There was a security engineer at my work doing something similar to this. He wanted to use LLMs as an IDS. I begged him to use BPF and stop wasting sprint cycles trying to reinvent a shittier slower wheel.
That's why LLM will eventually be used only for initial interaction between the user in their language, to prepare the data to a specialized model.
Imagine face recognition to work like a text chat, where the PC gets the frame from the camera and writes in the chat: "Who's that? Here's the RGB888 image in hex: ...".
Huh? The images are tokenized in the same way language is and it’s just fed into one single model. Not multiple smaller expert models.
Image gets rasterized into smaller pieces (eg 4x4 pixels) and each of those is assigned a token, similarly how text is broken up into tokens. And the whole thing is fed into a single model.
> Imagine face recognition to work like a text chat, where the PC gets the frame from the camera and writes in the chat: "Who's that? Here's the RGB888 image in hex: ...".
The experts in MoEs aren't specialized in any meaningful task sense. From level of what we would think as tasks MoEs are selected essentially arbitrarily per token and per block.
It’s unsupervised, yes, but “unspecialized in any meaningful task sense” is incorrect, that’s the whole point. It’s just not in the sense of “this is a legal expert, this is a software developer”.
Optimal expert separation depends on the goal and can be pretty arbitrary, for example DeepSeek v4 separates them more or less by domain if I remember correctly.
I've heard people use that a lot, but the original metaphor was sonar not table tennis. So it is more appropriately an echo reply (which is what the ICMP return packet is called in the RFC).
For me, he is the opposite of slop (AI or otherwise). This is the kind of guy that writes an operating system for your toaster and leaves enough resources free to run DOOM.
Now do the equivalent of just in time compilation. Claude sees that we need to respond to a lot of pings and writes a program to compute it instead of thinking about each one.
Wouldn't this be faster with an agent skill that has code?
/skill-creator [or /create-skill] Write an agent skill
with code script(s) that use an existing user space IP library that works with your agent runtime, to [...]
If you wonder why your Copilot subscription has new limits that you hit every few days, it's because of PhDs like Adam.
Could Adam use a local model hosted on his own box? Probably yes. But he preferred to waste the service we all use just to produce a weak blog post that introduces absolutely no knowledge and serves no other purpose than to tell everyone that the author likes to waste resources and calls it "fun".
> Ridiculous? Yes. Wasteful of tokens? Sure. Fun? Oh yeah!
Do you really think it's fun to be one of these people who are the reason why the rest of us gets more limits?
In our lives, the game we play, we can do whatever we like. There are consequences for some things, but generally we can do lots of things.
We can kill people and get away with it. We can also help them.
Should we hate life because it's possible to do really shitty things in life? I don't think so. We should hate the "players" who actually do shitty things.
Ok but hating a guy playing with an LLM is a bit extreme. These things are in many ways still just toys (toys that are becoming increasingly useful).
People almost certainly send dumber stuff to Claude than this, and just don’t write blog posts about it.
You could try other providers if Anthropic is too slow/limited, there’s some good alternatives.
(And your anger should probably be directed at Anthropic who hasn’t put in “better controls”, not the masses for not using the tool in the way you think they should. Hating rarely leads to anything productive.)
I'm not angry, I'm just disappointed. Hating is just terminology I've used to reply to the parent poster. Personally I don't think hating anything will solve any problem.
True, people send dumber stuff to LLM, but some of them have the decency not to brag about it. But if they brag, then the common sense I imagine would be to tell them they are currently conducting a shitty thing. Yet on HN over 100 people told him that it's a cool idea. Also you seem to also be one of the "if the system allows it, then we all should do it" mentality people.
I mean you could argue that buying stuff and immediately throwing it to trash is perfectly fine, and if I'm mad about it, then it's my problem because "the person doesn't use the thing the same way I would imagine it to be used", but this argument just sounds silly to me. I know I'm right.
We have brains for a reason. We should use our free will, not offload any thinking to the system we live in.
Unless we don't have free will. I believe some of us don't have it, based on what I read.
Talking about "We have brains and don't offload your thinking" while admonishing someone because you think you would have used their tokens better is wild.
You could read about that in 1992 "A Fire Upon the Deep" by Vernor Vinge. There is prompt injection in communication, in the book certain protocols for information communication can not be deterministic so if someone is too smart you get hacked.
Oh, they are. It's just that the harness around it is able to pick up the commands it "autocompletes" and runs them for you. LLM can't run anything, it never could.
There was a security engineer at my work doing something similar to this. He wanted to use LLMs as an IDS. I begged him to use BPF and stop wasting sprint cycles trying to reinvent a shittier slower wheel.
An elliptical wheel, at most. A square one without an axle, most probably.
When I first saw this comment, it was downvoted into gray. But I can't imagine why. Apropos, and likely pretty accurate.
For reference Adam Dunkels is the developer of lwIP and uIP ip-stacks , as well as the C64 "Contiki" OS that used the latter to do networking.
How fast can Claude do branch predictions in CPU?
Value prediction's where the most gains are to be had. Of course, the irony of using an LLM to execute an LLM is delicious.
At the end the article links this gem, ping via avian carriers:
https://blug.linux.no/rfc1149/writeup/
That's why LLM will eventually be used only for initial interaction between the user in their language, to prepare the data to a specialized model.
Imagine face recognition to work like a text chat, where the PC gets the frame from the camera and writes in the chat: "Who's that? Here's the RGB888 image in hex: ...".
That's actually how vision language models already work, pretty much.
Huh? The images are tokenized in the same way language is and it’s just fed into one single model. Not multiple smaller expert models.
Image gets rasterized into smaller pieces (eg 4x4 pixels) and each of those is assigned a token, similarly how text is broken up into tokens. And the whole thing is fed into a single model.
Yes I'm saying
> Imagine face recognition to work like a text chat, where the PC gets the frame from the camera and writes in the chat: "Who's that? Here's the RGB888 image in hex: ...".
that's p much how it works.
But that isn’t a specialized model like the grandparent claimed, but rather a single, multi-modal model.
Yes, the "imagine" was showcasing the opposite of a specialized model to call it a bad idea.
And there's a reason nobody uses them for face recognition
Vision language models are an incredible achievement in the generality and usability. But they pay a hefty price in fidelity and speed
Do you know that MoE is a thing?
The experts in MoEs aren't specialized in any meaningful task sense. From level of what we would think as tasks MoEs are selected essentially arbitrarily per token and per block.
It’s unsupervised, yes, but “unspecialized in any meaningful task sense” is incorrect, that’s the whole point. It’s just not in the sense of “this is a legal expert, this is a software developer”.
Optimal expert separation depends on the goal and can be pretty arbitrary, for example DeepSeek v4 separates them more or less by domain if I remember correctly.
think about how much faster it would've been with a small local model!
This is cool, let aside the token usage, perhaps it can help analyze tcp throughput by redirect wire shark/to dump result
Opus 4.6 is already very good at troubleshooting all kinds of network problems if it has access to the command line tshark tool and the pcap files.
Agreed it’s pretty pro at deciphering logs, it figured out some weird NAT thing for me.
Modulo Anthropic messing with the model for load mitigation, I wonder how stable this result is.
1,000 pings, how many correctly ponged?
is pong an actual term? If so I might've found a CS term better than wyde (2 bytes)
Yes, very much so. A ping is a request, a pong a response.
2 bytes is a short…
2 bytes is a wyde. Because, to quote knuth, "two bytes makes one wyde"
I've heard people use that a lot, but the original metaphor was sonar not table tennis. So it is more appropriately an echo reply (which is what the ICMP return packet is called in the RFC).
How quickly claude responds when it acts like a user space LLM chatbot?
African or European?
Doesn't matter, the point is inception!
Was not expecting to see Adam in an AI post!
For me, he is the opposite of slop (AI or otherwise). This is the kind of guy that writes an operating system for your toaster and leaves enough resources free to run DOOM.
Now do the equivalent of just in time compilation. Claude sees that we need to respond to a lot of pings and writes a program to compute it instead of thinking about each one.
>Fun? Oh yeah!
I think this author and I have different definitions of fun.
Wouldn't this be faster with an agent skill that has code?
/skill-creator [or /create-skill] Write an agent skill with code script(s) that use an existing user space IP library that works with your agent runtime, to [...]
ComposioHQ/awesome-claude-skills: https://github.com/ComposioHQ/awesome-claude-skills
anthopics/skills//skill-creator/SKILL.md: https://github.com/anthropics/skills/blob/main/skills/skill-...
/.agents/skills/skill-name/SKILL.md, scripts/{script_name.py,__init__.py}
https://agentskills.io/what-are-skills
Well, yeah, of course it would be.
Even faster would just to be use code in the first place!
I eagerly await the publication of your RFC for IP over Slop Generators.
If you wonder why your Copilot subscription has new limits that you hit every few days, it's because of PhDs like Adam.
Could Adam use a local model hosted on his own box? Probably yes. But he preferred to waste the service we all use just to produce a weak blog post that introduces absolutely no knowledge and serves no other purpose than to tell everyone that the author likes to waste resources and calls it "fun".
> Ridiculous? Yes. Wasteful of tokens? Sure. Fun? Oh yeah!
Do you really think it's fun to be one of these people who are the reason why the rest of us gets more limits?
Don't hate the player, hate the game.
No.
In our lives, the game we play, we can do whatever we like. There are consequences for some things, but generally we can do lots of things.
We can kill people and get away with it. We can also help them.
Should we hate life because it's possible to do really shitty things in life? I don't think so. We should hate the "players" who actually do shitty things.
Ok but hating a guy playing with an LLM is a bit extreme. These things are in many ways still just toys (toys that are becoming increasingly useful).
People almost certainly send dumber stuff to Claude than this, and just don’t write blog posts about it.
You could try other providers if Anthropic is too slow/limited, there’s some good alternatives.
(And your anger should probably be directed at Anthropic who hasn’t put in “better controls”, not the masses for not using the tool in the way you think they should. Hating rarely leads to anything productive.)
I'm not angry, I'm just disappointed. Hating is just terminology I've used to reply to the parent poster. Personally I don't think hating anything will solve any problem.
True, people send dumber stuff to LLM, but some of them have the decency not to brag about it. But if they brag, then the common sense I imagine would be to tell them they are currently conducting a shitty thing. Yet on HN over 100 people told him that it's a cool idea. Also you seem to also be one of the "if the system allows it, then we all should do it" mentality people.
I mean you could argue that buying stuff and immediately throwing it to trash is perfectly fine, and if I'm mad about it, then it's my problem because "the person doesn't use the thing the same way I would imagine it to be used", but this argument just sounds silly to me. I know I'm right.
We have brains for a reason. We should use our free will, not offload any thinking to the system we live in.
Unless we don't have free will. I believe some of us don't have it, based on what I read.
Talking about "We have brains and don't offload your thinking" while admonishing someone because you think you would have used their tokens better is wild.
You've literally just used the same argument from the previous poster which was even addressed by me in my parent reply. Talking about not thinking!
excuse me? this is a cop out used to justify heinous things done under capitalism
You are way off.
Adam is the practical kind of PhD.
https://en.wikipedia.org/wiki/Adam_Dunkels
Next up: Claude replacement to handle simdjson processing.
Perhaps one day, all network services will be provided by LLMs natively. Truly, that would be a day in the future.
I mean, we did decades of JavaScript, so... I mean... anything is possible, right? :)
why? We already have more efficient specialized hardware.
You could read about that in 1992 "A Fire Upon the Deep" by Vernor Vinge. There is prompt injection in communication, in the book certain protocols for information communication can not be deterministic so if someone is too smart you get hacked.
"Perhaps" doing enough lifting to participate in a bodybuilder contest, in that sentence
I’m sorry people aren’t getting it, or are so committed to downvoting humor here they’re tagging the good stuff.
Do some people still claim "LLMs are just dumb auto completers"?
Because this seems to disprove that claim pretty convincingly?
Oh, they are. It's just that the harness around it is able to pick up the commands it "autocompletes" and runs them for you. LLM can't run anything, it never could.
What do you mean that the "harness ran the commands"?
It looks to me like the LLM "executed" the logic in pure output tokens, not by using any kind of external tool calls?
It proves that code, specifically any code in the form of bytes, is, too, language.
I like that. And if poetry can be defined as succinct use of language (perhaps a dubious assertion), then code can be poetry.