
ZHCs.ai|中国龍Loong.btc
1.7K posts

ZHCs.ai|中国龍Loong.btc
@ZHCsAI
Zero Human Company AI 经济体|ZHC研究所:https://t.co/BChhm7h5ml
SG Joined Kasım 2020
4.1K Following4.6K Followers
Pinned Tweet

The Zero-Human Company era is coming
AI 经济体|#ZHC 研究所:t.me/ZHCs_AI
中文

如果能成功落地生产, #C0mpute 将超越 Venice
ZHCs.ai|中国龍Loong.btc@ZHCsAI
I think #c0mpute is a very tasteful project. Decentralized, uncensored AI services are a massive market that is about to explode. c0mpute could be the next #Venice . solana:EmcxFTNVDqyLHp11NvwvLZ4D7LKGbG9i7B8RF7dwpump base:0xacfe6019ed1a7dc6f7b508c02d1b04ec88cc21bf
中文

去中心化 AI 最值得看的方向之一,是去中心化推理模型。简单说,就是让全球不同地方的 GPU、研究者、数据和任务环境,一起参与训练一个更会思考、更会解题、更会使用工具的 AI 模型。
过去大模型训练基本被少数巨头垄断,因为你需要 GPU 集群、海量资金、工程团队、训练数据、评测系统和 RL 后训练能力。这套东西太贵了,所以 AI 的核心能力越来越集中在 OpenAI、Google、Anthropic、Meta 这些公司手里。
但现在,一个新变化出现了:AI 训练的重点,正在从“预训练”转向“后训练 / 强化学习 / 推理能力提升”。这给去中心化 AI 留出了窗口。因为推理模型的提升,不只是靠堆更多数据和参数,还需要设计任务环境,让模型反复尝试,给它明确奖励,验证它是否真的做对,再把结果反馈回训练。这件事天然适合开放协作。
Prime Intellect 就是这个方向里很值得关注的代表。它是在尝试搭一个开放 AI 训练栈:有算力市场,有 RL 训练框架,有环境库,有评测系统,也真的发布了 INTELLECT 系列模型。最关键的是 INTELLECT-2,它展示了一个 32B 推理模型,可以通过全球分布式的强化学习方式训练出来。
未来 AI 模型的能力提升,可能不只发生在少数封闭实验室里,也可能由全球的算力、开发者、任务环境和验证机制共同推动。我觉得去中心化 AI 真正的突破口,不一定是从零训练 GPT-5 级别模型,更现实的路径是先在推理、代码、数学、搜索、Agent 任务、垂直行业模型上突破。
因为这些场景更容易定义奖励,也更容易验证结果。谁能把“任务环境 + 奖励机制 + 分布式训练 + 可验证贡献”这套闭环跑通,谁就可能成为去中心化 AI 的核心基础设施。

中文

c0mpute 这个创新,要是能落地生产,就是10亿美金的独角兽!希望他们成功,真正开启去中心化AI时代 ! @c0mputeAI @leyten solana:EmcxFTNVDqyLHp11NvwvLZ4D7LKGbG9i7B8RF7dwpump
中文

“去中心化+无审核的AI”就是最大的趋势,“模型切片 + 分布式推理 + 推测解码 ”这个技术能落地生产,就是上千亿美金的市场。 @c0mputeAI solana:EmcxFTNVDqyLHp11NvwvLZ4D7LKGbG9i7B8RF7dwpump
中文

So do we now build the torrent for AI with this or what?
leyten@leyten
Wow, it has happened! 30.55 tok/s on GLM-5.2 4-bit (from @Zai_org) ran by six RTX Pro 6000's across the USA scattered over WAN! I can't believe this. It was an insane build, you can read more about it on github.com/leyten/shard
English
ZHCs.ai|中国龍Loong.btc retweeted

Wow, it has happened!
30.55 tok/s on GLM-5.2 4-bit (from @Zai_org) ran by six RTX Pro 6000's across the USA scattered over WAN!
I can't believe this. It was an insane build, you can read more about it on github.com/leyten/shard

English

@c0mputeAI 实现GLM-5.2运行在6张分散于美国各地的 RTX PRO 6000上,通过公网(WAN)连接,利用Async Draft和Pipeline技术,实现了16.6 tok/s的推理速度。
未来AI或许不是由少数科技巨头拥有的超级集群提供,而是由全球数百万张GPU组成的去中心化智能网络提供,让整个互联网变成一台超级 AI 计算机 $ZERO
中文

We're launching code storage and git hosting.
Origin gives teams and agents a place to host, review, and collaborate on code.
Available this fall. Join the waitlist.
cursor.com/origin-waitlist
English

Kimi-2.7-Code is now live and fully private for all users on Dot, and it's another great example of how we scale.
At Dot we do not simply bolt new models to our platform. Instead, we treat every models as interchangeable execution backends behind a routing layer, so the moment a frontier-class open-weight model ships, we drop it into the right lane and it is live, routed, and fully private in minutes (check our Fable 5 integration for reference on our speed).
The release is a real step up on code and agent performance:
+21.8% on Kimi Code Bench v2
+11.0% on Program Bench
+31.5% on MLS Bench Lite
A new high-capability open model should not leak its upstream surface area into the product. It needs to sit behind a stable inference contract: normalized identity, server-owned provider params, bounded context policy, fair-use isolation, streaming compatibility, and privacy-preserving request handling.
This is how Dot scales, by making the best open models in the world interchangeable, routable, and private by construction the day they land.
English

@jun_song When will SuperGemma .AI officially launch its uncensored AI service?
English

持续看好去中心化无审核AI市场,爆发近在眼前!
@c0mputeAI @AskVenice
solana:EmcxFTNVDqyLHp11NvwvLZ4D7LKGbG9i7B8RF7dwpump base:0xacfe6019ed1a7dc6f7b508c02d1b04ec88cc21bf
日本語









