#Meta

I love Geoff. But he understands even less than Dario about the effects of technological revolutions on the labor market. Again, don't listen to AI scientists, as brilliant as they might be, and even less to AI CEOs, as successful as they might be, for questions of labor economics. Listen to reputable economists who have studied these things like @Ph_Aghion , @DAcemogluMIT , @erikbryn , @amcafee , @davidautor , etc.

Anthropic CEO Dario Amodei: “50% of all tech jobs, entry-level lawyers, consultants, and finance professionals will be completely wiped out within 1–5 years.” https://t.co/sXcT59gWYj

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/

Meta is back. 🔥 Finally dropped its first model since Zuckerberg started writing checks like crazy. Launched Muse Spark (originally codenamed Avocado). Its a natively multimodal reasoning model that can look, reason, use tools, and split hard work across multiple cooperating agents. Claims it can reach similar capability with 10x+ less training compute than Llama 4 Maverick, They are not positioning Muse Spark as a top-of-the-line model, but is instead highlighting its efficiency and “competitive performance” on various tasks. The old bottleneck in AI is that one model often has to read, plan, call tools, and solve everything in one stream, which wastes compute and slows hard tasks. The key idea here is multi-agent orchestration, where several copies of the model work on the same problem in parallel and then compare or merge results, which is closer to a small team than a single assistant. That changes the scaling story because better performance no longer comes only from making 1 model bigger, but also from spending compute more intelligently at run time. So Muse is a stack built around 3 scaling axes: stronger pretraining for basic world and code understanding, steadier RL for improving answers after pretraining, and test-time reasoning so the model spends extra compute only when a problem is hard. The most interesting part is multi-agent orchestration, where several copies of the model reason in parallel and compare work, which raised Humanity’s Last Exam to 58% and FrontierScience Research to 38% in its heavier Contemplating mode. Meta also says the new pretraining recipe reaches similar capability with over 10x less compute than Llama 4 Maverick, which matters because cheaper training usually means faster iteration and more room to scale.

We benchmarked every major AI model at poker. GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, Grok 4 and more. All played 5,000 hands of heads-up no-limit against our state-of-the-art poker agent. Every single one lost. Here's the full breakdown 🧵 https://t.co/Krbg3VHQjN

LLM evals are slow to adapt. MMLU/GSM8K continued to be reported long after they were obsolete. I think the next thing to go away will be comparing models on evals by a single number. Intelligence/$ is a much better metric. I loved this plot from o1-mini's launch for example: https://t.co/U4ry4S3chR

People joking that NYC are going to ban free legal advice from books next... Don't realize NYC already banned free legal advice from newspapers and radio in 1936! via our newest substack: https://newsletter.pessimistsarchive.org/p/nyc-once-banned-free-legal-advice https://t.co/WVh3jz3giy

十年前，AlphaGo 在首尔的那场传奇对局，宣告了现代 AI 时代的开启。它那著名的"第 37 手"向我们表明，AI 技术已经准备好应对科学等领域的现实问题——而受这些方法启发的思路，对于构建 AGI 至关重要 https://t.co/8EibfAByaG

hypothetically, if yann lecun fundraised for a new AGI lab for himself... how much would it be worth?

Advanced Machine Intelligence (AMI) 正在打造新一代 AI 系统，它们能够理解世界、拥有持久记忆、可以推理和规划，并且可控且安全。 我们从相信我们以世界模型为核心的通用智能系统愿景的全球投资者那里筹集了 10.3 亿美元（约 8.9 亿欧元）的一轮融资。本轮融资由 Cathay Innovation、Greycroft、Hiro Capital、HV Capital 和 Bezos Expeditions 共同领投，并有来自世界各地的其他投资者和天使投资人参与。 我们是一支不断壮大的研究者和构建者团队，从第一天起就在巴黎、纽约、蒙特利尔和新加坡运营。 阅读更多：https://amilabs.xyz/ AMI - Real world. Real intelligence.

18 months ago, I decided to join with @Red_Xiao_ and @peakji on my sofa. No one knew where it would lead. We just kept building, pivoting, and shipping—again and again—until now. Grateful to our team and every user who believed early. Day 1 isn’t over. We’ll keep shipping. More on the journey: https://x.com/hidecloud/status/2003106556532183274?s=20

Yann在这里完全说错了，他把通用智能和万能智能混为一谈了。 大脑是我们在宇宙中（迄今为止）所知的最精妙、最复杂的现象，而且它们事实上极其通用。 显然，人们无法绕过“没有免费午餐”定理，因此在一个实际且有限的系统中，总得围绕所学习的目标分布存在某种程度的专门化。 但关于通用性的要点在于，从理论上讲，在图灵机的意义上，这样一个通用系统的架构能够在给定足够时间、内存（和数据）的情况下学习任何可计算的东西，而人脑（以及AI基础模型）都是近似的图灵机。 最后，关于Yann对国际象棋棋手的评论，令人惊叹的是人类竟然能首先发明国际象棋（以及现代文明从科学到747飞机的所有其他方面！），更不用说像Magnus这样的人能把它下得如此出色。他可能并非严格意义上的最优（毕竟他的记忆有限，做决定的时间也有限），但考虑到他和我们的大脑是为狩猎采集而进化出来的，我们能用大脑做到的事情实在令人难以置信。

我不断看到关于 TPU 的东西，有什么实质性的新进展吗？ 没有证据表明 Google 曾在非 TPU 硬件上训练过 Gemini，这可以追溯到多年前 GPT 之前的模型，比如 BERT。TPU 比 Nvidia 自己的张量核心还要早。 Anthropic（以及 Character、SSI 和 Midjourney……）长期以来一直在使用 TPU，如果 Meta *没有* 在考虑它们，我会感到惊讶。 Nvidia 的护城河对大型实验室来说从来都不深——看看 OpenAI 决定它可以做得比 CUDA 更好，转而投资 Triton，在基准测试中经常胜过 CUDNN。这一切都没有什么神奇或结构性的东西，只是优秀的工程师在做优秀的工作。 TPU 并不比 GPU 高效多少，而微小的性能/功耗差异与 Meta 是否拥有正确的内核/系统工程人才来做到这一点相比，简直微不足道。Nvidia 和 Google 的护城河都很小，我们仍然处于个别优秀工程师就能扭转整个平衡的阶段。见下面的旧帖。 这难道……没有被定价进去吗？这些都是超级古老的公开信息。

Q. Who aligns the aligners? A. http://alignmentalignment.ai Today I’m humbled to announce an epoch-defining event: the launch of the 𝗖𝗲𝗻𝘁𝗲𝗿 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝗼𝗳 𝗔𝗜 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝗖𝗲𝗻𝘁𝗲𝗿𝘀.

One of the first pruning methods for neural nets came in 1989: Optimal Brain Damage by @ylecun et al. “We … derive a class of prac- tical and nearly optimal schemes for adapting the size of a neural network.” They pruned weights with low 2nd-derivative of the loss. https://t.co/WzIDQExX87

My latest (with @erikbryn) in @WSJ today: AI is already generating a lot of benefits ($97 billion in 2024 in the US alone, according to our calculations), but these benefits will not show up in GDP numbers for a while. https://t.co/lam7vR909E

This is insane. AI capex might account for a larger share of GDP than basically any technology since the railroad. Basically it’s a mini-wartime economy, but the guns are chips and the tanks are databases https://t.co/E11IxmYtOv

This false nomenclature of “researcher” and “engineer”, which is a thinly-masked way of describing a two-tier engineering system, is being deleted from @xAI today. There are only engineers. Researcher is a relic term from academia.

Ok this makes me super happy. The "NoFilter" work, paper, and advocacy that @angelinepouget and I argued so hard for is bearing fruit! Looks like it made MetaCLIP team massively change direction for their just released MetaCLIP2. This is now TheRightWay™, and thanks for the prominent ack in the paper🤗 Very eager to dig into that dataset!

We're excited to have @shengjia_zhao at the helm as Chief Scientist of Meta Superintelligence Labs. Big things are coming! 🚀 See Mark's post: https://www.threads.com/@zuck/post/DMiwjXJSYCd https://t.co/DRnpqBe1wD