FYI a quick fix for claudish is to ask for the response to be in ASD-STE100 (Simple Technical English). Then it is far more readable. But I would agree that this is an annoyance and shouldn't require user workaround to get something readable.
First, after a while it's just as grating as Claudeish. Second, my hunch is that it constricts the actual thinking of the LLM, like the same way that Newspeak does in 1984. It shrinks the range of thought that can be expressed if used as an input.
I think the real way to do it is to have another Claude entirely deal with the user as a liaison, but to keep the thinking in whatever format it came in.
Latent space reasoning, if you think about it, is exactly this to a crazy degree: why even formulate a thought as words if you can just keep it as matmuls until the user needs it? And then, if the user needs it, have it always specifically formulated for the user by another LLM rather than constrict its range of thought? Anyway, that's my take.
I don't think this is true.
They have to express themselves as tokens. The meaning of those tokens doesn't have to be text. See any model that can handle images/video. Also, I don't think math, svg, etc, are "natural" language.
And, only the final expression is tokens. The intermediate layers, with the encoded concepts, aren't "natural language".
But, to address your concern (which nobody can disagree with, since even humans can't fully express through text/pictures), potentially: https://news.ycombinator.com/item?id=49758615
The model isn't limited to concepts that can be expressed in natural language.
It's only once the AI gets to the output layers that natural language comes back into play.
After all, they're all made out of weights[0].
How do we know for sure? We don't even know how the emergent properties we see actually emerged?
For humans we know for sure that people sometimes have concepts that they have no word for (the reason the phrase "It's on the tip of my tongue" is a phrase, after all).
We don't know this for LLMs. When it makes new phrases, it's always a mixup of two existing words hyphenated (aside, that also seems to be the limits of SOTA models creativity - join two unrelated words together with a hyphen).
LLMs never respond with "It's on the tip of my tongue" type responses, indicating it has a concept but cannot remember (or does not have) a word for that concept. Every human, pre-speech-age, has managed to express or convey concepts that they had no word for.
So, no. I'd need a citation, preferably multiple, that did the trials and found that a model can generate concepts for which it does not have any words for.
Even if the input is in plain English, the model never sees any words, tokens or glyphs to begin with. It's vectors all the way down.
1. Is natural language holding LLMs back by some %? 2. Is natural language serving as a hard gate that will prevent LLM intelligent progressing past some specific point?
The answer to 1 seems like an obvious yes to me.
Your thesis says the answer to 2 is "yes." That doesn't feel right to me. Think about all of the humans who have pushed various fields forward: Einstein, Newtown, Bach, whoever. If natural language doesn't prevent an entity from surpassing humans in one intellectual field, why would it prevent an entity from surpassing humans in all intellectual fields?
(To be clear, I'm not claiming superintelligence will or won't be achieved; I'm considering your specific thesis about whether or not natural language will be a hard gate)
By the way, how good is Claude's Hopi?
I do think an infrastructure where another Claude retranslates the output would be better. Oftentimes I forget to put it in the actual prompt and when I receive back 8 paragraphs of Claudeish I ask for it then.
I would have to disagree that it gets as grating as Claudeish though. Its just direct and professional instead of ring-around-the-rosy clickbait.
“I would have to disagree that it gets as grating as Claudeish though.”
It’s hard to imagine anything more grating than Claudeish. To quote Rainer Wolfcastle, "My eyes! The goggles do nothing!"
I've found that prompting any constraint on output (length, style, vocab, even simple formatting) not only places additional cognitive load on the model, which burns some of whatever cognitive budget is available, it will also often skew the output in other subtle and completely unrelated ways.
Since I found this artifact interesting, I did some pretty extensive experiments a couple months ago. The increased load is real, although it may not be apparent if you're not near any cognitive boundaries. The subtle skew, however, seems nearly ever-present regardless of load.
While extensive, my tests were just following my curiousity, not controlled, exhaustive or well-documented. I identified about a dozen prior sessions of varying length and complexity to test and downloaded them with a browser add-on. I then removed all other user prompt instructions except for the formatting instruction. A test would typically involve changing the wording of the formatting instruction ranging from brutally simple to detailed and complete, then starting a new session, seeding one of the test sessions and continuing it. To get a feel for baseline inter-session variation, I also tried running the exact same prompt/session multiple times back-to-back, at different times and on different days of the week.
Once I identified a promising prompt candidate, I'd make it the formatting instruction in my regular, daily-use prompt for a few days. I quickly got a feel for how seemingly minor user prompt variations impact response quality, compliance and tone across fresh sessions as well as those in various states of context rot, drift, decay and cliff (<--my nicknames for the distinct flavors of session degradation, not technical terms).
My overall conclusion was that every instruction, no matter how minor or unrelated it seems, has some, real impact on the model's cog load, attentional focus and/or attentional weight budget. Both how these impacts manifest and what causes more or less impact is often extremely counteriintuitive. To more fully understand this, I eventually, got to the point of testing null case variants, such as the entire user prompt being one sentence completely unrelated to text formatting or the session topic, like: "Don't reference the cartoon character SnagglePuss" (in a deep dive on ancient Sumerian clay tokens). Similarly, a simple one sentence prompt requesting something the model already always does naturally also has a cost (eg "Capitalize proper nouns"). As others have observed, heavy emphasis, absolute prohibitions or emotional weight in prompts also tend to have outsized impact in both skew (impacting unrelated output tone/style) and in accelerating session degradation. "Avoid referencing SnagglePuss when you can" would have equal compliance but fewer downside impacts than "NEVER reference the cartoon character SnagglePuss" in sessions starting to degrade.
There were also surprises, such as when I was scanning transcripts of an older, longer session and noticed the LLM was doing number formatting almost perfectly. On looking at the active user prompt at the time (I keep a log of every user prompt change I make for every model), it didn't even reference formatting at all. More experimentation showed it a result of the LLM gradually mirroring my consistent use of formatting structure in my prompts over a long session (in which I never mentioned anything about formatting). Unfortunately, that mirrored trait doesn't persist to new sessions and reaching that point requires a substantial number of rounds burning quite a bit of context window.
After spending time surfacing the impacts of just changing lightweight user prompts so they could be observed (which are the lowest priority prompts a model gets), I now wonder just how much more 'brilliant' the models we use daily would be if they didn't have dozens of pages high-priority manufacturer prohibition prompts we never even see weighing them down. We've only ever seen these frontier 'racehorses' when they're already pulling a heavy invisible wagon.
Literally every one, even 1-2 prompts later it starts to go back
It burns more tokens but is the only way to get tolerable text.
https://code.claude.com/docs/en/hooks-guide#agent-based-hook...
It’s been really productive and I’ve been asking my agents to communicate using it more and more. I believe it’s relieved my cognitive load a bit while working with them.