与「MOON」相关的搜索结果

Elon Musk: “Path to Petawatts is Mass drivers on Moon” SpaceX’s Lunar Mass Driver is a planned electromagnetic launch system on the Moon, similar to a giant coilgun-style catapult, designed to launch solar-powered AI compute satellites into orbit or deep space at extremely low cost

Sam Altman asked Elon Musk what he'd do if he were 22 again. Not "build Tesla." Not "start a rocket company." Two words: be useful That's the whole framework. It sits on two questions almost nobody asks themselves. How much better is your solution than what already exists? How many people can it actually help? Multiply the two. That's your real impact. Every company that has ever changed the planet started by solving a useful problem for a specific group. Nobody who set out to "change the world" ever did. The companies that transformed anything started by making one boring, useful thing 10x better. 22-year-olds looking for a moonshot miss the actual moonshot sitting in the boring problem next to them. Be useful. Ship it to a small group. See if they can't live without it. Repeat. What's the most important problem you think still doesn't have a real solution? Drop it in the replies.

On January 15, 2025, Falcon 9 successfully deployed Firefly Aerospace’s Blue Ghost Mission 1 and ispace’s RESILIENCE lunar landers on a trajectory to the Moon from pad 39A in Florida. For most of our missions, we plan a controlled deorbit of the Falcon second stage so it safely reenters over the ocean. For higher energy missions like those to a lunar transfer orbit, nearly all performance on the vehicle is devoted to successfully placing the payload in the intended orbit, and a controlled disposal maneuver is not always possible. We actively work to be as responsible as possible with hardware left in space to ensure space safety, including for more complex missions. In this case, over time, solar activity and gravity led the second stage toward the Moon. Impacts like this are rare, but they can happen with objects in these types of orbits, and we worked with NASA on the optimal disposal solution. Our focus remains on advancing reliable access to space while working toward even more sustainable operations with Starship in the future. As a fully reusable vehicle, Starship is designed to eliminate expendable upper stages entirely, reducing hardware left in orbit and enabling even more missions to the Moon, Mars, and beyond.

Kimi 开源后，真的意味着 AI 不收费了吗？ 白线主播 Minta 认为，Kimi K3 的开放权重正在改变 AI 模型竞争规则，但“免费模型”并不代表“免费 AI”。 2.8 万亿参数模型背后，需要 GPU、网络和工程团队支撑； 同时 Moonshot 也通过许可证保留大型商业分发场景的议价权。 未来 AI 竞争的关键，不只是拥有模型权重，而是谁能用更高的 GPU 利用率和更低的推理成本实现规模化服务。 内容不构成任何投资建议，请严格遵循当地法律法规。 《白线 WhiteLine》由吴说团队出品，从 Crypto 走向更广阔的资本市场，关注 AI 时代下的趋势变化与交易机会。

转一下苏剑林老师对 K3 架构的复盘。 一句话概括，K3 = KDA + MLA + Stable LatentMoE + AttnRes。整套设计没什么炫技，核心就是在模型效果、计算效率和训练稳定性之间做取舍。 这里稍微解释一下： KDA，一种线性 Attention MLA，一种 KV Cache 较小的 Full Attention 变体 Stable LatentMoE，对 LatentMoE 做过稳定性改造的 MoE 方案 AttnRes，用可学习的跨层注意力，替代固定等权的残差累加 几个比较有意思的点： ① K3 同时使用了 KDA 和 MLA。另外，训练仍然用 Moonlight 版本的 Muon 优化器，Attention 权重改成 Per-Head Muon，每个 Head 独立优化。苏神说这不会直接提高效果，主要是数学和结构上更合理：各个 Head 本来就相对独立，不应该在优化时耦合在一起。 ② 解决MoE“容易炸”的问题。 LatentMoE 会先降维，再使用更多专家，最后升维，在训推成本大致相同的情况下效果略好。不过连续的矩阵投影也让训练更容易出现数值不稳定。 K3 把 SwiGLU 换成了 SiTU-GLU，用 softcap 压住异常激活；又在 LatentMoE 升维前加了一层 RMS Norm。这个 Norm 不只是让训练更稳定，在 Valid Loss 差不多的情况下，不加它，某些 Benchmark 会稳定变差。 （这里有超级多的技术细节，感兴趣的可以看苏神原文） ③ 关于MLA DSV4 看上去都换设计了，K3 怎么还在用 MLA？苏神的答案是，目前 MLA 仍然很难被全面击败。它在训练阶段是 MHA 形态，推理时 KV Cache 较小；在固定训练成本和 KV Cache 大小时，MLA 依然近乎最优。 它的问题是对 MTP（推测解码）不够友好。MTP 的思路是“用计算换速度”，而 MLA 在 Decoding 阶段本身就比较吃计算，再叠一个 MTP，两边就容易抢算力。 换别的方案也有代价。比如换成 128+128 的 GQA8，效果很难打赢 MLA，KV Cache 还是 MLA 的三倍多；换成 256+256 的 MFA（本质上是 MQA），训练和 Prefill 成本又会上去。 目前还没有一个简单的 Attention 设计，能同时占住效果、训练成本、Prefill、KV Cache 和 Decoding 计算量。在 KDA+MLA 的混合架构下，MLA 的部分问题得到缓解，所以 K3 最后还是选了 MLA。 ④ 苏神觉得，DSV4 也不算真正“抛弃 MLA”。 DSV4 看上去重新设计了 Attention，但底层仍然有 MLA 的影子。它采用的是 head_dims=512、K=V 的 MQA，这正是 MLA 在 Decoding 阶段的形态；再加上 Sparse + Compress，前者减少计算，后者进一步压缩 KV Cache，同时减少计算。 方向很激进，不过 Infra 也更复杂。 ⑤ K3 的 MLA 可以去掉 RoPE，是因为 KDA 已经隐含提供了一种广义的位置信息。 这只适用于 KDA+MLA 的混合结构，像 K2 那种全 MLA 模型，直接去掉 RoPE 还是会明显掉效果。 感觉苏神真正的观点是：大模型架构里很少有免费的升级。K3 的设计哲学不是找一个碾压所有方案的新架构，而是在效果、训练成本、Prefill、Decoding、KV Cache 和稳定性之间，找一个当前更合适的组合。 苏神原文里还有很多技术内容，非常值得一看。 原文地址：

NASA recently recognized one of America’s outstanding engineers. At @SpaceX’s Rocket Development Facility in McGregor, Texas, I joined Congressman @PeteSessions and McGregor Mayor Jim Lilley to present Andrew Berkowitz with NASA’s Exceptional Public Achievement Award. Andrew’s pioneering research to better understand liquid methane and oxygen explosions is helping enable the next generation of rockets that will return Americans to the Moon, establish a sustained U.S. presence on the lunar surface, and prepare us for Mars. America’s greatest achievements in space are built by extraordinary people whose work often happens behind the scenes. Andrew’s contributions will help shape the future of exploration for years to come. Congratulations on this well-deserved honor, Andrew, and thank you for your contributions to NASA’s mission and America’s future in space.

NASA recently recognized one of America’s outstanding engineers.   At @SpaceX’s Rocket Development Facility in McGregor, Texas, I joined Congressman @PeteSessions and McGregor Mayor Jim Moody to present Andrew Berkowitz with NASA’s Exceptional Public Achievement Award.   Andrew’s pioneering research to better understand liquid methane and oxygen explosions is helping enable the next generation of rockets that will return Americans to the Moon, establish a sustained U.S. presence on the lunar surface, and prepare us for Mars.   America’s greatest achievements in space are built by extraordinary people whose work often happens behind the scenes. Andrew’s contributions will help shape the future of exploration for years to come.   Congratulations on this well-deserved honor, Andrew, and thank you for your contributions to NASA’s mission and America’s future in space.