🚨 AI News | TestingCatalog@testingcatalog · 33分钟前25

Meta is working on support for scheduled tasks for Meta AI on the web. Closing feature gaps 👀

译Meta 正在为网页版 Meta AI 开发定时任务支持。 关闭功能差距 👀

SemiAnalysis@SemiAnalysis_ · 56分钟前27

Meta Compute: Everyone Wants To Be A Cloud Zuck Takes Plan B? SpaceX 2.0, Bedrock 2.0, MSL Isn't Giving Up, Scaling RecSys by 10x... ClusterMAX ranking coming soon? https://newsletter.semianalysis.com/p/meta-compute-everyone-wants-to-be

译Meta计算：人人都想成为云 扎克伯格选择B计划？ SpaceX 2.0、Bedrock 2.0、 MSL并未放弃，将推荐系统扩展10倍……ClusterMAX排名即将推出？ https://newsletter.semianalysis.com/p/meta-compute-everyone-wants-to-be

Rohan Paul@rohanpaul_ai · 1天前42

Paper from Meta shows Quantized reasoning models often lose because they keep doubting a correct answer instead of finishing. Many of them reason well enough, but compression makes them hesitate at the wrong time. The problem is that post-training quantization, a way to shrink models after training, can make reasoning models cheaper to run but worse at finishing cleanly. The authors found that strong quantization does not only make models less capable, since in many failures the model already reached the right answer but then second-guessed itself. Their core idea is that quantization adds noise at uncertain word choices, so the model becomes more likely to pick words like “wait,” “but,” or “alternatively” that reopen the problem. They tested this across math, coding, and science tasks using 5 reasoning models, several quantization methods, and model sizes from 1.5B to 32B. The main result is that aggressive quantization raised overthinking failures up to 52%, while a small penalty on 50 hesitation words cut reasoning length by 12% to 23% and often kept or improved accuracy. Given compressed models are widely used to save memory and cost, very important to know that a very small decoding fix can stop many of them from wasting tokens and losing answers they already had. ---- Link – arxiv. org/abs/2606.00206 Title: "Quantized Reasoning Models Think They Need to Think Longer, but They Do Not"

译Meta 新论文发现，后训练量化虽能缩小推理模型、降低部署成本，但会导致模型在已得出正确答案后反复自我怀疑，浪费 token。量化在不确定的词选择上引入噪声，使模型更倾向使用“wait”“but”“alternatively”等词重新开启推理。在 5 个推理模型（1.5B-32B）的数学、编程和科学任务上，激进量化使过度思考失败率最高达 52%。通过给 50 个犹豫词施以小惩罚，可剪掉 12%-23% 的推理长度，同时保持甚至提升准确率。

Rohan Paul@rohanpaul_ai · 1天前74

Mandeep Singh from Bloomberg on Meta's move to cloud computing Meta’s $150B CapEx now needs ROI; renting compute may bring revenue. Selling compute can fund AI, but it looks more like fallback revenue than frontier AI leadership.

译Meta 计划将自身用于模型、广告等任务的过剩 AI 算力转化为云业务，允许开发者租用数据中心内的模型访问（类似 AWS Bedrock），也可能出租原始算力，旨在为高达 1500 亿美元的资本支出寻找回报，减少对广告收入的依赖。消息公布后 Meta 股价上涨超 10%，而 AI 云公司 CoreWeave 和 Nebius 分别下跌 10.8% 和 12.4%。Zuckerberg 透露几乎每周都有外部公司向 Meta 请求算力，但此举更像为支出过高担忧提供财务安全阀；Meta 要成为 AWS、Azure 级别的云服务商仍需应对计费、安全、开发工具等挑战。

Rohan Paul@rohanpaul_ai · 1天前66

Meta is turning excess AI compute into a cloud business after shares jumped more than 10%. Meta built huge AI infrastructure for its own models, ads, feeds, and assistants. That created a familiar cloud problem, because expensive chips cannot sit idle for long. The new plan would let developers rent model access hosted inside Meta’s data centers. This looks like AWS Bedrock, where customers call models without managing the hardware. Meta may also rent raw compute, which hits CoreWeave and Nebius more directly. Shares of CoreWeave fell 10.8% and Nebius fell 12.4% after the Meta cloud report, while Meta shares rose more than 10%. CoreWeave and Nebius are AI cloud companies that rent computing power to customers, and Meta is already a major customer for them. This news scared neo-cloud (CoreWeave, Nebius etc) investors because if Meta now rents out its own AI computing power, that will mean it buys less capacity from CoreWeave and Nebius while also competing with them for other customers. Zuckerberg had already said outside companies ask Meta for compute almost every week. That comment now reads like a financial escape valve for AI overspending fears. Meta could reduce its ad dependence while proving its AI buildout has outside value. The catch is that cloud is not just racks, chips, and cheap power. Customers expect billing, uptime, security, support, migration help, and stable developer tools. Meta can rent compute faster than it can become AWS, Azure, or Google Cloud. Still, the signal is serious because AI infrastructure is becoming a tradable commodity. --- bloomberg. com/news/articles/2026-07-01/meta-is-building-a-cloud-business-to-sell-excess-ai-compute

译Meta 利用为自有模型、广告和助手建设的大型 AI 基础设施产生的过剩算力，计划向开发者出租模型访问（类似 AWS Bedrock）及原始算力。消息引发股价剧烈反应：Meta 涨超 10%，而 AI 云公司 CoreWeave 跌 10.8%、Nebius 跌 12.4%。Zuckerberg 此前透露外部公司几乎每周都向 Meta 请求算力。此举既降低 Meta 对广告收入的依赖，也证明其 AI 建设的外部价值，但云业务涉及计费、安全、工具支持等复杂环节，Meta 难以快速成为 AWS 级别的云服务商。

Chubby♨️@kimmonismus · 1天前66

Meta is spending hundreds of billions on AI compute. Selling the excess may be the ROI plan. Via Bloomberg: Bloomberg reports Meta is planning to sell access to excess AI compute and hosted models from its own infrastructure. That would move Meta into a crowded but lucrative lane: AWS Bedrock-style model access on one side, CoreWeave-style raw GPU capacity on the other. Meta has committed hundreds of billions to AI data centers and chips, while investors keep asking how that spend turns into revenue. We got another big hyperscale player incoming.

译Meta 已投入数百亿美元建设 AI 数据中心和芯片，现计划通过出售多余 AI 算力和托管模型来变现。据 Bloomberg 报道，Meta 的商业模式将覆盖两个方向：类似 AWS Bedrock 的模型托管服务，以及类似 CoreWeave 的裸 GPU 算力租赁。这一举动将 Meta 推入拥挤但利润丰厚的赛道，同时也回应了投资者对巨额支出如何转化为收入的持续质疑。

Rohan Paul@rohanpaul_ai · 2天前51

Meta open-sourced a brain-to-text system that reaches 78% word accuracy without surgery. Brain2Qwerty v2 converts non-invasive brain recordings into text with 61% average word accuracy and 78% for its strongest participant. The system reads MEG signals from a helmet, not electrodes placed inside brain tissue. 9 volunteers typed about 22,000 sentences while researchers recorded 10 hours of neural activity each. Brain2Qwerty v1 mostly mapped brain signals to single typed characters. It tries to recover characters, words, and full sentence meaning together. The system studies those brain signals and tries to turn them into the words you wanted to type. - 61% average word accuracy across all participants - 78% word accuracy for the top participant - 50%+ of sentences decoded with no more than 1 word error Performance improves as the data pile grows Raw brain signals are messy because many mental and physical processes fire at once. Deep learning handles that mess by learning patterns directly from the original recordings. A fine-tuned LLM then uses language context to repair likely word and sentence errors. This explains why the system beats earlier non-invasive methods reporting 8% word accuracy. More than half of sentences from the strongest participant had one word error or less. Accuracy also improved as training data grew, suggesting more recordings may close more of the gap.

译Meta开源非侵入式脑机接口系统Brain2Qwerty v2，通过读取MEG头盔采集的脑信号实现文字输出，无需植入电极。9名志愿者每人录入约10小时、共约2.2万句神经活动数据。系统平均词准确率61%，最强参与者达78%；超50%句子解码误差不超过1个词。v2版本联合映射脑信号到字符、词汇及完整句义，深度学习直接从原始信号学习模式，再经微调LLM利用上下文修正错误。相比早期非侵入方法8%的准确率显著提升，且准确率随训练数据量增加而提高。

小互@xiaohu · 2天前75

Meta 发布 Brain2Qwerty v2 你帮你脑子里在想的什么，实时转换成文字 不需要任何植入，仅需佩戴 MEG（脑磁图）头盔就能把你大脑产生的磁信号实时解码成连贯句子，全程不需要任何手术 字词准确率达 61%，约是其他无创脑机接口方法（8%）的 7.6 倍；最佳参与者达 78%，超半数句子只差一个词。 这是目前性能最高的非侵入式脑机接口系统....

译Meta 发布 Brain2Qwerty v2，无需手术植入，仅佩戴 MEG（脑磁图）头盔即可将大脑磁信号实时解码为连贯句子。字词准确率达 61%，约为其他无创脑机接口方法（8%）的 7.6 倍；最佳参与者达 78%，超半数句子只差一个词。Meta 称这是目前性能最高的非侵入式脑机接口系统。

Chubby♨️@kimmonismus · 3天前56

Meta says Brain2Qwerty v2 can decode natural sentences from non-invasive brain recordings in real time, reaching 61% word accuracy. The system was trained on about 22,000 sentences from 9 volunteers, each recorded for 10 hours with MEG while typing. Meta compares that with 8% word accuracy from prior non-invasive methods. Its best participant reached 78%, with more than half of sentences decoded with one word error or less. This is still controlled lab research: small participant pool, MEG hardware, active typing data, and company-reported results. Not a clinical communication device yet. Meta is releasing the training code, while BCBL is releasing the v1 dataset, pushing brain-to-text research further into open neuroscience infrastructure. I am so hyped for the future.

译Meta发布Brain2Qwerty v2，一种非侵入式脑机接口系统，能从实时脑信号解码完整自然句子，单词准确率达61%。系统基于约22000个句子训练，9名志愿者每人使用MEG记录10小时。相比此前非侵入方法8%的准确率大幅提升，最佳参与者达78%，超半数解码句子仅错一个词或更少。该端到端管线能实时将原始脑信号解码为单词和语义。但研究仍在受控实验室阶段：参与者样本小、依赖MEG硬件、数据来自主动打字、结果由公司报告，尚未成为临床通信设备。Meta已开源训练代码，BCBL发布v1数据集。

宝玉@dotey · 3天前79

Meta 今天同时放出两个大动作：Brain2Qwerty v1 论文正式登上 Nature Neuroscience，v2 同日发布。v1 去年以预印本形式公开时，能从脑电信号里逐字母还原打字内容，字符错误率 32%。v2 跳过了字母这一层，直接做到句子级别的实时解码，平均单词准确率 61%，表现最好的被试达到 78%，超过一半的句子解码误差在一个词以内。 作为参照，此前非侵入式方法的单词准确率只有 8%。 这里说的“非侵入式”，就是不需要开颅手术、不需要往脑子里植入电极。被试戴的是 MEG（脑磁图）设备，通过头皮外的传感器捕捉大脑活动产生的微弱磁场。相比之下，Neuralink 那类侵入式脑机接口准确率能到 90% 以上，但代价是一台开颅手术。 v2 的训练数据来自 9 名志愿者，每人戴着 MEG 设备打字 10 小时，总共录了约 22,000 个句子。系统用端到端深度学习直接处理原始脑信号，再通过微调大语言模型来利用语义上下文，把嘈杂的神经数据“翻译”成连贯的语言。Meta 还提到他们用 AI Agent 来探索解码流程的优化方案，最终的训练配置由工程师人工选定。 一个有意思的发现：解码准确率随数据量呈对数线性提升。也就是说，单靠增加训练数据就有可能继续缩小和侵入式方法之间的差距。 Meta 开源了 v1 和 v2 的全部训练代码，合作方 BCBL（巴斯克认知、大脑与语言中心）则开放了 v1 的数据集。 离实用还有多远？ MEG 设备体积大、造价数百万美元、需要磁屏蔽房间，目前只能在实验室环境下运行。而且这次的被试都是健康人，能否在真正需要帮助的脑损伤患者身上复现效果，还没有验证。便携式 MEG 替代方案（基于光泵磁力计）正在研发中，但离消费级产品还有相当距离。 不过，把非侵入式脑机接口的句子解码能力从“几乎不能用”拉到“大致能沟通“，这一步本身的意义在于：它证明了不开刀也有可能做到接近开刀的效果，剩下的是工程问题而非原理问题。 对全球数百万因脑损伤而丧失沟通能力的人来说，一条不需要手术的路径，哪怕还很远，还是很值得期待。 官方介绍：https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/

译Meta 在 Nature Neuroscience 发表 Brain2Qwerty v1 论文，同日发布 v2。v1 从脑电信号逐字母解码，字符错误率 32%。v2 实现句子级实时解码，平均单词准确率 61%，最优 78%，过半句子误差一个词内。此前非侵入式准确率仅 8%。v2 用 MEG 设备采集 9 名志愿者各约 10 小时打字数据（约 2.2 万句子），结合端到端深度学习与微调大语言模型。准确率随数据量对数线性提升。Meta 开源 v1、v2 全部训练代码。MEG 设备仍体积大、成本高，但该成果为脑损伤患者提供了无需开颅的可行路径。

Rohan Paul@rohanpaul_ai · 3天前53

The Information: Meta has reportedly limited engineer use of Claude Code and Codex because rival model outputs could contaminate Meta’s own AI training data and create contractual trouble with Anthropic and OpenAI. Distillation risk starts when a new model of Meta learns from another model’s outputs (from OpenAI or Anthropic), so even accidental reuse of Claude or Codex answers could look like Meta extracted capability from competitors rather than built it alone. OpenAI’s terms bar using output to develop competing models, and Anthropic says its terms do not allow Claude outputs to train models competitive with Anthropic’s own systems. Both OpenAI’s and Anthropic's terms bar using output to develop competing models. IMO, the safest strategy could be ingredient tracking: use rival tools for ordinary productivity only when outputs are barred from model-training pipelines, evaluation sets, benchmark generation, post-training data, reward-model data, and internal datasets that later feed model development. Of course a strong lawsuit usually needs much more ugly facts like: mass scraping, fake accounts, rate-limit evasion, automated extraction, direct use of outputs as training labels, or internal records showing the buyer knew it was cloning a rival system. In this situation, som of the typical safeguards are clean-room rules, approved enterprise accounts, no consumer accounts for sensitive work, training-data provenance logs, dataset quarantine, prompt and output retention, automated scanners for “AI-generated by vendor X” material, and access controls separating coding-agent work from model-training data.

译The Information报道，Meta已限制工程师使用Anthropic的Claude Code和OpenAI的Codex，原因是为防止竞争对手模型输出污染Meta自身AI训练数据，并引发合同纠纷。OpenAI和Anthropic的服务条款均禁止使用其输出来开发竞争模型。知识蒸馏风险在于即使意外复用竞品输出也可能被视为从竞争对手提取能力。建议的策略包括成分追踪：仅在不用于模型训练管线、评测集、基准生成、后训练数据、奖励模型数据及内部数据集时才使用竞品工具。典型防护措施有隔离规则、企业账户审批、训练数据溯源日志、数据集隔离及自动扫描“AI生成”标记等。

AYi@AYi_AInotes · 3天前71

扎克伯格在憋大活啊， 非侵入式脑机解码已经干到单词级实时输出了， Nature 打底，这一步比所有人预想的都快

译Meta（扎克伯格团队）在非侵入式脑机接口研究上取得重大突破，推出 Brain2Qwerty v2。该模型基于同日发表在 Nature 上的 v1，是目前最高性能的端到端管道，能从原始脑信号实时解码句子，将解码能力从字符级提升至单词和语义级，显著提升整体通信准确性。这一进展比预期更快，有望帮助数百万因脑损伤或疾病无法交流的患者。

elvis@omarsar0 · 3天前77

Highly recommended reading. What an impressive use of LLMs and deep learning. Achieves "real-time sentence decoding from non-invasive brain recordings, approaching levels of accuracy previously exclusive to techniques that require brain surgery."

译Meta AI 发布 Brain2Qwerty v2，非侵入性脑信号编码器最新里程碑，论文同日发表于《Nature》。该模型能从原始脑信号实时解码完整句子，准确度逼近需开颅手术的侵入式技术；从 v1 的字符级解码升级为词语及语义级解码，显著提升通信精度，有望帮助因脑损伤或障碍无法交流的数百万患者。

Chubby♨️@kimmonismus · 3天前50

Meta is now facing the exact problem every AI company will soon face. It wants to replace expensive external coding tools like Claude Code and Codex with its own internal system, MetaCode. But to build a better coding model, Meta has to make sure it is not accidentally training or evaluating on outputs from rival models. That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

译Meta 正面临每个 AI 公司都会遇到的难题：想用内部系统 MetaCode 取代 Claude Code、Codex 等昂贵的外部编码工具，但在构建更好的编码模型时，必须确保不意外地使用竞争对手模型的输出进行训练或评估。这就是知识蒸馏陷阱——公司越依赖前沿模型建设内部 AI 基础设施，就越难证明智能来源的独立性。

meng shao@shao__meng · 3天前16

逗死我了，美国政府 BAN 了 Llama，是因为太强太危险了吗？当然不是，因为太烂了。。拿出来，丢人！

Rohan Paul@rohanpaul_ai · 4天前52

FT: Google capped Meta’s use of Gemini after Meta asked for more model compute capacity than Google could supply. Meta’s problem is that it uses Gemini inside safety automation, customer support, ad tools, coding, and internal workflows. Google’s problem is different because it has paying cloud customers, its own Gemini products, and limited data center capacity all competing for the same chips, power, and networking. Google Cloud’s March-quarter revenue rose to $20 billion, but Sundar Pichai said a shortage of compute capacity kept growth lower and helped backlog nearly double versus the previous quarter. --- ft .com/content/c5d52f72-71ef-40bc-bad3-61afdba8b378?syn-25a6b1a6=1

译Google限制了Meta对Gemini模型的使用，原因是Meta要求的计算容量超出Google供应能力。Meta在安全自动化、客服、广告工具、编程及内部工作流中均依赖Gemini。Google面临自身云客户、Gemini产品与有限数据中心容量之间的资源竞争。Google Cloud 3月季度收入增至200亿美元，CEO Sundar Pichai表示计算容量短缺制约了增长，并导致未交付订单较前一季度近乎翻倍。

🚨 AI News | TestingCatalog@testingcatalog · 4天前52

Google vs Meta 🤖 > Google introduces restrictions on Meta's use on Gemini amid capacity shortage, according to the Financial Times. > Reportedly, this negatively affected internal projects at Meta related to customer support and content moderation, causing delays. I bet token efficiency will be a huge market in the long run, with a very transparent and predictable business model.

译Google vs Meta 🤖 > 据《金融时报》报道，Google因容量短缺对Meta使用Gemini施加限制。 > 据报道，这负面影响了Meta内部与客户支持和内容审核相关的项目，导致项目延期。 我敢打赌，从长远来看，token效率将成为一个巨大的市场，其商业模式非常透明且可预测。

🚨 AI News | TestingCatalog@testingcatalog · 4天前60

Meta AI app for iOS got incognito chats and a new look for the Glasses page. The updated page has shortcuts for all the primary toggles, including live translation and conversation focus.

译Meta AI app for iOS 新增了隐身聊天功能，并为 Glasses 页面提供了新外观。 更新后的页面包含所有主要开关的快捷键，包括实时翻译和对话焦点。

SemiAnalysis@SemiAnalysis_ · 6天前60

BREAKING NEWS: OpenAI’s Technical Lead for Compute has returned to OpenAI after joining Meta in April 2026. What is going on at Meta that people are quitting after only a couple of months? Did Anuj get reorged into a data-labeling role at Meta, and is that why he left?

译突发新闻：OpenAI的计算技术负责人在2026年4月加入Meta后，现已重返OpenAI。Meta发生了什么，让人们只待了几个月就离职？ Anuj在Meta被重组到了数据标注岗位，这就是他离开的原因吗？

elvis@omarsar0 · 7天前41

New research from Meta. Building synthetic training data has stayed a fixed pipeline that you hand-tune and then freeze. Autodata casts an AI agent as a data scientist that builds training and evaluation data, with an implementation called Agentic Self-Instruct that extends classic Self-Instruct with agentic planning and tool use. Think of it as meta-optimization, where the data scientist agent is itself trained to produce stronger data, so the pipeline keeps improving instead of staying static. Across computer science research, legal reasoning, and reasoning over mathematical objects, it beats classical synthetic-data methods, and meta-optimizing the agent delivers an even larger uplift. Paper: https://arxiv.org/abs/2606.25996 Learn to build effective AI agents in our academy: https://academy.dair.ai/

译Meta 发布新研究 Autodata，提出 Agentic Self-Instruct 方法。该方法将 AI 智能体视为数据科学家，通过智能体规划与工具使用，替代传统手工调优后固定的合成数据流水线。该智能体自身可通过元优化持续改进，从而生成更强训练数据。实验在计算机科学、法律推理、数学对象推理三个领域均超越经典合成数据方法，且元优化带来更大提升。论文见 arxiv。

Rohan Paul@rohanpaul_ai · 7天前47

Very important Meta paper brings Autodata, an agentic data scientist to create high quality synthetic data. The main result is that agent-made data usually trained models better than standard synthetic data, and in legal tasks a trained 4B model beat a much larger 397B baseline. Treats synthetic data generation as a job for an agentic data scientist, not a prompt template. “Agentic Self-Instruct,” makes AI agents generate and meta-optimize synthetic training and evaluation data, improving performance over classical synthetic data methods across CS, legal, and math benchmarks. Autodata’s loop is simple: generate an example, let a weak model and a strong model try it, judge the results, then revise the recipe until the example sits in the useful zone. This is the best idea in the paper: difficulty is not a virtue by itself. A task should not just be “hard”; it should be hard in a way that teaches the weaker model something. If the weak model always gets it right, there is nothing to learn; if it always gets zero, there is also nothing to learn. --- The direction feels important because it reframes synthetic data from bulk imitation into curriculum design. The next frontier may not be models writing more examples, but models learning what makes an example worth learning from. ---- Link – arxiv. org/abs/2606.25996v1 Title: "Autodata: An agentic data scientist to create high quality synthetic data"

译Meta提出Autodata，将合成数据生成视为智能体数据科学家的任务。核心方法“Agentic Self-Instruct”让AI智能体生成并元优化合成训练与评估数据。循环流程：生成示例→弱模型与强模型分别尝试→判断结果→修订配方直至示例处于有用区间。论文强调难度不是美德，示例应针对弱模型的学习点。关键结果：在法律任务上，4B模型训练后超越了更大的397B基线。

Yuchen Jin@Yuchenj_UW · 7天前44

I didn’t realize Denny Zhou, who led the Gemini Reasoning Team, left Google 4 months ago for Meta’s TBD Lab. A lot of people left Google recently. I’m still waiting for Gemini to catch up in coding. Time for Sergey to pull a Code Red.

译我没意识到Denny Zhou——曾领导Gemini推理团队——已在4个月前离开Google，加入Meta的TBD Lab。 最近很多人离开了Google。我仍在等待Gemini在编码方面赶上。是时候让Sergey启动Code Red了。

SemiAnalysis@SemiAnalysis_ · 6月24日62

Meta leadership voting on a motion to re-allocate 7k engineers to data labelling org

译Meta领导层正在投票一项动议，将7000名工程师重新分配至数据标注部门。

Rohan Paul@rohanpaul_ai · 6月24日59

NYT: The Trump administration is pressuring Meta to submit its AI models to government review before public release. And now, Meta the only major U.S. AI lab still outside the voluntary review system. OpenAI, Anthropic, Google, xAI, and Microsoft have already agreed to share models with the government’s AI safety group, according to NYT reporting. The Govt review process is meant to test whether advanced models can help with sensitive cyber tasks, expose security weaknesses, or create national security risks before millions of people can use them. --- nytimes .com/2026/06/23/business/meta-ai-government-reviews-security.html

译据NYT 2026年6月23日报道，特朗普政府正施压Meta，要求其AI模型在公开发布前提交政府审查。Meta目前是美国唯一未加入该自愿审查系统的主要AI实验室。OpenAI、Anthropic、Google、xAI和微软均已同意与政府AI安全小组共享模型。审查目的旨在测试先进模型是否可用于敏感网络任务、暴露安全漏洞或构成国家安全风险。

🚨 AI News | TestingCatalog@testingcatalog · 6月24日45

Meta announced a new series of Meta Glasses in partnership with EssilorLuxottica. > Compatible with prescription lenses. > 26 styles across a range of colors, lenses, and frames. > Launching with Meta AI powered by Muse Spark from day one. My Meta HSTN still didn't get Muse Spark, tho 👀

译Meta宣布与EssilorLuxottica合作推出新系列Meta Glasses。 > 兼容处方镜片。 > 26种款式，涵盖多种颜色、镜片和镜框。 > 从第一天起即搭载由Muse Spark驱动的Meta AI。 不过我的Meta HSTN还没收到Muse Spark呢👀

Chubby♨️@kimmonismus · 6月24日31

Re: Meta Mythos rumors. A Meta Mythos would be fascinating. I just think the strategic need for it is much less obvious than it is for OpenAI or Anthropic. First of all, I still stand by my view that this would certainly be an exciting development for Meta, but fundamentally not nearly as important for Meta as comparable frontier-level progress is for labs like Anthropic or OpenAI. Why? Because Meta already has revenue and is pursuing a different path. Its LLM only needs to be good enough for consumers to keep using it. In practice, that means good enough for everyday use, simple daily questions, and somewhat more complex tasks. And for that, its current model is already sufficient, while clearly continuing to improve. A Meta Mythos would definitely be interesting, and I am happy to be surprised. But unless Meta actually plans to move into areas like autonomous scientific research, I still find myself asking: what is the real purpose?

译Kim 评论 Meta Mythos 传闻，认为其固然令人兴奋，但对 Meta 的战略意义远不及对 OpenAI 或 Anthropic 那样关键。原因是 Meta 已有稳定营收并走不同路线，其 LLM 只需足够好以维持消费者日常使用（简单问答及稍复杂任务），当前模型已胜任且持续改进。除非 Meta 计划切入自主科研等领域，否则 Mythos 级模型的真正目的何在仍存疑问。

🚨 AI News | TestingCatalog@testingcatalog · 6月21日22

Meta AI is getting a new Artifacts tab on the web. All the presentations, docs, web pages and other creations would be stored over there. Bridging the gap 👀

译Meta AI 网页版将新增一个 Artifacts 标签页。所有演示文稿、文档、网页及其他创作内容都将存储在此处。 缩小差距 👀

Chubby♨️@kimmonismus · 6月21日48

No more tokenmaxxing at Meta Meta is preparing to curb internal AI usage after employee token consumption surged so sharply that the company now expects internal AI costs alone to reach billions of dollars in 2026 (looking at you Claude). The move marks a sharp reversal from Meta’s earlier push to reward “AI-driven impact,” as the company now builds an AI Gateway to track spending, impose token budgets, and shift employees toward in-house tools like MetaCode.

译Meta 内部不再 token 拉满了。 Meta 正准备限制内部 AI 的使用，原因是员工 token 消耗激增，以至于公司预计仅内部 AI 成本到 2026 年就将达到数十亿美元（说的就是你，Claude）。 这一举措标志着 Meta 此前鼓励“AI 驱动影响力”的立场出现急剧反转，公司目前正在构建一个 AI Gateway 来追踪开支、设定 token 预算，并引导员工转向 MetaCode 等内部工具。

Rohan Paul@rohanpaul_ai · 6月20日40

Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/openai-just-moved-frontier-level 🗞️ OpenAI just moved frontier-level health AI from premium reasoning models into the free GPT-5.5 Instant model. 🗞️ The article that went super viral - Satya Nadella on organizational economics of AI and “token capital” 🗞️ Anthropic just showed Claude Opus 4.7 program a robodog in 12:07 mint, about 20x faster than last year’s Claude-aided human team on the tested tasks. 🗞️ This was long needed for AI in finance - Making SEC filings readable for machines without flattening the accounting logic. 🗞️ Mark Zuckerberg is trying to restart Meta’s hacker culture after 8,000 layoffs but employees are pushing back. 🗞️ Anthropic just rolled out a major Claude Design update, adding design system imports, code round-trips, and a fix for its heavy token usage issue.

译OpenAI将前沿级健康AI从高级推理模型移至免费的GPT-5.5 Instant模型。Satya Nadella关于AI组织经济学和“token资本”的文章走红。Anthropic展示Claude Opus 4.7用12分07秒编程机器人狗，比去年人类团队快20倍，并推出Claude Design大更新（支持设计系统导入、代码往返、修复高token用量）。AI金融领域推动SEC文件可机器读取但不简化会计逻辑。扎克伯格裁员8000人后试图重启Meta黑客文化，员工抵制。

Deedy@deedydas · 6月18日60

I thought this was a joke. Meta now has made 30-50% of software engineers on core teams become data labelers. Their job is "giving human feedback on AI-generated Github repos" in an org called Agent Data Optimization. Maybe we are all training data generators after all.

译我以为这是个玩笑。 Meta现在让核心团队中30-50%的软件工程师变成了数据标注员。 他们的工作是在一个名为Agent Data Optimization的部门中"对AI生成的GitHub仓库提供人类反馈"。 也许我们终究都是训练数据生成器。

Rohan Paul@rohanpaul_ai · 6月18日58

Meta’s CTO says that morale is near its lowest point in 20 years, as layoffs and AI labor shifts strain the company from inside. Meta CTO Andrew Bosworth compared the mood to the Cambridge Analytica era. Meta, over the recent past, has cut 10% of staff, moved roughly another 10% into work supporting AI model training, and faced backlash over tracking mouse movements and keystrokes for AI improvement. --- businessinsider .com/meta-cto-andrew-bosworth-addresses-morale-after-layoffs-ai-shift-2026-6

译Meta CTO Andrew Bosworth表示公司士气跌至20年来最低点，堪比剑桥分析事件时期。过去一段时间Meta已裁员10%，又将约10%员工调岗支持AI模型训练，因追踪鼠标移动和键盘敲击用于AI改进而引发争议。CEO Mark Zuckerberg在8000人裁员后试图重启黑客文化，但遭员工抵制；他承诺7月举办全公司AI黑客松，员工反应冷淡。

Rohan Paul@rohanpaul_ai · 6月13日67

Reuters: Meta just admitted its AI workforce rebuild moved faster than its organization could absorb. 10% of workers were cut, 7,000 were moved into AI workflow roles, and Zuckerberg is now telling staff some people may need to be moved back. Zuckerberg’s memo admits the company moved fast enough to create bad fits, especially after wider manager spans reportedly reached 50:1 in the new Applied AI Engineering unit. Meta is still spending aggressively, with annual capital spending raised to $125 B-$145B, mostly into huge compute, data centers, networking, and power. --- reuters .com/business/metas-zuckerberg-admits-mistakes-made-ai-transformation-2026-06-12/

译路透社报道，Meta在重建AI团队时动作过快。10%员工被裁，7000人转入AI工作流岗位，扎克伯格在内部备忘录中承认部分安排不匹配，可能需将部分员工调回。新成立的Applied AI Engineering单元管理跨度达50:1。Meta仍在大力投入，年度资本支出上调至$125B-$145B，主要用于算力、数据中心、网络和电力。

🚨 AI News | TestingCatalog@testingcatalog · 6月12日32

Meta is preparing to add 3 new modes for Meta AI: Deep Research, Presentations, and Social. All these features are already working, but users can now explicitly select what they want. Closing gaps 👀

译Meta 正准备为 Meta AI 新增三种模式：深度研究、演示文稿和社交。 所有这些功能已在运行，但用户现在可以明确选择他们想要的了。 正在缩小差距 👀

🚨 AI News | TestingCatalog@testingcatalog · 6月11日40

Meta is working on support for Custom Instructions for Meta AI on the web.

译Meta 正在为网络版 Meta AI 开发自定义指令支持。

Deedy@deedydas · 6月8日64

Meta AI has shockingly grown 2.5x in the last 2mos and is poised to be the #3 AI consumer app in the world behind Gemini and ChatGPT. Sadly, this growth is very likely inorganic given it has by far the worst retention by a mile: only 4.5% users stay in 30 days.

译Meta AI 在过去两个月内惊人地增长了 2.5 倍，有望成为仅次于 Gemini 和 ChatGPT 的全球第三大 AI 消费级应用。遗憾的是，这种增长很可能是非有机的，因为它的留存率迄今最差：只有 4.5% 的用户会在 30 天后继续使用。

Yann LeCun@ylecun · 6月6日10

Reminder.

译提醒一下。 （网友 @JosephJacks_：我们干脆让 Yann LeCun 当 AI 总统，然后收工吧？）

AI at Meta@AIatMeta · 6月5日64

Big congrats to our SAM 3D team for receiving a Best Paper Honorable Mention at #CVPR26! This prestigious recognition underscores their incredible work pushing the boundaries of computer vision. Read the paper here: https://arxiv.org/abs/2511.16624

译热烈祝贺我们的 SAM 3D 团队在 #CVPR26 获得最佳论文荣誉提名！这项殊荣凸显了他们在推动计算机视觉边界方面的杰出工作。 论文链接：https://arxiv.org/abs/2511.16624

Rohan Paul@rohanpaul_ai · 5月30日74

The information: Meta is preparing its biggest AI wearable push yet, with a AI pendant, more AI glasses, and a business service called Wearables for Work. Meta’s bet is that the next AI interface is not a chat box, but a sensor-rich device tied to an assistant that can remember meetings, summarize conversations, answer visual questions, and trigger actions. The reported target is huge: 10M wearable sales in the second half of 2026 and 6.8M monthly active wearable users by year-end. The software layer may matter more than the hardware, because Meta AI Assistant, Hatch, subscriptions, and wearable apps turn a device sale into recurring AI revenue. The pressure is obvious: Reality Labs posted a $4.03B operating loss on only $402M revenue in Q1-26, so Meta needs wearables to become a platform, not another expensive gadget line. --- theinformation .com/briefings/meta-plans-ai-pendant-part-ambitious-wearables-expansion

译Meta正准备迄今规模最大的AI可穿戴设备推进，包括AI项链、更多AI眼镜以及企业服务“Wearables for Work”。其押注下一代AI交互界面不是聊天框，而是具备丰富传感器、能记住会议、总结对话、回答视觉问题并触发操作的AI助手设备。报道的销售目标宏大：2026年下半年销量目标1000万台，年底月活用户目标680万。软件层被视作关键，可将设备销售转化为持续性AI收入。此举背后的财务压力明显：Reality Labs在2026年第一季度录得40.3亿美元运营亏损，营收仅为4.02亿美元，因此Meta亟需将可穿戴设备发展成一个平台，而非又一条昂贵的硬件产品线。

Rohan Paul@rohanpaul_ai · 5月29日65

Yann LeCun's new paper asks when LeJEPA truly learns hidden world variables, and finds Gaussian structure is the key. Means LeJEPA can only reliably learn the real hidden causes behind what it sees when those causes are shaped like a balanced Gaussian cloud. The paper proves that, when the true hidden variables are independent Gaussian variables and the paired views come from a stable noisy process, the best LeJEPA solution must recover those variables up to a rotation or flip. The paper gives a math reason for when a self-supervised AI model is really learning the structure of the world, not just making useful features that happen to work on a test. ---- Link – arxiv. org/abs/2605.26379 Title: "When Does LeJEPA Learn a World Model?"

译Yann LeCun团队的新论文探讨了LeJEPA模型学习真实世界隐藏变量的条件。其核心结论是，LeJEPA只有在真实的隐藏变量呈现高斯云结构时，才能可靠地学习它们。论文通过数学证明，当这些隐藏变量是独立高斯变量，并且配对视图由一个稳定的噪声过程生成时，LeJEPA的最优解能够以旋转或翻转等价的形式恢复这些变量。这项研究为自监督AI模型究竟在何时能真正理解世界结构（而不仅仅是提取在测试集上有效的特征）提供了理论解释。

SemiAnalysis@SemiAnalysis_ · 5月27日25

PoV: 70% of New Grad SWE at Meta being reassigned to apply their engineering talent to this RL task

译视角：Meta 70%的新入职软件工程师被重新分配，运用其工程才能参与这项强化学习任务。