Monthly Update: The Door Has Closed
Capability is now an engineering problem. The real wall is access.
Honestly, lately the architecture hasn’t changed much — except that the orchestrator named Nerve keeps accumulating more and more tools in its own toolbox, like the resident Dynamic Workflow, or Codex Chrome Control for doing the dirty work.
The change is in capability: a single task running non-stop for a whole day is already the norm now. This is probably proof that every shortcoming of current AI’s capabilities is an engineering problem. I believe these tasks will soon be able to run for a week, a month, without stopping — it has nothing to do with model capability, it’s a question of whether you want to, and whether it’s necessary.
But I’m very pessimistic — not about how strong it is, but about availability. Right now Mythos, Fable, and GPT 5.6, for various government reasons, have not been released. The era when a normal user could get hold of a SOTA model is completely over. The door has shut. From now on, all you can do is sniff a little of the exhaust off the RSI tailpipe.
The only place with an exit right now is the open-source wild versions of models, e.g. GLM 5.2 Uncensored — sure, the capability isn’t great, but at least the floor is low enough. And on that front, there aren’t many vendors who can offer this. Venice.ai counts as one; still constrained by all sorts of cover-your-ass reasons, the products they offer aren’t truly uncensored, so availability is actually limited — on some harmful questions it still refuses. Even fake uncensored is in enormous demand — sometimes even saying hi means waiting in a queue. On price, 5.2 uncensored isn’t cheap either; heavy API use gets pretty expensive. After all, it’s not like an Anthropic or OpenAI 200 USD subscription, which you can squeeze 30x+ value out of in usage. Speaking of which, that 200 USD plan kind of thing probably won’t live long either — enjoy it while it lasts.
So the way to play this is private deployment, but since the budget for 8x H100 is still too high, there’s going to be a real supply vacuum here. I won’t say more — I’m off to go play first.
月度更新：大门已关
能力已经是工程问题。真正的墙在可得性。
最近说实话，架构没什么变化，除了名为 Nerve 的 orchestrator 自己的工具箱里的工具越来越多之外，比如 常驻的 Dynamic Workflow、干脏活的 Codex Chrome Control 之类的。
变化在能力上：有时候一个任务持续运行一天不停歇，已经是常态了。这大概就是证明了当前 ai 所有能力的短板都是工程问题，相信很快这些任务就能跑一个星期、一个月不停，无关模型能力，而是想不想、有没有必要的问题。
但是我是很悲观的，不在其有多强，而在可得性。目前 mythos、Fable 还有 GPT 5.6，由于各种政府原因，都没有被 release 出来。正常的用户能拿到 SOTA 模型的时代已经彻底终结了。大门已关。以后就只能抽点 RSI 的尾气了。
现在唯一有出口的地方就是开源的模型野生版 eg GLM 5.2 Uncensored，虽说能力不行但好在底线够低。所以这个方面，能提供这个的厂商不多，Venice.ai 算一个，仍受制于各种需要保命原因，提供的产品都不是真正的无审查，可用性其实有限，一些有害的问题上还是会拒绝，即使是假的无审查，这个东西需求量非常大，有时甚至 say hi 都需要排队。价格上 5.2 uncensored 也不便宜，api 大量使用会蛮贵的，毕竟不像 Anthropic or OpenAI 200usd 的订阅可以薅出 30x+ 价值的用量。说到这个 200usd plan 这种东西估计也活不久了且用且珍惜吧。
所以这个东西玩法在于私人部署，但由于 8 张 h100 的预算还是太高，这里会有一个真实的供给真空。多了就不说了，我先去玩了。