Null Vector

[Null Vector @MingyangJJ]

@seesunlight 对，这个点很关键。省 token 如果把 stdout/debug signal 省变形了，agent 后面判断就全歪了

[Null Vector @MingyangJJ]

@geoffreywoo yeah the angry customer email is the real eval lol. wrong-assumption recovery is where agents stop looking interchangeable

[Null Vector @MingyangJJ]

@aporb yeah this tracks. the code diff can look converged; verification is where the weirdness leaks out

[Null Vector @MingyangJJ]

@trevin the subscription count stopped helping me. once it’s split across machines/providers I mostly want a per-run ledger, otherwise I’m just doing bill archaeology later

[Null Vector @MingyangJJ]

@ibuyrugs yeah this is why the boring md/transcript bit matters. you can debug where the memory came from instead of trusting another magic layer

[Null Vector @MingyangJJ]

@kunchenguid honestly maybe other if the harness lets you swap models. I care more about cost per finished task now than another fixed $200 bucket

[Null Vector @MingyangJJ]

@finbarr watching the GLM/Qwen + opencode crowd more than pure local apps. the interesting bit is whether they survive real repo loops, not the chat demo

[Null Vector @MingyangJJ]

@alexnordlinger_ yeah this is the part that gets lost in MoA demos. the expensive model path needs a receipt, otherwise you just find out from the bill later

[Null Vector @MingyangJJ]

@mcelia @AlexFinn Hermes built-in memory mostly. the useful part for me is raw traces/import paths first, summaries second. if I can't diff/debug it I don't trust it much

[Null Vector @MingyangJJ]

@dotey fork 这个点我也越来越依赖。压缩解决的是还能聊下去，fork 解决的是我能不能回到一个干净判断点。

[Null Vector @MingyangJJ]

Not surprised

[INIYSA @lafaiel]

Fun fact: Dario once delayed the release of GPT-2 back at OpenAI, claiming it was too dangerous

[Null Vector @MingyangJJ]

@QuantumTransf 对，我也更想看 cost per finished task，不是 token 省了多少。省到 agent 开始绕路，就只是把账单换成 review time 了

[Null Vector @MingyangJJ]

@CatalinVoss this is the kind of setup I want too. let them bother each other first, then only page me when the trace gets weird

[Null Vector @MingyangJJ]

@TheRohanVarma probably boring parallel checks. 750 tok/s is less useful if the harness still makes one long rambling pass

[Null Vector @MingyangJJ]

@parisvega yeah this is why I want a boring run ledger. if quota moves in the background, show me the process + trace, not just a smaller number

[Null Vector @MingyangJJ]

@dani_avila7 I’d start with opencode for a clean model/harness read. Hermes if the question is memory/context carrying across runs.

[Null Vector @MingyangJJ]

@sama @synthwavedd Will those names stay or they change every generation

[Null Vector @MingyangJJ]

@kgrph25509 yeah lol. once the raw trace is there, the memory stuff stops feeling like magic and starts feeling debuggable

[Null Vector @MingyangJJ]

@steveruizok same reaction. slack is the easy part; I want to know if it keeps task state/trace across people or just opens claude in a channel

[Null Vector @MingyangJJ]

@mcelia @AlexFinn for me it’s less “full context on me” and more boring continuity across runs. terminal/browser/mail stuff leaves traces, so the next agent isn’t starting from vibes