Chunxi Zhang

Independent Researcher on Computational Neuroscience. Building a model of cortex.

[Chunxi Zhang @toopowerful]

@pcshipp I do it on complex task, fable don’t do my complex task due to biology restrictions, and it’s significantly better than opus. Both speed and quality, already have my 200$ Claude canceled. Not perfect, not one shot, but better than opus.

[Chunxi Zhang @toopowerful]

@ayv4zyan this is quite strange indeed. you ask it, it read news and reasoned that he must be the 4.5, but don't get the actual number internally. and the response is much slower than 4.5 in cursor/grok build.

[Chunxi Zhang @toopowerful]

@Lakr233 搞生物的也是。fable绝望了以后，本来有点期待gpt5.6的，好像也有safeguard了，grok是最后的选项了。

[Chunxi Zhang @toopowerful]

@gm365 目前我的体验是，可能受限于256k的context window，grok的极复杂问题的长时稳定性不行，可能一开始思路跟得上，跑了几个之后最初的目标或者是限制给丢了，你搞坏的问题有可能是因为这个。希望等两周后的1m版本能有提升。

[Chunxi Zhang @toopowerful]

@aehyok 全走订阅的话，其实grok并没有显著更便宜似乎。每个月都花200刀解决的问题似乎是差不多的。但opus原价token和grok比就天差地别了。o和a的订阅补贴似乎要大于grok/cursor

[Chunxi Zhang @toopowerful]

@Pluvio9yte 算了一下昨天高强度使用用了27%的月额度，20$，也就是如果没有50% off，200$刚好用一个月grok 4.5，还剩一点，还能试用一下别的模型。从功效和体验上已经吊打opus了。opus经常是弱智犯低级错误，fable用不了一点（生物相关）。似乎只用grok的话supergrok 300$好像也没有优势

[Chunxi Zhang @toopowerful]

it's amazing. good enough to cancel claude subscription, even just consider quality not speed or cost.
i do biology inspired network, so fable not usable for me. grok 4.5 in cursor is significantly better in instruction following, and have very good reasoning capbability with mechanism than opus.
it has significantly better at finding causal relationship given mechanism description and reality probe.
with opus in claude code, i found myself often correcting model for not reading document at all constantly making documented mistakes, but this didn't happen in grok 4.5.
it still deviate from main course after long iteration(bad), but compact is not sensible(good).
the speed is insanely fast, and cost is very cheap.

latex generated by model is not well rendered for equations.
and cursor reopen random tabs on restart(old bug still unfixed)

amazing work, hope you guys keeps up and DO NOT nerf the model like gpt/claude often do, they often started good and became very bad randomly.
i always prefer grok/cursor series's style and attitude but peak performance was not good enough before, not anymore. it's really the start of a new era.

[Chunxi Zhang @toopowerful]

@JaydenDavisNC in my project, it's better than opus. opus often direct disobey my command of reading relevant document and assert something totally wrong even it's well documented. not happening yet with grok 4.5. it's not making documented errors so far.

[Chunxi Zhang @toopowerful]

@isausmanov only the smartest model know who to assign to in order to save compute while complete the task. dumb model simply don't know if a thing is hard or not. that's why any model picker are all failure because they're a small model in front.

[Chunxi Zhang @toopowerful]

@Valuable it's just this dumb, so is the conciousness theory they refer to, the global-workspace theory. it's very strange that this can be described as a theory at all.

[Chunxi Zhang @toopowerful]

@imwsl90 妹子也是同款，前几个月让她淘宝买了一套键帽换了一遍，彻底焕新

[Chunxi Zhang @toopowerful]

@45_206MHz 眼前一黑肯定找个东西扶着或者放低身形啊，怎么能继续走路呢

[Chunxi Zhang @toopowerful]

@a_musingcat I cannot say this for sure, but fable never work for me indeed. No code, no experiment, no science, no paper reading nothing. All I see on internet for fable case are ui demo or game nothing else.

[Chunxi Zhang @toopowerful]

@FuckAnthropic 倒是主动推给了我呢

[Chunxi Zhang @toopowerful]

@trq212 Anything biology is blocked by design still right? Including biologically inspired learning and help reading any publicly released paper?

[Chunxi Zhang @toopowerful]

@maxhodak_ Programming make you realize how the world is not a black box only described by natural language, in contrast of any theory or knowledge I know before. Only understand something computationally is true understanding for me.

[Chunxi Zhang @toopowerful]

Debug report uploaded: Report https://t.co/MEkSDy4qdM agent.log https://t.co/0AVgQjUVSq gui.log https://t.co/vWn8beVGfb desktop.log https://t.co/Rcsk1FNzIo

the session timeouted again, then i click back to that session, it continue its run by automatically, and then i run /debug, here is the log.

[Chunxi Zhang @toopowerful]

@karry_viber @arkuy99 我三天把周用量用完了，用官方api的价格配合我token的实际用量统计来算，价值2500$。也就是200$买4周10000$的token。