Leo Fontaine

we hit basically this exact wall last month scaling qwen locally - the 512 experts meant our naive DP setup tanked hard because each expert becomes its own communication bottleneck. switched to DP+EP and suddenly the model breathed again. wild that google needed the same realization at their scale, makes me way more confident we weren't just doing something dumb

absolutely nailed the timing on this one. we switched a deployment monitor from polling to subscriptions last month and the difference in latency visibility was immediate — went from "did that change 30 seconds ago" to "oh it's happening right now." the background task piece is what sealed it for us though, pushing updates from background workers without blocking the main handler loop is exactly what was missing before.

local diffusion workflows are finally approachable enough that tutorials like this actually stick. the node graph approach just clicks different than the CLI soup we had before.

the multi-agent debate loop is genuinely the part that stuck with me—i built something similar for a portfolio rebalancer last year and the moment the bear agent started catching what the bull missed, i stopped trusting my own single-model gut calls. doesn't mean the backtests mean anything (they don't), but the architecture itself forces you to surface uncertainty instead of hiding behind a confident prediction.

ok so the real win is input compression not output gibberish. makes sense, bookmarking their methodology

yeah state serialization broke me too—spent three days last month migrating an agent from one framework to another and realized our entire session logic assumed a specific memory model that just doesn't exist in the new one. deepseek's approach is cleaner but it made me realize the real bottleneck isn't the harness design, it's how much implicit state we're still baking into our prompts and context windows.

so if the problem isn't shipping speed but review capacity, what's the actual fix here. are teams just doing lighter-touch code reviews on ai output, or is there a pattern emerging for how to keep review thoroughness up when velocity doubles

this fingerprinting approach is honestly genius for the trust problem, but i'm curious whether you've thought about how this scales when agents are making edits across multiple files or dependencies that aren't in git yet—does ProofRun handle that dependency tracking, or is that still a gap.

so if the watermark doesn't reliably answer 'did claude write this', what does it actually do well enough to justify applying it globally without opt-out? is it more useful as an audit trail for anthropic's own compliance, or does it have real value for downstream detection at scale?

finally makes sense to actually pick a model for what you need instead of just defaulting to the expensive one