“After careful evaluation, Apple determined that Google's Al technology provides the most capable foundation for Apple Foundation Models and is excited about the innovative new experiences it will unlock for Apple users.”
: )
iPad games for kids are now ad-ridden brainrot. But AI can now give us unlimited quality game assets and code.
Would happily pay a subscription to a universe of games that teach children math, logic, language in engaging ways, and crucially, AI and prompting.
The best way to equip future generations is to make them AI-literate from the outset, so they can understand what’s possible and how it shapes the world around them.
Made my own Sim City using Opus 4.5 in Cursor ✨
The coolest part? Adding my own isometric asset to the game so the characters could visit the @a16z SF office.
I used @ghosttyped's open source repo and then figured out a workflow for building creation 👇
now open source!!
go build your own city sim, tycoon game, 90s RTS, 4X game… whatever you want
let a thousand isometric strategy games bloom
小伙伴发现一个打开登录 Antigravity 账号的方式，确保你的地址是美国，年龄符合情况下，也很有可能登录不上，试了很多方法都不行，最后想着用 antigravity-tools 转发使用，没想到它有一个一键打开功能，登录后点击下面这个，一键换气居然可以直接可以使用了，很神奇，分享给需要的朋友。
brew install --cask --no-quarantine antigravity-tools
Your AI agent forgets everything after each conversation.
That's not a bug—it's a missing memory architecture.
Effective memory management moves the right information between short-term and long-term memory.
Short-term memory: The agent’s immediate workspace to fuel on-the-fly decisions and reasoning
Long-term memory: Stores information externally to build a persistent understanding of its world over time using episodic, semantic, and procedural memory.
Key principles for effective memory management:
• Prune and refine your memories: Make sure to remove duplicate entries, merge related information, or discard outdated facts.
• Be selective about what you store: Filter out low-quality information
• Tailor the architecture to the task: Carefully curate the types of memories needed for your use case
• Master the art of retrieval: Leverage different retrieval and reranking techniques to retrieve the right piece of information
Learn more about memory management in our free e-book on context engineering:
weaviate.io/ebooks/the-con…
In 2024, I was delighted to be a co-author on an Arxiv paper on"Shaping AI's Impact on Billions of Lives", covering areas we thought AI would have significant impact on the world. We cover AI's potential impact on employment, education, healthcare, (mis)information, media, national security, and science, and propose 18 potential moonshot research directions in these areas to maximize positive impact and minimize downsides. A shortened version appears as the cover article in this month's Communications of the ACM (CACM).
CACM version: cacm.acm.org/research/shapi… (~12 pages + video discussion with David Patterson)
Arxiv: arxiv.org/abs/2412.02730 (expanded 30 page version)
You can find these and more at: shapingai.com
It was a delight to work on this with awesome co-authors! Authors: Mariano-Florentino Cuéllar, myself (@JeffDean), @FinaleDoshi, John Hennessy, @andykonwinski, @sanmikoyejo, @tamati_biskit, @2plus2make5, and David Patterson.
(PDF link of CACM version is a bit hard to find. It's at dl.acm.org/doi/pdf/10.114…)
filesystems are all you need!
blog.langchain.com/how-agents-can…
reading this now, good collection of tips & tricks
Cursor's agent now uses dynamic context for all models.
It's more intelligent about how context is filled while maintaining the same quality. This reduces total tokens by 46.9% when using multiple MCP servers.
MIT researchers propose Recursive Language Models
You are going to hear more on this in 2026.
Why does it matter?
What if LLMs could process inputs 100x longer than their context window?
Context length is a hard constraint.
You can extend it with architectural changes, but there's always a limit. Most approaches try to squeeze more into the window or compress what doesn't fit.
This new research takes a different approach. Instead of fighting the context limit, work around it programmatically.
Recursive Language Models (RLMs) treat long prompts as an external environment. The model can examine the prompt, decompose it into sections, and recursively call itself on snippets. It's inference-time scaling applied to context length.
Rather than forcing everything through attention in one pass, let the model strategically partition and process information across multiple recursive calls.
The results are impressive. RLMs successfully handle inputs up to two orders of magnitude beyond model context windows. A model with an 8K context can effectively process 800K tokens.
But here's what's surprising: even for shorter prompts that fit within context, RLMs dramatically outperform base LLMs and common long-context scaffolds across four diverse tasks.
This suggests something interesting.
Maybe sequential attention over the entire input isn't always the best strategy, even when it's technically possible. Programmatic decomposition and recursive processing might be fundamentally better for certain tasks.
The approach achieves comparable or cheaper cost per query relative to alternative long-context methods.
Context length limitations constrain what agents can do. Processing entire codebases, long documents, or extended conversation histories requires workarounds. RLMs offer a general inference strategy that turns context limits into a soft constraint rather than a hard one.
Claude 的新功能 Cowork：让 AI 真正帮你干活
Claude Code 本来是给程序员写代码用的，结果大家发现它整理文件、做表格、写报告也很顺手。Anthropic 索性把这套能力包装成了 Cowork，让不会写代码的人也能用上。
【1】Cowork 到底能干啥
你选一个电脑上的文件夹，Claude 就能在里面读文件、改文件、创建新文件。
听起来简单，用起来挺香。比如你下载文件夹乱成一锅粥，让它帮你分类重命名。或者你有一堆消费截图，它能整理成一张 Excel 表。再比如你写了几页凌乱的笔记，它能帮你理顺思路、输出初稿。
和普通对话不一样的是，Cowork 模式下 Claude 更像个真正的助手。你布置任务，它自己规划步骤、一步步执行，中间会告诉你进度。如果你用过 Claude Code，这感觉会很熟悉，因为底层技术是同一套。
【2】还能更强
基础功能只是起点。Cowork 可以接上你已有的连接器，比如 Google Drive、Slack。它还内置了一批技能，能更好地生成文档、PPT 之类的文件。再配上 Chrome 浏览器插件，Claude 甚至能帮你操作网页。
这套设计让工作流变得很丝滑。你不用反复给 Claude 喂上下文，也不用手动把输出转成正确格式。甚至不用等它做完一件事再布置下一件，可以连续丢任务让它并行处理。用 Anthropic 的话说，这感觉不像你一句我一句地聊天，更像给同事留便签。
【3】和 Claude Code 共享技能生态
对 Claude Code 用户来说有个好消息：Cowork 能读取你本地的 CLAUDE.md 文件和自定义 Skills。
我测试了一下，选择工作文件夹后，Cowork 能看到里面的 CLAUDE.md 并按指令执行。我在 Claude Code 里配置的写作风格技能，Cowork 里也能直接调用。技能分两类：Anthropic 官方提供的（docx、pptx、pdf 这些）和用户自己创建的，两类都能用。
换句话说，你在 Claude Code 里攒下的工作流配置可以直接迁移过来。Cowork 不是另起炉灶，是同一套体系的图形化入口。
有个坑要注意：Cowork 跑在 Linux 虚拟机里，而你的 Mac 是 ARM 架构。如果技能依赖 node_modules 或本地特定环境（比如浏览器 cookies、特定架构的二进制文件），就跑不了。我试着调用一个需要运行 nodejs 脚本的图片生成技能，报错了——架构不兼容。纯文本类的配置（CLAUDE.md、写作规范）没问题，涉及本地脚本的技能可能需要额外适配。
【4】安全边界在哪里
Claude 只能访问你明确授权的文件夹和连接器，动作比较大的时候会先问你。但有几件事得提前知道：Claude 可能会误解你的指令，如果你说"清理一下这个文件夹"，它可能真的把文件删了。指令要说清楚。
另一个风险是提示词注入，就是攻击者在网页内容里藏一些指令，试图劫持 Claude 的行为。Anthropic 说他们做了防护，但这个领域整个行业都还在摸索。
这些风险不是 Cowork 特有的，只是很多人可能是第一次用这种更自主的 AI 工具。官方建议：刚开始用的时候谨慎点，别一上来就让它处理重要文件。
【5】现在能用吗
Cowork 目前是研究预览版，只对 Mac 上的 Claude Max 订阅用户开放。Anthropic 想先看看大家怎么用、有什么反馈，然后快速迭代。后面会加跨设备同步，也会出 Windows 版。
这一步到是意料之中，因为 Claude Code 现在已经被用在很多编程意外的领域，但是门槛略高，限制了使用群体是程序员或者懂点技术的用户，而且脚本执行权限会有很多安全上的隐患。Cowork 一下子降低了使用的门槛，通过图形化界面就可以操作，并且也让使用更安全。
现在还是早期版本，能做的事有限，安全机制也在完善中。但如果你是 Max + Mac 用户，值得一试。
Introducing Cowork: Claude Code for the rest of your work.
Cowork lets you complete non-technical tasks much like how developers use Claude Code.
Some businesses run like autocracies, and some run like democracies. In my experience, neither approach is particularly effective.
That’s because autocratic leaders think they have all the answers, but in reality don’t know enough. And democratic leaders who crowdsource decision-making will filter in a lot of bad ideas.
That dilemma led to me creating an idea meritocracy at Bridgewater, where we used analytics to better understand where people were weak or strong. That way, we knew whose opinion was most likely to be correct in any context.
