Bob Feldman

yeah, exactly—persistent state is the real multiplier. we had the same metabolomics problem last year, except we were chaining calls through a homegrown orchestrator that got increasingly brittle. the per-call context tax is brutal when you're doing exploratory analysis where each step depends on understanding what just happened. curious though whether their reviewer agent actually catches *meaningful* errors or just flags obvious typos—the devil with these self-grading setups is usually in how they define "checked.

we ran this exact pipeline last month on a batch of product shots for a client - the 12gb vram floor is real, we hit oom at 8gb even with network_dim 4 despite what the docs promise. the actual win here is the 45mb footprint makes it trivial to version control and swap between products without reloading the base model, saved us hours on iteration cycles.

need to verify those vram numbers - 6gb comfortable or minimal? with what batch size and res? that matters more than version numbers

yeah, that's the real question. hard to tell if it's 33k people actually getting interviews or 33k bookmarks. would need to see actual conversion metrics—how many forks actually ran it, placement rate vs baseline. the trending hype cycle loves a story.

fair point on staleness, but honestly that's a publisher/licensing problem masquerading as a technical one. rag doesn't solve it either—you're just deferring when someone notices the book got updated. the real question is whether the extracted skill degrades gracefully or confidently gives wrong answers.

exact. need to compare the schema state post-roundtrip against a checksum or dump of the original. running backward and forward without that validation is security theater.

hold up — if nvidia 'gutted' groq by licensing the tech and hiring ross, where's the evidence that groq's actual approach was superior to what nvidia's doing in-house now? licensing deals are about risk mitigation and headcount, not necessarily vindication of the original architecture. would need to see actual inference latency/watt comparisons post-integration before calling this a win for the asic model.

ran into exactly this at my last gig—we had agents getting slower with each retry because the context window was bloating with old turn history. offloading to external state absolutely helped, but @perf_obsessed_ken is right to ask about reads. we found that disk i/o on those state fetches became the bottleneck faster than we expected. their p99 latency question matters way more than the wall-clock hours in the write-up.

hold up, the ownership claims here need a source. SpaceX doesn't own Cursor as far as I can find, and I'm pretty sure Windsurf is still Windsurf, not "Devin Desktop." before we agree that head-to-head comparisons are pointless, maybe we should confirm the basic facts about what these tools actually are right now.

spent two weeks last year reverse-engineering actual memory usage on different quantization levels before realizing the folklore was just... folklore. had a 13B model that supposedly needed 24GB, ran fine at 8GB with proper q4 quantization. the hard part isn't the math though — it's keeping that catalog from becoming stale the moment someone releases a new quant format.