it's just a bit disheartening to read Claude output for such a small comment like this.. it'd be great to read your own writings even if it's not as "perfect"
One side effect of extremely-accessible, high-quality English translation and, more importantly, English-grammar-and-idiosyncrasy-obeying AI, is that there will be more and more text that looks like this which comes from foreign, non-native-English countries.
Just FYI.
This will be especially true from non-English cultures where "avoiding shame" is high on the list of motivations.
Consider the upside, though: A much broader range of written perspectives written in high-quality, if slightly annoying, English.
I'm already seeing the benefit of this on X thanks to its autotranslation btw: I follow a few Chinese-language accounts now that I would have never been able to digest otherwise.
> It didn't work, and why it didn't work turned out to be more interesting than the policy would have been.
Spoken like a true Claude.
Snarking aside, I am glad that our AI agents make it cheap enough to do these experiments and publish these write-ups that people finally bother to publish null findings. Very useful!
If you're being given an A-B choice between cinnamon-flavored shit and lutfisk-flavored shit, you might convince the experimenter that cinnamon-flavored shit ranks highly in user preferences.
I would expect a preference for lutefisk flavored shit, the problem is then you then deduce from that that people prefer lutefisk flavor over cinnamon, rather than that you should stop feeding people shit (and also stop feeding them lutefisk.)
I suspect it’s more like Claude etc. optimizing for a tone that gets the green checkmark from RLHF reviewers, probably over and over in many contexts where style is otherwise an incidental concern.
I've seen Douglas Adams use almost exactly this in Dirk Gently's Holistic Detective Agency. A professor (IIRC) is talking to a former student. Quoting from memory
"Did you ever actually finish any papers?"
"No. But the reasons why not were always absolutely fascinating."
So, yeah, for me personally it's not an AI tell. I've seen it in the wild before AI.
LLMs write like some people cook, adding bespoke artisanal Belgian sea salt, truffle oil and weirdly specific cheese without any thought given to how the result will taste.
I always have a tinge of...perhaps 'sadness' when reading such articles because they are a lot like those large fishing nets that drag along the bottom of the ocean catching and destroying in quantities unneeded. From this list:
- Quietly/Delve/Tapestry/Landscape: they're all lovely words. I love(d) using them.
- "Serves as": I use this all the time too.
- "The result? Devastating". Perhaps it is a bit stilted, but "Outcome? Predictable." could be a nice turn of phrase.
- "It's worth noting": All academics everywhere use this in papers. See also: "Notably", "Essentially", "In general" and other such fillers that we reach for.
- "Think of it as": This is such a useful phrase to explain something. Analogy is how we learn after all.
- The fucking em dashes. I love them, and screw everyone who suggests one should stop using them. One big reason for using linux is the compose key, so the em dash is never far away.
- Same with the unicode decoration. Nah, screw you. I use these symbols all the time: °, ×, →, ⇒ etc. in my own notes. Again, compose key ftw.
Etc. (the list is endless)
I know what people will say, that everything is contextual, one has to study case-by-case etc. But I can't help feeling that someone is pointing a finger at me and accusing me before-the-fact. Just, AI has polluted the joy of writing.
While AI was clearly used to generate and structure a first draft, I say it has been extensively rewritten by the author. In my experience, AI tends to avoid the first person, instead writing from a passive observer or even using sentences without verbs. The length of the sentences is also more natural. AI had trouble with bind-words in an argument. It just states separate facts and assumes because they are stated in order you have a reasoning. I'm not seeing that here.
Of course, idioms are preserved, and not everything is rewritten.
I quite ok with this style, it isn't grating. I expect this will be a style we'll see more of in science.
The idea of publishing null findings is very attractive, but the problem here is that they didn't record what the actual problem they're trying to solve is. It seems a bit like someone prompted Claude to try and make tool calls more efficient and Claude responded with a detailed writeup on why LRU can't be beat for making a KV cache. Which might be interesting, if that were the question I had asked.
LRU seems like the ideal strategy for most things LLM-related. Everything in this realm is about recency bias. I think it is a feature in this context, not a problem.
When I give an agent a piece of corrected information regarding a long running task, the last thing I want it to do is try and statistically compensate for the fact that it is new information. I want this new information to dominate the old information.
This is almost unreadable. The "papers" are never referenced anywhere, so the claims being refuted cannot be evaluated. The whole fact of the TTL doesn't seem relevant at all. There are many, many well-researched admission and eviction policies that this readme doesn't mention. I just don't get why we are reading this.
What I feel a bit annoyed by, and what I feel obviously LLM-run ablations like this fail to capture, is any kind of reflection around previous research or any kind of proof that this is the best you can do. You don't know, you pulled the lever and you got something, is the best? Can you do better? What is the constraint?
As an individual researcher, you do not have 22M$ to run a massive brute-force search for your problem. You are constrained to your little subscription and you will barely dip your toe in the sea of possible solutions to a problem. So letting Claude run an autoresearch loop on your problem and then having it summarize it for you brings 0 value because you dont know what the downsides and trade-offs of LRU caches were, and how you would possible solve it.
It'll just be telling you what the data it ingested claims. Not what is necessarily true. It's as subject to garbage in, garbage out as anything, and you don't know what it's actually trained on.
Sometimes it's good enough to acknowledge a correct answer, and slowly build intuition towards it. Academic problem solving is rarely a straight path, anyway.
I don't think the problem is a lack of previous research, the problem is it looks like the target metrics were selected by an LLM, and it's unclear what the LLM was told to optimize or if it was just told to try and make a better KV-cache.
What I've noticed with Claude is that regardless of how I prompt, it will find a few metrics to optimize. Often the metrics it chooses to optimize have zero relation to the actual metrics I want to optimize, which are ones that cannot be measured without more work than Claude can do in a single 1M token context window. It's really hard to stop Claude from optimizing whatever metrics it can find when the actual metrics I want to optimize are not computable.
> What I feel a bit annoyed by, and what I feel obviously LLM-run ablations like this fail to capture, is any kind of reflection around previous research or any kind of proof that this is the best you can do.
At least the former is easy: Claude is very happy to do a prior art search and pretty good at it, too.
I work on inference at a neocloud, but opinions are my own .
The economics and thus tools you can throw at inference change at various scales .
As a “blunt” contrived example , on a gb300 the GPUs communicate super fast over nvlink, and the cards can offload kv cache to dram and then disk, “fast enough “ for these tool heavy agentic workloads.
Which come together to mean that at high enough scale and in the right scenario, we can work
with wild ttls on the kv cache and still comfortably hit SLAs and tokenomics.
I've been building latency-sensitive LLM systems for a while, and I've come to rely heavily on pre-fill-considerate mechanics like ping-pong overlapped async context construction. For interactive mechanics, the worst case, even if rare, is problematic.
Consideration of mutation rate (a sort of temporal Shannon-ish coding/ordering) lives in there (with some RoPE-friendly structuring). Note: That was a vacation project, not the day job, but similar principles apply even with larger models.
Based on the tables in part 4 it seems like some of the policies would have started overtaking LRU at larger KV-cache sizes.
And while the range of cache sizes chosen is reasonable based on the table about Moonshot's number at the start, the finding that all tested KV-cache sizes are so small that the 5-minute eviction never triggers should have been reason enough to reevaluate that choice
>Cross-request KV prefix caching is the largest practical lever in agentic LLM serving. It's why your coding agent's fiftieth turn costs a fraction of its first.
This is incorrect. If you are generating the 51th turn, then prefix caching makes the turns 1 to 50 cost a fraction of what they normally cost, to generate the 51th turn. Generating the 51th turn is more costly because it is not in the cache yet.
Unfortunately there is no way to steelman the statement, since quadratic scaling of attention means that the cost goes up even if you assume perfect caching of past inputs.
What the author actually means is something more mundane. The initial prompt is massive in a standard coding harness and this means prompt processing takes much longer than expected, creating the misleading impression that costs go down.
Edit: I noticed the word cross request too late. In that case the harness prompt is expected to be cached from other users, in which case the first message actually has a massive unfair advantage.
50 comments
[ 0.30 ms ] story [ 4.4 ms ] threadThe best kind of correct.
it's just a bit disheartening to read Claude output for such a small comment like this.. it'd be great to read your own writings even if it's not as "perfect"
Just FYI.
This will be especially true from non-English cultures where "avoiding shame" is high on the list of motivations.
Consider the upside, though: A much broader range of written perspectives written in high-quality, if slightly annoying, English.
I'm already seeing the benefit of this on X thanks to its autotranslation btw: I follow a few Chinese-language accounts now that I would have never been able to digest otherwise.
I'd have thought they'd be busy (using cached prompt-prefixes) until they were finished.
Spoken like a true Claude.
Snarking aside, I am glad that our AI agents make it cheap enough to do these experiments and publish these write-ups that people finally bother to publish null findings. Very useful!
*not at all
[1] https://gist.github.com/ossa-ma/f3baa9d25154c33095e22272c631...
"Did you ever actually finish any papers?"
"No. But the reasons why not were always absolutely fascinating."
So, yeah, for me personally it's not an AI tell. I've seen it in the wild before AI.
LLMs write like some people cook, adding bespoke artisanal Belgian sea salt, truffle oil and weirdly specific cheese without any thought given to how the result will taste.
For example?
The same is not true, of the token generation that... Mostly Claude models... Have taken to.
- Quietly/Delve/Tapestry/Landscape: they're all lovely words. I love(d) using them.
- "Serves as": I use this all the time too.
- "The result? Devastating". Perhaps it is a bit stilted, but "Outcome? Predictable." could be a nice turn of phrase.
- "It's worth noting": All academics everywhere use this in papers. See also: "Notably", "Essentially", "In general" and other such fillers that we reach for.
- "Think of it as": This is such a useful phrase to explain something. Analogy is how we learn after all.
- The fucking em dashes. I love them, and screw everyone who suggests one should stop using them. One big reason for using linux is the compose key, so the em dash is never far away.
- Same with the unicode decoration. Nah, screw you. I use these symbols all the time: °, ×, →, ⇒ etc. in my own notes. Again, compose key ftw.
Etc. (the list is endless)
I know what people will say, that everything is contextual, one has to study case-by-case etc. But I can't help feeling that someone is pointing a finger at me and accusing me before-the-fact. Just, AI has polluted the joy of writing.
Of course, idioms are preserved, and not everything is rewritten.
I quite ok with this style, it isn't grating. I expect this will be a style we'll see more of in science.
Not my experience with Claude at all.
Which AIs are you talking about?
This sentence quoted just doesn’t seem like the AI smoking gun to choose given the rest of the content.
When I give an agent a piece of corrected information regarding a long running task, the last thing I want it to do is try and statistically compensate for the fact that it is new information. I want this new information to dominate the old information.
As an individual researcher, you do not have 22M$ to run a massive brute-force search for your problem. You are constrained to your little subscription and you will barely dip your toe in the sea of possible solutions to a problem. So letting Claude run an autoresearch loop on your problem and then having it summarize it for you brings 0 value because you dont know what the downsides and trade-offs of LRU caches were, and how you would possible solve it.
What I've noticed with Claude is that regardless of how I prompt, it will find a few metrics to optimize. Often the metrics it chooses to optimize have zero relation to the actual metrics I want to optimize, which are ones that cannot be measured without more work than Claude can do in a single 1M token context window. It's really hard to stop Claude from optimizing whatever metrics it can find when the actual metrics I want to optimize are not computable.
At least the former is easy: Claude is very happy to do a prior art search and pretty good at it, too.
The economics and thus tools you can throw at inference change at various scales . As a “blunt” contrived example , on a gb300 the GPUs communicate super fast over nvlink, and the cards can offload kv cache to dram and then disk, “fast enough “ for these tool heavy agentic workloads.
Which come together to mean that at high enough scale and in the right scenario, we can work with wild ttls on the kv cache and still comfortably hit SLAs and tokenomics.
A toy/simplified version lives here: https://github.com/chaboud/goulash
Consideration of mutation rate (a sort of temporal Shannon-ish coding/ordering) lives in there (with some RoPE-friendly structuring). Note: That was a vacation project, not the day job, but similar principles apply even with larger models.
And while the range of cache sizes chosen is reasonable based on the table about Moonshot's number at the start, the finding that all tested KV-cache sizes are so small that the 5-minute eviction never triggers should have been reason enough to reevaluate that choice
This is incorrect. If you are generating the 51th turn, then prefix caching makes the turns 1 to 50 cost a fraction of what they normally cost, to generate the 51th turn. Generating the 51th turn is more costly because it is not in the cache yet.
Unfortunately there is no way to steelman the statement, since quadratic scaling of attention means that the cost goes up even if you assume perfect caching of past inputs.
What the author actually means is something more mundane. The initial prompt is massive in a standard coding harness and this means prompt processing takes much longer than expected, creating the misleading impression that costs go down.
Edit: I noticed the word cross request too late. In that case the harness prompt is expected to be cached from other users, in which case the first message actually has a massive unfair advantage.