Taking advantage of today's favorable market conditions, I filled in the pit dug two days ago - the final chapter of decentralized reasoning Where is the PMF for decentralized reasoning? I see one, not necessarily the only one, but at least I think it is the current Make Sense one, which allows civilian grade graphics cards such as the 4090/MacBook Pro to participate in "big model" reasoning, rather than reasoning for small and medium-sized models This solves the hidden line problem I mentioned last time The remaining visible lines are mostly technical engineering issues. In the previous comment section, the owner @ thecryptoskanda revealed the grassroots nature of IO technology outsourcing, but this problem is not unsolvable. A reliable technical team is likely to be able to solve the problems of visible lines (stability, dropout rate, etc.) And the matter of the secret line requires black technology, which is no longer a matter of technical expertise or unprofessionalism, nor is it a so-called 'better' thing http://IO.NET ”Achievable Who created this black technology? - @ Gradient.HQ, the AI project team I talked about in my first post Why did you ask them to chat? Two reasons One reason is that my former colleague @ JoySong_J, a beauty from ABCDE, gave up the opportunity to join a big unicorn public chain after disbanding and chose Gradient, which was far less famous than that public chain in my opinion at the time. I was a little curious at the time, but I never delved into the project's purpose. All I knew was that it was a DeAI Until a few days ago, I saw Qwen and Kimi reposting a post about the Gradient inference engine Parallax, and a Web3 AI project was officially liked and reposted by two top AI models in China? What the hell? This kind of thing has hardly happened in our circle before You can see Kimi's original post - "I'm glad to collaborate with Parallax @ Gradient.HQ to bring the open-source model Kimi K2 to more people. With Parallax, you can run Kimi K2's trillion level MoE on Mac and PC clusters. No need for enterprise grade GPU Do you know how awesome Kimi K2 is? Just released news today - K2's just released Thinking version punches GPT5, kicks Anthropic, training costs only $4.6 million, AI circle explodes again So I found Founder @ 0xEricYang through Joy, talked for two hours, and finally figured out what they were doing. However, I felt that it was not something that could be explained in just a few words to make everyone understand what they were doing. Coincidentally, the privacy race took off a few days ago, which laid the foundation for the first two rounds Technically speaking, it is quite complex. You can simply understand it this way - the sharding technology that ETH wanted to do back then was approximately implemented by Parallax in large-scale model inference. Slice the inference process of the large model into multiple small "micro tasks", with each graphics card only responsible for computing certain layers. Finally aggregated together There are too many black technologies inside - how to cut through the process? How to schedule and allocate after cutting? How to optimize node upload communication (20-50M bandwidth environment) after calculating? How to aggregate the results? How to prove validity through encrypted signature and zero knowledge proof (ZKP) after aggregation is complete Without elaborating on the details, those interested can conduct their own research. The recognition of Qwen (Alibaba) and Kimi (Dark Side of the Moon) basically indicates everything, and they should be the first in the industry that I know of. Compared to the leading investors in Multicoin, Pantera seems less important (look at IO's investment institutions, don't be superstitious about VC) Finally, let me say three things 1. The advantages of decentralized reasoning like Parallax - Kimi's post makes it clear - Bring open source big models to more people. This gives decentralization vs. centralization various advantages such as autonomy, censorship resistance, privacy, single point of failure, and of course, a greater price advantage 2. The disadvantages of decentralized reasoning like Parallax are equally evident, and the efficiency and speed are inevitably not as fast as centralization, which is simply certain. But considering the issue of power shortage that I mentioned in my previous post - in the next few years, there may be a situation where there is not enough electricity, and centralized big model inference at the data center level will inevitably make resources scarce and more expensive in the case of insufficient electricity. The price advantage of decentralized inference has become a very attractive trade off. In the past, you had no choice but to have centralization, but now you have a choice There is one more thing you need to consider - this is also one of the logical foundations for Huang Renxun's three scaling laws, which suggest that there will be a 1 billion fold increase in inference in the future. The future may be an era dominated by Agent Economy, where all actions and decisions of Agents are based on inference, with a frequency and density that is hundreds or thousands of times higher than the current use of GPT by humans. The SEO of search engine optimization is shifting towards GEO (Generative Engine Optimization) for LLM and agents. The current payment system is designed for humans, so we have x402 for agents. Will the sensitivity of inference for agents in the future be much lower than that of humans in terms of performance? There may be a perceptual difference between humans waiting for a reasoning result for 1 second and 2-3 seconds. An agent that runs continuously 24 hours a day may not care so much. If the results are the same, and decentralized reasoning costs can be 70% or even cheaper, the agent is unlikely to care about waiting for an additional 2-3 seconds for each reasoning result? 3. The Parallax inference engine is only half of Gradient's black technology, and its other half's black technology in decentralized training is no worse than Nous and Prime Intellect that I mentioned earlier. It is very likely that you will see neutral models trained by Gradient distributed training appear next year (I feel that Gradient is most likely focused on the post training RL or SFT stage, rather than pre training, as I mentioned in the second bullet of the previous post). But this is not within today's discussion topic. I am not as clear about the PMF and business prospects of decentralized training as I am about decentralized reasoning. If I see something different next year, I will share it with everyone In summary, after watching and discussing over a hundred so-called DeAI projects in the past three years, Gradient is one of the few that I think or hope they can keep running. To be honest, I don't care too much about how many B FDVs they will open in the future. Instead, in the cryptocurrency industry, where 99% of projects either fall into narrative or can only be considered high within the industry, I can see a Web3 AI project that has gained recognition from top Web2 companies in terms of technology and has the potential to find a combination of traditional world and Crypto PMF in the future. It's really difficult Just one word - live on!