre­dictable, leav­ing a fin­ger­print that in­formed read­ers rec­og­nize and ma­chines can mea­sure. One recent clas­si­fier at­tributed text to one of six model fam­i­lies using lex­i­cal, syn­tac­tic and struc­tural fea­tures. According to Pangram, these pat­terns sur­vive rewrit­ing, trans­la­tion and other AI-as­sisted edits. You can ban cer­tain tokens, but the un­der­ly­ing token dis­tri­bu­tion re­mains. And that dis­tri­bu­tion has been col­laps­ing as fron­tier models im­prove.

Dumb and Dumber

As post-train­ing datasets grow, it be­comes in­tractable for humans to write or judge every ex­am­ple. Luckily, models are pretty good these days, right?!

Much of model post-train­ing now relies on other models to syn­the­size data, pro­vide pro­cess su­per­vi­sion and judge out­puts. This cre­ates a feed­back loop where models prefer out­puts that adhere to their own dis­tri­bu­tion. Familiar forms beget fa­mil­iar forms: models favor what they would've writ­ten them­selves and what they deem good writ­ing. As judges, they suffer from the same issue as our ear­lier writ­ing skill. Take GRPO driven by an LLM judge, for ex­am­ple: we can give a fron­tier judge model a list of rules to pe­nal­ize, and it is per­fectly ca­pa­ble of en­forc­ing them. But when it makes those pat­terns less likely, they are not nec­es­sar­ily re­placed with higher-en­tropy, more cre­ative out­puts! The un­der­ly­ing dis­tri­bu­tion hasn't just shifted, it's shrunk! (I promise that sen­tence con­struc­tion was ironic.) The judge model simply does not have taste that the model being trained can dis­till.

I gen­uinely be­lieve this is a core ten­sion in fron­tier model de­vel­op­ment. I've spoken to people at fron­tier labs who main­tain that model prose is simply an align­ment prob­lem. My gut dis­agrees: model in­tel­li­gence might be in­trin­si­cally linked to slop output. But I also don't think we should de­pri­or­i­tize in­tel­li­gence or ca­pa­bil­i­ties for the sake of the dis­cern­ing reader. So what can we do?

An Expensive Compiler

Earlier this year, I gave up. I hate read­ing AI slop and didn't want to in­flict that feel­ing on anyone else. This wasn't pure mag­na­nim­ity: I also could­n't trust it to com­mu­ni­cate my intent to others with­out either weak­en­ing my ar­gu­ment or risk­ing my re­la­tion­ships and rep­u­ta­tion. But I still wanted to rely on AI to brain­storm, reason and re­search before writ­ing it myself! Not be­cause I'm lazy (well I am, but that's not why!), but be­cause it's gen­uinely good at this. This is not an easy circle to square, be­cause once you have the AI output, it's fin­ger­printed!

My idea to solve this ten­sion was as fol­lows: work with the model as normal to ideate and reach a con­sen­sus, then make it ex­press the result as a struc­tured brief that was as terse and se­man­ti­cally dense as pos­si­ble. In GCSE English, we learnt PEEL: Point, Evidence, Explanation and Link. I always thought it was rigid and stupid: I could­n't pic­ture Hemingway PEELing his para­graphs! It hadn't served me in GCSE English, but when I needed to com­press model output as far as pos­si­ble, I landed on PEEL in cave­man English.

I am an ex­pen­sive com­piler. I take this brief, which con­tains the dis­tilled essence of the ar­gu­ment I wish to ex­press, and I com­pile it into prose. I con­struct a nar­ra­tive. I weave to­gether the input with a nat­u­ral rhythm. Most im­por­tantly, I apply my taste: I choose what to pri­or­i­tize and where to add humor. Much as I did when writ­ing songs, I get to think about the emo­tional re­sponse I want to elicit from my reader, sen­tence by sen­tence! This system works great. But like I said, I'm a little lazy. I also work in AI, and when you're a hammer, ev­ery­thing looks like a nail! So of course there's a part of me that sees this prob­lem as fol­lows:

frontier-authored PEEL brief expensive compiler (me?) sexy prose

Let's train a god­damn model, baby! But before I get too ex­cited, I should ac­knowl­edge the Sisyphean task I've set myself. I've de­scribed myself as the ex­pen­sive com­piler: an ar­biter of taste and a con­struc­tor of nar­ra­tive. I try to read my work through my read­ers' eyes. These are dis­tinctly human fac­ul­ties. Have you ever seen an LLM with a sense of humor? How the hell do you teach a model taste?!

Madison

I'm not a sci­en­tist, and this is an opin­ion piece, so let's get even more sub­jec­tive! To my mind, Taste == Love. Whether we cul­ti­vate our taste de­lib­er­ately or ac­quire it pas­sively, I think of it as the weighted sum of our emo­tional re­sponses to every input we en­counter through­out our lives. Each input leaves an im­pres­sion with its own weight and rate of decay. When I hear a song I like, it leaves an im­pres­sion on my taste. When I read a pas­sage I hate, that in­forms my taste. When I fall in love... you can almost pic­ture the weights shift­ing in­ter­nally, much like a mod­el's pa­ram­e­ters during train­ing. Continuous learn­ing, baby!

But the learn­ing signal for this in­ter­nal taste model comes from our 'heart': our emo­tions. And ma­chines don't have that! Instead of trying to give Sol a Soul, I'm going to accept that and try to work around it. Maybe it would­n't matter that the model can't have taste if it could mimic a snap­shot of mine per­fectly.

When I was in high school, I dis­cov­ered Mad Men. To this day, it's my all-time fa­vorite TV show. It left more than just an im­pres­sion on me: I binged Ogilvy's books. My first in­tern­ship was at JCDecaux. My first part-time job was as a copy­writer (I'm rusty, I know!).

Through in­tro­spec­tion, I dis­tilled my own taste into a set of char­ac­ter­is­tics: good writ­ing is, in my opin­ion, cre­ative, per­sua­sive, in­spir­ing and un­pre­ten­tious. Most im­por­tantly, it com­mu­ni­cates clearly. Advertising copy from the 1960s and 70s em­bod­ies these char­ac­ter­is­tics. A san­i­tized (read: de-misog­y­nized) corpus would supply most of Madison's con­tin­ued pre-train­ing data and su­per­vised fine-tuning ex­am­ples.

David Ogilvy's 1959 Rolls-Royce advertisement, led by the headline about the electric clock being the loudest noise at sixty miles an hour.

I de­cided to call my model 'Madison' after Madison Avenue; the female name also fit be­cause I orig­i­nally fine-tuned Gemma (but I do not rec­om­mend Gemma!). Madison is now fine-tuned from a Qwen base model. Crucially, I'm not op­ti­miz­ing for gen­eral in­struc­tion fol­low­ing, which is why I use a base model rather than an in­struct model. Instead, I'm teach­ing it one narrow task: take a PEEL brief and com­pile it into cre­ative, high-en­tropy, limpid prose.

When I'm work­ing on a prob­lem with a model, I find it useful to ask, 'Could I do this with­out a model?' I was par­tic­u­larly in­ter­ested in struc­tural en­tropy: after a basic CPT and SFT, I found that Madison pro­duced sen­tences that varied more nat­u­rally than those of fron­tier models. It made better vo­cab­u­lary choices and had more per­son­al­ity. But on a para­graph level, it was still too pre­dictable. What if I could inject cre­ativ­ity into the para­graph scaf­fold­ing al­go­rith­mi­cally, then get Madison to fill in the gaps? This has shown re­mark­able po­ten­tial so far.

The last chal­lenge is the hard­est: who grades the output? I know it has to be me. Madison's ob­jec­tive is to write well, and in this con­text, 'well' means ac­cord­ing to my taste! So the only eval that mat­ters is this: do I love it? That's not easy to ex­press in Python! My hy­poth­e­sis is that Madison can only cap­ture my taste if I give every piece of feed­back myself. Oof. That's a lot of time when your com­bined GPU rental and por­ridge oats budget is ~$100!

The Split

If Madison can com­pile a slop brief into good English, then this opens up an­other op­por­tu­nity: in­stead of con­strain­ing our fron­tier models to out­putting some­thing that re­sem­bles human-read­able prose, we can set them free and op­ti­mize them purely for their in­tel­li­gence and token ef­fi­ciency, tics be damned! If a small, spe­cial­ized Madison-class model can take their de­gen­er­ate slop and trans­late it into nat­u­ral prose, we get better com­mu­ni­ca­tion with our agents while fron­tier rea­son­ing models can be post-trained even more ag­gres­sively.

To be fair, I doubt I'm the first person to sug­gest this split. And we know ver­sions of it al­ready exist in pro­duc­tion: for ex­am­ple, in the Claude app, the dis­played 'think­ing' traces are ac­tu­ally sum­maries of the mod­el's true rea­son­ing by a small model. I don't know whether that model is specif­i­cally trained for the task, but it's not ex­actly Charles Dickens! Just be­cause a model is trained for a tightly cir­cum­scribed task, such as com­pil­ing briefs into English in Madison's case, does­n't mean it won't suffer from dif­fer­ent symp­toms of the same fail­ure modes as other models. The split alone isn't the silver bullet, but it does allow me the flex­i­bil­ity to ag­gres­sively fine-tune a model to my taste with­out having to worry about its broader in­tel­li­gence or ca­pa­bil­i­ties. From what I've seen, this can be pushed very far, and could get Madison as close to being a great copy­writer as pos­si­ble with­out ac­tu­ally making it an al­co­holic wom­an­izer (if you take one thing away from this ar­ti­cle, watch Mad Men!).

What's Next?

This is a gen­uinely ex­cit­ing prob­lem to work on, and it feels urgent. If the qual­ity of fron­tier model text output con­tin­ues to de­grade, we may be forced to make dif­fi­cult de­ci­sions: de­cel­er­at­ing, com­pro­mis­ing, sac­ri­fic­ing our abil­ity to mon­i­tor them. They'll also gen­er­ate far less eco­nomic value than an­tic­i­pated in many of the in­dus­tries we're count­ing on them to trans­form. We in­vented LLMs to serve us. As long as we don't pa­per­clip fac­tory our­selves, we're going to be living with some­thing ap­proach­ing su­per­in­tel­li­gence, and it needs to speak our lan­guage(s), be­cause I don't speak fluent linear al­ge­bra!

I'd love to talk with other people in­ter­ested in this prob­lem, so feel free to email me at sean@mozeik.com. Thanks for read­ing, and thanks to ev­ery­one whose ear I chewed off talk­ing about this!

Worth passing on? Share on X