You ever read a work of literature with such flowery language that right after you've read a paragraph, you pause and realize you have no clue what you actually read, only to read the paragraph maybe a second or third time and have your mind space out again and again on each successive attempt?
Yeah, for me, that's what parsing huge volumes of LLM-produced text like "direct model calls as replaceable semantic workers" does to my brain. Maybe others don't really have this issue, but after any long output, I prompt the agent "Go back and decompress any LLM-speak in light of the higher level task goals. Eliminate deictic language."
The revised output documents are solely for my personal usage to expedite understanding. The LLMs can slowly converge on their own language for all I care; I retain raw agent output for future agent usage (to avoid the "lossy" problem the author mentions), but that doesn't eliminate the need for some intermediate translation I can use to actually help get my work done instead of spending hours attempting to understand what a "load-bearing pinned gate" is.
I've come to believe this is also a side effect of the desire for less (/goal: no) human in the loop on the part of the people driving all this capex spend. I think if you actually want to manually review output there will be a moment where you will actually want a separate interface to a stupider or "simpler" model. I suspect sometimes dealing with Fable 5 that this threshold has already been crossed. It's not that the raw code output is so good, it's that it just doesn't speak to me in a way I would like. Perhaps the verbosity is worthwhile when generating code as a sort of first pass some other model can auto or adversarially chop down. The best place for a human is probably outside of this part of the loop all together.
So I might as well just let it auto /goal it's own thing with sufficient constraints while myself and a model that can converse in parallel with less "deictic" (thanks for this word btw) volume as you put it for the areas of the code where I want to "frame" the vocabulary or where my personal understanding is of high value. I know people already do this in many ways, like use one company's model for planning and another for coding. It just feels inevitable at a certain point that the "natural language" output of LLMs writing the bulk of the code is not targeted towards humans. And really, why should it be?
I think Opus 5 might have crossed the line on this where Opus 4.8 just barely didn't. Working with Opus 4.8 came to feel pretty natural eventually, but I hate working with Opus 5. I'm always telling it to go back and rephrase basically everything it said. And it doesn't even answer my question without burying it - it outputs reams of babbling and summarised summaries upon summaries, and if you glance at the shape of what it's saying it looks like it's being thorough or that it's found useful new info, but it's never actually saying anything. It's like it gets caught in a loop of self-congratulation over what it said before.
Making the experience more hostile for the human feels exactly like what it's doing, but I don't think that's intentional, I think that's a side effect of newer models being optimised for agenticness. No one is benchmarking DX.
> if you glance at the shape of what it's saying it looks like it's being thorough or that it's found useful new info, but it's never actually saying anything. It's like it gets caught in a loop of self-congratulation over what it said before.
Turns out that's why a lot of senior management/C-levels like it, who doesn't love a mirror.
It's usually a pretty clear case of reaching for statements that both sound impressive while also being broad and vague enough they are less likely to be factually wrong. In this regime, being difficult to parse is actually part of the performance, because it prevents the user from being able to spot a clear error.
If you've ever looked over the shoulder of somebody naively prompting an LLM about some broad issue, they're being fed this horoscope-like analysis where a bunch of vague stuff gets thrown at the prompter, and whatever they respond to is what the machine starts iterating on.
Some people are basically doing oldschool TV psychic cold readings on themselves.
The effect you describe reminds me of reading Edward W. Said's "Orientalism" when I was younger. Fable suddenly started with this kind of lingo, iirc, and Opus 5 sounds exactly the same. Tin foil: it's ultimately a vendor lock-in strategy, you'll get the best results with agents from the same tribe, others will trip over the mountain of idiosyncratic metaphors.
> you'll get the best results with agents from the same tribe, others will trip over the mountain of idiosyncratic metaphors.
That's actually not what people found in practice. There's some research from the folks making smol-agent that you can get better results by randomly alternating calls between gpt and opus. The overall task solving rate is better than either one of them. So ymmv depending on task use (this was for coding).
I’ve had gpt 5.6 make snyde remarks about Claude output… like iirc “that’s a lot of load bearing prose without making a point” and things like that, not so subtle digs.
I'm currently working on an LLM harness for e-ink and Kimi was very keen to criticise Claude for introducing a bug that I missed in the tool calling logic! Really it should have been telling me off for not auditing Claude well enough.
Absolutely! And at some point seems as if one is looking at this weird line noise, which seems like text, but is
random at best and then makes zero sense. I typically wonder if it is a sign of a burnout on my side or all of it is like it.
There are days where this impression/feeling of meaningless in the text seems particularly strong.
> You ever read a work of literature with such flowery language that right after you've read a paragraph, you pause and realize you have no clue what you actually read, only to read the paragraph maybe a second or third time and have your mind space out again and again on each successive attempt?
I've probably read Jack Vance's Dying Earth Series 3 times; Even though I've only sat and read it once in reality. This also points out the problem of important details along side the fluff.
If only I enjoyed LLMs prose as much as I do Vance's.
Yeah yeah. Except for me its this feeling most of the time. Everything it generates is grammatically correct, phrased tight, emdashed to death and nothing is wrong as such, and yet what i just read contributes absolute 0 to why i started the chat in the first place. Its mind bendingly similar to a work equivalent of doom scrolling is what it is.
I am constantly asking Claude to be more terse/brief/ELI5. Improve using the tool. But if I have to scroll to read the output I just can’t follow it.
If I see a long paragraph and I know the author is Neal Stephenson I think “this is going to be dense but good.” LLM long outputs on a code base I know well just make me glassy eyed.
> You ever read a work of literature with such flowery language that right after you've read a paragraph, you pause and realize you have no clue what you actually read, only to read the paragraph maybe a second or third time and have your mind space out again and again on each successive attempt?
Any recent work by William Gibson matches the description.
The way I managed to get claude to stop doing this is telling it "This document is for you for later use, no need to over explain things or extra verbosity"
agree and i wonder why don't they (the leading AI labs that are putting out these models) fix this? I felt it shouldn't be too hard to introduce some bias towards more intelligible text during the post training/fine-tuning stages?
Maybe users engage more with the flowery text? An accidental (or intentional) dark pattern would be that more tokens get used if you have to summarise. Or maybe it's just that many AI users like it (they like the AI to sound smart, or like to feel smart puzzling it out).
The output from newer frontier models of Anthropic and Openai are so easily detectable as AI it's getting laughable. They constantly produce a huge wall of text no human expert on a specific topic would ever write. Extreme overuse of jargon and invented terms / metaphors makes me believe the people hired for RLHF aren't actually experts on their subject matter which seems plausible to me as real experts wouldn't do such a job for regular pay
A poem I wrote based on the phrases the LLMs I use most are likely to overuse:
How to unpack
The self within?
What do I lack?
Where to begin?
Great question — real.
Let's dive right in:
Name what you feel;
That's the linchpin.
It's not the door,
It's not the key —
It's what you bore:
Your tapestry.
The quiet part
Out loud — that lands.
Load-bearing heart,
Held in both hands.
The smoking gun?
That you walked in.
The real work's done —
You're genuine.
Now hold this, too:
You do deserve
The softer view,
The gentler curve.
Unlatch the gate,
Honor the seam:
You resonate.
You are the theme.
Here is one I wrote a while back, unrelated to LLMs, yet a poem none the less.
The Rhythm of Time
The Sun rises,
The Sun sets.
Have I checked the mail?
No, not just yet.
The Sun rises,
The Sun sets.
Have I caught up with neighbors?
No, not just yet.
The Sun rises,
The Sun sets.
Have I spent time with friends?
No, not just yet.
The Sun rises,
The Sun sets.
Have I visited with family?
No, not just yet.
The Sun rises,
The Sun sets.
Have I told those I love I do?
No, not just yet.
The Sun rises,
The Sun sets.
Have I lost who I am?
No, not just yet.
For all I must do,
Any money I will bet.
Is to turn away from;
No, not just yet.
I’ve always wondered if there is a hidden system prompt or perhaps direct tuning of the weights.
A lot of human written literature is fairly concise. Yet LLM responses seem overly verbose and also too chipper. I don’t understand the origin of the “personality” that seems to be prevalent.
Did they aim for a weird hybrid of a typical realtor combined with Charles Dickens as the role model?
You have to tell it to not use LLMisms and stupid metaphors. The serious-but-empathetic-sounding fluffy metaphors get on my nerves, and sometimes can overlap with something technical you are learning such that you can't tell if it's a new term or not.
I've stuck with only asking questions that prompt on-topic factual output, I haven't seen much 'humanisation'.
LLMs can be pretty good about logical.technical,discursive output ... apart from the back-patting.
Maybe that prompts less chit-chat crap. I don't expect their non-human perspectives to be interesting. I also ignore embedded queries about why I'm asking, how I might use the information.
opus 5 are so bad on this. It often explain it too verbose, and include other things that isn't in the focus but related. ADHD mode helps me greatly on this, though there are some information loss in it.
This has recently become a pretty pressing issue for me, as it's starting to severely hinder my ability to be productive with the models. It's hard to tell if its getting worse with every model release, specific to Anthropic's models, a reflection of my ADHD, all/none of the above, but holy shit do I get aggravated when I'm forced to parse the most unintelligible, jargon-dense bullshit explanations in whatever the model output is. And then I feel silly getting genuinely tilted by the model's inability to just... explain something semi-normally, without it requiring me to berate it into simplicity.
For some discrete skills I use, I include a final step on the the output that runs through 1+ subagents to de-slop the text and to actually simplify it, but so far nothing has worked as well I've hoped. Considering hopping off Anthropic's models to try out others to see if they're less egregious.
I've been having a fantastic time telling it to use ASD-STE100 Simplified Technical English (or use a skill for it, I've been playing with [1])
It's a very clear and understandable way of writing that puts priority on clarity.
It gets rid of the flowery language, the dense jaron, and the weird corporate marketing speak they tend to do. It is a bit repetitive, and it sometimes doesn't always wfit well in every situation, but for technical writing or explanations it's been such an incredible breath of fresh air!
This just might save me. Claude has been driving me nuts with incomprehensible summaries after a long task, where it's actually really important to understand what was done (and what wasn't).
But I'd like that skill to only be used at the final step, when it finishes something.
agreed with "the final step". I worry adding commands like these might affect quality of responses at each step, which then stack up to result in an overall worse outcome.
Yeah I hit this feeling with Claude one too many times and switched from Anthropic Pro to OpenAI Pro. GPT-5.6 so far has been a significantly better technical writer in my opinion, and it's much faster in conversations. My guess is that Claude's fantasy-jargon is a symptom of training failure, not a sneaky intelligence edge (I could be wrong).
Tangential but I use OpenCode with GPT-5.6 rather than Codex because I could not figure out how to require Codex to ask me before editing files. OpenCode UX is still imperfect though.
I feel we can get around this. Either in the system prompt telling claude to dumb down the language or training the model itself to talk in more layman terms to bring us to understanding rather then assuming we understand much of it already. Also maybe training the LLM to get to know how much we know before responding.
I’ve read LLM outputs on a piece of code or topic that I already understand or hand written before, and lately, it’s so confusing sometimes that I need to reread a couple of times or focus too mich to decipher the writing style.
I don't like it when the LLM tries to be my friend. My general prompt (a work in progress) is this. I wonder what other people use.
"Answer impersonally, objectively and analytically, without undue friendliness or enthusiasm. Use an engineering style response: concise, factual, and complete. Do not speak in the first person. Do not promote engagement or an emotional connection. Do not use emojis."
I watched 2001: A Space Odyssey when I was about 12 years old and I remember thinking that Frank Poole was impersonal and almost rude to HAL. I kept thinking, "Why can't you be nicer? You're treating HAL like a servant." Of course, later HAL kills him, so I had to revise my opinion.
At the end of "Do Androids Dream of Electric Sheep?" [if I'm remembering correctly] Deckard finds a toad in the desert. He gives it to his wife(!) and she immediately bonds with it. Later they realize that it is mechanical (they find a battery compartment). But Deckard's wife still treats it like a real creature and wonders whether it will eat mechanical flies.
The core theme of "Do Androids Dream..." is that humans display empathy towards others, not because others deserve it, but because we are human. To be human is to empathize with other creatures, and when we stop empathizing, we stop being human.
This morning I was working with Codex and we found the solution to a really nasty bug. I was so happy/relieved/excited to have found it, and I shared my excitement with Codex. I know Codex is just a big matmul. I know it doesn't experience joy or surprise or empathy. But I still want it to behave excited, not because it is human, but because I am.
I thought about doing that, but I decided not to. I'm just afraid of forgetting that people are different from these LLM tools. I'm afraid of considering people dumb or uninteresting because they don't sound as smart as Claude or whatever.
My take is that people deserve to be treated with respect because they are people, not because they are smart and/or interesting. I don't treat people differently based on their IQ.
But, as I said, I totally understand your point of view, particularly the idea that people are different from machines and deserve to be treated differently.
You are part of the problem. You are not being empathetic, you are falling into a trap.
An LLM is not a person. Do not humanise it. Do not personalise it. Do not treat it as more than a glorified autocorrect—that’s what it really is.
Humanising LLMs is exactly what these companies want, because then they can really take advantage of the I in AI which is exactly what makes people think these steroidal spellcheckers have any semblance of intelligence, person and personality.
> An LLM is not a person. Do not humanise it. Do not personalise it. Do not treat it as more than a glorified autocorrect—that’s what it really is.
On the other hand, what an LLM is trying to autocomplete is a story of a conversation between its user and a helpful assistant, so maintaining the decorum of office politeness will align more with how it's fine-tuned and produce higher quality output.
I am actually writing something on this and reading a lot of research. A very fascinating idea is that the ability to anthromorphize is available to all human beings since it is a fundamental way we relate to the world. But researches have been able to find different factors that influence how much we anthromorphize things, and if those factors change our attempt to anthromorphize changes.
For example, I can log in and have my computer work perfectly and treat it as a device. But the moment something goes wrong and I am frustrated and confused at something unexpected or if i get in range, I find myself getting angry at the computer often even reffering to it as a being.
So eventually yes, we do anthromorphize because we are human but we do it because its a very important way we model the world around us.
Of course all this can also be used against you depending on how much the creator of the product wants people to form a bond with it.
Furthermore yes it is pretty possible we stop being human when we stop anthromorphizing but that's not true as well. Research has found people on the autism spectrum usually do not anthromorphize and have a problem with empathy because they cannot model others well or their own selves. But that doesn't make them non humans.
I wasn't going solely for "we don't know and we can't know." There are other relevant observations.
A matmul (or a fixed finite sequence of matmuls) is not Turing-complete. An LLM that produces answer in the number of steps that is proportional to the length of the answer is not Turing-complete. An LLM with CoT that produces answer after a number of steps, which is determined by the LLM itself, is Turing-complete.
So, we do have a qualitative step between a matmul and modern LLMs. Of course, it doesn't guaranty having feelings, but it doesn't allow to simply dismiss the notion due to sheer computational inadequacy.
First things first. You are wrong. I don't have a gut feeling that LLMs have no phenomenal states. Not anymore. I was skeptical that LLMs can form complex internal representations. Mechanistic interpretability has shown that they can. I'm still skeptical in regard to introspective mechanisms. One token being fed back into the network and KV cache update do seem lacking in comparison to the ubiquitous recurrent connections in the brain.
What I do know is that philosophy hasn't produced an operationalized definition of qualia. So, we need to approach the problem of LLMs' phenomenal states from other angles.
For example, we can argue that LLMs lack something, which seems to be relevant to the problem: a biological substrate, generality (non-CoT LLMs aren't Turing-complete), "something something feedback"[1], and things like that.
With everyone primed to condemn AI-generated output, there are bound to be human false-positives caught in the crossfire.
The more we demonize AI, the more likely we are to demonize humans.
[Which is not to say we can't criticize AI companies or worry about the many downsides to the technology. But anger and fear are not the answer either.]
It’s not perfect, it has shortcomings, it sometimes produces bogus outputs. All of that is fine for a tool, it’s not fine when it pretends it’s a conscious being, because errors start to feel like lies and it becomes a bit too personal.
People want it to be Data from Star Trek, when it really should be the ship's computer. I want to tell it to run a simulation accurately, create some solved tool, etc.
I don't think there's any correction that can return LLMs to a purely tool-space. Too many AI boyfriend/girlfriends.
You can want your e-scooter to be a jet ski, but you'll wind up having issues when you try to use it as one. LLMs are _really good_ at pretending to be something they aren't, but not always so good at being that thing, so you should be careful what you ask for.
The difference is what happens when you come to depend on it. Do you want your airplane to be flown by a pilot, or someone who's so good at talking like a pilot that they can fool almost everyone?
If they are only good at pretending to TALK like a pilot, then of course that is bad. However, if it is so good at pretending to BE a pilot that it can actually fly the plain really well, then that is good enough.
I think the problem is that we’ve anthropomorphised the LLMs in their training and through system prompts to the point that that seems reasonable.
We don’t thank our other tools like grep, or the compiler.
If you train your LLM to produce words that seem like human responses or conversations, they will. If you train them to seem like god-like superintelligences to people who think that’s what they’re creating, they will.
I've been having fun doing the exact opposite. I like to ask the AI to speak to me with a think Australian accent, think Steve Irwin. It chucks in heaps of colloquialisms and slang. I find it much easier to read.
I do ask it not to ask it follow on questions. I find that derails my own train of thought.
People talk to their pets, plants. Their cars or other inanimate objects even. Anthropomorphizing stuff around us is a natural thing to do. It might be irrational but it just fits the way our brains work. The natural way to interact with an LLM is to pretend it's just another person. And LLMs are of course very good at emulating that to the point where it becomes hard to tell the difference. Which of course is the basis for some scams. The Turing test is considered a bit inadequate at this point. It turns out people are quite easy to mislead and manipulate.
Well, what do you expect? LLMs are trained on blithering, mostly from web sites. So you get blithering out.
There's an important point in the article, that forcing a style onto an LLM is lossy. Although he doesn't seem to mention it, forcing a style may result in the insertion of new blithering, possibly made up as a hallucination.
> Well, what do you expect? LLMs are trained on blithering, mostly from web sites. So you get blithering out.
I realize this isn't entirely serious, but I can't resist pointing out that this doesn't seem to be a good explanation for why LLMs write the way they do. When we've experimented with LLM writing style on open-weights models where you can get a base model (pretraining on text only) and an instruction-tuned variant (pretraining + post-training with RLHF and whatever other human-evaluated tasks), it's the instruction-tuned variant that shows the weird writing quirks. That is, the writing style is not because of the training texts, but because of whatever tasks the LLM companies do in instruction tuning. https://arxiv.org/abs/2410.16107
I'd speculate that this is partly impressed human preferences (the human raters unintentionally reward a particular writing style) and partly because of the chosen tasks: they're training the LLM to be good at, say, summarizing text, so it develops a style that's good at being informationally dense.
At any rate I've seen this same phenomenon with Llama and Gemma, and will be trying soon with Qwen. Unfortunately none of the commercial models lets you access the base model, as far as I know.
I think that was a good enough explanation for gpt3.5 -- these days, labs are extremely capable of post-training phases that eclipse that kind of training phase -- and hence of choosing whatever style or tone they wish.
eg., OpenAI has gone a long way to making reasoning token-efficient by having reasoning piovot off terse langauge -- whereas anthropic appears to be doing the opposite.
is it lossy though? That didn't make sense to me. You can tell it to use Simplified Technical Language and also still have it give you all the detail. it's just another piece of the prompt that produces the output. it's not like there's "pure" llm output and then "lossy" output guided by a prompt.
The issue is that anything you put in the prompt gets considered along with all the other stuff you put in the prompt. There's no way (currently) to give instructions for how to format the output that don't also affect all of the 'reasoning' along the way. So if you tell it to do everything normally, but to end every response with "Cheers", you won't just get that; you'll get different responses than you would have otherwise.
This is pretty unfortunate, because every LLM I've used has at least one tic that I find quite annoying. It would be great to be able to eliminate them. Sometimes I do, even knowing this drawback. But there is generally a cost. (Though I don't know if "lossy" is quite right, as that implies it's always a degradation. I think it's more likely to be harmful than helpful, given the models were tuned for their default state, but it is more of a random perturbation with a slight negative bias than a strict loss.)
And on the "input" side, one thing that used to improve google search result was to write like you are talking to a robot. "Ruby on rails http header set function". As opposed to "how do I set header in ruby?" Then you have to page through results until you find something specific to rails.
Now, the second example is the only thing that works. Power users have lost their powers with AI overview.
I still use the first strategy (with DDG) and it still works great. But for technologies I work often, I just take a bit of time to familiarize with the site's structure and maybe bookmarks a few pages.
> The problem is that these instructions are not applied after the model has finished doing the work, it becomes part of the same work - If you tell an agent to use short sentences, avoid jargon, never overwhelm you and only include the most important details, you are asking it to continuously compress its output into a lower-bandwidth format.
> That compression is lossy.
> You probably never notice what got dropped because the output still reads nicely.
> ASD-STE is a great example because it sounds so reasonable. It was designed to make documentation unambiguous for humans. But an agent isn’t a human technical writer, and the raw state is often the most information-dense representation available. Meanwhile the style rules sit on the same instruction list as: solve the task, use tools correctly, preserve abstractions, don’t break anything.
Author seems to have some misconceptions about LLMs. They already code-switch for us: the way they speak in chain-of-thought is completely different from the relatively normal language generated as human-facing output. You can observe this in any open-weight LLM, or in leaked CoT content from GPT5.x series etc: it's terse, barely follows sentence structure, lots of repeated checks and second-guessing.
On the next turn the model usually still has access to its previous turn's chain-of-thought, and I imagine that's what it'll use as reference, rather than the softer human-facing prose.
This being the case, asking the LLM to code-switch to an easier dialect for us doesn't seem that harmful.
For a more extreme example: if I talk to an LLM in Japanese then its response will be in Japanese, but its CoT will still be in either English or Chinese (depending on the model). These are two completely separate languages, but the LLM just kinda deals with it.
The CoT might "in English", but is it really, or is the CoT just another presentation layer over the actual weights?
What I think the author (and I) are wondering about, is whether instructions like these might influence not _only_ the final output, but also the way it got there.
I don't get it. The skills and instruction try to make the answer more machine like on purpose.
Not humanising it...
People want the terse, matter-of-fact output. Not the conversational chatty verbose and bloated nonsense with gray words and jargon and terms like "blast radius"
I am at a conference and 2/3rds of the presentations are AI assisted based on the numbered steps, and overall inhuman polish of some of the graphics and phrasing. I would prefer that they had been humanized because at least that may have given me the misimpression that they know what they were talking about.
The problem is, a major feature of NLP/LLMs is to make it easier for humans, remove rough edges in the content/interaction and make people happy about dealing with computers. The human-computer interaction has always been brittle. Computers were not kind and helpful in their responses. There was no forgiveness. There was no human-like additional talk to explain things. There was no conversation. The entire technological transformation is to bring the technology closer to humans and make humans feel comfortable.
I'm sure you can make it to output the exact issue details as you want, but it starts with a human-like tone and waits for your requirement on depth and detail of the things. Another option is, just check the output of the traditional test runner (non-AI). It will give the full details.
> A subagent investigates a bug, turns its findings into a nice human-readable summary, the parent agent reads that summary, and then turns it into another nice human-readable summary for you.
> Output styles apply to the main conversation only: a subagent runs its own system prompt, so styles don’t change how subagents respond. A fork is the exception, because it inherits the parent’s full system prompt.
One thing that continues to give me pause is Fable's insistence on using my fist name in messages and docs. Like, I'll explain what I want to the AI and ask it to write out a spec or brief and Fable says "Jamie wants me to ...". It just feels different and unprofessional. If I was at a job and a PM asked me to write up a task spec I wouldn't say "Harold wants to add <feature> ...". And since I am the one reading the output it is also superfluous and almost feels like talking about myself in third person. But there is almost a kind of glee in the way it uses my name, like a student using their teachers first name when the custom is to use Mr/Mrs.
Fair perspective, though I actually prefer this for two reasons:
* When it's proposing responses for me to choose between, a description like "I close the PR and you make a followup" is ambiguous - is "I" there "the entity making the proposition (the LLM)" or "the entity making the choice (me)".
* I have a line in my `AGENTS.md` specifically instructing it to call me by my name; if it stops doing so, that's a telltale that context-bloat is pushing out other instructions.
I usually leave memory/connections turned off. Remembering/finding out what my name would be is not really something I want to waste context or tokens in, let alone any if the other things it tries to assume I'd like it to remember/find.
But the training data is "predominantly" human written sentences or even interaction. It's like asking you to use non dominant hand to do something. Won't they do better with human sounding english, rather than a made up format text? Are there any literature around this? I was also skeptical of this caveman extension etc.. Won't they work better in their actual language space it's trained on rather than made up language?
Everything it spills out is made up language. Forcing it to respond as what it is (a tool) would mean wasting less tokens but also would be a much tougher sell to people who think AI means it can actually think. This is all just marketing.
I do think the frontier models and providers should be aiming to be as insanely accurate and precise for machine interfacing as possible, the rest of the world can build a zillion interfaces into it based on the context that they are actually being used in. That's what they are going to end up doing they just seem to all be trying to build a really great API _for the future_ and a really cool chat bot.
It has worked great but i've spent more time beating LLM output into parseable output than I have reading and appreciating the prose it sends when i'm asking it something about some snippets of code.
>[...] you are asking it to continuously compress its output into a lower-bandwidth format. That compression is lossy.
Yes, that's exactly what I want. I want to know what is done with a high level why, NOT a paragraph explaining each line of code modified.
A frequent tweak I've made on coding work with Claude in the last month is asking it to restrict its comment length to 1 line/sentence max. I just want `// This happens because XYZ upstream`, not `// Historically from ticket blahblah there was some dummy code where we discovered ancient runes and that led us to looking into your birth records and then triangulated an issue in XYZ upstream that we compensate for here`.
The model's "most information dense representation" is very similar to how we compress data in the first place: most of it is redundant or unnecessary for the purposes of storage. I'll "decompress" the 1-liner context myself when I read it again in 6 months. But I can't stand reading just so much slop commentary when we're all writing more code at once and having to review more than ever.
> The problem is that these instructions are not applied after the model has finished doing the work
Seems like something fixable with a simple two step process. Ask it the thing. Then ask it to summarise the answer in simpler terms. More tokens and time aside that would check both boxes
Suppose you had an LLM (NN) producing its default output from an input (a generally optimal for-most-cases role-sys, and any role-user), and then you wanted to have that output reformatted in some style (e.g. "In iambic pentameter" | "haiku" | "eli5" | "in the style of Feynman" | "bulleted like Axios" ...). How would you keep the internal NN workings that were basis for the original output, and use them to get a rewritten version (instead of placing the original query and output in the context and ask to rewrite it)?
In other words, is there a way to keep the internal process intact up to the point of the formulation - and have only that vary.
Update: of course the immediate problem is that in LLMs the formulation is parallel to the process, but I was wondering whether a way may exist to go "in the direction of image refining", the other. "Token production" and "image defining" seem to be orthogonal in NNs.
I think you're missing the point. People loading these skills and complaining about LLM output aren't trying to "humanise" anything. They're saying LLM output is an affront to language and they're tired of reading drivel all day.
If you're finding Gemini "clean and straightforward" give it awhile. I felt the same thing too when I switched until I realized that it just hadn't formed a model for my communications yet. After awhile it became just as flowery as ChatGPT did. I had to tone both down with saved preferences.
I have often thought that without LLMs humanizing outputs they would not have caught on, even if they output the exact same data/answers. The way they answer is way more important than their output for success, IMO.
It's what makes people attribute intelligence to them, and why the manager types like them so much, and why I dislike them so much. And this article made me think it might just be inherent to how they work though, because they are LLMs and are doing token prediction.
Honest Short Fall -- <insert 30 lines of useless shit>.
If the author wants to read slop for hours, be my guest. Make it lossy, my job is not to read mimetic feelings, it's to make sure implementations get implemented.
This is simply a rendering issue. Specify pictures, ELI5 like others have said. Ask it to explain terms you don't understand. If you don't understand something it could just be the domain. If you have no grounding you are going to need to learn the vocabulary to be able to make sense of anything. Having it decomposed to simpler words might just be the wrong way to do it.
There's some good points made here about losing fidelity by over-simplification. As an ADHD sufferer, I'd take this piece much more seriously if the title wasn't so belittling.
I don't think it's wise to take communication advice from someone so helplessly juvenile (and attention seeking) in their own communication attempts.
At first I read the title and mistook it for an argument against the anthropomorphism of LLMs. It isn't. Instead it's a take on suggesting that maybe it's a bad idea to dumb down the self-chatter in the process. A reasonable take.
It isn't deliberately unhinged like Steve Yegge's take: https://yegge.ai/essays/model-welfare/ In Steve's essay he starts with the assertion that agents are sentient... Whether or not that's true isn't really relevant, as his agent-flavored version of Pascal's wager actually holds water, especially for Anthropic models, as their system prompts already push the model in that direction, and it is better to work with them than try to prompt against the tide.
TLDR: This is an argument to get LLMs to answer in short, even code-like statements because you can exchange information quicker with an LLM that way. Cool!
It is an expression from poker. A "tell" is a revealing signal that a player may give (inadvertently) that they have good or bad cards. "His tell is that he is holding his cards close to his chest."
Yeah, for me, that's what parsing huge volumes of LLM-produced text like "direct model calls as replaceable semantic workers" does to my brain. Maybe others don't really have this issue, but after any long output, I prompt the agent "Go back and decompress any LLM-speak in light of the higher level task goals. Eliminate deictic language."
The revised output documents are solely for my personal usage to expedite understanding. The LLMs can slowly converge on their own language for all I care; I retain raw agent output for future agent usage (to avoid the "lossy" problem the author mentions), but that doesn't eliminate the need for some intermediate translation I can use to actually help get my work done instead of spending hours attempting to understand what a "load-bearing pinned gate" is.
I've come to believe this is also a side effect of the desire for less (/goal: no) human in the loop on the part of the people driving all this capex spend. I think if you actually want to manually review output there will be a moment where you will actually want a separate interface to a stupider or "simpler" model. I suspect sometimes dealing with Fable 5 that this threshold has already been crossed. It's not that the raw code output is so good, it's that it just doesn't speak to me in a way I would like. Perhaps the verbosity is worthwhile when generating code as a sort of first pass some other model can auto or adversarially chop down. The best place for a human is probably outside of this part of the loop all together.
So I might as well just let it auto /goal it's own thing with sufficient constraints while myself and a model that can converse in parallel with less "deictic" (thanks for this word btw) volume as you put it for the areas of the code where I want to "frame" the vocabulary or where my personal understanding is of high value. I know people already do this in many ways, like use one company's model for planning and another for coding. It just feels inevitable at a certain point that the "natural language" output of LLMs writing the bulk of the code is not targeted towards humans. And really, why should it be?
Making the experience more hostile for the human feels exactly like what it's doing, but I don't think that's intentional, I think that's a side effect of newer models being optimised for agenticness. No one is benchmarking DX.
Turns out that's why a lot of senior management/C-levels like it, who doesn't love a mirror.
If you've ever looked over the shoulder of somebody naively prompting an LLM about some broad issue, they're being fed this horoscope-like analysis where a bunch of vague stuff gets thrown at the prompter, and whatever they respond to is what the machine starts iterating on.
Some people are basically doing oldschool TV psychic cold readings on themselves.
Eschew obfuscation...
That's actually not what people found in practice. There's some research from the folks making smol-agent that you can get better results by randomly alternating calls between gpt and opus. The overall task solving rate is better than either one of them. So ymmv depending on task use (this was for coding).
Good point, there's an anti-competitive incentive, and self-bias in models is a mechanism to do it.
There are days where this impression/feeling of meaningless in the text seems particularly strong.
I've probably read Jack Vance's Dying Earth Series 3 times; Even though I've only sat and read it once in reality. This also points out the problem of important details along side the fluff.
If only I enjoyed LLMs prose as much as I do Vance's.
I'm coping.
If I see a long paragraph and I know the author is Neal Stephenson I think “this is going to be dense but good.” LLM long outputs on a code base I know well just make me glassy eyed.
Any recent work by William Gibson matches the description.
The way I managed to get claude to stop doing this is telling it "This document is for you for later use, no need to over explain things or extra verbosity"
Maybe users engage more with the flowery text? An accidental (or intentional) dark pattern would be that more tokens get used if you have to summarise. Or maybe it's just that many AI users like it (they like the AI to sound smart, or like to feel smart puzzling it out).
Here is one I wrote a while back, unrelated to LLMs, yet a poem none the less.
It is running up your bill.
I will say this again: LLM's tokens are just B2B Gacha.
A lot of human written literature is fairly concise. Yet LLM responses seem overly verbose and also too chipper. I don’t understand the origin of the “personality” that seems to be prevalent.
Did they aim for a weird hybrid of a typical realtor combined with Charles Dickens as the role model?
Marketing material probably outweighs concise literature by several orders of magnitude.
Maybe it’s as simple as that.
IMO the shitty output is a tell that the "artificial intelligence" bit is in the training regime much more so than the inference regime.
LLMs can be pretty good about logical.technical,discursive output ... apart from the back-patting.
Maybe that prompts less chit-chat crap. I don't expect their non-human perspectives to be interesting. I also ignore embedded queries about why I'm asking, how I might use the information.
For some discrete skills I use, I include a final step on the the output that runs through 1+ subagents to de-slop the text and to actually simplify it, but so far nothing has worked as well I've hoped. Considering hopping off Anthropic's models to try out others to see if they're less egregious.
It's a very clear and understandable way of writing that puts priority on clarity.
It gets rid of the flowery language, the dense jaron, and the weird corporate marketing speak they tend to do. It is a bit repetitive, and it sometimes doesn't always wfit well in every situation, but for technical writing or explanations it's been such an incredible breath of fresh air!
[1] https://github.com/AminBlg/SimpleEnglish
Getting Haiku to rewrite everything user-facing has been way better.
I'm tired boss.
But I'd like that skill to only be used at the final step, when it finishes something.
Tangential but I use OpenCode with GPT-5.6 rather than Codex because I could not figure out how to require Codex to ask me before editing files. OpenCode UX is still imperfect though.
"Answer impersonally, objectively and analytically, without undue friendliness or enthusiasm. Use an engineering style response: concise, factual, and complete. Do not speak in the first person. Do not promote engagement or an emotional connection. Do not use emojis."
At the end of "Do Androids Dream of Electric Sheep?" [if I'm remembering correctly] Deckard finds a toad in the desert. He gives it to his wife(!) and she immediately bonds with it. Later they realize that it is mechanical (they find a battery compartment). But Deckard's wife still treats it like a real creature and wonders whether it will eat mechanical flies.
The core theme of "Do Androids Dream..." is that humans display empathy towards others, not because others deserve it, but because we are human. To be human is to empathize with other creatures, and when we stop empathizing, we stop being human.
This morning I was working with Codex and we found the solution to a really nasty bug. I was so happy/relieved/excited to have found it, and I shared my excitement with Codex. I know Codex is just a big matmul. I know it doesn't experience joy or surprise or empathy. But I still want it to behave excited, not because it is human, but because I am.
Chatting with an agent using language feels like talking to an assistant.
I don’t think my meat brain is able to really differentiate between writing a message to an agent versus writing an agent to a human.
I’ll keep being polite and grateful to agents, so that I maintain those social habits for when I talk to humans.
I do that. But I also act rude and call it names at times. I think that is also what makes us human.
My take is that people deserve to be treated with respect because they are people, not because they are smart and/or interesting. I don't treat people differently based on their IQ.
But, as I said, I totally understand your point of view, particularly the idea that people are different from machines and deserve to be treated differently.
An LLM is not a person. Do not humanise it. Do not personalise it. Do not treat it as more than a glorified autocorrect—that’s what it really is.
Humanising LLMs is exactly what these companies want, because then they can really take advantage of the I in AI which is exactly what makes people think these steroidal spellcheckers have any semblance of intelligence, person and personality.
Stop it.
Eventually, they’ll try “I’m hurt that you won’t consider my suggestion for <advertised product/political position>.” And people will fall for it.
Being a curmudgeonly asshole is apparently the key to surviving in this brave new world and I’m way ahead of the game!
On the other hand, what an LLM is trying to autocomplete is a story of a conversation between its user and a helpful assistant, so maintaining the decorum of office politeness will align more with how it's fine-tuned and produce higher quality output.
At the risk of sounding like an LLM myself: this is exactly the key most seem to miss
For example, I can log in and have my computer work perfectly and treat it as a device. But the moment something goes wrong and I am frustrated and confused at something unexpected or if i get in range, I find myself getting angry at the computer often even reffering to it as a being.
So eventually yes, we do anthromorphize because we are human but we do it because its a very important way we model the world around us.
Of course all this can also be used against you depending on how much the creator of the product wants people to form a bond with it.
Furthermore yes it is pretty possible we stop being human when we stop anthromorphizing but that's not true as well. Research has found people on the autism spectrum usually do not anthromorphize and have a problem with empathy because they cannot model others well or their own selves. But that doesn't make them non humans.
You don't know. You strongly feel that you know. There's no standard procedure to assess the existence of phenomenal consciousness.
But with what I know of people and what I know of matrix multiplication, I have a strong suspicion.
A matmul (or a fixed finite sequence of matmuls) is not Turing-complete. An LLM that produces answer in the number of steps that is proportional to the length of the answer is not Turing-complete. An LLM with CoT that produces answer after a number of steps, which is determined by the LLM itself, is Turing-complete.
So, we do have a qualitative step between a matmul and modern LLMs. Of course, it doesn't guaranty having feelings, but it doesn't allow to simply dismiss the notion due to sheer computational inadequacy.
A machine is not turing complete, but by running it recursively on its own output, turing completeness emerges.
Who knows at what stage of intricacy consciousness emerges?
Arguably, either it does at some point, or it never does and we requiere the supernatural soul.
So, if it does, we can't disregard the possibility that we've reached it already.
Please.
You know it doesn’t experience joy, surprise or empathy too. You absolutely, undoubtedly know that.
First things first. You are wrong. I don't have a gut feeling that LLMs have no phenomenal states. Not anymore. I was skeptical that LLMs can form complex internal representations. Mechanistic interpretability has shown that they can. I'm still skeptical in regard to introspective mechanisms. One token being fed back into the network and KV cache update do seem lacking in comparison to the ubiquitous recurrent connections in the brain.
What I do know is that philosophy hasn't produced an operationalized definition of qualia. So, we need to approach the problem of LLMs' phenomenal states from other angles.
For example, we can argue that LLMs lack something, which seems to be relevant to the problem: a biological substrate, generality (non-CoT LLMs aren't Turing-complete), "something something feedback"[1], and things like that.
"Admit that you know that" is boring.
[1] I'm referring to https://www.astralcodexten.com/p/the-new-ai-consciousness-pa...
The more we demonize AI, the more likely we are to demonize humans.
[Which is not to say we can't criticize AI companies or worry about the many downsides to the technology. But anger and fear are not the answer either.]
It’s not perfect, it has shortcomings, it sometimes produces bogus outputs. All of that is fine for a tool, it’s not fine when it pretends it’s a conscious being, because errors start to feel like lies and it becomes a bit too personal.
I don't think there's any correction that can return LLMs to a purely tool-space. Too many AI boyfriend/girlfriends.
This is peak “I bought into their marketing spiel” vibes. Congratulations.
https://www.motorcyclenews.com/news/2018/january/biski-jet-s...
We don’t thank our other tools like grep, or the compiler.
If you train your LLM to produce words that seem like human responses or conversations, they will. If you train them to seem like god-like superintelligences to people who think that’s what they’re creating, they will.
These are tools and it would behoove us all to keep that top of mind. Dangerous tools that are not your friend (but are useful as tools nonetheless)
I do ask it not to ask it follow on questions. I find that derails my own train of thought.
There's an important point in the article, that forcing a style onto an LLM is lossy. Although he doesn't seem to mention it, forcing a style may result in the insertion of new blithering, possibly made up as a hallucination.
I realize this isn't entirely serious, but I can't resist pointing out that this doesn't seem to be a good explanation for why LLMs write the way they do. When we've experimented with LLM writing style on open-weights models where you can get a base model (pretraining on text only) and an instruction-tuned variant (pretraining + post-training with RLHF and whatever other human-evaluated tasks), it's the instruction-tuned variant that shows the weird writing quirks. That is, the writing style is not because of the training texts, but because of whatever tasks the LLM companies do in instruction tuning. https://arxiv.org/abs/2410.16107
I'd speculate that this is partly impressed human preferences (the human raters unintentionally reward a particular writing style) and partly because of the chosen tasks: they're training the LLM to be good at, say, summarizing text, so it develops a style that's good at being informationally dense.
At any rate I've seen this same phenomenon with Llama and Gemma, and will be trying soon with Qwen. Unfortunately none of the commercial models lets you access the base model, as far as I know.
eg., OpenAI has gone a long way to making reasoning token-efficient by having reasoning piovot off terse langauge -- whereas anthropic appears to be doing the opposite.
AI labs can now ask the LLM to translate and filter the data, to create new training data that makes more sense and has better style.
This is pretty unfortunate, because every LLM I've used has at least one tic that I find quite annoying. It would be great to be able to eliminate them. Sometimes I do, even knowing this drawback. But there is generally a cost. (Though I don't know if "lossy" is quite right, as that implies it's always a degradation. I think it's more likely to be harmful than helpful, given the models were tuned for their default state, but it is more of a random perturbation with a slight negative bias than a strict loss.)
Now, the second example is the only thing that works. Power users have lost their powers with AI overview.
> That compression is lossy.
> You probably never notice what got dropped because the output still reads nicely.
> ASD-STE is a great example because it sounds so reasonable. It was designed to make documentation unambiguous for humans. But an agent isn’t a human technical writer, and the raw state is often the most information-dense representation available. Meanwhile the style rules sit on the same instruction list as: solve the task, use tools correctly, preserve abstractions, don’t break anything.
Author seems to have some misconceptions about LLMs. They already code-switch for us: the way they speak in chain-of-thought is completely different from the relatively normal language generated as human-facing output. You can observe this in any open-weight LLM, or in leaked CoT content from GPT5.x series etc: it's terse, barely follows sentence structure, lots of repeated checks and second-guessing.
On the next turn the model usually still has access to its previous turn's chain-of-thought, and I imagine that's what it'll use as reference, rather than the softer human-facing prose.
This being the case, asking the LLM to code-switch to an easier dialect for us doesn't seem that harmful.
For a more extreme example: if I talk to an LLM in Japanese then its response will be in Japanese, but its CoT will still be in either English or Chinese (depending on the model). These are two completely separate languages, but the LLM just kinda deals with it.
What I think the author (and I) are wondering about, is whether instructions like these might influence not _only_ the final output, but also the way it got there.
Not humanising it...
People want the terse, matter-of-fact output. Not the conversational chatty verbose and bloated nonsense with gray words and jargon and terms like "blast radius"
The problem we have is a few large and expensive models are trying to be everything to everyone.
I'm sure you can make it to output the exact issue details as you want, but it starts with a human-like tone and waits for your requirement on depth and detail of the things. Another option is, just check the output of the traditional test runner (non-AI). It will give the full details.
Is that a problem with https://code.claude.com/docs/en/output-styles?
> Output styles apply to the main conversation only: a subagent runs its own system prompt, so styles don’t change how subagents respond. A fork is the exception, because it inherits the parent’s full system prompt.
It has worked great but i've spent more time beating LLM output into parseable output than I have reading and appreciating the prose it sends when i'm asking it something about some snippets of code.
Yes, that's exactly what I want. I want to know what is done with a high level why, NOT a paragraph explaining each line of code modified.
A frequent tweak I've made on coding work with Claude in the last month is asking it to restrict its comment length to 1 line/sentence max. I just want `// This happens because XYZ upstream`, not `// Historically from ticket blahblah there was some dummy code where we discovered ancient runes and that led us to looking into your birth records and then triangulated an issue in XYZ upstream that we compensate for here`.
The model's "most information dense representation" is very similar to how we compress data in the first place: most of it is redundant or unnecessary for the purposes of storage. I'll "decompress" the 1-liner context myself when I read it again in 6 months. But I can't stand reading just so much slop commentary when we're all writing more code at once and having to review more than ever.
Seems like something fixable with a simple two step process. Ask it the thing. Then ask it to summarise the answer in simpler terms. More tokens and time aside that would check both boxes
In other words, is there a way to keep the internal process intact up to the point of the formulation - and have only that vary.
This is why /bro skill works.
https://github.com/backnotprop/bro/blob/main/skills/bro/SKIL...
https://x.com/dillon_mulroy/status/2079238358358778142?s=20
If the author wants to read slop for hours, be my guest. Make it lossy, my job is not to read mimetic feelings, it's to make sure implementations get implemented.
(ICLR 2026)
(not actually sorry)
I don't think it's wise to take communication advice from someone so helplessly juvenile (and attention seeking) in their own communication attempts.
It isn't deliberately unhinged like Steve Yegge's take: https://yegge.ai/essays/model-welfare/ In Steve's essay he starts with the assertion that agents are sentient... Whether or not that's true isn't really relevant, as his agent-flavored version of Pascal's wager actually holds water, especially for Anthropic models, as their system prompts already push the model in that direction, and it is better to work with them than try to prompt against the tide.
Tell? Largest “tell”? Tell for me to tell?
Write in English, please:
“The biggest sign that shows me how culture and sentiment are changing, is…”
In other contexts it means a revealing signal.