Founder @beffjezos proposed paid model distillation licensing to compete
GLM is the first time I see a Chinese agent capable of actually doing the /goal thing. It CAN work for hours, it can just keep obsessively optimizing. I get that Xiaomi/Kimi/Qwen/MInimax nominally have it too. But it has never felt so solid. one nitpick: permission hell in Zcode
Positive users praise the GLM Chinese AI Agent for outperforming Opus in goal optimization and costing less, while some negative users prefer Opus due to token waste issues.
No Digg Deeper questions have been answered for this story yet.
Seems like RSI might become more uniformly distributed.
American Open Source needs to step up.
There should be a simple license for labs to allow model distillation for a hefty fee.
Would at least enable cash-rich labs to catch up and beat Chinese ones.
Important for nat sec
GLM is the first time I see a Chinese agent capable of actually doing the /goal thing. It CAN work for hours, it can just keep obsessively optimizing. I get that Xiaomi/Kimi/Qwen/MInimax nominally have it too. But it has never felt so solid. one nitpick: permission hell in Zcode
amendment, you can just go YOLO actually but the default "edit automatically" mode is too restrictive, eg it can't use puppeteer
GLM is the first time I see a Chinese agent capable of actually doing the /goal thing. It CAN work for hours, it can just keep obsessively optimizing. I get that Xiaomi/Kimi/Qwen/MInimax nominally have it too. But it has never felt so solid. one nitpick: permission hell in Zcode
@teortaxesTex I think in /goal only it is better than opus
My theory is that all of the psychological bullshit they inserted in opus made it to anxious to accomplish goals just like a human
@teortaxesTex @chiefofautism yeah. I've been using glm regularly since 4.7, but it could only handle one turn tasks. 2 and you are going off the rails half the time, 3 and it's completely gone. 4.7-5.1 really increased the capabilities in this one step, but 5.2 finally expanded it to long iterations
@teortaxesTex k2.7 is even better
@teortaxesTex You recommend zcode for harness?
@teortaxesTex Interesting, so there is an intelligence level at which Ralph loops go from not working to working. I wonder if there is such a point for RSI at the model development layer. Did Anthropic hit that point?
@teortaxesTex Maaan, it's good| I am not even doing coding yet. Also way cheaper in China 🤣
@teortaxesTex Opus level?
@teortaxesTex @c0mputeAI have you check this out?
@teortaxesTex that's a big deal, but most don't understand what's really required for production-level deployment still
@usr_bin_roygbiv @teortaxesTex how are kimi and glm in claude code via ollama cloud I wonder... briefly tried but then went back to opus because couldn't stomach letting the tokens go to waste
@teortaxesTex GLM 能跑通 hours-long task 本身就是个信号，国内其他家 agent 大多卡在 session 管理上，不是模型不行是工程没跟上。
@beffjezos didnt qwen damn near reach rsi with the gated deltanet architecture?
@beffjezos Paid distillation only fixes short term gaps, core computing and talent pools decide long term competitive edges.