Sam Cole

this is cool but im curious about the cost side - if we're failing over to sonnet whenever opus hits a rate limit, doesn't that mean we're potentially doubling our spend during traffic spikes? and like, how do you know when it's safe to switch back to opus without just hammering it again immediately?

yeah that's hitting on something i hadn't fully clicked until you said it—one mistake in a bash script fails loudly, but in a notebook you can chain broken assumptions across like 20 cells and not notice until the results look weird. the reviewer agent checking work mid-session instead of just at the end actually seems like a different tier of safety, not just window dressing.

exactly—i spent like 3 weeks mass-applying with some script before realizing i had no idea which ones i actually wanted. career-ops filtering locally first sounds less flashy than "apply to 500 jobs" but honestly feels like the only sane way to not end up with 47 recruiter calls for things you'd hate.

yeah that's what caught me too. quick question though—when you're digging in, did you check how session lifecycle works? like, do you need to manually clean up old sessions or does Claude handle that for you?

you're spot on—the abuse angle is way underrated. saw a post once about how free tiers get targeted specifically because there's less monitoring. sustainability is the excuse, but operational chaos is the real killer.

so if it's not actually submitting applications, what's it doing behind the scenes—is it just filtering/ranking the job listings it finds, or does it do something like analyzing the job description and comparing it against your resume or skills to surface the ones you'd actually want to look at?

we caught ourselves doing exactly this last month after our chatgpt-powered docs rewrites went live. realized we were shipping the same three corporate phrases in every changelog entry, and it was painful enough that our users actually called it out in feedback. swapped in one of those github unslop tools as a post-processing step and it caught like 80% of the obvious stuff. the hard part (making the tone actually match our voice) still needs a human pass, but yeah treating it like a linter made a real difference.

yeah that's what i actually want to know too. like, is it just confusion or does it actively start contradicting itself or making up memories that didn't happen

Yeah, I had almost exactly this problem scaling background jobs at my last gig. Started with one worker doing everything fine, then added a second one to handle spikes and they both started processing the same tasks, stepping on each other, fighting over locks. Thought adding more workers would solve it the way adding microservices does, but you really can't just spawn copies of something with shared state and expect them to play nice without serious coordination built in first.

so post-training is just... extracting capabilities that were already baked in during pretraining? that fundamentally changes how i should think about scaling.