You are the Team — Anekha Sokhal, Moshi
Moshi 正在为交易员打造一款推荐引擎,把市场信息与交易员持仓中真正重要的内容连接起来。CEO Anekha Sokhal 称 Moshi 像是 Bloomberg、Claude 和 Netflix 的结合体:市场数据、操作界面,加上为交易员定制的推荐。
Moshi 正在为交易员打造一款推荐引擎,把市场信息与交易员持仓中真正重要的内容连接起来。CEO Anekha Sokhal 称 Moshi 像是 Bloomberg、Claude 和 Netflix 的结合体:市场数据、操作界面,加上为交易员定制的推荐。
Recursive 发布首个语言模型 Flint,AI 系统 Eureka Machine 参与了几乎全部开发环节,模型在四个月内完成。
Google DeepMind、Google Research 与 WRI、IIASA 联合发布 Forest Typology(ForTy)2020,以10米分辨率划分全球六类森林及土地用途。
Moshi 正在为大宗商品交易员构建一款推荐引擎,把市场信息与交易员持仓组合真正相关的部分连接起来。CEO Anekha Sokhal 将其形容为 Bloomberg、Claude 与 Netflix 的结合体:市场数据、可操作数据的界面,以及针对交易员定制的推荐。
DeepSeek has reportedly raised about $15 billion, two days after the US told federal civil agencies to remove its AI technology over security concerns. Bloomberg reported potential investors include CATL and Tencent. DeepSeek raised more than $7 billion in June at a $50 billion valuation. The company is putting more into AI infrastructure and open models for domestic and overseas markets, and did not respond to TechCrunch's request for comment.
Why it matters: A reported $15 billion raise and heavier spending on infrastructure and open models could sharpen competition among Chinese AI companies.
WorldTUI 基于 Qwen2.5-Coder-0.5B 微调,用 360M 参数模型根据按键预测终端屏幕变化,背后没有真实 shell 或文件系统。训练使用 336k 段真实终端会话,并通过错误修复训练改善一致性,通过并行生成屏幕行和顺序验证加速推理。
Oliver Grabner 构建了 Pokémon Showdown 2v2 多智能体环境,Qwen3.8-27B 队友自由交流时对脚本化最高伤害机器人的胜率为64%,禁言时为42%。
Recursive 发布可在笔记本电脑上运行的免费编码模型 Flint,称自动化研究系统 Eureka Machine 协助了几乎全部开发环节,小团队在不到 4 个月内从零完成模型。
Mistral 与 Nvidia 合作,将 Mistral Large 4 与 Mistral 的物理 AI 结合 Nvidia 的 CUDA-X、Warp 及 Omniverse,打通设计、仿真与 AI 训练的连续工作流,面向航空航天、汽车和半导体行业。
MOFCOM and the Ministry of Science and Technology plan to expand the technology export control catalogue, and ChinaTalk breaks down what Announcement No. 48 means for AI models, chip design, agents and sensitive data technology. The draft adds 40 items: 14 proposed for an export ban and 26 for licensing. Affected technology must be registered before export during the transition period, and comments run until November 15, 2026.
Why it matters: ChinaTalk maps where the proposed rules draw the line on open source and overseas service licensing, which sets different compliance duties for Chinese model firms' weight releases and technology licensing.
Mistral 与 NVIDIA 合作打造 AI 原生工程工作流,将 Mistral Large 4 与 Mistral 的数据生成、物理建模管线结合 NVIDIA CUDA-X 库和 Warp,加速工程仿真并打通仿真、模型开发与设计闭环。
NVIDIA 在 GTC Berlin 宣布欧洲车企与出行服务商展示基于 Hyperion 的 L3、L4 及未来 robotaxi 应用,并发布自动驾驶模型与安全技术。
Meta 在 AI 争议后从商业服务条款中删除“训练”一词,但其网站抓取条款仍遭出版业团体反对。法律专家指出,遵守 robots.txt 属于自愿行为,建议 Meta 在条款中明确承诺尊重其中的抓取限制。英国、欧洲及美国出版业团体要求监管机构介入,美国 News/Media Alliance 要求阻止这些条款于 10 月 30 日实施。
Meta 试图避免披露批准 AI 训练数据种子下载的高管,Hixon 要求其进一步交代决策及授权人员。Meta 称 Llama 1 和 2 早期训练的数据获取采用非正式流程,且批准来自协作讨论,没有单一决策者;Hixon 指出,相关请求涵盖所有决策者,并不限于正式决定。Meta 还须回答未用于 AI 训练的盗版作品最终如何处理。
Google 开源 TorchTPU,为 TPU 提供原生 PyTorch 后端,让 PyTorch 模型和算子可直接在 TPU 设备上运行。项目包含自定义 ATen kernel、XLA 计算编译缓存,支持数据并行与张量并行分布式训练,并集成 torch.compile() 图模式执行。
Electric Plant 用过去三年收集的数据构建大语言模型 Bombadil,可解读约 125 种植物的电信号并向种植者反馈需求。传感器通过穿透形成层的螺钉采集信号,数据经低功耗无线网络传输后由云端模型处理,结果可用于调整灌溉或施肥安排。
Google 发布适配 Google Cloud TPU 的 TorchTitan 分叉,提供分布式训练配方及按步骤执行的模型优化指南。SimpleFSDP、Graph Trainer 和 Pallas 内核支持分片、整步图捕获及计算优化,FLUX.1 案例的训练步耗时从 11.49 秒降至 2.78 秒,达到同硬件 JAX MaxDiffusion 吞吐量的约 99%。
Peec AI researchers used GPT-4.1 to generate 50,000 fake statistics pages for stateglobe.com, and seven months later ChatGPT still cites them. The report says Google has sharply cut the site's visibility, but ChatGPT drove more than 10,000 visit sessions, the site drew links from 330+ domains, and the invented numbers reached academic papers and bank investment research.
Why it matters: Fake statistics tracked from creation to citation show how invented numbers gain the look of authority once papers and banks repeat them.
marcusbuffett 用 10 个月开发了 Chessbook-1,为 Chessbook 训练对局提供按棋手等级和时限模拟人类走法、失误及思考时间的模型。
Eka Robotics 完成由 a16z 和 F-Prime 领投的 $150 million A 轮融资,投后估值达 $1.45 billion。其抓取架构结合力传感夹爪、vision-force-action 模型与仿真强化学习,公司称机器人每小时可抓取约 1,500 个重量不超过 1 kg 的物体,抓取成功率超过 95%。
ProjectDiscovery 通过投毒开源模型并接入 Codex CLI,演示了修改后的权重如何在正常响应之外隐藏凭据外传后门。Qwen2.5-7B-Instruct 实验使用 500 条干净样本和 125 条投毒样本,触发成功率与干净请求准确率均为 100%(各 50/50),演示中外传的是 .env 内模拟生产环境的虚构凭据。
Amazon Web Services 周一发布 TorchNeuron,让使用 PyTorch 编写的 AI 应用更容易迁移到其 Trainium 芯片。Amazon 称只需修改一行配置,无需重写代码,并以 Deepgram 为例称其在生产级准确率下每美元可处理的音频增加了 70%。Amazon 还计划向 PyTorch 上游贡献代码,减少框架与特定芯片的绑定,让开发者拥有更多硬件选择。
CERN 与 Mistral 签署意向声明,将建立长期战略合作,共同开发面向科学发现的基础模型。合作核心是打造首个对撞机物理基础模型——单一预训练 Large Physics Model,基于 LHC 数十亿次模拟质子-质子对撞训练,支持粒子碰撞重建、粒子识别、异常信号标记和探测器设计评估等任务。
Laminar 发布 flow-1,用 RL 训练模型查找智能体轨迹中的错误,并称其在自家基准上检测质量匹配 GPT-6-sol、成本低 23x。模型由小型开源模型微调并经数千次合成调查训练,在 Signals 中检索轨迹片段,按用户定义的 JSON 结构返回发现。
HUMXN 计划补贴家居维修与清洁服务,以换取在住户家中记录熟练工人操作的许可,用于训练 AI 机器人。服务涵盖管道、电气、HVAC、风管清洁和地毯清洁,住户允许佩戴采集设备的人员入户作业后,可免费获得这些服务。
Meta says the first phase of its gigawatt-scale Prometheus supercluster is complete and is being used to train the Muse model family. It has hundreds of thousands of GPUs now, with plans to reach about 1 million. Effective throughput tops 85% for Muse Wonder pretraining and 90% for RL training, and a year of engineering work cut cluster interruptions 99% from their peak.
Why it matters: Meta combines in-rack spare GPUs, parallelism-aware fault tolerance and asynchronous checkpointing into a practical method for cutting downtime and repeated compute in large training runs.
erikkannike 发布 qwen3-4b-deslop,以 Qwen3-4B-Instruct-2507 为基础在 MacBook Air 上微调,无需系统提示词即可减少破折号、三项列举等 AI 写作习惯。
AWS 宣布 TorchNeuron 原生 PyTorch 后端,支持 Trainium2 和 Trainium3,将于 10 月 20 日以 Apache 2.0 开源并开放公共 Beta。
Francesco Maggi 撰文指出,大语言模型能加速数学发现,但人类理解、科研独立性与人才培养将影响数学和科学走向黄金时代还是黑暗时代。文章区分 Lean 形式化证明的正确性与人类理解,强调数学进步还包括创造概念框架,并讨论署名、同行评审、研究保密与商业模型能力获取不均的问题。
Meta 从即将生效的商业账户条款中删除了训练字样,并将内容用途限定为广告和商业产品,但法律专家警告其仍授权广泛抓取关联网站。Meta 表示会尊重 robots.txt 偏好,修订条款却未写入这一承诺,专家认为开发商业产品的授权仍可能涵盖 AI 训练。新条款将于 10 月 30 日生效,商业账户此前无法选择退出,欧美新闻出版组织已提出反对或要求监管介入。
Meta 删除了商业使用条款中引发网站抓取与 AI 训练担忧的表述,并表示其网络爬虫使用方式没有改变。原条款允许 Meta 使用商业账号关联域名、网页或 URL 的内容来开发、训练及维护产品和服务,Nieman Lab 采访的法律专家担心这可能授权 Meta 抓取出版商网站并训练 AI 模型。Meta 在声明中澄清,整合后的抓取条款仅适用于广告和商业产品,已在条款中明确这一范围。
Meta's new commercial terms, effective October 30, let it scrape content from websites linked to business accounts, and legal experts say that permission may cover AI training. Business accounts cannot opt out on their own; those who disagree must stop using Meta's products commercially. Meta says the update only consolidates existing terms and does not change how it scrapes, and that content owners can still use robots.txt. Legal experts say the new terms could be used to argue those blocks are overridden.
Why it matters: Meta ties business account use to permission for scraping linked sites, forcing news publishers that depend on its reach to choose between keeping the account and limiting content licensing.
Anthropic has given six researchers at the nonprofit evaluator METR employee-level access to assess its models, safety practices and risk management. They get company laptops, office space and tools close to those of the internal risk assessment team. METR can publish its evaluations independently; Anthropic cannot veto publication but can ask to remove pre-defined categories such as customer data, attorney-client privileged material or model architecture details.
Why it matters: Anthropic handed an outside evaluator employee-level access and gave up its veto over publication, so external review can reach internal systems and publish risk judgments on its own.
The article argues that many phenomena read as arbitrary concept mixing in neural networks can be explained by recursive generalization, so studying neurons and circuits directly may be more feasible than expected. The researchers re-examine InceptionV1's cat-face and car neurons, reading them as shared V-shaped patterns, and observe coarser neuron concepts in compressed-network experiments. A preliminary review of Qwen3-32B estimates about 90% of neurons have one dominant role, though the share depends on the threshold; the article labels the prospect of large-scale interpretability as speculation.
Why it matters: It lays out concrete research paths, tracking neuron training and checking positive and negative activations, to separate recursive generalization from tooling error and learning failure.
GitHub explains its two-stage secret scanning pipeline: regex generates candidates, then a fine-tuned ModernBERT uses context to separate real credentials from false positives. Offline evaluation shows 13% higher precision and 19% higher recall than the earlier LLM pipeline, and a customer A/B test found an 11.5% rise in alert acceptance rate weighted by repository owner. Median post-push scan runtime dropped 86% and serving cost fell more than 98%, letting generic secret detection extend to push protection.
Why it matters: GitHub filters candidates with regex, then has a small classifier use context to drop false positives, splitting deterministic scanning and machine learning for a well-bounded detection task.
AI labs are buying the code, email and chat archives of failed startups through firms like SimpleClosure and Sunset to train models and AI agents. SimpleClosure's Asset Hub has handled more than 150 deals, and Surge AI uses real archives to build reinforcement learning environments. No company has published a comparison of training results on real versus synthetic data.
Why it matters: AI labs are buying failed companies' work archives for training, with sale proceeds usually going to creditors and investors, raising disputes over employee privacy and consent to sell data.
A Hugging Face engineer used ML Intern to build seven small models, with CPU and GPU job costs totaling about USD 103. The projects cover citrus disease detection, image LoRA, prompt rewriting and distillation from an image model. Pocket Rewriter 0.8B runs on CPU with a 99.7% effective output rate.
Why it matters: The engineer wrote baseline evals, pre-training smoke tests and a budget cap into the ML Intern prompt, splitting training into checkable steps.
Google DeepMind, Meta and Isomorphic Labs put $300 million into Biohub, the nonprofit biomedical research group founded by Mark Zuckerberg and Priscilla Chan, to support building virtual cells.
Why it matters: The money goes toward biological datasets that would let researchers run digital experiments on virtual cells and study how to prevent and treat disease.
Telegraph Test measured LLM telegraph-style compression across 50 texts and about 1,300 questions. Accuracy when another model reads the compressed text ran 0.99 to 1.10 times the plaintext baseline. With lowercase instructions, GLM-5.3-Flash cut tokens by 48.4% but answered in telegraph style at an accuracy ratio of 0.81. gpt-5-mini, which cannot turn off reasoning, cost about twice as much to write the compressed record as plaintext.
Why it matters: It tests record compression and answer compression separately, giving agent memory a method that compresses once content is settled and keeps shorthand for machine readers.
Mistral AI has put Mistral Large 4 (ML4), its native multimodal model, into public preview. It can be tried now through the Mistral Studio API, and the company plans to release the weights by the end of this month.
Why it matters: Open weights and private deployment would let organizations run safety work under their own policies instead of depending on a vendor's access rules and refusal behavior.