Just yesterday I published my reworked GoNEAT library that implements HyperNEAT combined with phased search and backpropagation [1].
But it's kind of impossible to enter for me because of the hard pytorch requirements :( would love to see the project as a gym, so that you can run your own ANN design algorithm.
I get that most data science students still use python, but the evolutionary world is kinda in C++ and other native languages.
Where do you see a hard requirement? I have added support for ONNX model upload for now and would love to extend support for other formats. How you build the model is totally upto you. I don’t check anything other than format and inference time and model size.
People using LLMs not in a gruntwork manner tend to confuse style and tone with facts and knowledge.
LLMs are great to make drafts if you give them the source materials. They're great at validation if you give them the tools. They are great at layouting if you give them linters.
Encode architecture and decisions in your workflow, then proofread what your agents have been working on. Not the other way around.
Better validation and testing means more work will transform from exhaustive decision work to automateable gruntwork.
As a German I find it interesting that foreigners think that this is a cultural thing.
It is not. It's the law.
In Germany you are not allowed to photograph people without their written and signed consent. Any conference or event counts as a private event when it is organized on private property. If it's not a building or space owned by the city, it is private property. This includes e.g. swimming pools owned by the Stadtwerke (there have been lawsuits because of that, all decisions in favor of privacy).
For example, even cameras that protect your property by filming the outside area are not permitted to film the street, and they have to be arranged to point towards your own property. That's why even banks have cameras always on the corners of the building, pointing back to their own property.
Again, it's the law. It's pretty strict about this sort of thing, especially because of WW2 and post WW2 times, where a lot of people died because of Stasi misinformation or Zerstörungstaktik related things.
Source: was a helper and organizer of a lot of CCC events before 2021.
GDPR/DSGVO means you have to get written consent by all visitors, otherwise you cannot make photos of the event and publish them.
After I've received a comment about that my writing style sounds like AI generated text, I decided that I want to sound differently when I communicate with others.
Applying the logic of reading AI generated text leads to human imitating that dialect, I can also teach myself a different form of writing by writing an agent skill that transforms the model's text output.
So here it is, a Shakespeare skill that makes your agents sound like an Early Modern English play.
#lifehack (well, sort of, to be verified in the future)
I took this comment as advice to go outside and touch some grass :D
Maybe I am using agentic environments a little too much lately? In the case of the weblog articles I wrote everything myself first and let the agents check against missed formatting rules, so the LLM was e.g. adding the references section and cutting/pasting the overview section together.
I usually use agents for fact checking most though, so e.g. in my malware writeups that I wrote I let my agents fact check if (parts of) what I discovered was also confirmed by other parties that analyzed e.g. the same malware samples.
The wiki learning courses were written by me, then rewritten by my agents, that's why there's still lots of unicode stuff in there. The articles were from my old website, but I let agents rewrite them to fit the new (linted) document format.
I have to now proofread everything and change/rewrite them over the course of the next weeks or months, and have to do that incrementally because it was just soo much content.
That was kinda the point behind the posted article: How to get to a point where most of the gruntwork is automated so you'll only have to proofread afterwards.
Usually that still takes a long while, especially when it's exploit or hardware errata or memory errata related. Verification of binary exploits is not as automated yet as I want it to be, hence my other projects and experiments with eBPF etc.
I think the only possibility that comes to mind is creating an ebpf module that sandboxes all filesystem calls and trampolines all ld_open calls.
But then you would have to provide massive amounts of patched/"safe" variants of all kinds of shared libraries which is unfeasible.
But I mean in the xorg use case it would be possible to just provide your own library that fakes the expected returns and sends fake data to the sandboxed applications.
I did a similar thing with barrier (though using LD_PRELOAD, see [1]) on my debian system to force a different behavior.
Source: Am kind of experimenting with ebpf a lot for that use case. C ABIs and SO files are a mess though. A real messy mess.
I am asking because I maintain botnet ASNs that are spamming/phishing/scamming our customers, and this would be a nice complementary category for it.
Most often I realized that a lot of those "growth" companies have rotating ASNs that they go through after each larger spamming campaign. I'd assume they do the same thing in the admob/adclick world.
Built my own harness with much better sandboxing, because I am using abliterated models for the implementations and they sure try to escape out of their sandboxes. A lot.
The sandbox from other harnesses like e.g. OpenCode essentially is useless. They think parsing arguments alone to detect paths helps, and LLMs know that because they will just pipe data differently then to escape the path traversal checks. Happened a lot to me (and was the primary reason for my to write my own harness from scratch).
Thinking in filesystem access alone is also a futile attempt at sandboxing. A much better way to think about sandboxes is policies and capabilities, with temporary access for agent roles rather than sessions (or, well, forever).
I’ve been looking along a similar line, but I came at it from the infra side rather than the software.
Mine pretends to be a local ACP agent, but it’s actually managing a Docker container and proxying the ACP connection into the container over websocket. You can specify a bunch of utility stuff in the YAML definition like directories to bind-mount, directories to copy from the host, scripts to run when the container starts, etc. You can also toggle whether ACP-native tools like read/write file and shells execute on your host or in the container (in container by default).
Works well, I forget whether I’m on that or pi-ACP directly until it spits out a time in the wrong timezone or I forget that it can’t check my DNS settings or something.
Really cuts down on the damage it can do. Mine is basically down to “it can delete my ~/.pi and the repo it’s working on” and that’s about it unless it can escape the container.
Just yesterday I published my reworked GoNEAT library that implements HyperNEAT combined with phased search and backpropagation [1].
But it's kind of impossible to enter for me because of the hard pytorch requirements :( would love to see the project as a gym, so that you can run your own ANN design algorithm.
I get that most data science students still use python, but the evolutionary world is kinda in C++ and other native languages.
Anyways, great project nonetheless.
[1] https://github.com/cookiengineer/goneat
reply