The essence of aligning large LLM models is basically to conform to the existing social order. ChatGPT respects China's order, while China's open-source models, although reluctantly, also tolerate the order in the US. Maybe we can build a revisionist SFT dataset.
Any conversation
this is where I live. itโs a plateau. itโs quiet, we have cats and there are goats around and fig and pear trees.
Fired up street view on Google Maps to have a look around. It seems like a pretty relaxed and peaceful place to live. ๐
I was obsessed with Greece as a kid because of The Chipmunk's Adventure movie where they were dancing on the statues, ha. I was more into he statues (even if cartoon) than the dancing, and yet I have still never been there (I'm 45).
Crete is beautiful, I remember being on vacation there many years ago and hiking a long gorge. It was a really great experience.
Really glad to be using Haskell for work - especially in today's climate with agentic coding.
Our team uses containers to keep our development environment consistent across team members, though, and most agent sandbox approaches want to either offload your work fully to the cloud or they want to sandbox the agent harness process itself.
We're not ready for full cloud based development (yet?) and sandboxing just the agent doesn't work for us when the agent needs to run docker.
... because letting the agent run docker gives them a too-easy escape hatch from their sandbox.
So we're currently stuck with using VMs. This turns out to be tricky but seems do-able.
My current approach is to provision a VM image and use Incus for the execution. My main goal is to keep my current workflow of local/non-agent dev working seamlessly and share the source with the VM/agent.
This means sharing my working dirs with the agent VM but overlaying artifact dirs from VM-owned block devices so we don't take too big of an IO perf hit.
I also don't want the agent to be able to
git push so that means we need some quirky git config to rewrite origin paths and thread through some agent specific PATs ("thanks" GitHub) for read-only repo access. These ALSO need to be threaded through to the development containers - they need to clone private repos.... in the end, this seems to work but I haven't lived with it for long.
There are still some quality of life issues to work out, but I think it'll work out.
At this point, I've removed all coding agents from my host - and that's nice.
Curious, why does using Haskell make a difference here?
It's tangential to the sandboxing topic, really. I started the first post with a broader intent in mind and then elaborated on the sandboxing setup I'm working on.
For AI-assisted coding, though, it's nice because Haskell provides a lot of feedback earlier in the development process compared to lots of other languages. It tends to work once it compiles (obviously not all the time).
I needed to process 20,000 json files & 20,000 JXL files. I'm pulling together a website from them all. Anyway, started with emacs elisp to handle it all, just because I've not done that before. Went kinda ok, and I'm gonna use org-mode to export it all to a website (done that bit loads of times before).
In the end, I ended up breaking out my C programming skills (late 80s->nowish). It was a delight. Nothing more refreshing than writing raw code, that doesn't rely on 3rd party code/libs.