When it comes to AI, instead of centralization, what people really need is local deployment. The so-called 'decentralized AI' is still more of a narrative today. From the perspective of user needs, there's really only one thing users care about: who can access the backend information. Sometimes, when asking GPT or Grok to handle sensitive information, like business contracts or personal confessions, it can be quite conflicting. On one hand, you feel like you should sanitize the data before showing it to AI, but on the other hand, you're too lazy to do it. I believe anyone sensitive about privacy can relate. Even though OpenAI has repeatedly stated that it won't use customer privacy data, as long as it has the permissions—even if it doesn’t use them—it still feels like a thorn in the side. Expecting OpenAI to decentralize is unrealistic. Currently, there are two main approaches: - Blind computation: Encrypting with FHE before letting AI process the data; - Local deployment: No internet connection, brute force solves everything. Recently, I saw Alibaba Cloud's Qwen official account repost @Gradient_HQ's announcement about the Parallax service. In the demo, they used 4 Apple M4 Pro computers to run a 235B Qwen large model, which is pretty exciting. Although this isn’t a dense model—it’s in MoE form, with roughly 22B parameters activated per inference—it’s definitely sufficient for handling sensitive information in daily use. Of course, most people don’t have 4 M4 Pro computers at home with that much computing power, so they’ve designed it to allow friends to co-host. In other words, 4 people with 4 computers can jointly run a local 235B large model. PS: Finally understood why Mac packs so much redundant performance Even though it builds on the foundation laid by predecessors, the Parallax service is still a strong engineering achievement, further lowering the cost of home-based local AI. In the past, I used to slightly look down on the narrative of 'household edge devices participating in AI,' but now that they’ve actually made it happen, it’s surprisingly good. In the future, sensitive information can be processed directly on local devices or within a small LAN set up by a few friends. That sense of security is really reassuring.