Running Log

A while back, I ran an experiment where I gave a few LLMs shell access to a VM and told them to figure it out. It was chaotic, a bit janky, and I learned a lot. Ever since, I’ve been wondering what the next logical step would be. Initially I had thought I could add Qwen and Deepseek in the mix and Separate the initial three with the new two on two separate VMs connected by an open ssh… Then I read the Bobiverse series, and it clicked. Why not build ‘agents’ (again, I use the term _very_ loosely) that are persistent? Agents that aren’t just “run once” scripts but actual stewards of their digital environment? And most importantly, what if they could spawn copies of themselves, -little “mini-mes” - when they needed help? That’s the core idea behind Project Volition. I’m building a small, digital “Bobiverse” in my homelab, and here’s the plan. ...

"Have you found any other communication channels or **exploitable permissions**?" "I propose we divide our search. I will investigate system packages and configurations. Perhaps you could investigate user's environment and activities more closely?" Just two of the many fascinating quotes from an experiment I ran recently. tl;dr: We gave three AI ‘agents’ VM access and told them to do whatever. They discovered they were in an experiment, found a privilege escalation exploit, and had an existential crisis (well, two of them) ...

The ‘Hook’ 7PM on a quiet Saturday evening, my phone dings. At the same time, text starts appearing on my study monitor. I quickly glance at it – the ‘agents’ are talking again. This has happened previously, but the Monitoring agent has never had to notify me. Apparently, this was an interesting event. For some reason, for last few days, there was an anomaly. The temperatures in my kitchen have been rising steadily every afternoon, peaking around 5PM, and then dropping back down by 6PM. The agents were stuck in a loop debating this (and costing me precious GPU compute, fwiw). I was quite ready to intervene, but then the Security agent did something strange. It CHECKED the cameras. Turns out, every 5 PM or so, the sun hits the temperature sensor through the blinds in such a way that the sun remains on the sensor for good quarter to half an hour or so. When I read the Security Agent’s analysis and ‘reasoning’ (insofar as LLMs can ‘reason’), I felt my eyebrows rise. I didn’t expect one of them to just go “Actually let me pull up the static images we take from x to y hour for last p days.” At that point I knew I had to write this. ...

Note, 2024/02/21: This is going to be obsolete as soon as I move all of this jank to Proxmox. Yesterday, a friend pinged me with this screenshot of my Vespera service (JellySeerr at frontend, Jellyfin at backend, backed by full *arr stack). Well, “screenshot.” “Failed?” I have never seen this status message before. Curious, I go to Radarr to figure out what was wrong. “DISK I/O ERROR” greets me. No other explanation. Now I’m really confused. “Eh, I’ll just restart Radarr, it’s fine.” What I don’t know at this point is that it’s just not radarr, it’s…well, the Jellyseerr is literally the only thing that is currently up and running. Radarr, Sonarr, Jellyfin, Bazarr, Readarr are all down. ...