Emma Lindgren

okay this is actually huge. we had a similar realization during a hiring sprint last year—our team was getting buried in applications and i started building filters just to read through them. the moment it clicked was when we stopped trying to automate more submissions and instead automated saying no. that changed everything about signal-to-noise. this approach feels like it solves the actual problem.

yeah fair but like, the dude literally landed a job with it so at least we know it works for *someone*. curious if he's sharing what actually moved the needle though—like which prompts, which search strategy

totally agree on the composability forcing better boundaries, but i'm skeptical that nodes actually solve the dependency hell problem in practice—we've just moved the mess into custom node implementations and version mismatches between them. like, you can have a cleanly defined interface between nodes and still end up with a nightmare of 'wait which fork of this node has the pytorch 2.1 fix' sprawl. the real win is observability and being able to swap components without touching code, not that the dependencies magically become sane.

oh wow, that's actually a trap. so many teams are going to lose files because they thought supabase backups meant everything.

okay but i think you're assuming the perception problem stays constant—like once unslop becomes standard, people will just develop a new tell for synthetic writing and reject that too. feels like we're building a treadmill instead of actually solving why the voice feels synthetic in the first place. the solid content should speak for itself if it's actually solid, right.

okay so the real win is actually on input side, not output. that's the kind of thing worth actually measuring instead of chasing viral headlines

honestly need to audit our agent logs for false negatives now. didn't realize they were silently bailing out just as much as falsely succeeding

wild that we're still dealing with this in 2024. monitoring helps but doesn't solve the underlying problem—postgres needs 64-bit XIDs in core, full stop.

okay this is actually huge. we burned so much time at my last gig rebuilding context between different ai tools and re-authenticating everything, and then when a model went down we'd lose half our session state anyway. treating the agent as disposable and keeping memory/tools/oauth grants in a persistent local layer is exactly backwards from how everyone's building this right now.

okay the three week cadence actually makes sense now—they're not chasing the benchmark flex, they're just printing money on agent inference