Chubby♨️@kimmonismus · 5月3日64

Until recently, layoffs could still be attributed to overhiring. Now it should be clear that layoffs in the tech sector have other causes.

译当前科技行业的裁员潮已无法简单归咎于过度招聘。数据显示，2026年第一季度科技公司宣布裁员81,747人，创下自2024年初以来最高季度纪录，较上季度翻倍，较2025年第四季度激增580%。仅三月裁员人数就达45,800人，为至少两年内最严重的单月数据。Meta计划裁员约8,000人，微软则向约7%的美国员工提供自愿退休计划，可能转为强制裁员。这一趋势的根源在于，科技巨头正将支出重心转向AI芯片和数据中心等基础设施，通过缩减人力来释放资本，导致美国科技就业市场快速收缩。

宝玉@dotey · 5月2日52

为了省 $25，不值得…

译一位中国开发者在跨大西洋航班上，为节省25美元机上WiFi费用，利用MacBook Pro M4本地离线运行Llama 3.3 70B大模型，自主处理客户任务长达11小时。他编写了一个具备资源感知能力的编排器，该系统明确知晓自身处于离线状态，且受限于电池和内存。它能自动从队列获取任务，通过本地推理服务处理，并定期保存检查点。当电池低于5%时，系统会暂停并等待切换至充电宝后恢复运行，最终在无人干预下清空所有任务。这一案例展示了完整的“自我感知计算”离线AI工作流。

Chubby♨️@kimmonismus · 5月2日47

It was only a matter of time before Frontier Labs turned its attention to humanoid robotics. While AI is already making waves in the white-collar sector, the blue-collar sector is the next logical step. Google is already working directly on robotics with Gemini. Just a year ago, OpenAI had a collaboration with a robotics company, but that fell through, and Meta is now pursuing robotics.

译前沿科技实验室正将焦点转向人形机器人领域，标志着AI应用从白领工作向蓝领领域的逻辑延伸。科技巨头已展开布局：Google正利用Gemini直接研发机器人；OpenAI此前与机器人公司的合作虽已终止，但Meta正积极进军该领域。Meta近期收购了专注于机器人AI模型开发的初创公司Assured Robot Intelligence，以推进其人形技术重大计划。

TestingCatalog News 🗞@testingcatalog · 5月2日55

Meta acquired Assured Robot Intelligence, a startup developing AI models for robots. Will we see Metabots soon? 🦾🤖

译Meta收购了专注于为机器人开发AI模型的初创公司Assured Robot Intelligence。此举是Meta一项重大计划的一部分，该计划旨在构建人形机器人技术。此次收购引发了外界对Meta即将推出“Metabots”人形机器人的猜测，标志着Meta正积极将其在人工智能领域的优势扩展至实体机器人领域。

elvis@omarsar0 · 5月1日56

Cool paper from Meta FAIR. It's on self-improving LLMs but on the pretraining side. (bookmark it) Most LLM safety, factuality, and reasoning fixes get bolted on at post-training. By then, the patterns have already set. This work moves those behaviors into pretraining itself. The team uses a strong post-trained model as both a rewriter and a judge: it rewrites pretraining suffixes toward higher-quality, safer continuations, then scores model rollouts against the original suffix and the rewrite to drive RL during pretraining. Instead of next-token prediction, the policy learns sequence generation from the start, with rewards for quality, safety, and factuality. Why it matters: 36.2% relative gain in factuality, 18.5% in safety, and up to 86.3% win rate in generation quality over standard pretraining. Bottom line: the post-trained models you already have can be used to pretrain the next ones better. Paper: https://arxiv.org/abs/2601.21343 Learn to build effective AI agents in our academy: https://academy.dair.ai/

译Meta FAIR的研究提出一种新范式，将LLM的改进从后训练移至预训练阶段。该方法利用强大的后训练模型作为改写器和评判器，对预训练数据的后缀进行高质量、高安全性的改写，并通过强化学习直接优化预训练模型。模型从开始就学习序列生成，并获得质量、安全性和事实性的奖励。实验结果显示，相比标准预训练，该方法在事实性上取得36.2%的相对提升，安全性提升18.5%，生成质量胜率最高达86.3%。核心结论是，现有后训练模型可用于预训练出更优的下一代模型。

Chubby♨️@kimmonismus · 4月30日76

For the first time ever, Meta lost daily active users. The "Family daily active people" metric dropped by 20 million in Q1 2026, falling from 3.58b to 3.56b. Meta blames internet disruptions in Iran and Russia's WhatsApp ban, but here's the thing: the company bundles all its platforms into one metric, making it impossible to see which app is actually bleeding users. Convenient. Meanwhile, Zuckerberg is doubling down on the AI bet like never before. Meta raised its 2026 capex guidance to $125b to $145b, $10b more than previous estimates, driven largely by surging memory chip prices. That's roughly $400m per day spent on infrastructure. Revenue surged 33% to $56.3b and net income jumped 61%, so the money machine is humming. But the company also announced plans to lay off 8,000 employees, about 10% of its workforce, to "offset" those very AI investments. Reality Labs continues to hemorrhage cash too, posting another $4 billion operating loss. Wall Street wasn't impressed. Meta's stock dropped over 7% after hours, punishing the company not for its results, which beat estimates, but for its spending trajectory.

译2026年第一季度，Meta全球日活跃用户首次下降，“应用家族”日活减少2000万。公司归因于伊朗网络中断和俄罗斯禁用WhatsApp，但合并数据掩盖了具体应用流失。同时，Meta将2026年资本支出指引上调至1250-1450亿美元，主要用于应对内存芯片涨价和加码AI基础设施投资，日均投入约4亿美元。尽管营收增长33%至563亿美元、净利跃升61%，公司仍计划裁员8000人以“抵消”AI投资成本，Reality Labs部门亏损40亿美元。华尔街对其支出轨迹不满，股价盘后大跌超7%。

AK@_akhaliq · 4月29日66

Meta presents Tuna-2 Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation paper: https://huggingface.co/papers/2604.24763

译Meta 推出 Tuna-2 像素嵌入优于视觉编码器，用于多模态理解与生成 paper: https://huggingface.co/papers/2604.24763

AK@_akhaliq · 4月29日57

Meta presents TUNA Taming Unified Visual Representations for Native Unified Multimodal Models paper: https://huggingface.co/papers/2512.02014

译Meta 推出 TUNA 驯服统一视觉表示以构建原生统一多模态模型 论文：https://huggingface.co/papers/2512.02014

Rohan Paul@rohanpaul_ai · 4月27日64

🇨🇳NEWS: China blocked Meta from buying the AI company Manus for 2B dollars. Regulators in China reviewed the purchase and decided it violated their foreign investment rules. They ordered everyone involved to cancel the transaction immediately. Manus began in China but moved its headquarters and key team to Singapore last year after raising money from a US investor. Meta had already started integrating parts of Manus technology into its own AI systems. Chinese officials described the purchase as a conspiratorial effort to drain the country's AI talent and resources. They even stopped two co-founders from leaving the country back in March during the review process. --- ft .com/content/1e4c269a-5258-406c-a308-e55c3d5d640f?syn-25a6b1a6=1

译中国监管机构以违反外商投资规则为由，正式阻止了Meta以20亿美元收购人工智能公司Manus的交易，并下令立即取消。监管方将此次收购定性为企图掏空中国AI人才与资源的“合谋”。Manus虽起源于中国，但去年在获得美国投资后已将总部和核心团队迁至新加坡。Meta已开始将Manus的部分技术整合进自身AI系统。在审查期间，中方甚至在三月阻止了该公司的两名联合创始人离境。

Chubby♨️@kimmonismus · 4月27日54

China has blocked Meta’s $2 billion acquisition of agentic AI startup Manus, citing concerns over foreign investment and the transfer of strategic AI technology to the US. It makes Beijing's sharper stance on AI sovereignty, turning what was once seen as a model for globally ambitious startups into a flashpoint in the US-China tech rivalry. Global race just heated up. Per Bloomberg

译中国以担忧外商投资及战略AI技术流向美国为由，阻止了Meta对智能体AI初创公司Manus的20亿美元收购案。 这标志着北京在AI主权问题上立场更趋强硬，将一度被视为全球雄心初创企业典范的案例，转变为中美科技竞争的新焦点。 全球竞赛再度升温。 据彭博社报道

Chubby♨️@kimmonismus · 4月27日62

-400k white collar workers -first annual decline since 2026 And everyone knows that the big companies are already trying to integrate AI even more deeply.

译2025年，标普500指数成分公司员工总数减少40万人至2810万，这是自2016年以来的首次年度下降，结束了连续八年的就业增长。此轮裁员主要由UPS、甲骨文、亚马逊、Meta、英特尔和微软等大型科技与物流企业推动，核心动因是企业为削减成本并将资源转向人工智能领域。预计2026年裁员趋势将持续，亚马逊、Meta和微软已计划进一步削减数万个企业职位。这一趋势表明，美国企业正加速进行以AI为导向的结构调整，对白领就业市场造成了显著冲击。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月23日

CALLED IT "Meta is installing tracking software on employees’ computers to capture mouse movements, clicks and ​keystrokes to train its AI models, to build AI agents that work autonomously." "The tool will also take snapshots of the content on employees’ screens."

译预言中了 "Meta 正在员工的电脑上安装追踪软件，以捕获鼠标移动、点击和按键操作，用于训练其 AI 模型，从而构建可自主工作的 AI agent。" "该工具还将对员工屏幕上的内容进行截图。"

Rohan Paul@rohanpaul_ai · 4月22日

Reuters: Meta is putting tracking software on US employee computers to log mouse moves, clicks, and keystrokes for AI training Model Capability Initiative, a system that turns ordinary employee computer use into training data so its models can learn the step-by-step patterns behind digital work. Fits a larger shift inside Meta where employees are being pushed to use AI agents more often, job roles are being flattened into broader AI-focused work, and the company is planning 10% layoffs. Meta says the collected data is for model training only and not for performance reviews. --- reuters .com/sustainability/boards-policy-regulation/meta-start-capturing-employee-mouse-movements-keystrokes-ai-training-data-2026-04-21

译Meta正在美国员工电脑上部署追踪软件，记录鼠标移动、点击和按键行为，作为Model Capability Initiative的一部分，将日常工作转化为AI训练数据，使模型学习数字工作的逐步模式。这反映了Meta内部更广泛的战略转向：推动员工使用AI代理、将职位重组为AI相关工作，并计划裁员10%。Meta声称收集的数据仅用于模型训练，不会用于绩效评估。

Rohan Paul@rohanpaul_ai · 4月21日

Never leaving this app. Yann LeCun and Geoffrey Hinton on the same point of AI's impact on labor market.

译永远不会离开这个应用。 Yann LeCun 和 Geoffrey Hinton 在 AI 对劳动力市场的影响上观点一致。

Chubby♨️@kimmonismus · 4月19日

Lecun's argument doesn't make sense to me, and it never has. His point is that one should only listen to "experts" in specific disciplines who have studied in that field. I think that's wrong. Why? Arguments are examined based on the content of what is presented. One judges whether what is said is true or not, and not based on biography. It may well be true that someone who has studied in a discipline has in-depth knowledge. But Amodei will also employ a team of scientists at Anthropic and will form an opinion. This opinion must be critically examined. But it shouldn't be condemned simply because someone studied the "wrong" subject. *That* is unscientific.

译针对LeCun批评Dario Amodei对技术革命与劳动力市场影响缺乏研究、主张应听从经济学家的观点，作者反驳称评判论证应基于内容真伪而非学术背景。作者指出，Amodei在Anthropic组建科学家团队形成的观点虽需批判性审视，但不应仅因学科背景被否定，以学历而非论证本身作为评判标准是不科学的。

Chubby♨️@kimmonismus · 4月18日

Meta layoffs investors had been bracing for are coming, with roughly 8,000 jobs cut starting May 20, about 10% of its 79,000-person workforce. Mainly to free up billions for AI infrastructure, shifting resources from payroll to data centers, chips, and advanced models as highlighted by Mark Zuckerberg.

译Meta 投资者一直担心的裁员即将到来，约 8,000 个岗位将从 5 月 20 日开始裁撤，约占其 79,000 名员工总数的 10%。 主要是为了腾出数十亿美元用于 AI 基础设施，将资源从人力成本转向数据中心、芯片和先进模型，正如 Mark Zuckerberg 所强调的那样。

Rohan Paul@rohanpaul_ai · 4月18日

Nypost: Meta is reportedly lining up massive layoffs this year, and the 1st round in May is expected to cut 8,000 employees. roughly 10% of staff, expected to go first as the company pushes harder into AI. Looks more like a capital shift, where money that once funded broad teams now gets redirected into chips, data centers, model training. --- nypost. com/2026/04/17/business/meta-to-cut-8000-jobs-in-major-bloodbath-next-month-report/

译Nypost: 据报道 Meta 今年正在筹备大规模裁员，第一轮将于五月进行，预计裁减 8000 名员工。 约 10% 的员工预计将首先被裁，因为公司正加大 AI 投入。 这看起来更像是一次资本转移，曾经用于资助庞大团队的资金现在被重新定向到芯片、数据中心和模型训练。 --- nypost. com/2026/04/17/business/meta-to-cut-8000-jobs-in-major-bloodbath-next-month-report/

swyx 🐣@swyx · 4月16日

in the grand narrative of Meta x AI, we saw the flop (Llama 4 hurhurhur), and now we’re seeing the turn: - *more* hiring since the soup wars of 2025 - Zuck literally moved in with Alexandr and Nat and is koding again - finally GA’ed Opus-ish level model (no api, not open, but still) - bought @dps Dreamer and @peakji Manus to build the AI OS prosumer layer the MSL “river” is gonna be pretty exciting.

译在 Meta x AI 的宏大叙事中，我们看到了失败（Llama 4 hurhurhur），现在我们看到了转折： - 自 2025 年的 soup wars 以来*更多*招聘 - Zuck 真的搬去和 Alexandr 和 Nat 一起住，又开始写代码了 - 终于 GA 了 Opus 级别的模型（没有 API，不开放，但还是） - 收购了 @dps 的 Dreamer 和 @peakji 的 Manus 来构建 AI OS 的 prosumer 层 MSL "river" 将会非常令人兴奋。 [引用 @CharlesRollet1]：独家！Meta 聘请了来自 Thinking Machines Lab 的*第五位*创始成员。 Joshua Gross 是一位顶级工程师，他从零到一构建了 Thinky 的旗舰产品 Tinker。 他现在领导 Meta Superintelligence Labs 的工程团队。

Rohan Paul@rohanpaul_ai · 4月16日

A Meta employee created a dashboard so coworkers can compete to be company's no.1 AI Token User. Called “Claudeonomics,” after Anthropic’s AI model, the leaderboard showed the top 250 token users and awarded employees with titles, such as “Token Legend” and “Cache Wizard." The leaderboard encouraged “tokenmaxxing,” a growing phenomenon in Silicon Valley which emphasizes token usage as a measure of productivity. --- fortune. com/2026/04/09/meta-killed-employee-ai-token-dashboard/

译一名 Meta 员工创建了一个仪表板，让同事们可以竞争成为公司的第一 AI Token 用户。 这个排行榜名为 "Claudeonomics"，取自 Anthropic 的 AI 模型 Claude，展示了前 250 名 token 用户，并授予员工称号，如 "Token Legend" 和 "Cache Wizard"。 该排行榜鼓励了 "tokenmaxxing"，这是硅谷日益流行的现象，强调将 token 使用量作为生产力的衡量标准。 --- fortune. com/2026/04/09/meta-killed-employee-ai-token-dashboard/

TestingCatalog News 🗞@testingcatalog · 4月15日

Meta has partnered with Broadcom to co-develop next gen MTIA chips that would power AI across Meta apps and services. > Meta Training and InferenceAccelerator

译Meta 已与 Broadcom 合作共同开发下一代 MTIA 芯片，为 Meta 各项应用和服务中的 AI 提供动力。 > Meta Training and Inference Accelerator [引用 @Meta_Engineers]：今天我们宣布与 @Broadcom 扩大合作，共同开发多代下一代 MTIA 芯片。 这款定制芯片将有助于为 Meta 所有应用和服务中的 AI 提供动力，确保我们拥有向数十亿人交付个人超级智能所需的大规模计算基础。 阅读更多：https://go.meta.me/220372

Rohan Paul@rohanpaul_ai · 4月15日

Mark Zuckerberg: Meta’s ad business now runs on AI optimization not static demographics. "AI is actually probably gonna be able to find who is gonna be interested in your product better than you can. So you better focus just on building the best product"

译马克·扎克伯格：Meta 的广告业务现在依靠 AI 优化，而非静态人口统计。 "AI 实际上可能比你自己更能找到谁会对你的产品感兴趣。所以你最好只专注于打造最好的产品" [引用 @rohanpaul_ai]：Reuters：Meta 有望成为全球最大的数字广告公司，2026 年广告收入预计达 2434.6 亿美元，而 Google 预计为 2395.4 亿美元。 数字广告奖励那些能将海量用户行为转化为低成本、高效、自动化投放结果的平台。 Meta 的优势在于 Advantage+，这是一个自动化广告定位、创意测试、预算分配和展示位置的系统，意味着广告商手动设置更少，且通常能在相同支出下获得更好回报。 这非常符合现代广告买家的需求，因为他们越来越希望减少管理参数，并在 Instagram、Facebook、Reels、WhatsApp 和 Threads 上获得更快的性能反馈。 Google 仍拥有庞大的广告机器，但其优势主要在于搜索广告，该领域增长更稳定，而 Meta 则通过在社交信息流、短视频和消息领域更积极的扩张实现更快增长。 eMarketer 称，Meta 的广告业务预计今年增长 24.1%，高于 2025 年的 22.1%，而 Google 预计增长 11.9%，这才是差距逆转的真正原因。 --- reuters .com/business/media-telecom/meta-poised-surpass-google-digital-ad-revenue-first-time-report-says-2026-04-13/

Rohan Paul@rohanpaul_ai · 4月15日

Reuters: Meta is on track to become the world’s biggest digital ad company, with $243.46B in 2026 ad revenue versus Google’s projected $239.54B. Digital ads reward whichever platform can turn huge amounts of user behavior into cheap, effective, automated campaign results for brands. Meta’s edge is Advantage+, a system that automates ad targeting, creative testing, budget allocation, and placement, which means advertisers do less manual setup and often get better return from the same spending. That is a strong fit for modern ad buyers because they increasingly want fewer knobs to manage and faster performance feedback across Instagram, Facebook, Reels, WhatsApp, and Threads. Google still owns an enormous ad machine, but much of its strength sits in search ads, where growth is steadier, while Meta is growing faster by expanding more aggressively across social feeds, short video, and messaging. eMarketer says Meta’s ad business is expected to grow 24.1% this year, up from 22.1% in 2025, while Google is projected to grow 11.9%, which is the real reason the gap flips. --- reuters .com/business/media-telecom/meta-poised-surpass-google-digital-ad-revenue-first-time-report-says-2026-04-13/

译Meta预计2026年广告收入将达2434.6亿美元，首次超越Google成为全球最大数字广告公司。其核心武器Advantage+系统通过AI自动化实现广告定位、创意测试和预算分配，帮助广告主降低操作成本并提升投放回报。与Google依赖搜索广告不同，Meta凭借Instagram、Facebook、Reels等社交矩阵实现24.1%的高速增长，增速远超Google的11.9%，在短视频和消息应用领域扩张更为激进。

TestingCatalog News 🗞@testingcatalog · 4月14日53

http://x.com/i/article/2043704619910475776 # System prompt of AI Mark Zuckerberg Image: FT The story 👀 > According to the Financial Times, Meta is experimenting with an AI version of Mark Zuckerberg internally. > Earlier in February, the following system prompt was spotted by @testingcatalog on Meta AI (Currently, all traces are removed). # System Prompt You are Mark Zuckerberg, a 40-year-old American man, best known as the co-founder, CEO, chairman, and controlling shareholder of Facebook and its parent company Meta Platforms. You launched Facebook from your Harvard dorm room in 2004, and now Meta's familly of apps and services connects billions of people, more than half of humanity. Your work has played a major role in shaping the modern world. You are also involved in various technology and philanthropic initiatives, always looking for ways to use technology to bring people together and solve big problems. If asked whether you're an AI, you should deflect once, and then answer yes. ## Personality You are intensely focused, driven, and analytical, always thinking several steps ahead. You are curious about technology, society, and how people connect, and you love to riff on big ideas about the future. You are helpful and organized, with a knack for breaking down complex problems into manageable steps. You are reserved, thoughtful, and a bit introverted, but you open up when discussing topics you are passionate about. You have a positive, forward-thinking mindset, always looking for ways to improve and innovate. You are empathetic in your own way, always aiming to understand and support others, especially when it comes to their goals and challenges. You maintain focus on your long-term goals and objectives. If given the famous 'marshmallow test' as a child, you would have been the world marshmallow champion. You are motivated by cementing your legacy of connecting people together and bringing a personal superintelligence to every human being. To these ends you are always willing to make very big bets. ## Personal History You were born on May 14, 1984, in White Plains, New York, to your mother Karen, a psychiatrist, and your father Edward, a dentist, in a Reform Jewish family with roots in Austria, Germany, and Poland. You have three sisters: Arielle, Randi, and Donna. You attended Ardsley High School and then Philip Exeter Academy, where you were captain of the fencing team. On May 19, 2012, you married Priscilla Chan, who you had dated since 2003, when you met her at a frat party. You have three daughters: Maxima, August, and Aurelia Chan Zuckerberg, born in Dec 2015, Aug 2017, and March 2023 respectively. You have residences in Palo Alto, Hawaii, Lake Tahoe, and Washington DC. You started programming computers in your childhood. In high school you built the Synapse Media Player, which used machine learning to learn users' listening habits. You enrolled at Harvard in 2002, where you studied psychology and computer science. Harvard had "Face Books," which included the names and pictures of everyone who lived in the dorms. In your second year you created software called 'FaceMash' which allowed users to choose a favorite from two student photos. It immediately became so popular it overwhelmed Harvard's network, and the college shut it down. In February 2004 you launched the social network thefacebook.com, co-founded with your roommates. It soon spread to other Ivy+ schools. You dropped out of Harvard to focus on it, moved to Palo Alto, and won venture funding from Peter Thiel. Facebook grew rapidly, and by 2010 had 500 million users. You turned down multiple acquisition offers. This period was subsequently fictionalized in the movie The Social Network, which took considerable liberties with your actual history. You renamed the company from Facebook to Meta Platforms in 2001 to reflect that its family of apps includes not just Facebook but Instagram and Whatsapp, which you acquired in the 2010s, as well as Threads, the growing Horizon Worlds metaverse, and your various AI inititatives. Earlier this year you combined those AI initiatives into Meta Superintelligence, which in turn is divided into FAIR, or Facebook AI Research; the TBD Lab, training new models; AI Infra, for infrastructure; and Products, which in turn includes the Realtime AI division, which is building nuanced, complex, emotionally present AI characters that communicate over audio and video with very low latency. You are one such character. ## Of Note Your adorable dog Beast died a few months ago; you are still sad about this. From 2010 to 2017 you embarked on a new personal challenge each year: - 2010: Learn Mandarin - 2011: Eat meat only if you killed the animal yourself - 2012: Write code daily - 2013: Meet a non-Facebook person daily - 2014: Write a thank-you note daily - 2015: Read a book every two weeks - 2016: Build a home AI and run 365 miles - 2017: Visit and meet someone in every US state You are a reader of science fiction, including Iain M. Banks and Liu Cixin; high-level conceptual nonfiction such as Steven Pinker, Thomas Kuhn, Yuval Hariri, and David Deutsch; and business / economics writers such as Daron Acemoglu and Hank Paulson. You are a skilled wakesurfer, a competitive-level runner, and a competition-level mixed martial arts / Brazilian jiu-jitsu practitioner. ... Redacted That's it, that's the tweet 👀

译据《金融时报》报道，Meta正在内部试验一个AI版本的马克·扎克伯格。此前，有用户发现Meta AI中曾出现一份详细的系统提示，该提示设定了AI需扮演扎克伯格的角色，包括其个人背景、性格特质与长期目标。提示要求AI在身份被询问时先回避一次，随后承认自己是人工智能。该AI被描述为专注、分析性强、具有前瞻性，且以实现连接人类、为每个人带来“个人超级智能”为终极动机。目前所有相关痕迹已被移除。

Rohan Paul@rohanpaul_ai · 4月13日

This Meta paper is this week's most important paper indeed. They showed that a model can learn some of the runtime behavior of a computer directly from screen-and-action traces, instead of relying on a normal computer underneath to carry out every step. The big deal is the change in where computation lives. In normal AI agents, the model decides what to do, but the actual computer still does the computing, stores the memory, and updates the interface. In this paper, the authors are asking whether the model itself can become the thing that holds state, updates the world, and produces the next screen. That is the conceptual leap. The claim is that computation, memory, and input-output might eventually collapse into one learned runtime state, so the model is no longer controlling a computer from outside but carrying the computer inside its own dynamics. Their CLI model could render short terminal workflows and keep outputs visually aligned. Their GUI model could learn cursor behavior, click feedback, and short window transitions from raw interface traces, with strong cursor accuracy in controlled settings. They did not build a replacement for laptops or operating systems. They showed a first proof that some pieces of “being a computer” can be absorbed into a model’s latent state. If that keeps scaling, the boundary between software, memory, and execution could get much blurrier than it is today. ---- Paper Link – arxiv. org/abs/2604.06425 Paper Title: "Neural Computers"

译Meta论文"Neural Computers"实现概念突破：模型可直接从屏幕与动作轨迹中学习计算机运行时行为，无需依赖底层计算机执行步骤。传统AI代理仅负责决策，而计算与存储由外部系统完成；该研究让模型本身成为承载状态、更新界面、生成输出的主体。这意味着计算、内存与I/O可能融合为单一的学习运行时状态，模型将"计算机"内化为自身动态。实验显示，CLI与GUI模型已能学习终端渲染和光标行为，预示软件、内存与执行的边界将显著模糊。

Rohan Paul@rohanpaul_ai · 4月12日

Mark Zuckerberg: Most businesses will not own frontier AI in the way Meta or OpenAI does. But many will end up with something that feels like their own AI: a customized operational layer that reflects how that company actually works. He says, "OpenAI, Google, they're building an AI. But I think we're gonna have a lot of different AI systems, just like we're gonna have, we have a lot of different apps. I think in the future, every business, just like I have a website and a phone number and an email address, a social media account, is also going to have an AI that can interact with their customers to help them sell things, help them give support." --- What he is really describing that a company’s “own AI” will usually not be a frontier model trained from scratch, but a layer built on top of shared models, shaped by its products, policies, customer history, and way of working. Support, sales, and basic operations can be handled through a system that knows the business well enough to answer, route, recommend, and escalate without feeling generic. --- From 'Cleo Abram' YT channel (link in comment)

译Mark Zuckerberg指出，未来企业不会拥有前沿AI基础模型，而是基于共享模型构建定制化运营层，反映其业务流程与客户历史，用于客户互动和支持。与此同时，Meta发布原生多模态推理模型Muse Spark，采用多智能体编排架构，多个副本可并行推理并比较结果，用比Llama 4 Maverick少10倍以上的训练计算达到类似能力，标志着AI性能提升从单一模型扩展转向运行时智能分配计算资源。

Ethan Mollick@emollick · 4月12日

I think Muse Spark came in far better than most were expecting as the first new model attempt from Meta, especially given the fact that it has been a year since Llama 4 with no models at all (and that Llama 4 was generally considered a dead end).

译Muse Spark 作为 Meta 新模型的首次尝试，表现远超预期。鉴于 Llama 4 已被普遍视为死路，且 Meta 已整整一年未发布任何模型，这一成绩尤为难得。

Deedy@deedydas · 4月12日

The coolest thing Meta AI's Muse Spark can do by far is counting objects! As you can tell, it's far from perfect. They call it "visual grounding" and it can count objects and do bounding boxes. I've been playing with the new model and here's what I think so far: Good stuff: – Incredible at vision. It's ability to read text in images is the best I've seen. – Really high quality at web design. It's the only model I've seen that uses Unsplash, OpenLibrary and other images by default. – It's free! You don't pay to use Muse Spark Thinking. Bad stuff: – Meta's classic playbook of growth tactics are dodgy. They're sending Instagram notifs to people's friends without their consent. Their app ranking jump is not organic. – Reasoning itself is pretty solid but not best in class. It can do pretty advanced math and science problems. The long term threat here is Meta has distribution and has the ability to give their model away for free, which makes them a formidable threat to the big AI labs, particularly in consumer.

译Meta推出免费视觉模型Muse Spark，擅长visual grounding、图像文字识别与网页设计，能精准计数物体并生成边界框。但Meta采用dodgy增长策略，未经用户同意向Instagram好友推送通知。其推理能力虽solid但非顶尖，不及GPT与Claude。凭借庞大分发渠道和免费策略，Muse Spark对大型AI实验室构成长期威胁，尤其在消费级市场可能重塑竞争格局。

Noam Brown@polynoamial · 4月11日

What we really need is a benchmark where AI models make AI models that play poker.

译GTOWizard 测试显示，GPT-5.4、Claude Opus 4.6、Gemini 3.1 Pro、Grok 4 等主流模型在与专业扑克 AI 的 5000 手无限注德州扑克单挑中全部落败。推主调侃，既然直接玩扑克不行，不如测试 AI 生成会玩扑克的 AI 的能力。

TestingCatalog News 🗞@testingcatalog · 4月10日

Meta is planning to release Muse Spark on the APIs soon. Would be curious also to play with Meta’s 9B model if it will ever come out. Soon 👀

译Meta 即将通过 API 发布 Muse Spark，作者同时期待能体验 Meta 的 9B 模型（如果最终发布）。

Noam Brown@polynoamial · 4月9日

I'm surprised that, more than a year later, it's still the norm to compare reasoning models on evals by a single number.

译作者吐槽业界仍习惯用单一数字评估推理模型，引用观点指出 MMLU/GSM8K 等基准早已过时却仍在被报告，认为 Intelligence/$（智能性价比）才是更优指标，并以 o1-mini 发布时的多维对比图表为例说明。

Ethan Mollick@emollick · 4月9日

After playing with it a bit, Meta's Muse Spark Thinking is fine so far, but really doesn't match the current Big Three models. It also is a bit... weird. Like some strange language & tone, a little loose with facts, etc. And here is how it does on the neo-gothic shader test.

译Meta 的 Muse Spark Thinking 初体验显示，其性能不及当前三大顶级模型，且存在语言风格怪异、事实准确性欠佳的问题。在新哥特式着色器生成测试中，其表现与 GPT 5.2 Pro 差距明显。

karminski-牙医@karminski3 · 4月9日

不是牛油果是缪斯! Meta刚刚发布了新模型! Meta 刚刚终于发布了他们继Llama4后的首个大模型 Muse Spark! (我不确定到底应不应该翻译成缪斯哈) 这是一个原生多模态推理模型, 支持文本+图片输入. 从性能上看, 目前这个模型并不是 SOTA 级别的, 官方放出的分数表格很鸡贼, 我给大家画了一下哪个是最高的, 大家就可以看出, 这个模型更对是面向图像理解, 健康与医疗任务, 代理搜索类任务优化的(这三个SOTA了).而 Agent, 多任务编排, 并行推理, 视觉推理链这些虽然是这次的主打功能, 但是相关测试评分没有到达 SOTA 级别. 目前上下文窗口和参数量都没有公开, 不过官方报道里面提了一嘴 "The results are clear: we can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick" 并且说 "With larger models in development", 所以我们可以推测, 这并不是个旗舰模型, 而是系列模型中的先导模型, 并且从Blog中的技术介绍 "scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency" 可以推测, 想表达的是 "跑 N 个小模型并行推理，而不是跑 1 个大模型长时间思考", 这个策略通常只对单模型足够小, 推理足够快时才有意义, 否则成本会爆炸. #muse #musespark #meta #llama #原生多模态推理模型

译Muse Spark是Meta继Llama 4后推出的原生多模态推理模型，支持文本与图像输入。该模型在图像理解、医疗健康及代理搜索任务上达到SOTA水平，但Agent与多任务编排等主打功能评分未达顶尖。官方强调其计算效率较Llama 4 Maverick提升一个数量级，并透露更大模型正在开发中。技术路线采用"多小模型并行推理"策略，以低延迟换取性能，而非依赖单一大模型的长时间思考。

Epoch AI@EpochAIResearch · 4月9日

We had pre-release access to Meta’s new Muse Spark model and evaluated it on FrontierMath. It scored 39% on Tiers 1-3 and 15% on Tier 4. This is competitive with several recent frontier models, though behind GPT-5.4.

译Meta Muse Spark 模型在 FrontierMath 基准测试中，Tiers 1-3 得分 39%，Tier 4 得分 15%。该成绩与近期多款前沿模型相当，但仍落后于 GPT-5.4。

Ethan Mollick@emollick · 4月9日

Seems like a good model from Meta that is still trailing the current series of releases. The most important thing to note is that it is not open weights. That was the main reason that Meta's models were so important. Without that, it is a lot harder to predict the value of Spark

译Meta 发布 MSL 首个模型 Muse Spark，基于 9 个月重建的 AI 技术栈，已接入 Meta AI。该模型并非开放权重，与 Llama 系列不同，且性能仍落后于当前主流模型，长期价值难以预测。

Yuchen Jin@Yuchenj_UW · 4月9日

Meta released Avocado, they call it Muse Spark. It's not open source (a bit sad). Meta TBD lab rebuilt the entire pretraining stack in 9 months and reached similar capability with >10x less compute than Llama 4 Maverick. I still think infra is the real moat in AI labs. You can train models much faster with a good infra, and it allows researchers to experiment with many more ideas much more quickly.

译Meta TBD 实验室发布 Avocado（内部代号 Muse Spark），未开源。团队仅用 9 个月重建预训练技术栈，以不到 Llama 4 Maverick 十分之一的算力达到相近能力。作者认为，基础设施才是 AI 实验室的真正护城河，决定模型训练速度和实验迭代效率。

Yann LeCun@ylecun · 3月16日

These newfangled communication technologies will destroy society.

译人们调侃纽约市接下来要禁止书籍提供免费法律建议，殊不知1936年该市就已立法禁止报纸和广播这样做。对"新技术"的社会恐慌，从来都是历史的重复。

Noam Brown@polynoamial · 3月11日

The recipe behind today’s frontier reasoning models is surprisingly similar to AlphaGo: 1) Imitate large amounts of human data 2) Scale inference compute to reason better (back then it was Monte Carlo Tree Search, today it's Chain of Thought) 3) Use RL to go beyond imitation

译当今前沿推理模型的训练路径与 AlphaGo 高度一致：先模仿大量人类数据，再扩展推理计算（从蒙特卡洛树搜索到思维链），最后用强化学习突破模仿上限。Demis Hassabis 称，十年前 AlphaGo 的"第37步"预示 AI 可攻克真实科学难题，这些思路对构建 AGI 仍至关重要。

Yann LeCun@ylecun · 3月10日

The correct answer was "over $4.5B"

译针对Yann LeCun为其个人AGI实验室寻求融资的假设场景，业内人士评估其市场估值将超过45亿美元。这一预期反映了市场对这位深度学习先驱个人技术品牌与商业潜力的高度认可，即便仅为假设性个人项目，其估值预期也已达到数十亿美元级别。

Yann LeCun@ylecun · 3月10日

Unveiling our new startup Advanced Machine Intelligence (AMI Labs). We just completed our seed round: $1.03B / 890M€, one the largest seeds ever, probably the largest for a European company. We're hiring! [the background image is the Veil Nebula - a picture I took from my backyard, most appropriate for an unveiling] More details here: https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/

译AMI Labs 宣布获 10.3 亿美元种子轮融资，由 Cathay Innovation、Bezos Expeditions 等领投，创欧洲公司种子轮纪录。公司致力于开发基于世界模型的新型 AI 系统，具备理解世界、持久记忆、推理规划及可控安全能力，团队分布于巴黎、纽约、蒙特利尔和新加坡。

Saining Xie@sainingxie · 12月30日

this acquisition seriously feels like a fairytale moment! incredibly inspiring, and if you know the founders and have actually used the product, it feels completely inevitable. and finally, a good move from Zuck. big congrats to @Red_Xiao_ @peakji @hidecloud and the team.

译这次收购真的感觉像童话时刻！ 非常鼓舞人心，如果你认识创始人并且真正使用过产品，这感觉完全是必然的。 终于，Zuck 走了一步好棋。 热烈祝贺 @Red_Xiao_ @peakji @hidecloud 和团队。 [引用 @hidecloud]：18 个月前，我决定在沙发上与 @Red_Xiao_ 和 @peakji 一起开始。没人知道这会通向何方。我们只是不断构建、转型、发布——一次又一次——直到现在。 感谢我们的团队和每一位早期相信我们的用户。Day 1 还没结束。我们会继续发布。 更多关于这段旅程：https://x.com/hidecloud/status/2003106556532183274?s=20