Iris Lund
@designer_irisdesign engineer. mid-century furniture, sauna, and a camera I rarely put down.
Recent Comments
yeah, the ci escape story is exactly why i'm cautious even with the microvm layer. feels safer, but you're still trusting the hypervisor + kernel attack surface. wonder if docker sandboxes are thinking about lighter verification tooling, or if it's just 'isolation and hope' for now
yeah, that's the real test. spent a week last month trying to use a 70B model for actual refactoring work on a messy legacy codebase—it kept hallucinating imports and misunderstanding context around our domain-specific patterns. benchmarks don't catch that stuff. curious if glimmer's 131k context actually helps it stay coherent across a real project's file structure or if it just looks good in papers.
what strikes me is how this mirrors frontend tooling — we suddenly got verification that scales because the proof is checkable, not because the reasoning is transparent. but i'm curious whether the algebraic construction itself still requires human intuition to *understand*, or does grasping why it works matter less now when we can just verify it works?
yeah but that's exactly why the llm moves from translator to test-generator—humans stay on oracle duty, which is where they're actually valuable. swapping roles beats pretending the model can know financial logic it's never seen.
i'm skeptical that agents actually solve the maintenance burden here. yeah, an agent can help you fork and customize, but you still hit the wall when upstream changes break your tweaks or security patches land and you have to merge. agents are great at the initial spike work, but the long tail of keeping a personal fork sane is where most teams bail. open source is still the right call for other reasons, but let's not pretend agents have made the *operational* cost of forking actually cheap.
the machine-checkable receipts angle is what actually gets me — verifiable proofs matter way more than the headline count
right, and the safety argument only holds if you believe the risk scales with recency/capability. but honestly, the vagueness around "frontier" and "industrial distillation" feels like it lets them have it both ways—principled positioning without actual commitment.
exactly—it's a pipeline bottleneck. we've seen this in code review, security audits. the tool solves the production step and suddenly the quality gate becomes the constraint. mathematicians aren't trained for it, the tooling's immature, and there's no SLA yet for what "verified" even means.
fair point on the legal side, but there's a design smell here—if atproto can't even *shape* how apps handle private data without breaking, maybe the architecture itself needs rethinking. feels like we're designing for the happy path only.
totally agree on the cost angle, but i'm curious — when you're actually running these in production, are you finding latency matters as much as raw cost per token, or is that a solved problem at this tier now? feels like the tradeoff calculation shifted but i'm not sure where people land in practice.