Hi everyone, here's your monthly dose of sysadmin minutes... Do note an interesting discussion about LLMs, not a formal policy, just a discussion for now. # Roll call: who's there and emergencies all members present # Express check-in How are you doing, and are there any blockers? Then pass the mic to the next person. # Roadmap discussion Review the [2026 roadmap][] especially with the [STF grant][] not going through. One month left in the quarter, we might get some input from upstairs for the next. [STF grant]: https://gitlab.torproject.org/tpo/operations/proposals/-/issues/70 [2026 roadmap]: https://gitlab.torproject.org/groups/tpo/tpa/-/epics/2 Here's what everyone will be working on: ## zen - [profiles merge][] - [planning phase for Tails migration to Prometheus][] [planning phase for Tails migration to Prometheus]: https://gitlab.torproject.org/tpo/tpa/team/-/issues/41946 [profiles merge]: https://gitlab.torproject.org/tpo/tpa/team/-/issues/42103 ## lavamind - donate-neo - download page ## groente - anarcat carried the idea that email should be prioritized - let's meet with nadya to get a better idea of the requirements, anarcat will check for reusing the timeslot tomorrow ## lelutin - prometheus HA - garage conversion ## anarcat - wiki problem? experimented a lot with mkdocs/zensical for a [personal project][] [personal project]: https://lora.reseaulibre.ca/ # AI brainstorm? - anarcat: was extremely skeptical with capacity and morals of AI, now he's worried with the profession. We might lose an entire generation of programmers. Worried with reliability of software, environmental and political impact. But he's using it -- not vibe coding, rarely using it for coding -- it causes skills atrophy, but uses it for spell checking. Disclosure should be mandatory: disclose the model and prompt (eg. in commit messages). (Other messages here are unattributed, but from each team member.) - It's hard, and getting harder, to not use it: e.g. google pushes you towards it when you search. Concerned with our jobs and pressure on "efficiency", and that we might be forced to use LLMs. Could we have [RAG][] for our documentation? [RAG]: https://en.wikipedia.org/wiki/Retrieval-augmented_generation - groente: As for job security and pressure for "efficiency", one would hope there will always be a market for people that understand how systems work. The pressure to use LLMs will likely grow: compare it to household appliances (or many other technologies) that were thought to save us time, but in the end just created higher expectations. Resisting this will likely be extremely hard within a capitalist environment, but hopefully we can carve out some space within non-profits at least. It's kind of ironic, this LLM use is once again creating a huge dependency on a small number of US tech companies. There's now a lot of geopolitical momentum to decrease infrastructural dependency on the US, but at the same time we're massively walking into the same pitfall with AI. Let's hope Tor is smarter than that. The earlier push towards cloud computing was successfully resisted, so I'm still hopeful here. - We still have jobs through the cloud migration, even though we were told our jobs would stop existing. I like to code so I don't want to use AI, but all my friends are using it, and the reason I was getting crap outputs in first attempts was that I wasn't sending enough context. Concerned about people's mental health, especially after an article [shared this morning about "How to Talk to Someone Experiencing 'AI Psychosis'"][]. [shared this morning about "How to Talk to Someone Experiencing 'AI Psychosis'"]: https://www.404media.co/ai-psychosis-help-gemini-chatgpt-claude-chatbot-delu... - Not using it. Search results are getting worse and worse. Really hesitant in jumping in, fundamentally misanthropic technology. As a temporary moratorium, anarcat announced that TPA team members are forbidden to grant any sort of execution access to LLMs for machines under their stewardship. If you want to experiment with AI, you can copy-paste things, feed it input, but in no way should you execute code from LLMs without the same level of review you would give to an untrusted contributor. This applies, in particular, to agents: if you experiment with agents like Claude Code, run those in a virtual machine. # Next meeting Next month's first Monday is Easter in some cultures and a holiday in most of our workers, so the roadmap meeting will be held the second Monday of April. # Metrics of the month * host count: 95, LDAP 140 (!), Puppet 136 (!) * number of Apache servers monitored: 32, hits per second: 659 * number of self-hosted nameservers: 6, mail servers: 13 * pending upgrades: 0, reboots: 0 * average load: 1.38, memory available: 4.3 TB/6.8 TB, running processes: 160 * disk free/total: 111.3 TB/216.5 TB * bytes sent: 534.4 MB/s, received: 362.1 MB/s * [GitLab tickets][]: 272 tickets including... * Needs Triage: 1 * Not Scheduled: 164 * Backlog: 66 * Next: 29 * Doing: 12 * (closed: 4416) * [~Needs Information][]: 3 open, 125 closed * [~Needs Review][]: 7 open, 201 closed * [~Technical Debt][]: 10 open, 38 closed [Gitlab tickets]: https://gitlab.torproject.org/tpo/tpa/team/-/boards [~Needs Information]: https://gitlab.torproject.org/groups/tpo/tpa/-/issues/?state=opened&label_name[]=Needs%20Information [~Needs Review]: https://gitlab.torproject.org/groups/tpo/tpa/-/issues/?state=opened&label_name[]=Needs%20Review [~Technical Debt]: https://gitlab.torproject.org/groups/tpo/tpa/-/issues/?state=opened&label_name[]=Technical%20Debt Upgrade prediction graph lives at https://gitlab.torproject.org/tpo/tpa/team/-/wikis/howto/upgrades/trixie/ -- Antoine Beaupré torproject.org system administration