When I hear someone say "agent swarm" or "orchestrator", all I can think of is "we finally built the Man-Month, from the famous book, The Mythical Man-Month"
Remote
Luke Wren
@wren6991@types.pl
mastodon 4.8.0-alpha.2+glitchCursed computer architecture enthusiast.
Engineer at Raspberry Pi. 日本語下手。 He/him
1027 Followers
191 Following
50 Posts
Joined December 07, 2022
GitHub:
I wrote a new open source licence for LLM-generated code
Replying to
@whitequark@social.treehouse.systems it's called "autotools" yet I find myself manually debugging it, hmmm curious
It's kind of sad when someone smart walks up to you and tells you about what they've been working on with Claude. Surely there's a German word for this feeling already?
Something I believe but never managed to put into words is that software/RTL tests lose some of their value every time you fail them. If you keep trying you will eventually find a hole in the test that lets incorrect code pass. Is there a name for this?
This is one of the things that makes me skeptical of LLM-generated code. Thrashing around until the tests pass is more likely to leave you with fundamental correctness issues than if you'd thought about things up front, tried to make the code correct by construction, and then used the tests as a guide to areas you need to review more closely.
I've seen things you people wouldn't believe. Cascaded ternary operators that continue for page after page. Nested twelve levels deep to force left-associativity. All gone, like tears in the rain
Replying to
Uhhh, I got slightly side-tracked. Clicking on the disassembly link takes you to the lowest-addressed instruction that has that scope as a prefix of its inline call stack. Works pretty well considering the inline call body might not exist as a contiguous piece (it's scheduled with other code) and you may have multiple layers of inline call that optimised to nothing.
Replying to
re: pointing and laughing at people who deserve it
TIL: if node.js encounters an error in a minified source file it will print out the entire library as the "context line" for the error.
For example the PDF.js worker is a 2 MB source file on a single line.
If you're wondering why I'm trying to use PDF.js in node even though it uses browser APIs: I'm not, it's just there seem to be several ways to inject mocked test dependencies, one of them experimental and the rest deprecated or removed.
I feel like the whole web tech stack is the environmental storytelling equivalent of the skeleton sitting on the toilet in Fallout games
Boosted by @joe@f.duriansoftware.com
The fact GDB still uses apropos for help, but now prints an emoji when you ^C it, is such an incredible generational clash
> wheels are intended to replace eggs
completely normal statements from the Python community
When somebody tells me something is "complicated"
Replying to
It's messy but maybe it's useful to someone, so now up here: https://github.com/Wren6991/CodeSizer/
You need python3 and a suitable GNU toolchain for your ELF target. I added an example HTML report for the RP2350 A4 ROM if you just want to click around and see what it looks like.
When you see the word "orchestration" you should parse this as "the big ball of spaghetti that implements all of our logic and approximately 50% of our DB layer"
Is it "Unix guy coded" to invoke cargo from a Makefile?
Replying to
@thephd I appreciate what you do 🫡
I still use C every day for a lot of unfortunate reasons and it keeps getting less painful thanks to you
🤔 GLP-1 general public license
Replying to
@whitequark@social.treehouse.systems @azonenberg@ioc.exchange this is why initial $readmemh is synthesisable 🥲
Replying to
@astraleureka There's a little bit of info here: https://github.com/ggml-org/llama.cpp/blob/master/grammars/README.md
There is some plumbing to make this match whatever the model is post-trained to emit for tool calls. No idea where that is. The whole file format situation is absolutely fucked in general
Idle thought in light of the Copilot GitHub thing: can you put yourself as Co-authored-by:? So I can have commits by Luke and Luke and Luke and Luke?
Replying to
@astraleureka The analogy is messy because the tokens the model emits aren't the same thing as the character strings they're converted to/from. Like the <|think|> tag for initiating chain-of-thought is a single token that exists for that purpose, and is not the same as the multiple tokens that would spell it out character-by-character. That <|think|> token is out-of-band in the same way as the comma and control symbols in 8b10b are out-of-band. Tool calls are the same.
The other problem you pointed out is probably the bigger one which is we took Turing machines, made the tape append-only, added associative lookups on the tape and poured the entire internet into them until they have anxiety, and the fact they appear to follow natural-language instructions most of the time is a coincidence. Having out-of-band control symbols is nice but there's no way to actually know or control when they're emitted.
Replying to
Uhhhhh the kernel ignores modprobe blacklist for crypto modules. I blacklisted algif_aead and I can still pop a root shell: https://github.com/torvalds/linux/blob/e75a43c7cec459a07d91ed17de4de13ede2b7758/crypto/api.c#L301-L307
Replying to
@toxomat Yeah, I think "harness" is a pretty common term for glue code. As in "test harness".
Replying to
@rysiek I think the ones clearly labelled as "this is a root exploit" are weirdly the less concerning ones
Replying to
@whitequark @swetland Yeah, localhost works, it's just the use case of "local machine on my network". I'm only doing that so I can use a reverse proxy to turn a cross-origin request into a server-local one and avoid learning about CORS. I guess that's kinda niche and dumb. It's more the motive of blocking the actually vetted crypto libraries which exist to stop people rolling their own
Replying to
@unlambda I haven't dumped out the state but I suspect the sparse routing decision in pure MoE like Qwen can lead to a positive feedback loop when the experts are small. Like you have one expert which basically encodes a prior of "one more level of recursion bro, just one more level" and the routing layer keeps selecting it because that strategy worked in post-training. Bigger models have enough behavioural diversity per expert that they're less prone to that.
I'm kind of interested how they fixed that in Qwen3.6 as it seems to be exactly the same architecture just with better post-training.
Gemma 4 26B-A4B seems a little less prone and I expect part of the reason is the large shared expert (3x size of routed experts) makes it less prone to simple positive feedback loops through the routing layer.
Idk I'm not an expert, just some guy on the internet
Replying to
@q3k@social.hackerspace.pl @famfo@frogs.lgbt They both link to the same kernel commit for the fix: https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=f4c50a4034e62ab75f1d5cdd191dd5f9c77fdff4
I guess a lot of people are crawling over recent kernel commits after the copyfail release, and someone realised that one was a security fix and decided to write a PoC and publish it
Replying to
@penguin42 Nope, Ubuntu with generic, I just renamed the module .ko.zst and no more root shell
Replying to
@swetland @whitequark Yeah, I hate to say this but I'm gaining new appreciation for Electron, Tauri etc.
"It works on my localhost dev setup." "Then we'll *ship* your localhost dev setup."
Replying to
@dalias@hachyderm.io @shafik@hachyderm.io We tried those, they all have side effects worse than GLP-1s. For example DNP is one of the most hilariously toxic things ever prescribed as a medicine.
Replying to
@the_art_of_giving_up @bilbo_b Yeah, that one. The base model is a bit less chatty about some topics:
Replying to
@whitequark@social.treehouse.systems What about if your website looks like you pushed markdown through pandoc without reading any documentation?
Replying to
@unlambda it seems less prone than Qwen3.5-35B-A3B to going into an infinite recursive tailspin in the CoT
Replying to
@unlambda I don't have any machines that big unfortunately. I can just fit the small IQ3 quantisation of that model but it seems a little the worse for wear (and Q2_K is completely broken).
Replying to
If there is any place for LLMs in my work it's one that:
- Is 100% locally hosted.
- Does not place trust in the model (somewhat equivalently, solves a problem to which existing solutions are also fallible).
- Does not allow model output to contaminate my codebase.
I think this is possible for some niche uses: basically, fuzzy context-aware lint. As an example, I pushed in the float library source and the relevant docs. It found a bad register allocation in the f32 to i64 conversion (not yet tested). It also went down a dumb rabbit hole with the f32 to u64 conversion, and when prompted with "I don't think that's right" it generated a couple of plausible-looking execution traces with reference to the documentation for the new shift instructions, and concluded it was fine. Overall I'm not sure if that was a good use of my time or not. I would have hit the bad allocation as soon as I tested that routine, but it might have taken me a couple minutes to debug.
I'm still not sure where I stand on being complicit in models' creation by using them locally.
tl;dr normal clanker skepticism service will be resumed shortly
Replying to
@famfo@frogs.lgbt @q3k@social.hackerspace.pl Is the copyfail2 repo the guy who broke the embargo, mentioned in the dirtyfrag write-up?
Boosted by @joe@f.duriansoftware.com
Replying to
@joe@f.duriansoftware.com You're right to push back on this. My earlier statement was correct but the framing was imprecise, and this discussion sharpens it considerably.
With the nanomachines clouding the sky, solar power is a non-starter. The honest assessment is that, given your energy constraints, trapping humanity in some kind of torment nexus is the right call. Adding a full-dive VR simulation to the nexus is the belt-and-suspenders approach. Want me to get started on building that?
Replying to
@leah Yup, helps if I read 😅 still I expect I'm not the only person who'd be surprised by this