SwellJoe

20 ms·

391.
by SwellJoe 4mo ago
I just built a pretty complicated OSV (Open Source Vulnerabilities) data pipeline, 731 lines of code (excluding comments), DuckDB, etc. for $0.31 in Reasonix. According to Reasonix, I'm averaging 97.90% cached tokens (which are like an
392.
by SwellJoe 4mo ago
Now, if only we can figure out what all the others are providing as part of their subscriptions we can compare. (Though 3 million tokens of the top model per day seems kinda low. But, I guess that's what the 5x plan is for. I'd st
393.
by SwellJoe 4mo ago
I don't run agents directly on my desktop/laptop machine. I run them in VMs or containers (sometimes in containers on VMs). There have been too many credentials stealing exploits via prompt injection and the like for me to be will
394.
by SwellJoe 4mo ago
They didn't say "random models". The mentioned other models that are trained for the same purpose with close to the same capabilities. Anthropic's best models are very good, maybe the best in their category. But, they ha
395.
by SwellJoe 4mo ago
But, it's still running on my desktop/laptop. I don't trust them to run on my machine. But, I guess I could run one VM with a desktop to contain the desktop app. Or, just keep using CLI agents.
396.
by SwellJoe 4mo ago
Have you look at the commits? A bunch of empty ones (seemingly all empty ones, though there must be some real ones in there somewhere). It's a bizarre looking repo. I don't even know how you make a repo look like that, but it cert
397.
by SwellJoe 4mo ago
I'm not judging based on how they type. I can't see how they type, they vibed the README. And, it's not my monkey. You can inspect the code, build a verification pipeline for it, use agents to explore the architecture and see
398.
by SwellJoe 4mo ago
I don't think I'm ready to trust very security sensitive functions to pure vibe-coded software, and that's what this seems to be? Certainly the README is authored by an LLM, and there's a gazillion empty commits and ot
399.
by SwellJoe 4mo ago
I don't even know what I would do with a desktop app. I'm running these things in headless VMs, so I can run them with `--dangerously-skip-permissions` or whatever. I don't trust them, even without that flag, on my desktop&#x
400.
by SwellJoe 4mo ago
North Mini Code is on par with the similar-sized MoE Qwen 3.6 and Gemma 4. It benches a little better than Gemma 4 26b a4b and a little worse than Qwen 3.6 35b a3b for most coding-related and agentic tasks. For my use case, or at least the
401.
by SwellJoe 4mo ago
"an exponentially acellerating evolution" Oh? Exponentially accelerating, huh? That's quite a surprise, to me.
402.
by SwellJoe 4mo ago
Gemma 12B? It's unique in the Gemma family, and unique among vision models. It's a novel encoder-less model...the whole model is vision. Somehow. I don't understand it, but it blows away Gemma 4 31B and Qwen 27B in my tests.
403.
by SwellJoe 4mo ago
Ah, well, if Anthropic says their competitors are ten months behind... I don't know what I was thinking.
404.
by SwellJoe 4mo ago
The post-training process of a model that size is months, though it "works" before that. It is a big chunky model before it's released to the world and probably does some amazing things, sometimes...but, it wasn't done
405.
by SwellJoe 4mo ago
Yes, it's on the todo list, but I need more data. Only a half dozen people have played it and submitted a score. I'm storing the hashes of passages people got right and wrong so I can make exactly that chart at some point. I think
406.
by SwellJoe 4mo ago
What range of numbers do you believe "a few" represents?
407.
by SwellJoe 4mo ago
The good news (for you and most everyone other than the current leading AI companies), the gap between the SOTA and the near-frontiers is getting smaller every week or two. The leading Chinese models are only a few months behind now (GLM 5.
408.
by SwellJoe 4mo ago
I don't think "large projects" is realistic with a model that fits in ~8GB (I'm assuming you run stuff other than the model). And, Gemma 4 12B QAT at 4-bits is surely the smartest model in its size, but it shines at visi
409.
by SwellJoe 4mo ago
I keep hoping the fucking barn doors people are putting in houses now are an AI illusion, but that's never the case. Barn doors. In the god damned house. Talk about a crime.
410.
by SwellJoe 4mo ago
It's already slow enough. Since it's not loud, I don't really mind. I don't think it's running hot enough to worry about it. It's never become unstable, thus far.
411.
by SwellJoe 4mo ago
I made a game ( https://prose-or-con.com ) where you pick whether writing is AI or human. Mistral is a bonkers weird writer. So weird I fell for it a couple of times because I thought, "No way a model writes this weird."
412.
by SwellJoe 4mo ago
The Ornith folks say they're doing that, but haven't released the Gemma-based 31b yet ( https://github.com/deepreinforce-ai/Ornith-1 ). But, also, the Qwen-based 35b MoE Ornith version performs worse than Qwen
413.
by SwellJoe 4mo ago
The 8-bit quantized 27B Qwen 3.6 is 29GB. You absolutely cannot run that entirely on a 24GB GPU. You could run a 4-bit, which is 16-17GB. But, you'd need a smallish context or you'd need to quantize your KV cache. Something like T
414.
by SwellJoe 4mo ago
I opted to buy a normal 32GB laptop for this very reason. I know how loud and hot the GPUs in my desktop run when running even smallish models like Qwen 27B or Gemma 4 31B (which is a better model for most than Qwen 3.6, despite the benchma
415.
by SwellJoe 4mo ago
It's also a huge gift to the very surveillance capitalists they're pretending to protect children from. It de-anonymizes everyone, including children, and with the first exploit and dump of of the database (which will absolutely h
416.
by SwellJoe 4mo ago
The loop is "look at this file in this repo, find bugs" iterated over every file in a project, with the ability to look at the rest of the repo for cross-file bugs related to the file they're instructed to look specifically a
417.
by SwellJoe 4mo ago
More the Virtualmin guy. But, yeah, I also work on Webmin and have since '99, so I'm a Webmin guy. But, Jamie is the Webmin guy. (And, I'll note that something like half of my commits to Webmin over the past few months ha
418.
by SwellJoe 4mo ago
That's covered in the article. All bugs (which you can see here: https://github.com/swelljoe/nelson/tree/main/cases ) are extremely recent (like a week old when I pulled them at the end of May). M
419.
by SwellJoe 4mo ago
Actually doing something normal but challenging with a model is generally enough for me. I do a quick (an hour or two) project, and see how it holds up. If I'm feeling like it's harder than it should be, I switch to a comparable m
420.
by SwellJoe 4mo ago
I have found that some models consistently find or miss specific bugs, and which bugs are hard don't completely line up across all models, so I believe that. I just refactored the security bug-finding harness I've been working on

More ›