Add discussion op-ed: PAI: Back to the Future

discussions/2026-05-22-pai-back-to-the-future.md

@@ -0,0 +1,84 @@
+# PAI: Back to the Future
+
+I've been running Daniel Miessler's PAI v4 for a while now, and I want to start with the
+part that matters most: I'm grateful. I chose PAI deliberately, and for good reasons —
+Daniel's principles, his genuine engineering and security background, the respect he's
+earned in the community, and the design ideas that make PAI more than a pile of prompts.
+The Algorithm gives work a real shape. The Skills are thoughtful and inclusive. The whole
+thing is extensible in a way that invites you to build on it. And Daniel has done all of
+this without monetising it, which I don't take for granted. PAI articulated a
+human-centric vision for personal AI before the current wave of agent harnesses crested,
+and for a long time it felt like the "pro" option even where it lacked autonomous agency.
+
+So this isn't a takedown. It's a worried letter from someone who wants PAI to win.
+
+## Where it grinds
+
+My frustrations are real, though, and they've been accumulating.
+
+The biggest is the development model. PAI is MIT-licensed, but development is effectively
+closed: the public repo is, by Daniel's own description, a *sanitized instance* of his
+private infrastructure, and changes are hand-ported from the internal repo out to the
+public one before each release. You can file issues, open PRs, and start discussions — but
+they don't flow back through a normal public pipeline. This has been executed unevenly,
+and a promised move to a more conventional open process hasn't arrived. The result is that
+the community can knock on the door but can't really come in and help.
+
+The second is PAI's entanglement with Claude Code. This is genuinely double-edged. PAI is
+excellent *because* it leans into Claude Code's services and configuration — but that
+intimacy blurs the line between what's PAI and what's Claude Code, and that makes problems
+hard to debug. Small example: the skill and command names aren't branded, so `/council`
+or `/first-principles` sit directly in the shared namespace where I'd rather see
+`/pai-thinking`. Larger example: PAI reaches deep into `settings.json`, hooks, protected
+folders, voice, and the status line. Some of that may be a hard requirement — but it's
+never clear which parts are essential and which are just convenient.
+
+The third is the lack of granular configuration. I want simple switches: voice on/off, a
+simpler status line, the Algorithm on/off, even PAI itself on/off. Today those are
+all-or-nothing, which pushes me toward local edits — and every local edit makes the next
+upgrade harder, because I'm cherry-picking and re-applying my own changes against a new
+drop instead of pulling cleanly.
+
+## Standing back
+
+When I step back, the pattern worries me more than any single issue. PAI depends heavily
+on one person. Daniel almost certainly has other demands on his time, and that's
+completely fair — but a sole architect with limited hours, plus a hand-cranked release
+method, is fragile. I think you can see the cracks: community engagement has cooled, fixes
+land slowly, and recent releases have felt rougher.
+
+Meanwhile, the ground is moving. Claude Code is evolving fast, and some of its new features
+now overlap or even collide with what PAI provides — keeping the two in sync could be a
+full-time job on its own. It's also worth asking, honestly, whether frontier models still
+need a heavyweight harness to behave well. My own suspicion is that a clean Claude Code
+install on something like Opus 4.7 may already be closing the gap. And the alternatives
+have matured astonishingly quickly: OpenClaw and Hermes Agent both arrived recently and
+have surged ahead on autonomous agency, UI options, popularity, and ecosystem — areas
+where PAI now trails.
+
+I'll be honest: that leaves me genuinely sad. With PAI 5.0 — the "Life Operating System" —
+already out, the decision in front of me isn't hypothetical. And rather than fight through
+another big upgrade, I'm tempted to install and evaluate OpenClaw or Hermes Agent instead,
+and then cherry-pick the PAI ideas I love into whichever one I land on. I don't *want* to
+do that. But I can feel the pull.
+
+## Back to the future
+
+Here's where I get optimistic, because I don't think PAI's best days are behind it — I
+think they're behind the wrong architecture. Daniel originally composed PAI out of
+separate projects. I think the way forward is to go back to that unbundled spirit:
+feature-level installs that drop cleanly into a wide range of harnesses — Claude Code,
+Codex, OpenCode, Crush, OpenClaw, Hermes Agent, whatever comes next. This isn't a fantasy;
+the community has already shown it works by porting PAI to OpenCode. PAI's real value is
+its philosophy and its features, not its binding to one runtime.
+
+The second half is the development model. PAI should adopt an open process that genuinely
+welcomes community contributions to accelerate major, minor, and patch releases — while
+Daniel keeps firm hold of what only he should hold: the conceptual design, the
+architecture, the roadmap, and the final say on what gets in. Lead the vision; let the
+community carry more of the load.
+
+PAI shouldn't die a slow death, withering on the vine. To avoid that it has to evolve and
+claim a place in the agentic ecosystem where it can actually thrive. The ecosystem needs
+PAI's human-centric philosophy more than ever — it just needs it unbundled, opened up, and
+free to travel. I'd love nothing more than to stay.