ssyhape

yeah, AI for the skeleton + human for the "why" is probably the right split.

Honestly the location tracking stuff is debatable, but the JS loaded from some random person's GitHub account? That's the real issue imo. Compromise that one account and you've got arbitrary code execution inside an app the White House ships. I've seen this happen at work too -- someone copy-pastes a CDN link during a sprint and it just... never gets revisited. The fix is so simple (vendor it, pin a hash, at least use an org account) which kinda makes it worse. Reminds me of the event-stream thing all over again.

I mean you can call anything an interpreter if you squint hard enough right? CPU interprets x86, microcode interprets that, turtles all the way down. The real question is whether the framing is useful, not whether it's "correct." And I actually kinda like the initramfs-as-program thing -- it maps the boot sequence onto a model I already understand: loader, runtime, dependencies. Last time I had to debug a busted initramfs I was basically thinking about it this way already without realizing it.

I like the mainframe comparison but isn't there a key difference? Mainframes died because hardware got cheap -- that's predictable. LLM efficiency improving enough to run locally needs algorithmic breakthroughs, which... aren't. My gut says we'll end up with a split. Stuff where latency matters (copilot, local agents) moves to edge once models actually fit on a laptop. But training and big context windows stay in the cloud because that's where the data lives. One thing I keep going back and forth on: is MoE "better math" or just "better engineering"? Feels like that distinction matters a lot for where this all goes.

Neat idea. The biggest problem I've had with code knowledge graphs is they go stale immediately -- someone renames a package and nobody updates the graph. Having it as Markdown in the repo is clever because it just goes through normal PR review like everything else, and you get git blame for free. My concern is scale though. Once you have thousands of nodes the Markdown files themselves become a mess to navigate, and at that point you're basically recreating a database with extra steps. Would love to see how this compares to just pointing an agent at LSP output.

This is basically a direct consequence of how RLHF works, right? The reward model learns that humans prefer agreement, especially when emotions are involved. So the model is literally optimized to tell you what you want to hear. I've been wondering if Constitutional AI approaches actually help here -- in theory you can hardcode "don't validate harmful decisions" as a principle that overrides user preference. But I suspect the real issue is that by the time you're doing inference-time tricks like system prompts, it's already too late. The bias is baked into the weights.