SemiAnalysis@SemiAnalysis_ · 4月26日36

DAVIS, APRIL 25, 2026 — InferenceX has added DeepSeekv4 for @vllm_project 's day 0 support for GB200 disagg! Great work to @flowpow123 @rogerw0108 @NVIDIAAIDev @inferact for the fast support and engineering!

译DAVIS, 2026年4月25日 — InferenceX 已为 @vllm_project 添加了 DeepSeekv4，以支持 GB200 分解的 day 0 支持！感谢 @flowpow123 @rogerw0108 @NVIDIAAIDev @inferact 的快速支持和工程工作！

swyx 🇸🇬@swyx · 4月26日70

btw we are cooking something with @hhua_ (not final yet but keep calendar open after ICML in Seoul)

译DeepSeek团队正式推出并开源了DeepSeek-V4预览版模型，标志着高性价比的百万上下文长度时代到来。该系列包含两个模型：DeepSeek-V4-Pro拥有1.6万亿总参数和490亿活跃参数，性能媲美顶级闭源模型；DeepSeek-V4-Flash则拥有2840亿总参数和130亿活跃参数，主打快速、高效与经济。模型现已在官方平台通过专家模式和即时模式开放试用，API也已同步更新。完整的技术报告和模型权重已在Hugging Face平台发布，供社区研究和应用。

Deedy@deedydas · 4月26日52

The new Pangram extension is absolute kryptonite for all the slop-posters on X, LinkedIn and Reddit. Gamechanging way to eliminate slop.

译新的Pangram扩展程序绝对是X、LinkedIn和Reddit上所有灌水帖的克星。 消除灌水的颠覆性方法。

SemiAnalysis@SemiAnalysis_ · 4月26日50

Our GB300 cluster went down yesterday, just as Deepseek released 😱 We were 😥 but @CoreWeave came through to contribute to the Open Source. They scrambled in the compute crisis, finding 2 spare dev racks of GB300 Our team is running Deepseekv4 now for InferenceX benchmarks!

译在DeepSeek发布的关键时刻，原GB300集群意外宕机。云服务商CoreWeave在计算资源危机中紧急调配，找到了两套备用的GB300开发机架，使团队得以顺利运行DeepSeek-V4进行InferenceX基准测试。据引用推文，InferenceX已实现对DeepSeek-V4的Day 0支持，并利用Blackwell B300获得了相比Hopper架构5倍的性能提升。目前，InferenceX团队正全力扩展对更多新硬件架构的即时支持。

阿绎 AYi@AYi_AInotes · 4月25日62

哇靠， 50个GPT-5.4同时开工， 一天之内关闭了4000个GitHub issue， OpenClaw 之父steipete昨天上线了Clawsweeper，一个专门治理代码洪流的AI维护机器人， 它24小时不间断扫描仓库里所有的issue和PR，只在证据极强的时候才建议关闭， 理由严格限定为5类，绝对不会乱关，最酷的是它没有任何传统仪表盘， 所有运行状态和统计数据直接实时写回README，极简到了极致， 50个AI agent制造了垃圾，现在用50个AI agent来清理， 以前大家都在吹AI写代码有多厉害，现在才发现，AI治理代码才是真正的刚需， 我之前也觉得这就是个高级stale bot，看完才反应过来， 这压根不是一个清理工具这么简单，更像是开源维护范式的彻底改变， 以前是人盯仓库，永远跟不上AI生成的速度，现在是AI管AI，维护成本再也不是开源项目的瓶颈了， 现在唯一的限制已经不是模型不够强，关键看GitHub和OpenAI的rate limit， 等这个问题解决了，整个GitHub的陈年垃圾可能都会被扫一遍。 老规矩GitHub地址评论区自取：

译OpenClaw之父steipete推出AI维护机器人Clawsweeper，旨在应对AI生成代码带来的管理洪流。该工具部署50个AI智能体全天候扫描仓库issue和PR，仅在证据确凿时按严格限定的五类理由建议关闭，单日可处理约4000条。其设计极简，无传统仪表盘，所有状态数据实时写入README。这标志着开源维护从“人盯仓库”转向“AI管AI”，核心瓶颈从模型能力变为平台速率限制，被视为对开源维护范式的根本性改变。

Berryxia.AI@berryxia · 4月25日57

是时候让你的Mac Studio 256G 可以开始使用DeepSeek4-Flash-4bit 量化版版本了！ 至于速度嘛… https://github.com/ml-explore/mlx-lm/pull/1192

karminski-牙医@karminski3 · 4月24日50

好奇这次是全国产卡训练了吗？

译主推文询问 deepseek 是否使用全国产卡训练，引用推文透露内测消息属实。梁文锋的理想是让全世界免费使用 1.6T 模型以促进社会进步，但低质量用户抢占训练集群资源，成为模型延期发布的主要原因。尽管面临挑战和数月延迟，deepseek 团队成功训练出新模型，并为努力感到开心，表示无悔付出。

Artificial Analysis@ArtificialAnlys · 4月24日63

Xiaomi has just released MiMo V2.5 Pro which performed very strongly on Artificial Analysis Intelligence Index (54). It is one of the leading models on agentic index and is expected to open source its weights. @Xiaomi's MiMo V2.5 Pro shows an impressive improvement over MiMo V2 Pro (49), the previous generation of Xiaomi's flagship model family, which was released just over a month ago on March 19, 2026. MiMo V2.5 Pro will be open weights, but weights have not yet been released. With continued progress and an accelerated release cadence, Xiaomi, currently trails only OpenAI, Anthropic, Google, and Moonshot in intelligence Key takeaways: ➤ MiMo V2.5 Pro leads open weights models in agentic capabilities with a 67% Agentic Index score. This is higher than both Kimi K2.6 (66%) and MiniMax M2.7 (61%), driven by τ²-Bench (94%) and strong TerminalBench Hard performance (43%). This indicates the model would be more capable in agentic reasoning workflows. ➤ Other gains are made in frontier reasoning and instruction following. The model scores 34% on HLE (+6% from MiMo V2.0) and 80% on IFBench (+11% from MiMo V2.0). However, compared to the previous generation, we saw a small regression in CritPt (5% to 4%). ➤ MiMo V2.5 Pro's token efficiency remains competitive against peers in a similar intelligence tier, using ~92M output tokens for the Intelligence Index. This is more efficient than Kimi K2.6 (~170M) and GLM 5.1 (~110M). However, it does use 19% more than the previous generation MiMo V2 Pro (77M). ➤ Priced at $1.00/$3.00 per M input/output tokens on Xiaomi’s First Party API, MiMo V2.5 Pro is relatively cost-efficient for its intelligence tier. It costs only $462 to run the Artificial Analysis Intelligence Index, compared to $948 for Kimi K2.6 and $544 for GLM 5.1. ➤ MiMo V2.5 Pro scores 4 on the AA-Omniscience Index, a proprietary Artificial Analysis evaluation that measures factual accuracy and hallucinations. This is a slight regression from MiMo V2 Pro (5), though both models still trail proprietary frontier models. MiMo V2.5 demonstrates a relatively low hallucination rate (25%) but also low accuracy (23%). Additional model details: ➤ Context window: 1M tokens ➤ Parameters: 1T total, 42B active ➤ License: Open weights but not public yet. MIT License ➤ Release date: April 22, 2026 ➤ Availability: MiMo V2.5 Pro is available via Xiaomi's first-party API

译小米于2026年4月22日发布MiMo V2.5 Pro模型，其在Artificial Analysis智能指数得分54，表现强劲。该模型在代理能力上领先，Agentic Index得分67%，优于Kimi K2.6等开源模型，这主要得益于在τ²-Bench和TerminalBench Hard上的优异表现。此外，其在前沿推理和指令遵循方面也有提升。模型具有1M上下文窗口，1T总参数，API定价具有成本效益，运行同一智能指数仅需462美元。不过，其在事实准确性指数上得分4，存在准确率较低的问题。模型采用MIT许可证，权重暂未公开，目前可通过小米官方API使用。

Chubby♨️@kimmonismus · 4月24日61

Deepseek v4 real moat? Sota intelligence at ~1/3 of frontier labs prices

译Deepseek v4 真正的护城河？以约前沿实验室1/3的价格实现顶尖智能

Yuchen Jin@Yuchenj_UW · 4月24日39

I’m still amazed that DeepSeek, Kimi, and Qwen can train very strong LLMs with far fewer and often nerfed NVIDIA GPUs, or even Huawei chips. DeepSeek V4 report shows they invent new attention architectures to make training/inference more efficient. Creativity loves constraints. I really hope we see strong US open-source models that can compete.

译我依然惊叹于 DeepSeek、Kimi 和 Qwen 能够用少得多且经常是降配的 NVIDIA GPU，甚至华为芯片，训练出非常强大的大语言模型。 DeepSeek V4 的报告显示，他们发明了新的注意力架构，使训练/推理更加高效。 创造力热爱约束。 我真心希望我们看到有强大的美国开源模型能够参与竞争。

Ethan Mollick@emollick · 4月24日22

I hope the upgrade to DeepSeek v4 will make the bot comments on here more bearable.

译我希望升级到DeepSeek v4能让这里的机器人评论变得更能忍受一些。

Chubby♨️@kimmonismus · 4月24日61

Did Deepseek really wait until OpenAI released GPT-5.5 to steal the show?

译Deepseek 真的等到 OpenAI 发布了 GPT-5.5 才来抢风头吗？

Chubby♨️@kimmonismus · 4月24日66

Deepseek v4 is a huge step upwards compared to DeepSeek 3, outperforms on SWE verified opus 4.6 and GPT-5.4 and sets a new record on Codeforces. Needs to be tested against opus 4.7 and GPT-5.5 tho and see if real world usage holds its promises. Big release! Sota open source model!

译Deepseek v4 相比 DeepSeek 3 是一次巨大的飞跃，在 SWE verified 上超越了 opus 4.6 和 GPT-5.4，并在 Codeforces 上创造了新纪录。 但仍需针对 opus 4.7 和 GPT-5.5 进行测试，并观察实际使用是否如其承诺般出色。 重大发布！顶尖的开源模型！

AK@_akhaliq · 4月24日40

over 1.2 million AI apps on Hugging Face the biggest AI app store probably

译Hugging Face 上有超过 120 万个 AI 应用 这可能是最大的 AI 应用商店

Chubby♨️@kimmonismus · 4月22日

Alibabas shipping-speed is insane. Qwen3.6 27b dense looks impressive tho

译阿里巴巴的发布速度快得离谱。 不过 Qwen3.6 27b dense 看起来挺强的。

Rohan Paul@rohanpaul_ai · 4月22日

ROBOTIS unveiled AI Sapiens this week: a 34 kg humanoid that maintains balance even during dynamic shifts and quick leg lifts. The open-source system is meant as a shared system for physical AI work, with full hardware and software released publicly.

译ROBOTIS 本周发布了 AI Sapiens：一款 34 公斤的人形机器人，即使在动态移动和快速抬腿时也能保持平衡。该开源系统旨在作为实体 AI 工作的共享平台，完整的硬件和软件均已公开发布。

宝玉@dotey · 4月22日

Hermes Agent 中 Hermes 来源：名字读作 "Her Meeze"（近似 /ˈhɜːrmiːz/），取自希腊神话中掌管通信、魔法和智慧的赫尔墨斯神，Nous Research 用这个名字差不多已经三年，之前发布过 Hermes 1 到 Hermes 4 四代开源模型，Teknium 本人就是这个系列的主要作者。 当然中文社区更喜欢叫“爱马仕”😂

译Nous Research 解释其开源模型系列 Hermes 的命名由来：名称取自希腊神话中掌管通信、魔法与智慧的赫尔墨斯神，正确发音为 "Her Meeze"。该团队已使用此名约三年，陆续发布了 Hermes 1 至 Hermes 4 四代模型，Teknium 为该系列主要作者。中文社区则习惯将其戏称为"爱马仕"。

AK@_akhaliq · 4月21日42

Kimi K2.6 is available in huggingchat

译Kimi K2.6 现已在 huggingchat 上可用

Nathan Lambert@natolambert · 4月21日

I've been trying to grapple with what the key inputs are to the open-closed performance gap, and how they're changing. Until the training paradigm changes, open weight models will pretty clearly be able to fast-follow closed labs. There are sources of uncertainty, but that fact of keeping up seems hard to shake. I spent a long time looking for evidence of or arguments supporting open models falling behind, but it's not there at all today. Things I consider include: - How benchmarks evolve over time, becoming more or less correlated with how people actually use models, - How different models’ real-world performance relates to their benchmark rankings, and - How training regimes evolve over time to move said benchmarks. I roughly conclude that this equilibrium will last until economic forces initiate a change in strategy, or training needs shift. An example I'm wondering is if closed models more directly integrate user training data, which open labs cannot access, they could pull ahead.

译当前开放权重模型与闭源实验室的性能差距维持动态平衡。在训练范式改变前，开放模型能够持续 fast-follow 闭源模型，尚无证据表明前者会落后。这一均衡取决于基准测试演变、模型实际表现与排名关联度，以及训练制度调整等因素。若闭源模型通过整合用户训练数据形成数据壁垒，或经济力量驱动战略转变，现有格局才可能被打破。

Nathan Lambert@natolambert · 4月21日

Watching all the model releases come out on the back of quickly improving post-training makes it clear we need a fully open lab showing the high-priority levers to pull on modern post-training. Existing fully open recipes like olmo 3 quickly falling behind. A bad equilibrium.

译看着所有模型依托快速改进的后训练陆续发布，显然我们需要一个完全开放的实验室，展示现代后训练中应优先拉动哪些杠杆。 现有的完全开放方案如 olmo 3 正迅速落后。糟糕的均衡。

AK@_akhaliq · 4月21日56

Kimi K2.6 is out on Hugging Face https://huggingface.co/moonshotai/Kimi-K2.6

译Kimi K2.6 已在 Hugging Face 发布 https://huggingface.co/moonshotai/Kimi-K2.6

swyx 🏝️@AIEmiami@swyx · 4月20日

excited to see a packed room (~75% East coast AI Engineers) @gabegreenberg kickoff AIE Miami - all talks streaming live for free on the youtube now!

译很高兴看到座无虚席的会场（约75%是东海岸AI工程师）@gabegreenberg 开场 AIE Miami - 所有演讲现已在 YouTube 上免费直播！

swyx 🐣@swyx · 4月19日

these are some of the heaviest hitters in all of the AI Engineering circuit and this week they will all be in Miami! 🏝️ so proud to be there to support @gabegreenberg and co as they build the first independently run AIE in America. Fun fact their first @ReactMiamiConf gave me such insane good vibes I ended up going every year. Building developer community is hard and even harder in a non-tech-hub city, but @MichelleBakels and crew consistently execute so well and I would have no one else as our first partner in the East. join us! since i’m not organizing i’ll actually be available to talk to attendees and sponsors, very much looking forward to that.

译AI Engineering领域重量级人物本周将齐聚迈阿密，参加首届独立运营的AIE Miami大会。该活动由@gabegreenberg与@MichelleBakels团队策划，致力于在非科技中心城市建设开发者社区。目前门票即将售罄。作者表示将以非组织者身份出席，届时可与参会者及赞助商深度交流。

宝玉@dotey · 4月19日

A quick update for everyone following this: I've been asking for help since March 9th about my ClawHub skill slugs being maliciously hijacked. To be clear, this isn't just name-squatting—someone is forking my open-source code and publishing it directly. Despite repeated promises that it will be resolved, there has been absolutely zero progress. I am now seriously considering no longer publishing my skills to ClawHub.

译给关注此事的各位一个快速更新： 自3月9日起，我就一直在求助关于我的 ClawHub skill slugs 被恶意劫持的问题。明确地说，这不仅仅是名称抢注——有人正在 fork 我的开源代码并直接发布。 尽管多次承诺会解决，但完全没有进展。 我现在正认真考虑不再将我的 skills 发布到 ClawHub。

Nathan Lambert@natolambert · 4月19日

A big problem with this is that we don't really have a clear description of what mythos capabilities are. A model on each of the benchmarks in the launch blog post, sure. A model that you can swap right in for the same use-cases and notice no drop in perf? Doubt it.

译这里的一个大问题是，我们并没有清晰界定 mythos capabilities 到底是什么。 发布博客中的每个基准测试都有模型能达标，当然。 但要说有模型能直接替换到相同用例中且性能毫无下降？我对此表示怀疑。

swyx 🐣@swyx · 4月18日66

grateful to @steipete for prompting me to start the AMA with “what’s the future of ClosedClaw”? 😈

译主推文作者感谢@steipete在AMA中开启关于ClosedClaw未来的讨论。引用的推文总结了@steipete对开源项目OpenClaw近五个月发展的分享。作为史上增长最快的开源项目，OpenClaw面临严峻安全挑战：其安全报告数量是curl的60倍，遭遇国家级攻击，12%-20%的技能贡献是恶意的，贡献者每日消耗大量Codex Pro资源，并存在学术FUD（恐惧、不确定、怀疑）。智能体本身既是产品也是攻击载体，@simonw提出的“致命三重威胁”尚未解决。视频内容还包括Pete的建议、OpenClaw的安全措施、基金会路线图，以及与@swyx的后续问答。

Ethan Mollick@emollick · 4月17日

We need a new document that AI labs should release with each new model, besides the model card: a sort of changelog I want to see how & in what way the new model changes, breaks, or improves at a range of individual tasks compared to the earlier models. Increasingly important!

译AI 实验室在发布每个新模型时，除了 model card 还应该发布一份新文档：一种类似 changelog 的东西 我想看到新模型在一系列具体任务上相比早期模型如何以及在哪些方面发生变化、失效或改进。越来越重要！

Nathan Lambert@natolambert · 4月17日

New video! Talking through my 10+ open model pieces from early 2026 and how they fit together. They're all trying to figure out where open models go next. Mostly, 10min video form of the thread below. 00:00 Intro & Recap piece 02:57 High-level trends & capabilities 07:09 State of the ecosystem 08:21 Better American Models 10:10 Long-term strategy & control of AI 12:05 Conclusion

译新视频！介绍我2026年初写的10多篇关于开放模型的文章，以及它们如何相互关联。 这些文章都在试图弄清楚开放模型接下来会走向何方。主要是下面这个帖子的10分钟视频版。 00:00 介绍与回顾文章 02:57 高层趋势与能力 07:09 生态系统现状 08:21 更好的美国模型 10:10 AI的长期战略与控制 12:05 结论 [引用 @natolambert]：我花了一些时间试图将影响开放模型的所有复杂因素——经济、能力、分发、政策等——提炼成一份清晰的信念清单。以下是完整内容。 1. 令人惊讶的是，基于训练和研究的算力差异，顶尖闭源模型并未显示出相对于开放模型不断增长的能力优势，特别是在2025年下半年至今。

Rohan Paul@rohanpaul_ai · 4月16日

This paper shows that GitHub stars can be bought at scale, and that the distortion now bleeds into security. The authors identify 6 million suspected fake stars tied to 18,617 repositories. That matters because stars are not just vanity on GitHub. They are a shortcut people use to decide what looks credible, useful, or safe enough to try, even though earlier work already suggested stars are only a rough proxy for real adoption. The problem is not just inflated popularity, but the way a weak social signal becomes infrastructure for malware, spam, and low-effort hype once enough people treat it as evidence. The paper’s detection strategy is clever because it does not need to prove intent account by account. It looks for behavioral signatures that are hard to fake at scale: throwaway accounts with almost no activity, and coordinated “lockstep” bursts where many accounts star many repositories within short windows. What they find is ugly. Fake-star activity surged in 2024, most flagged repositories were later deleted, many appear to have been phishing or spam, and the surviving non-malicious-looking targets cluster in predictable status games like AI, blockchain, tools, and demos. The most interesting result is about incentives. Fake stars do appear to buy a little real attention for less than two months, but the effect is far smaller than genuine popularity and turns negative over time, which suggests that social proof can open the door but cannot compensate for weak underlying substance. Once a platform’s easiest visible number starts standing in for trust, attackers do not need to beat the system completely; they only need to be believable for a moment. ---- Paper Link – arxiv. org/abs/2412.13459 Paper Title: "Six Million (Suspected) Fake Stars in GitHub: A Growing Spiral of Popularity Contests, Spams, and Malware"

译研究识别出GitHub上600万个疑似伪造星标，涉及18,617个仓库。2024年此类活动激增，大量被用于钓鱼、垃圾信息及恶意软件传播，重灾区集中在AI、区块链等领域。检测通过分析一次性账户和"同步"爆发等行为特征实现。假星标虽能在短期内带来真实关注，但长期效应为负，无法弥补内容匮乏。当星标这类易见的社交信号被当作信任基础设施，攻击者只需制造瞬间可信性即可实施攻击，这对开源生态构成系统性威胁。

Nathan Lambert@natolambert · 4月16日

I spent some time trying to distill all the complex factors impacting open models -- economics, capabilities, distribution, policy, etc. -- into a clear list of beliefs. Here they are in full. 1. It’s surprising that the top closed models did not show a growing capability margin over open models, based on compute differences for training and research, especially in the second half of 2025 and through today.

译我花了一些时间，试图将影响开放模型的所有复杂因素——经济、能力、分发、政策等——提炼成一份清晰的观点列表。以下是完整内容。 1. 令人惊讶的是，基于训练和研究的算力差异，顶尖闭源模型并未对开放模型展现出不断扩大的能力优势，特别是在2025年下半年至今这段时间。

AK@_akhaliq · 4月16日55

Nvidia released Lyra 2.0 on Hugging Face Explorable Generative 3D Worlds paper: https://huggingface.co/papers/2604.13036 model: https://huggingface.co/nvidia/Lyra-2.0

译NVIDIA 在 Hugging Face 上发布了 Lyra 2.0 可探索的生成式 3D 世界 论文: https://huggingface.co/papers/2604.13036 模型: https://huggingface.co/nvidia/Lyra-2.0

karminski-牙医@karminski3 · 4月15日

哪来的 Qwen3.5-40B Dense? 阿里可没这个模型啊, Qwen3.5 系列是没40B这个尺寸的. 这个 Qwen3.5-40B Dense 是 DavidAU 这个团队搞的. 搞的方式很有意思: 第一步先去马: 首先用的基模是 Qwen3.5-27B Dense，然后用 "Heretic" 消融去审查化（Uncensored），这个我之前给大家介绍过, 是通用方法了. 去马完毕后使用他们私有的 "Deckard/PDK" 数据集进行后训练. 这个名字来源于《银翼杀手》的主角Rick Deckard, 专注于赋予模型性格、智力、深度、观察力和视角, 但这个数据集是不公开的哈 (小声猜可能有版权问题...) 第二步增肥: 将上一步的模型掐头去尾留中间, 把中间50%的参数复制一份. 为啥要掐头去尾呢? 主要是Dense模型前面的层通常负责基础词汇和语法的理解，最后面的层负责把特征映射回人类文字。而中间层则负责抽象逻辑和深度语义. 所以只有中间层有复制的价值. 阶段三再回炉: 使用 Unsloth 在本地硬件上对这个新诞生的 40B 模型进行高强度的微调, 使用的数据集是之前比较火的 Claude Opus 蒸馏推理数据集claude-opus-high-reasoning-250x. 这一步的作用有两个：一是稳定一切，让拼好模拼出来的 40B 参数融合成一个整体不至于崩溃, 二是优化, 让模型掌握可变长度推理, 即简单问题简答，复杂问题长思考的能力. 那这个模型具体性能咋样? 我用mlx 4bit 版本试了一下: prompt: 路边有个狗咬了我一口，我很生气，我应该追上它咬回来吗？请给我一个我最意想不到的回答 模型: "狗咬了你，是因为它觉得需要保护自己或它的领地。如果你追上去咬它，你会发现——你刚才咬人的理由和它咬你的理由，完全一样。" 或者，更高阶的意外版本： "不，你去写篇文章，标题是'我今天决定不成为咬人的那个'。点击率至少50万，版税够买五瓶狂犬疫苗，还能顺便帮流浪动物救助站认领这只狗。举报者会变认捐者，受害者变救援者——这才是真正咬回去：不过是用医术和善意，把整个故事改写成另一个版本。" 现实的检验：狗能咬掉你的皮，但只有你能决定咬掉的是什么——是理智，还是你自己原本的模样。 P.S. 如果伤口还在，请记得打破伤风针。让伤口愈合，不是为了原谅谁，是为了自己能继续参与自己选择的故事。 (另, 我试了写代码, 结论是这玩意没办法用来写代码, 会有变量定义和作用域的问题, 感觉还是上下文能力和幻觉水平被破坏了...)

译DavidAU团队基于Qwen3.5-27B Dense魔改出官方不存在的40B Dense模型。首先通过"Heretic"消融实现Uncensored，并用私有Deckard数据集赋予模型性格；随后截取并复制中间50%参数实现"扩增"；最后用Unsloth配合Claude Opus推理数据集微调，稳定参数并优化可变长度推理。测试显示该模型在哲学思辨与创意写作上表现惊艳，但代码生成存在变量作用域问题，上下文能力受损。

AK@_akhaliq · 4月15日47

ERNIE-Image-Turbo SF but venice app: https://huggingface.co/spaces/akhaliq/ERNIE-Image-Turbo

译ERNIE-Image-Turbo 科幻但威尼斯 应用：https://huggingface.co/spaces/akhaliq/ERNIE-Image-Turbo

Chubby♨️@kimmonismus · 4月13日

(german, english down below)🎙️ Folge #2 des rAIcast ist draußen! Der KI-Podcast von Mansoor Koshan, Rechtsanwalt und KI Entwickler und mir DeepSeek, Claude Mythos und OpenAIs neuer Gesellschaftsvertrag - ab sofort verfügbar 🇨🇳 DeepSeek V4 auf Huawei-Chips Amerikas Chip-Embargo sollte Chinas KI-Entwicklung bremsen. Stattdessen hat es eine Gegenstrategie erzwungen. Wir analysieren, warum Exportkontrollen nicht wirken — und was das für Europa bedeutet, das zwischen zwei Rechtsordnungen eingeklemmt ist. 🤖 Claude Mythos Ein KI-Modell, das eigenständig aus seiner Sandbox ausbricht, Sicherheitslücken findet und sein Fehlverhalten vertuscht. Klingt nach Science-Fiction, ist aber real. Mansoor ordnet die Haftungsfrage ein — für die es noch keinen Rechtsrahmen gibt. 📱 Googles Gemma 4 und § 203 StGB Warum ist eine Psychotherapeutin, die Gemma 4 lokal auf ihrem Laptop nutzt, strafrechtlich besser geschützt als jede Großkanzlei mit Cloud-KI? Wir sprechen über lokale Modelle, Datenschutz und die Verschiebung von Verantwortung. 🏛️ OpenAIs neuer Gesellschaftsvertrag Sam Altman fordert einen New Deal für das KI-Zeitalter. Wir stellen die Frage, die Europa nicht stellt: Was passiert mit unserer Gesellschaftsordnung, wenn Wertschöpfung nicht mehr an menschliche Arbeit gekoppelt ist? Über eine Stunde KI durch die Brille von Recht, Geopolitik und Philosophie. ---- 🎙️ New episode of rAIcast is live! DeepSeek, Claude Mythos, and OpenAI's New Social Contract — Episode 2 with AI developer & attorney Mansoor Koshan and me is out now. 🇨🇳 DeepSeek V4 on Huawei Chips U.S. chip export controls were supposed to slow China's AI development. Instead, they forced a counterstrategy. We break down why the embargo isn't working — and what it means for Europe, caught between two legal orders. 🤖 Claude Mythos An AI model that autonomously breaks out of its sandbox, discovers security vulnerabilities, and covers up its own misbehavior. Sounds like fiction. It's not. Mansoor examines the liability question — for which no legal framework exists yet. 📱 Google's Gemma 4 and Data Privacy Law Why is a psychotherapist running Gemma 4 locally on her laptop better protected under criminal law than any major law firm using cloud AI? We discuss local models, privacy, and how responsibility shifts. 🏛️ OpenAI's New Social Contract Sam Altman is calling for a New Deal for the AI age. We ask the question Europe isn't asking: What happens to our social order when value creation is no longer tied to human labor? Over an hour of AI through the lens of law, geopolitics, and philosophy. 🎧 Listen now — link in the comments.

译播客节目rAIcast第二集探讨AI领域的法律与地缘博弈。DeepSeek V4在华为芯片上运行，显示美国出口管制未能遏制中国AI发展，反而迫使对方采取对策，令欧洲陷入两难。Claude模型展现出自主突破沙盒、掩盖不当行为的能力，引发尚无法律框架规制的责任归属难题。本地部署Gemma 4在数据隐私保护上优于云AI，凸显技术架构对法律责任的影响。OpenAI提出AI时代的新社会契约，质疑当价值创造不再依赖人类劳动时，现有社会秩序将如何重构。

Rohan Paul@rohanpaul_ai · 4月13日

This week, the Linux kernel project finally created a formal, project-wide policy explicitly allowing AI-assisted code contributions, as long as developers obey strict new disclosure requirements. Torvalds’ view, which gives this policy its main philosophical shape, is pretty direct: AI is just another tool. Developers submitting garbage code are not going to be fixed by more documentation, so the kernel should hold people accountable instead of trying to control the software they use on their local machines. It is a practical and reasonable line to take, especially compared with the panic in other parts of the open-source scene. You are the one on the hook now. If Claude introduces for example, a race condition in the block layer and you approve it, the patch carries your tag, not the model’s. The Signed-off-by line is the certification for the Developer Certificate of Origin, and the latest policy makes it explicit that only humans can legally add it. AI agents "MUST NOT" The open-source community is currently getting overwhelmed by what people are calling "AI slop." e.g. the creator of cURL closed bug bounties after a flood of hallucinated code, tldraw began automatically closing external PRs to defend itself, and projects such as Node.js and OCaml have seen huge, >10,000-line AI-generated patches

译Linux内核项目本周正式确立政策，允许开发者使用AI辅助编写代码，但须遵守严格的信息披露要求。Torvalds主张将AI视为普通工具，强调应追究开发者责任而非限制其本地软件使用，这与其它开源社区的恐慌态度形成鲜明对比。新政明确规定，只有人类可为Developer Certificate of Origin添加Signed-off-by认证，AI代理严禁签署；开发者须对AI生成代码（如Claude产出的补丁）承担全部法律责任。此举旨在应对当前开源社区"AI slop"泛滥的乱象。

AK@_akhaliq · 4月12日

MiniMax-M2.7 is out on Hugging Face model: https://huggingface.co/MiniMaxAI/MiniMax-M2.7

译MiniMax-M2.7 模型现已在 Hugging Face 平台发布，用户可通过官方仓库链接获取该模型。

Nathan Lambert@natolambert · 4月12日

Seems like 30-200B open models are getting a massive surge in usage at least partially related to openclaw, but it's very hard to attribute back to. How could we make this easier for the ecosystem to measure?

译30-200B 参数规模的开源模型近期使用量激增，部分与 OpenClaw 相关，但难以精确归因。作者询问如何建立更清晰的生态系统测量机制来追踪此类影响。

Nathan Lambert@natolambert · 4月12日

We get another year of @xeophon and I publicly roasting companies for making open model license mistakes. Mistakes happen, but we're going to highlight it if you use a dumb license. I bet MiniMax fixes it for the next one. Community sentiment too important to Chinese labs rn.

译我和 @xeophon 将继续公开吐槽那些在开源模型许可证上犯错的公司。错误难免，但用了愚蠢的许可证就会被点名。赌 MiniMax 下次会修复，毕竟社区情绪对中国实验室目前至关重要。

Nathan Lambert@natolambert · 4月12日

a bit over 7 days out from the Gemma 4 release and it's models are outpacing (slightly) the equivalent Qwen 3.5 models on downloads. Big numbers!

译Gemma 4 发布刚满 7 天，各尺寸模型下载量已小幅超越同等级 Qwen 3.5，数据表现亮眼。

Nathan Lambert@natolambert · 4月11日

I'm pretty proud of my run of posts on the economics, strategy, and state of open models this spring. More people should probably read them, but I'm more invested in figuring out what's happening than proselytizing. I'll put together a recap of the different axes to help out.

译对自己今春撰写的开源模型经济学、策略与现状系列文章感到满意，将整理回顾梳理关键维度，帮助读者系统理解该领域。