与「Max_Ver」相关的搜索结果

Ethan Pimstone sets a new max vertical jump world record with an incredible 1.33m (52.5”) leap. 🚀 The hang time on this jump is absolutely unreal. 😳 (Via: ethanpimstone1/IG)

Exciting Claude Duet update: every Pro, Max and Ultra subscriber in the US who was on the waitlist as of yesterday has now been invited. Thank you for your patience! If you haven’t joined yet, the waitlist is here: We also wanted to share a note on usage. To give you room to play with Claude Duet, the research preview comes with limit resets: 3 for Pro, 1 for Max and Ultra. Please check your inbox, give Claude Duet a try, and keep the feedback coming. We’re aiming to open up to more countries outside the US very soon.

Huge AI announcements from OpenAI, Anthropic, Meta, and Google in the past 48 hours. Here's everything going on in AI right now: 1. The authors of AI 2027 just moved their timelines up. They now expect AI on par with the best human AI researchers to arrive in 2027, after METR estimated Anthropic is improving about 2x as fast thanks to AI. Daniel Kokotajlo gives it a 50-50 chance by May 2027 and says he's "scared, sad, and angry." 2. Anthropic opened the investor pitch for its IPO, with shares likely trading as soon as Thursday next week. An updated filing also says it has now crossed its own safety line for AI-driven progress. 3. OpenAI launched Businesses in ChatGPT, a beta hub where brands can set up AI agents for shoppers, claim a verified profile, and track how often ChatGPT mentions them. Paid ads can point users to those agents too. This is like the start of SEO for chatbots. 4. Anthropic rolled out Claude Sheets and Whiteboards in beta, made Claude Data generally available, and announced Claude Meetings, an assistant that sits in on your video calls. Add those to Slides, Design, and Docs. Seems like Claude is slowly rebuilding the entire office suite. 5. Anthropic also released Claude Fable 5.2 and Mythos 5.2 yesterday, with Fable debuting at #1# on Artificial Analysis' Intelligence Index, 5 points ahead of Google's Gemini 4 Carbon. It's also 20% cheaper and up to 30% faster than 5.1. Google lost the top spot before Carbon even shipped (!) 6. Meta released Muse Ember, its smallest model, built for high-volume agent work. It ties Claude Haiku 5.5 on the Intelligence Index at roughly one-third of the cost per task ($0.14 vs. $0.41 on an average test). Meta shipped its first Muse model in April and now has an answer for every tier OpenAI and Anthropic sell. 7. Google announced 2 new AI chips coming next year: the TPU 8tx, which keeps data encrypted even while a model is working on it, and the 8i Max, which joins 2 chips into one for agents that need to hold an entire codebase in view. Google already has nearly 100 TPU customers (including Anthropic and Apple) and every available chip is spoken for. 8. Nvidia's top lawyer reportedly emailed agents in August 2025 while they were investigating whether Singapore's Megaspeed smuggled its chips to China. The Commerce Department then halted the case and handed it to the Justice Department. Not a great week for Nvidia. 9. Adobe signed a multi-year deal with Disney to build custom AI models trained on Disney's characters and visual styles, with work already underway for theme parks, film, and ESPN. Pretty interesting to see one of the most protective IP owners in entertainment is now training its own AI models. 10. OpenAI acquired Odessia, a months-old AI trip-planning startup from Sonder founder Francis Davidson, who is joining OpenAI to work on travel. Between this and #3#, ChatGPT is coming for shopping and travel. 11. France unveiled the Paris Call for AI Safety and Security, with OpenAI, Anthropic, and Mistral committing to build emergency off switches into their most advanced models. Interesting timing given #1# on this list. Daily reminder that this is the worst AI will ever be.

Chart said $SOL at $116 is just price correction, down about 3.8% today. I wanted to know how traders were actually positioned behind that move, so I opened the Perps data in @tryquantio instead of just staring at the price movement. Funding sits at the neutral floor, about 0.03% a day. Longs aren't overpaying and shorts aren't pressing. Open interest is $661M against $245M in daily volume, which tells me a lot of capital is parked, not scalping. That's a coiled market, not a crowded one. Second question: is this move backed by positioning or is the market overcrowded? Verdict: not overcrowded. The drop from $120 came with no liquidation cascade and funding never turned negative. That reads as spot-led rebalancing, not panic. It's a coiled market, not a crowded one. Third question: what risk should I watch before taking the trade? Lose $109 on a daily close and the $661M in open interest turns vulnerable. Forced long liquidations could drag price toward $100 to $104. Reclaim $124.50 and trapped shorts fuel a squeeze toward $135 to $140. Other risks: chop between $112 and $118 can whip out tight stops, heavy supply sits at $120 and $124.50, and SOL still follows BTC. One BTC pullback can wipe out local support. What changed my view: the chart looked weak, but the positioning says patience, not panic. My plan: 2x to 5x leverage max, hard stop under $109, and partial profits at $124.50. Want to understand what's happening behind the chart? Check the Perps market inside Quant AI. @tryquantio #QuantAIPioneers#

NVIDIA 发布 Skill2Env：用“集体技能”强化智能体 NVIDIA 研究者们把社区公开的 Agent Skills 编译成可执行 RL 训练环境的数据流水线：3.4k 个 Skills 变成 8k 个带程序化测试和行为量规的终端任务；仅 300 步 RL 训练就让 Qwen3.8-27B 在 Terminal-Bench 2.1 上提升 4.7 个百分点，且模型行为显著向源 Skills 的方法论对齐。 开源项目： 核心洞察：公开 Agent Skills 是一个被忽视的监督来源 Agent Skills 是“教智能体做某件事”的文件夹：一个 SKILL.md 加上可选的脚本、参考资料和资产。论文指出，把公开 Skill 语料当作数据来读，它同时提供三样东西： · 任务分布的采样：人们真正想让智能体处理的任务分布（有人愿意花时间写下工作流，说明这活儿值得自动化）； · 真实世界的锚点：指向真实的仓库、数据集、工具和工件； · 结果测试表达不了的质量标准：领域专长、默认参数、常见坑、“好结果长什么样”。 # 数据流水线：四阶段编译，验证靠构造 1. Plan（分解）：容器化的 Codex 规划器读取完整 Skill 包、联网调研相关公共资产，把 Skill 拆解成若干可验证的 workflow，每个附带元计划（场景、初始世界、预埋缺陷、难点来源、解法草案、验证策略）、资产建议和“任务轴池”（任务原型 × 验证器模式 × 人物画像）。 2. Diversify（多样化）：宿主从轴池采样一组组合，加上复杂度、指令语气、请求者专业水平。关键设计是轴池以 workflow 为条件：研究型 workflow 配“证据可追溯”验证和研究者画像，而不是从全轴乘积空间乱抽，这让多样化保持 sensible。 3. Create（构造）：全新创建者 Codex agent 在 Docker 内工作，尽可能用真实素材（钉在特定 commit 的开源仓库、真实版本化文档、官方 API 规范）；需要联网服务的场景改造成本地替身（stub 服务器、录制回放 fixture、PATH 上的假 CLI、种子数据库），求解时绝不依赖网络。创建顺序被严格固定：先建世界 → 写指令 → 写测试 → 写量规 → 最后才写参考解，测试先于解法冻结，保证解法必须迁就评分契约而非反过来。 4. Verify（验证）：宿主端无模型参与的接收门：静态检查（布局、符号链接、Dockerfile 安全、基础镜像按内容摘要钉死）+ 两个容器内试跑：Oracle（参考解）必须全指标满分，NOP（什么都不做的 agent）必须全指标零分。任一失败即拒绝。 值得注意的一个反直觉选择：不做 teacher 模型预验证（不像部分工作用强模型试解、解不出就丢弃任务）。理由有二：这会把任务难度上限压到验证器能力，且成本翻倍；而 group-based RL 的在线动态过滤（rollout 无优势的 prompt 自动不产生梯度）天然淘汰过难/过易任务。 # 数据画像：广、贵、且忠实于源 规模与成本：7,971 个任务，用 GPT-5.6 Sol（xhigh 推理档）生成，API 花费超 9 万美元。（脚注：出于法律原因，公开发布的数据集改用 Kimi-K3-max 在同一流水线下生成。） 领域分布：13 个领域中，软件工程仅占 22.5%，AI/ML 10.5%，商业/金融/法律/HR 10.5%，营销 9.3%……论文对比了 TMax-15K、Terminal-Bench、DeepSWE 等，Skill2Env 是唯一全覆盖 13 域、且非技术知识工作占大头的语料。 忠实度探针（很聪明的设计）：用任务指令+量规作查询、对 3.4k 个 SKILL.md 做 TF-IDF 检索，73.2% 的任务 top-1 命中真实源 Skill，94.6% 进 top-10（随机 0.03%）。单用量规也有 68.5% top-1，证明量规携带的是 Skill 专属方法论而非泛泛建议。 SFT 数据：用 GLM-5.3 对每个任务 rollout 两次，得到 15,968 条轨迹，平均奖励 0.74，中位轨迹 19 次模型调用 + 23 次工具调用。 S2EBench：考虑到公开基准饱和，从 SkillHub 另外生成、逐条人工审核（指令无歧义、忠实于源 Skill、测试公允）后的 79 任务私有 held-out 基准。 # RL 实验：基础设施 + 极简配方 基础设施（论文明确说“现代 agentic RL 首先是基础设施挑战”）：Molt（PyTorch 原生全异步训练，Ray + vLLM + FSDP2）+ Polar（agent rollout 层：rootless Apptainer 沙箱、代理回传 token ID 和采样时 log-prob、prefix merging 把 harness 的多次补全缝合成训练轨迹）。 配方（刻意走“简单路线”）：GRPO 组归一优势 + DPPO 的 binary-KL 信任域掩码（δ=0.05，超出阈值的 token 直接丢弃，无需参考模型，还能防训练-推理失配）；G=8 rollouts/组，批 64，lr 1e-6 恒定，无 KL 惩罚、无熵奖励、无 SFT 热启动，每任务 65k 上下文。 量规校准奖励：开量规时，额外由 GPT-6 Astra 做 LLM-as-Judge（带“宪法”：惩罚无脑循环、reward hacking、答非所问；hacking 实证 = -5 分），总奖励 r = r_V + λs/5（λ=0.2），即 judge 最多把程序化奖励拉动 ±0.2。量规是校准可执行结果奖励，而非取代它，这是与“Rubrics as Rewards”一系的定位差异。 # 四项发现（论文最有信息量的部分） 发现 1：小规模 RL 即有跨域迁移。 仅 300 步、只用 2,400 任务子集训一个 epoch：S2EBench pass@1 +4.3（均分 +18.5），Terminal-Bench 2.1 +4.7（49.4→54.1）。训练集与 TB 无重叠（13-gram Jaccard < 0.8），且训练集从未针对 TB 调过，论文将其解读为规划、工具使用、收尾能力的通用提升而非任务族记忆。这让 27B 本地模型显著缩小了与云端前沿模型的差距。 发现 2：量规校准 RL 在基准上落后于纯结果 RL，一个诚实的负结果。 量规版在 TB 2.1 只有 50.1（纯结果版 54.1）；训练中量规版的程序化奖励长期停在 0.5–0.6，judge 分项从头到尾无上升趋势，两个奖励在训练分布上互相拉扯。论文不把它当作对量规奖励的终审判决（两者优化不同目标，而基准只考结果那一半），并给出两个疑因：λ=0.2 的加性形式让失败任务仍能拿正奖励、judge 看不到文件系统等设定均未调优；以及更本质的，Skill 写下的方法论可能本来就不是最大化基准通过率的分布。 发现 3：行为确实向 Skill 对齐，量规的价值所在。 200 个任务的成对偏好测试（judge 拿源 SKILL.md 当标准，比较匿名化的 base 与 RL 轨迹）：纯结果 RL 已被偏好 54.5% vs 33.5%；量规版被偏好 73.0% vs 24.0%。这说明量规奖励买到的东西在结果基准上看不见，但对“怎么做事”影响实质，对网页开发、报告综合、开放研究这类难验证任务尤其重要。 发现 4：GLM-5.3 蒸馏 SFT 反而伤害 Qwen。 在 GLM-5.3 轨迹上做 SFT：27B 上 TB 2.1 掉到 45.8；4B 上直接崩塌（TB 18.7→3.4，出现思维/工具调用死循环）。归因：教师的 interleaved-thinking + 工具调用风格与学生自身 post-training 不兼容，模仿覆盖了学生依赖的行为模式却带不来教师的能力。与 TMax 报告的“SFT 混合数据劣化已后训练的 Qwen”互相印证。因此论文所有 RL 结果都从未修改的原始 checkpoint 出发。

Grok 4.6 just tied for #1# on the Artificial Analysis Agentic Index • Grok 4.6 (high) — 59 • Claude Opus 5 (max) — 59 Outperforming Claude Fable 5 and GPT-5.6 Sol We’re entering the agentic era, and this is exactly the kind of benchmark that matters so much: tool use, planning, autonomy and complex problem solving Grok 4.6 is now sitting at the very top And that matters even more as Grok powers Grok Build and Grok Bot, where the model has to go beyond answering questions and actually take actions, use tools and complete real work Grok’s agentic capabilities are getting seriously powerful