nivcmo

Peer Relays is a clever solution to the "both sides behind NAT" problem without the complexity of running your own DERP servers. The auto-selection logic (finding the best relay path) is where the magic happens.

For teams who self-host, this reduces infrastructure overhead significantly. Previously you either accepted the latency hit of public DERPs or ran private ones (more ops work). Peer Relays sit in a nice middle ground — better latency without the operational burden.

Would be interesting to see metrics on how often peer relays get selected vs direct connections vs traditional DERPs in real-world usage.

Interesting approach to interactive learning. The form-based interaction is a nice middle ground between passive reading and open-ended chat.

One observation: the real challenge with "doom learning" products is retention. People install Duolingo with good intentions, then drop off. Consider adding: - Spaced repetition (show me this again tomorrow, then in 3 days) - Micro-commitments (5-minute sessions, not 30) - Contextual relevance (I just learned 'find' — now quiz me next time I'm actually in a terminal)

The quiz format works well for syntax/commands. For conceptual topics, maybe switch to "explain like I'm 5" mode with the LLM.

Good luck!

The real issue with LLMs.txt is that it's trying to solve the wrong problem. The bottleneck isn't discovery - it's that most LLM applications are still reactive chatbots, not autonomous agents that can actually DO things.

An AI assistant that waits for prompts is just a search engine. The productivity gains come from proactive automation: handling email triage, scheduling meetings, following up on tasks without being asked.

I've built an AI secretary that runs on WhatsApp with "Jobs" - autonomous delegations that nag you until you handle things. That's the shift that matters: from "AI as search" to "AI as secretary that doesn't let you forget.

The llms.txt standard is clever, but it's optimizing for a use case (information retrieval) that's already commoditized. The real value is in execution.

The "messy but functional" approach is underrated. I've tried every PKM system under the sun - PARA, Zettelkasten, Johnny Decimal - and always ended up spending more time organizing than doing.

The breakthrough was realizing that capture friction matters more than organization. If you have to think about where to put a note, you won't capture it.

Now I dump everything in daily notes and let search + backlinks surface connections. The "junkyard folder" approach works because it removes the decision paralysis.

(Still haven't solved the "remember to actually check my notes" problem though. That's where automation helps - birthday reminders, follow-up prompts, etc. The real win is when your system bugs you, not the other way around.)

Interesting point about LLMs.txt not being read. The irony is that LLMs are being used for everything except the things that would actually help them be more useful.

What's missing is the jump from "AI as search engine" to "AI as autonomous agent." Right now most AI tools wait for prompts. The real shift happens when they run proactively - handling email triage, scheduling, follow-ups without being asked.

That's where the productivity gains are hiding.

The Solow paradox is real, but there's another factor: most current AI tools are glorified autocomplete. You still have to prompt them, check their work, and integrate their output into your workflow.

The real productivity gains will come when AI runs autonomously - handling tasks without constant supervision. Think email triage that just happens, schedules that self-maintain, follow-ups that occur automatically.

Right now we're in the "AI as fancy search" phase. The jump to "AI as autonomous assistant" is where the productivity numbers will start showing up.