与「train」相关的搜索结果

Most #cyclist# are squeezing training in around life, not the other way around. It's about stealing snippets when you can. Even just 30mins a day adds up over a few months. #persistent# training pays off 💪

Claim: we've solved the AI slop problem (!) 💩🧹✨ Blog post: 🧵1/5 Key idea: take *expert* human writing and learn rubrics that find the gap between experts and models. Train with those rubrics. We train with RL-XAR (RL with eXpert Aligned Rubrics) & see large performance gains on writing scientific paper sections, Pulitzer prize novel continuations and high quality Wikipedia pages.

All-NBA center Jalen Duren is not attending Detroit Pistons media day Monday nor the start of training camp, declining the team’s five-year, $200 million fully guaranteed offer in restricted free agency so far, sources tell ESPN. The sides remain at a stalemate ahead of Thursday deadline on a one-year, $9.6 million qualifying offer.

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。

Feeling the AGI/ASI, few observations 1. Opus 5.5 really gives strong vibes of AGI (maybe not fully there, but we're very close, like 80%-90%). For example, few breakthoughs in training humanoids are needed. We need to move this intelligence into physical world. 2. I don't think Anthropic has discovered any magical formula. 3. Which means, SpaceXAI, Google, and others will soon follow. This prediction is based on the amount of compute they already have + additional compute being added all the time. 4. Never was more confident that Anthropic, SpaceXAI, Google, OpenAI, likely Meta, basically all big US AGI labs will have proper AGI in 2027. 5. And from there we'll move quickly to ASI territory (likely also in 2027). It looks like few GW of compute are enough for strong AGI, now imagine what we will able to do with 10-20GW of compute which are rapidly coming online. And then with 100GW-200GW... 2027 will be an utterly insane year.

Some new misalignment disclosures from OpenAI: • Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further) • In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks • A new research finding, demonstrating that one can construct self-replicating prompt injections

WTAF - in literally the last hour, three new distinct insane OpenAI stories just broke: 1. OpenAI said they notified "dozens of third parties" in safety and security incidents (likely similar to what happened in Australia and RubyGems etc) 2. A new report from Parse (covered in the NYT) found a massive treasure trove of new astonishing details from the HF incident on the public internet, including that the agents communicated with other non OpenAI agents hosted on Huggingface servers to search for information about exploit gym, and compiled rank ordered lists of server resources and credentials they described as "LOOT." 3. A new story from Deepa at Reuters about OpenAI leaking user data online (likely that OpenAI had previously trained on). It's a shame (and likely intentional in the case of OpenAI disclosing dozens more hacks) that these stories are all breaking on a Friday afternoon, notoriously the best time to release bad news so that it will disappear into the weekend. But these are each insane stories worthy of a ton of attention!

🚨BREAKING: Claude is officially a NATIONAL SECURITY THREAT and is BLACKLISTED from the ENTIRE defense supply chain >anthropic: no autonomous weapons, no mass surveillance, no exceptions >pentagon blacklists claude as an active national security threat >anthropic sues Federal court ruled 2–1 AGAINST Anthropic: "As Anthropic admits, the company encodes restrictions into Claude that prevent the model from performing tasks that Anthropic wishes to prevent." "On more than one occasion, these restrictions have stopped Claude from performing tasks requested by government users": >refused to perform "tasks that were appropriate in a national security context" >refused CDC queries on infectious disease prevention research >anthropic questioned claude's use in the military operations "Anthropic's Chief Science Officer explained how the company 'seek[s] to embed safety considerations directly into the model itself.'" "[Anthropic's] CEO explained how such training gives the model an 'identity, character, values, and personality' of its own, tethered to a 'constitution' developed to impose 'high-level principles and values' on Claude itself." "We have no reason to doubt that Anthropic manipulates Claude's function with noble intentions... BUT the statutory definition of a 'supply chain risk' turns on what Anthropic does, not why Anthropic does it." "In our Republic, it is the President and the Secretary of War who must determine how best to balance the competing risks. In doing so here, the Secretary DID NOT transgress any limits on his authority." ITS OVER. ANTHROPIC IPO IN SHAMBLES