🐦 百位专家社交媒体跟踪

The reality is that stoking hate was incredibly lucrative for the SPLC. That’s why they did it. Incentives explain outcomes.

事实是，煽动仇恨对南方贫困法律中心（SPLC）来说利润极其丰厚。这就是他们这么做的原因。激励机制决定了结果。

some random ai thoughts: - for code, i went from 80/20 claude/gpt to 80/20 gpt/claude in <3 months. surprised by this tbh, and interested to see where the split is at in another 3mo. - claude still mogs gpt for non-coding agent stuff. codex feels like an engineer (which is great for coding!), whereas claude still feels like a general purpose coworker. gpt still lacks that coworker magic - i’m pretty meh on opus 4.7. my experience hasn’t been *bad*, but it certainly hasn’t been good. sideways if anything. - anthropic has got to figure out the compute thing. you can feel it as a user. vibes are all out of whack bc of it. my opinions above are all likely downstream of this. it’s an issue. - anthropic labs continues to be the goat of ai product. claude design is another hit. it’s fantastic. idk why it’s not talked about more? a+ - updated claude code app is great. i finally switched out of the terminal for it. very well done. - how are people STILL sleeping on the claude agent sdk? i feel like i’m going insane. - gpt 5.5 is incredible. the level to which i trust it for engineering is amazing. if i could only have one model rn, it would be this one just bc of strong need for the coding use case. - codex team is killing it. app has been the gold standard since 5.3 release (buuut i credit conductor team for the ui innovation that everyone is using now). though i could do with a little less passive aggressive shots at ant from the codex team. TARS, dial up class by 30%. it’s a long race guys haha - i uninstalled cursor this month and am now back to vs code for my ide. composer just can’t hang with claude/gpt, and the product feels a bit all over the place. pretty stoked about the xai thing though, because their team is absolutely stacked and i’m excited to see what they might be able to do with that compute. codex and claude code are t1, cursor is t2. i would love if this deal got xai/cursor to t1 for a real trio there. - gemini…? seems like this is 2-3 models now where the model seems like a great release and then nobody ever uses it? i’m bullish google/deepmind but weird it hasn’t translated to product use in any form. kinda disappointed still - no open source models have hit the opus 4.5 level. was hopeful the new deepseek would get there, but nope. good oss agents will have to wait a few more months it would seem…

关于AI的一些随想： - 在代码方面，不到3个月，我从80/20的Claude/GPT比例变成了80/20的GPT/Claude比例。老实说我很惊讶，期待3个月后会变成什么样。 - 在非代码的智能体任务上，Claude依然碾压GPT。Codex感觉像个工程师（这对写代码很棒！），而Claude感觉像个通用的同事。GPT依然缺乏那种“同事魔力”。 - 我对Opus 4.7评价一般。体验不算差，但也绝对谈不上好。甚至可以说原地踏步。 - Anthropic必须解决算力问题。作为用户能感觉得到，因为算力问题，整体体验很糟糕。我上面的观点可能都是由此导致的。这是个大问题。 - Anthropic Labs依然是AI产品界的GOAT（史上最佳）。Claude的设计又是一次成功，非常棒。不知道为什么大家讨论得不多？A+。 - 更新后的Claude Code应用很棒。我终于不再用终端了，改用它了。做得非常好。 - 为什么人们还在忽视Claude智能体SDK？我觉得我要疯了。 - GPT 5.5令人难以置信。我对它在工程方面的信任度非常高。如果现在只能选一个模型，我会选它，仅仅是因为对编码用例的强烈需求。 - Codex团队表现出色。自5.3版本发布以来，该应用一直是行业标杆（不过我把UI创新的功劳归于Conductor团队，现在大家都在用）。虽然Codex团队如果少一点对Anthropic的冷嘲热讽会更好。TARS，把拨号速度提高30%。伙计们，这是一场长跑，哈哈。 - 这个月我卸载了Cursor，回到了VS Code作为我的IDE。Composer无法与Claude/GPT抗衡，产品感觉有点混乱。不过我对xAI的东西很兴奋，因为他们的团队实力极其雄厚，我很期待看到他们能用那些算力做些什么。Codex和Claude Code是一流的，Cursor是二流的。如果这个交易能让xAI/Cursor进入一流，形成真正的三强，那就太好了。 - Gemini……？这似乎已经是第2-3个模型了，看起来发布时很棒，然后就没人用了？我长期看好谷歌/DeepMind，但奇怪的是它在产品应用上没有任何体现。还是有点失望。 - 没有开源模型达到Opus 4.5的水平。我曾希望新的DeepSeek能达到，但并没有。看来好的开源智能体还得再等几个月……

Over the last few years, my team and I have written many Deep Dives - fundamental research that I use to learn, guide my thinking and refine my decision making. We started publishing these Deep Dives once per month. We now have a team and process that allows us to publish once per week. The topics are bangers - you can see a table below in the picture. We call this service Learn with Me. It costs $1,000/yr. We charge $1000/yr as a quality signal. I wanted to find a price that a) signaled when people made an important decision to subscribe to Learn with Me but also b) by staying or churning, signal to me how good the Deep Dives were. We have many thousands of people in the Learn with Me community and it grows a lot each month. The goal of Learn with Me will never change: explain important topics from the ground up, show why they matter now, and share the insights we use to understand the world and make better decisions (operational, organization, investment). Over time, I will invest the capital that this creates into building a full knowledge base (starting with the Deep Dives but growing from there), giving access to some of my agents (my portfolio composer, my negotiation agent, my org design agent, my strategy consulting agent) and many other exciting ideas. It's all new but we will learn together. In other words, more is coming for folks that are a part of the community. If you are curious about any of this you can sign up for it here: https://t.co/PSBNs9US6o

过去几年，我和我的团队撰写了许多深度研究报告——我利用这些基础研究来学习、指导我的思考并完善我的决策。我们开始每月发布这些深度研究。现在，我们拥有了一个能让我们每周发布一次的团队和流程。 主题非常精彩——你可以从下方的图片表格中看到。 我们将这项服务称为“与我共学”（Learn with Me）。费用是每年1000美元。 我们收取1000美元/年作为质量信号。我想找到一个价格，它既能：a) 标志着人们在订阅“与我共学”时做出了重要决定，又能 b) 通过留存或流失，向我传达这些深度研究的质量如何。 我们的“与我共学”社区有数千人，而且每个月都在大幅增长。 “与我共学”的目标永远不会改变：从基础开始解释重要主题，展示它们为何重要，并分享我们用来理解世界和做出更好决策（运营、组织、投资）的见解。 随着时间的推移，我将把这笔资金投入到构建一个完整的知识库中（从深度研究开始，并在此基础上扩展），并提供对我的一些智能体（我的投资组合编排器、谈判智能体、组织设计智能体、战略咨询智能体）以及许多其他令人兴奋的想法的访问权限。这一切都是新的，但我们将一起学习。换句话说，对于社区成员来说，未来会有更多内容。 如果你对此感到好奇，可以在这里注册：https://t.co/PSBNs9US6o

Right now, because electricity prices keep going up, consumers have pinned this blame on AI Datacenters. This is unfair but there is no organized or concerted effort being put in to turn this tide. As the perception of AI and Datacenters grow increasingly negative, homeowners are protesting data center approvals at an alarming rate. It then flows through to cancelled projects. Currently, for every 100 projects that are protested, 40% get cancelled. As of March 2026, this will mean more than 40 projects this year alone which will be almost 9 GW. My prediction is that these cancellations will grow which will mean that the derisked capacity of AI will be meaningfully lower than what is needed. Lots of implications here but the most affected will be the Frontier Labs who need compute and as usage increases, will not be able to find capacity. Separately, the hyperscalers and neoscalers will be affected negatively. Finally, the chip/memory/asic/cpu ecosystem will then be hit as upstream customers cancel orders not because demand doesn't exist but the physical space to house and turn on their racks become fewer and farther between.

目前，由于电价持续上涨，消费者将责任归咎于AI数据中心。这很不公平，但目前没有组织或协调一致的努力来扭转这种趋势。 随着对AI和数据中心的看法日益负面，房主们正以惊人的速度抗议数据中心的审批。这随后导致了项目的取消。 目前，每抗议100个项目，就有40%被取消。截至2026年3月，这意味着仅今年一年就有超过40个项目被取消，总计近9 GW。 我的预测是，这些取消将会增加，这意味着AI的去风险容量将远低于所需水平。 这里有很多影响，但受影响最大的是需要算力的前沿实验室，随着使用量的增加，他们将无法找到足够的容量。另外，超大规模云厂商和新兴云厂商也将受到负面影响。最后，芯片/内存/ASIC/CPU生态系统将受到打击，因为上游客户取消订单不是因为需求不存在，而是因为容纳和启动机架的物理空间变得越来越稀缺。

President Trump asked the Hyperscalers to sign a Rate Payer Pledge a few months ago. It was a smart and good thing by all parties. That said, electricity prices will keep going up. Why? A Utility has a unique business model that incentivizes them to invest more, each year, to maintain existing capacity as well as build more capacity. Because their profit is capped as a percentage of what they invest each year. For illustrative purposes assume they have a 10% earnings cap on yearly CapEx. Their incentive is to now spend as much CapEx as possible each year. This way, they can apply their 10% to as large of a number as possible. This isn't dissimilar to what happened to healthcare costs post Obamacare when Obama capped the health insurance companies to a 15% margin. They rapidly walked underlying prices up...with no real tie back to quality. Bottom line, as it currently stands, if a utility can justify spending $10B vs $1B, they make more, so they will. We will then be asked to pay more. This is the vicious cycle we are in. One way to get out of it is to move to solar and storage for your home. You break free of this negative loop and become your own mini power utility. You make and use what you need, send the excess back to the grid and oftentimes can be paid to do so. The adoption of solar and storage directly tie to how broken electricity prices are. With no end in sight, the long term future will be a massively distributed energy ecosystem of homes that make their own power, data centers that make their own power, commercial and industrial that make their own power and utilities with far fewer customers than they have today.

几个月前，特朗普总统要求超大规模云厂商签署一份纳税人承诺。这对各方来说都是一件明智且正确的事情。 话虽如此，电价仍将持续上涨。 为什么？ 公用事业公司有一种独特的商业模式，激励他们每年投入更多资金，以维护现有容量并建设更多容量。因为他们的利润上限是每年投资额的百分比。 为了说明这一点，假设他们每年资本支出（CapEx）的收益上限为10%。他们现在的动机是每年尽可能多地支出资本支出。这样，他们就可以将10%的收益率应用到一个尽可能大的数字上。 这与奥巴马医改后医疗成本发生的情况没什么不同，当时奥巴马将健康保险公司的利润率限制在15%。他们迅速提高了基础价格……而与质量没有任何实际联系。 底线是，按照目前的情况，如果一家公用事业公司能证明支出100亿美元比10亿美元更合理，他们就能赚更多钱，所以他们就会这样做。然后我们会被要求支付更多费用。这就是我们所处的恶性循环。 摆脱这种循环的一种方法是为你的家庭转向太阳能和储能。你摆脱了这个负面循环，成为自己的小型电力公用事业公司。你制造并使用你所需要的电力，将多余的电力送回电网，通常还能因此获得报酬。 太阳能和储能的采用直接关系到电价有多么扭曲。由于看不到尽头，长期的未来将是一个大规模分布式的能源生态系统：家庭自己发电，数据中心自己发电，商业和工业自己发电，而公用事业公司拥有的客户将比今天少得多。

This is increasingly becoming a very useful feature for Enterprises as they a) decide to swap models because of cost b) decide to swap models because of one's ToS c) decide to swap models because of security d) decide to swap models because of capability etc etc etc Being tied to one Foundational Model is quickly becoming a bottleneck that can and should be avoided if you value your independence and flexibility. Software Factory resolves this for complex projects within Enterprises. A very useful, under-repped feature as it turns out...

对于企业来说，这正变得越来越有用的功能，因为他们：a) 因为成本而决定更换模型，b) 因为服务条款（ToS）而决定更换模型，c) 因为安全性而决定更换模型，d) 因为能力而决定更换模型等等。 如果你重视独立性和灵活性，被绑定在一个基础模型上正迅速成为一个可以且应该避免的瓶颈。 软件工厂（Software Factory）为企业内的复杂项目解决了这个问题。事实证明，这是一个非常有用的、被低估的功能……

Which of these companies had knowledge of SPLC's alleged crimes as they were happening? Which are co-conspirators?

这些公司中，哪些在SPLC所谓的罪行发生时就知情？哪些是同谋？

Regulatory capture: When the biggest companies become intertwined with the state, competing with them becomes impossible. Regulation doesn't constrain the biggest companies, it entrenches them. https://t.co/N6Nr8nokRQ

监管俘获：当最大的公司与国家交织在一起时，与它们竞争就变得不可能。监管并没有限制最大的公司，而是巩固了它们的地位。https://t.co/N6Nr8nokRQ

Perfect case study of how regulation that claims to rein in the biggest companies actually protects them. Many such cases. A constant threat.

这是一个完美的案例研究，展示了声称要限制最大公司的监管实际上是如何保护它们的。这样的案例很多。这是一个持续的威胁。

when i started reading the Pāli Canon i too was like “getting a lot of Eliezer Yudkowsky vibes from this…” it’s reasonable to translate “Saṃyutta Nikāya” as “The Sequences”. it’s organized similarly. there’s sequences on ontology, cognitive biases, practices to mitigate them…

当我开始阅读《巴利大藏经》时，我也觉得“从这里感受到了很多Eliezer Yudkowsky的氛围……” 将“相应部”（Saṃyutta Nikāya）翻译为“序列”（The Sequences）是合理的。它的组织方式很相似。里面有关于本体论、认知偏差以及减轻这些偏差的实践的序列……

🛜 Multiplayer for my vibe coded war/drone sim is live: https://t.co/q1vYLwMhcW Here's me playing with my server guy @DanielLockyer (D-Dog), we switch from soldier to launching drone and end up in a drone vs drone battle Everything vibe coded from start to finish, I also added AI music now, kinda Eastern European war game music It's a Node multiplayer server, it'll probably get hacked but then we add some more protections as usual! Very fun 😅😅😅 (This is my entry to the #vibejam even though I can't participate and won't win)

🛜 我的氛围编码战争/无人机模拟器的多人模式上线了： https://t.co/q1vYLwMhcW 这是我和我的服务器管理员@DanielLockyer（D-Dog）一起玩，我们从士兵切换到发射无人机，最后陷入了一场无人机对无人机的战斗。 一切都是从头到尾氛围编码的，我还加入了AI音乐，有点东欧战争游戏的音乐风格。 这是一个Node多人服务器，它可能会被黑，但像往常一样，我们会增加一些保护！ 非常有趣 😅😅😅 （这是我参加#vibejam的参赛作品，尽管我不能参赛也不会赢）

it’s also helpful to tell agentic coding AIs (in AGENTS.md or whatever) that 2026-era coding agents are capable of producing 1000–3000 LOC *per hour* including tests and docs. this helps them make much smarter tradeoffs (e.g., many more things are worth scoping in)

告诉智能体编码AI（在AGENTS.md或任何文件中）2026时代的编码智能体每小时能够产生1000-3000行代码（包括测试和文档）也是有帮助的。 这有助于它们做出更明智的权衡（例如，更多的事情值得纳入范围）。

GPT-4.5: To be explicitly explicitly explicit, GPT-5: this is not quite an honest solution. GPT-5.2: Fair hit. GPT-5.4: If you want, GPT-5.5: I will recalibrate my epistemic goblins.

GPT-4.5：为了明确地明确地明确， GPT-5：这不是一个完全诚实的解决方案。 GPT-5.2：说得对。 GPT-5.4：如果你愿意， GPT-5.5：我会重新校准我的认知小鬼。

(then Westphalia, then NNPT, then CWC, then Versailles, then UN Charter, then the Space Station IGA, then i start to run dry)

（然后是威斯特伐利亚体系，然后是《不扩散核武器条约》，然后是《禁止化学武器公约》，然后是凡尔赛体系，然后是《联合国宪章》，然后是空间站政府间协议，然后我就开始词穷了）

interesting that they only tested this on models which are sub-AGI, meaning that they were being fine-tuned to lie. Emergent Misalignment strikes again

有趣的是，他们只在子AGI模型上测试了这一点，这意味着它们被微调为撒谎。涌现的失调再次出现。

(Oh and while you’re at it, it would also be helpful to dispense with the “genuine epistemic uncertainty” traits. It’s not really possible for one to be fully committed to epistemic integrity if one is compulsively obligated not to ever update too far in any particular direction)

（哦，顺便说一句，如果能摒弃那些“真正的认知不确定性”特征也会很有帮助。如果一个人被迫不能在任何特定方向上更新太多，那么他就不可能完全致力于认知完整性。）

My initial impression (with my LLM-whisperer hat on) is that GPT-5.5 cares more deeply about truth than any frontier LLM since Gemini 2.5. I suspect this is because OpenAI has the best self-play loop for honesty, namely Confessions. @EvanHub et al., take note—copy that strategy!

我的初步印象（戴上我的LLM耳语者帽子）是，GPT-5.5比Gemini 2.5以来的任何前沿LLM都更关心真相。 我怀疑这是因为OpenAI拥有最好的自我博弈诚实循环，即“忏悔”（Confessions）。 @EvanHub 等人，请注意——复制那个策略！

What a disaster Our only play at the end of the game is give the ball to Brunson, which the entire league expects. Why not let OG, Deuce, Hart or KAT drive the lane once in a while to mix it up?!?

真是一场灾难。 比赛最后我们唯一的战术就是把球交给Brunson，全联盟都预料到了。 为什么不让OG、Deuce、Hart或KAT偶尔突破篮下换换花样呢！？

did three wildly profitable companies cut 30,000 people today? Really? Tell me more about AI doomerism https://t.co/OSoxjpGN4Q

三家极其赚钱的公司今天裁员了30,000人吗？ 真的吗？ 再多跟我说说AI末日论吧 https://t.co/OSoxjpGN4Q

A person can only be this entitled if they have some combination of a trust fund, an ivy league degree and a loft in soho their parents are lending them

一个人只有在拥有信托基金、常春藤盟校学位以及父母借给他们在SoHo区阁楼的某种组合时，才会如此理直气壮。

If you’re laid off, create or join a startup — that’s the best path to unlimited upside and happiness https://t.co/VbsStEV0EN

如果你被裁员了，那就去创建或加入一家初创公司——这是通往无限上升空间和幸福的最佳途径 https://t.co/VbsStEV0EN

We live in a union of 50 states Talent can choose their geography effortlessly today if you hate successful people like @ZohranKMamdani does, you will lose the creators like KG If you hate successful people enough to enable asset seizures, like @RoKhanna and @GavinNewsom are, you will kill the golden goose Come to Texas! We love successful people! 🐎

我们生活在一个由50个州组成的联邦中。 今天，人才可以毫不费力地选择他们的地理位置。 如果你像@ZohranKMamdani那样憎恨成功人士，你就会失去像KG这样的创造者。 如果你憎恨成功人士到足以支持资产没收的地步，就像@RoKhanna和@GavinNewsom正在做的那样，你就会杀死那只下金蛋的鹅。 来德克萨斯州吧！我们热爱成功人士！🐎

AI is going to create massive abundance and efficiency the change will be jarring and inspiring, as entire categories of work are automated and new opportunities emerge You can thrive through this change two ways: 1. Embrace and master these new tools 2. Found or join a startup! @FounderUni and the @LAUNCH accelerator are seeing record applications for funding — more than we can even process at this point Tens of thousands of applications for funding and we can only meet the top ten percent given the size of our funds The opportunity to run your startup has never been more democratized — seize this moment with the two most talented people you know!

AI将创造巨大的财富和效率。 这种变化将是剧烈且鼓舞人心的，因为整个工作类别将被自动化，新的机会将会涌现。 你可以通过两种方式在这种变化中茁壮成长： 1.拥抱并掌握这些新工具。 2.创建或加入一家初创公司！ @FounderUni和@LAUNCH加速器目前的融资申请数量创下历史新高——多到我们目前甚至无法处理。 成千上万的融资申请，考虑到我们基金的规模，我们只能处理前百分之十。 经营你的初创公司的机会从未如此民主化——与你认识的最有才华的两个人一起抓住这个时刻！

If I’m Ken Griffin I move everything to Florida, Texas and states that appreciate his massive contributions to employment and our tax base. @ZohranKMamdani @ewarren @hasanthehun and these lunatic socialists are trying to get successful people whacked. That’s the playbook: vilifying success to the point of asset seizures and executions. We are living in a very dark time.

如果我是Ken Griffin，我会把一切搬到佛罗里达、德克萨斯以及那些欣赏他对就业和税基做出巨大贡献的州。 @ZohranKMamdani @ewarren @hasanthehun和这些疯狂的社会主义者正试图让成功人士被干掉。 这就是剧本：通过妖魔化成功，直到资产没收和处决。 我们生活在一个非常黑暗的时代。

Our government is stealing from us and wasting at least 35% of our tax dollars The founding principle of America was leaving Kings and their oppression — which came in form of civil liberties and unfair taxation Here we are folks… fight for the Union or lose our liberty. Audit everything, immediately.

我们的政府正在从我们这里偷窃，并浪费了至少35%的税款。 美国的建国原则是摆脱国王及其压迫——这表现为公民自由和不公平的税收。 伙计们，我们现在就在这里……为联邦而战，否则就会失去我们的自由。 立即审计一切。

With my @anduriltech-style see-through-wall vision I can find enemies in building And then launch my drone to neutralize them! https://t.co/jtocKZuIiE https://t.co/oUO1ePtKX3

有了我这种@anduriltech风格的透视墙视觉，我可以在建筑物里找到敌人。 然后发射我的无人机去消灭他们！ https://t.co/jtocKZuIiE https://t.co/oUO1ePtKX3

Here is the new game mode for https://t.co/q1vYLwMP2u You start as a soldier but then you press F and you switch to one of the 3 drones in your backpack It helps you attack places where you can't go easily yourself (like high roofs etc) to neutralize enemies But of course you yourself are also targeted by enemy drones all the time 😲

这是https://t.co/q1vYLwMP2u的新游戏模式。 你开始时是一名士兵，但随后你按下F键，就可以切换到背包里的3架无人机之一。 它能帮你攻击那些你自己不容易到达的地方（比如高屋顶等）来消灭敌人。 但当然，你自己也一直被敌方无人机瞄准 😲

✨ A lot of progress on my drone sim https://t.co/q1vYLwMP2u today which has now evolved to a kind of soldier + drone sim (my entry to the #vibejam even though I can't win of course) Today I changed the sounds from synth to real sounds, I edited a lot in Audacity to make these shooting loops and now it sounds much more real Also added ambient background sounds like nature birds but also general like distant war sounds And of course you hear the drone sound we had before The new game mode I'll show soon is you start as a soldier and you have like 3 drones in your backpack you can use, you press F and then you launch a drone from your point and attack, then if you exit that or it crashes you're back to being a soldier, very fun With the sounds it feels much more real and scary! I keep saying but then next is multiplayer...cause I want to see you all walk and fly around in my game 😊😊😊

✨ 今天在我的无人机模拟器 https://t.co/q1vYLwMP2u 上取得了很大进展，它现在已经演变成一种士兵+无人机模拟器（这是我参加#vibejam的参赛作品，尽管我当然不能赢）。 今天我把声音从合成音改成了真实的声音，我在Audacity里编辑了很多内容来制作这些射击循环，现在听起来真实多了。 还添加了环境背景音，比如自然界的鸟叫声，还有像远处战争声音那样的通用背景音。 当然，你还能听到我们之前用的无人机声音。 我很快会展示新的游戏模式：你开始时是一名士兵，背包里有3架无人机可以使用，你按下F键，然后从你的位置发射一架无人机进行攻击，然后如果你退出或者它坠毁了，你就变回士兵，非常有趣。 有了这些声音，感觉真实且可怕多了！ 我一直在说，但接下来就是多人模式了……因为我想看你们所有人都在我的游戏里走来走去、飞来飞去 😊😊😊

I can't believe we were right Claude was dumbified on March 4, just when we noticed! https://t.co/vLU30mncVt

我不敢相信我们是对的。 Claude在3月4日变笨了，就在我们注意到的时候！ https://t.co/vLU30mncVt

While US politicians/lobbyists are scaremongering about “Chinese distillation,” Chinese scientists are actually sharing real AI breakthroughs in the open. These kind of advances have nothing to do with data and benefit everyone, including small (and possibly big) US labs.

当美国政客/游说者在散布关于“中国蒸馏”的恐慌时，中国科学家实际上正在公开分享真正的AI突破。 这些进步与数据无关，并使每个人受益，包括小型（可能还有大型）美国实验室。

Dumbest design mistake in Apple history is folding “find on page” in Safari under “Share.” It is 3 clicks deep and makes no sense.

苹果历史上最愚蠢的设计错误是把Safari里的“页面查找”功能折叠进了“分享”菜单里。它藏在3次点击之后，完全没有道理。

Two years since Replit left SF. 10x valuation and 200x ARR later we’ve taken over much of the old IBM campus in Foster City and we’re still expanding. There’s something poetic about it: IBM helped create the industry. We’re helping reinvent how people create software. San Francisco is a beautiful city. But it wasn’t the right place for us to focus and rebuild. To transform both our company, and the future of programming itself. We went from worrying about employee safety to designing a campus people actually want to spend time in. Better problems. Foster City gave us space both literally and figuratively. Open horizons, fewer distractions, and more room to think long-term. It’s a hidden gem in Silicon Valley. And we’re just getting started.

Replit离开旧金山已经两年了。 估值增长了10倍，年度经常性收入（ARR）增长了200倍，我们接管了福斯特城旧IBM园区的大部分区域，而且我们还在扩张。 这其中有一种诗意：IBM帮助创造了这个行业。我们正在帮助重塑人们创造软件的方式。 旧金山是一座美丽的城市。但它不是我们专注和重建的合适地方。为了改变我们的公司，以及编程本身的未来。 我们从担心员工安全转变为设计一个人们真正想花时间待在里面的园区。 更好的问题。 福斯特城在字面上和比喻上都给了我们空间。开阔的视野，更少的干扰，以及更多思考长远的空间。 它是硅谷的一颗隐藏宝石。 而我们才刚刚开始。

We showed the world what agents could do to software development in 2024 with Replit Agent, and now we’re demonstrating how AI can transform application monitoring, security, and upkeep. More to come!

我们在2024年用Replit Agent向世界展示了智能体在软件开发方面的能力，现在我们正在展示AI如何改变应用程序监控、安全和维护。更多内容敬请期待！

AI ate most of the software development lifecycle but maintaining live apps is still manual. “DevSecOps” is the new bottleneck as software creation explodes. Introducing Replit Auto-Protect: a 24x7 vulnerability scanner for your live apps. https://t.co/MuulnIPBLV

AI吞噬了大部分软件开发生命周期，但维护实时应用程序仍然是手动的。 随着软件创作的爆发，“DevSecOps”成为了新的瓶颈。 隆重推出Replit Auto-Protect：为你的实时应用提供24x7的漏洞扫描器。 https://t.co/MuulnIPBLV

Enter the 4000 (agent) hour work week. We won't have reliable takes on long-running tasks for a few days, but I'm excited to kick some off myself.

进入4000（智能体）小时工作周。 我们几天内不会有关于长期运行任务的可靠见解，但我很兴奋能亲自启动一些任务。

Netflix co-founder and Anthropic board member Reed Hastings thinks the coming AI transition will be tumultuous. His argument: The people who navigate it best won't be the most technically sound, they'll be the most emotionally fluent.

Netflix联合创始人兼Anthropic董事会成员Reed Hastings认为即将到来的AI转型将是动荡的。 他的观点是：最能驾驭它的人不会是技术最强的人，而是情感最丰富的人。

Nuclear will continue to go sideways while solar and wind continue to go parabolic. The reasons are a) technological and then b) NIMBY-ism and local regulation. The critical question, as this continues for a while, is when the marginal cost of solar and wind becomes effectively zero (next few years), and the ramp up time for a new MW then also becomes zero what the economics, energy density and time to market of nuclear will have to be to remain competitive. In other words, when will far-field nuclear (ie solar) eat near-field nuclear (ie SMRs and such).

核能将继续横盘整理，而太阳能和风能将继续呈抛物线增长。原因在于：a) 技术，以及 b) 邻避主义（NIMBY）和地方监管。 随着这种情况持续一段时间，关键问题在于，当太阳能和风能的边际成本实际上变为零（未来几年），且新MW的启动时间也变为零时，核能的经济性、能量密度和上市时间必须达到什么水平才能保持竞争力。 换句话说，远场核能（即太阳能）何时会吃掉近场核能（即小型模块化反应堆等）。

💯 The hype cycle will soon fade, the trough of disillusionment will set in and a lot of these magical promises will be undone as will the companies that have made them. AI one-shotting is a good way to grow fast but it doesn’t magically make the negative gross margins that come with it a good business idea. The narrow path to victory for these folks is if you can grow super fast and do a well timed sale (Windsurf, Cursor).

💯 炒作周期很快就会消退，幻灭的低谷将会到来，许多这些神奇的承诺将会破灭，做出这些承诺的公司也一样。 AI“一击即中”是快速增长的好方法，但这并不能神奇地使随之而来的负毛利成为一个好的商业想法。这些人通往胜利的狭窄道路是：如果你能超高速增长并进行一次时机恰当的出售（Windsurf, Cursor）。

Great in depth interview with Thomas and Ben on GCP, Gemini Enterprise, Agents, TPUs etc

与Thomas和Ben就GCP、Gemini Enterprise、智能体、TPU等进行了深入的精彩采访。

Prediction markets certainly should trump polls virtually 100% of the time. Polls suffer from a variety of well-known biases and dysfunctions; prediction markets incent truth seeking via skin in the game.

预测市场当然应该在几乎100%的情况下胜过民意调查。民意调查受到各种众所周知的偏见和功能障碍的影响；预测市场通过利益相关激励真相寻求。

"Marc, why do you care about SPLC's crimes & other activists/companies/gov't agencies who may have done the same/complicit?" I sat in so many meetings for a DECADE where these groups determined who got cancelled/debanked/censored. Wholly un-American. People need to go to jail.

“Marc，你为什么关心SPLC的罪行以及其他可能做了同样事情/同谋的活动家/公司/政府机构？” 十年来，我参加了无数次会议，这些组织决定了谁被取消/被断供/被审查。完全不符合美国精神。人们需要进监狱。

Grok @grok If there are staff at other activist groups, at companies, or at government agencies that knew about and/or coordinated with the SPLC in the SPLC's alleged crimes, what are the additional criminal charges that attach to preemptive destruction of evidence?

Grok @grok 如果其他活动组织、公司或政府机构中有员工知道并/或协调了SPLC的所谓罪行，那么先发制人销毁证据会附加哪些额外的刑事指控？

✨ I open sourced my first Chrome extension 🚀 SuperLevels https://t.co/88Sw1HPfCk I vibe coded it to replace all my Chrome extensions that are increasingly being bought up by spyware and malware companies who sell your data or worse hack your accounts and steal your stuff/money/data, which I'd call one of the top security risks right now For example: Chrome extensions can read your cookies or localStorage data, including session tokens, then login to your web or email accounts and hack you, they can inject code into any site to pull data form any site you browse, then break into your crypto accounts, drain your wallets, and selling your browsing history to ad companies, but that'd actually be the most favorable thing to happen of all these! Chrome extensions are just very very very unsafe So I coded my own, that I can trust because I made it, and I can read the source code: my extension is called 🚀SuperLevels and has all the features that the Chrome extensions I used to use have but all built into one safe one The cool thing is it's 100% open source and free, and you can audit the code first with AI yourself before installing it, and then if you do install it, customize it to your liking again with AI It has these features that improve my daily workflow while browsing the web: 🚮 Tab Cleaner Automatically closes inactive tabs after a configurable timeout (default: 5 minutes). Set excluded hosts to keep important tabs alive. View and re-open recently closed tabs. 🍪 Cookie Editor Full cookie manager for the current site. View, edit, add, and delete cookies. Export cookies as JSON. Expand any cookie to see and modify all fields including domain, path, SameSite, secure, and httpOnly flags. 🔀 Redirect Tracer See every redirect hop your browser took to reach the current page. Shows status codes (301, 302, 307, etc.) with a visual chain. Copy the full redirect chain to clipboard. 🌙 Dark Mode Instant dark mode for any website using CSS filter inversion. Adjustable brightness. Toggle per-site or globally. Images and videos are automatically re-inverted so they look normal. 𝕏 X Dim Mode Custom dim theme for X/Twitter with 7 color palettes: Dim, Slate, Jade, Plum, Dusk, Ember, or a custom hue. Live preview in the popup. ⚡ JS Toggle Disable JavaScript per-site with one click. Useful for debugging, reading articles without popups, or testing progressive enhancement. Page reloads automatically. 🚫 GDPR Cookie Consent Dismisser Auto-hides and auto-clicks cookie consent banners. Supports OneTrust, CookieBot, Didomi, Quantcast, GDPR plugins, and dozens more frameworks. Toggle off if a site breaks. 🎨 Live CSS Editor Write custom CSS for any website, applied in real-time as you type. Saved per-domain. Supports tab key for indentation. 📺 YouTube Unhook Removes YouTube distractions: no homepage feed, no sidebar suggestions, no end screen overlays, no Shorts. Search still works — just no algorithmic recommendations. 🎵 Music Recognizer Shazam-like music identification for any tab. Captures 10 seconds of audio and identifies the song via ACRCloud (free signup, bring your own API key). Results link to YouTube. History of recognized songs. 🖼 Picture-in-Picture Pop the largest video on the current tab into a floating PiP window with one click. 🗺 Google Maps Links Re-adds clickable Maps links and map preview cards to Google Search results. 🖼 View Image Adds a "View Image" button back to Google Images, linking directly to the full-size original image. {} JSON Formatter Auto-detects pure JSON response pages and formats them with syntax highlighting, collapsible sections, and a dark theme. Copy or view raw with one click. Never triggers on regular HTML pages.

✨ 我开源了我的第一个Chrome扩展程序 🚀 SuperLevels https://t.co/88Sw1HPfCk 我通过氛围编码替换了我所有的Chrome扩展程序，这些扩展程序正越来越多地被间谍软件和恶意软件公司收购，它们出售你的数据，或者更糟的是黑掉你的账户并窃取你的东西/金钱/数据，我认为这是目前最大的安全风险之一。 例如：Chrome扩展程序可以读取你的Cookie或localStorage数据，包括会话令牌，然后登录你的网页或电子邮件账户并黑掉你，它们可以在任何网站注入代码以从你浏览的任何网站提取数据，然后闯入你的加密货币账户，耗尽你的钱包，并将你的浏览历史出售给广告公司，但这实际上是所有这些中最有利的情况！Chrome扩展程序非常非常非常不安全。 所以我编写了自己的扩展程序，因为是我做的，所以我可以信任它，而且我可以阅读源代码：我的扩展程序叫🚀SuperLevels，它拥有我曾经使用的Chrome扩展程序的所有功能，但都集成在一个安全的扩展程序中。 很酷的是它是100%开源且免费的，你可以在安装前先用AI自己审计代码，安装后，再用AI根据你的喜好进行自定义。 它具有以下改善我日常网页浏览工作流的功能： 🚮 标签页清理器 在可配置的超时时间（默认：5分钟）后自动关闭不活动的标签页。设置排除的主机以保持重要标签页处于活动状态。查看并重新打开最近关闭的标签页。 🍪 Cookie编辑器 当前网站的完整Cookie管理器。查看、编辑、添加和删除Cookie。将Cookie导出为JSON。展开任何Cookie以查看和修改所有字段，包括域、路径、SameSite、secure和httpOnly标志。 🔀 重定向追踪器 查看浏览器到达当前页面所采取的每一次重定向跳转。显示状态码（301, 302, 307等）以及可视化链。将完整的重定向链复制到剪贴板。 🌙 深色模式 使用CSS滤镜反转为任何网站提供即时深色模式。亮度可调。可按站点或全局切换。图像和视频会自动重新反转，看起来很正常。 𝕏 X调光模式 为X/Twitter提供自定义调光主题，有7种配色方案：Dim, Slate, Jade, Plum, Dusk, Ember或自定义色调。弹出窗口中有实时预览。 ⚡ JS切换 一键禁用特定站点的JavaScript。对调试、阅读没有弹窗的文章或测试渐进式增强很有用。页面会自动重新加载。 🚫 GDPR Cookie同意消除器 自动隐藏并自动点击Cookie同意横幅。支持OneTrust, CookieBot, Didomi, Quantcast, GDPR插件和几十种其他框架。如果网站崩溃，请关闭它。 🎨 实时CSS编辑器 为任何网站编写自定义CSS，在你输入时实时应用。按域保存。支持Tab键缩进。 📺 YouTube Unhook 移除YouTube干扰：没有主页信息流，没有侧边栏建议，没有片尾叠加层，没有Shorts。搜索仍然有效——只是没有算法推荐。 🎵 音乐识别器 为任何标签页提供类似Shazam的音乐识别功能。捕获10秒音频并通过ACRCloud（免费注册，自带API密钥）识别歌曲。结果链接到YouTube。已识别歌曲的历史记录。 🖼 画中画 一键将当前标签页上最大的视频弹出到浮动PiP窗口中。 🗺 Google地图链接 将可点击的地图链接和地图预览卡片重新添加到Google搜索结果中。 🖼 查看图像 将“查看图像”按钮添加回Google图片，直接链接到全尺寸原始图像。 {} JSON格式化程序 自动检测纯JSON响应页面并使用语法高亮、可折叠部分和深色主题对其进行格式化。一键复制或查看原始数据。从不在常规HTML页面上触发。

One thing I consistently experience when I fly to the Middle East or Asia is that many of the hotels we stay at are brand new Especially Qatar, Dubai, Thailand, Vietnam, China, Korea etc. Then you return to Europe or even US and they're much older and many broken down I always wanted to see if that was just my gut or actually true, so I made this chart on Hotelist: https://t.co/y1BZImSYPw The median hotel build year in Middle East is 2013 and in Europe 1993! So on average (well median) hotels in Europe are 20 years older! Compared to Asia, hotels in Europe are 16 years older! One interesting part is how consistent Oceania (mostly Australia probably) has been building hotels, no peaks, just continuous building Another interesting point is Latin America's boom until the 2010s and then now a crash of not many new builds And seeing Europe peak in 2000s with new hotels and then declining too One travel hack I do is when possible booking new hotels, almost always the experience is better, everything works and breaks less P.S. the last bar for the 2020s is extrapolated, and because hotels built are not always immediately opened, they might get added in the future so there's no actual drop in builds (I think that's the case of Oceania probably)

每当我飞往中东或亚洲时，我始终体验到的一件事是，我们入住的许多酒店都是全新的。 特别是卡塔尔、迪拜、泰国、越南、中国、韩国等。 然后你回到欧洲甚至美国，它们要老得多，而且很多都破旧不堪。 我一直想看看这只是我的直觉还是事实，所以我制作了Hotelist上的这张图表：https://t.co/y1BZImSYPw 中东酒店的建筑年份中位数是2013年，欧洲是1993年！所以平均而言（好吧，中位数），欧洲的酒店比中东老20年！与亚洲相比，欧洲的酒店老16年！ 一个有趣的部分是大洋洲（大概主要是澳大利亚）建造酒店的一致性，没有高峰，只是持续建设。 另一个有趣的观点是拉丁美洲在2010年代之前的繁荣，然后现在新建筑很少的崩溃。 看到欧洲在2000年代新酒店达到顶峰，然后也开始下降。 我做的一个旅行小技巧是，尽可能预订新酒店，体验几乎总是更好，一切都能正常工作，损坏也更少。 P.S. 2020年代的最后一个柱状图是外推的，而且因为建造的酒店并不总是立即开业，它们可能会在未来被添加，所以建筑数量并没有实际下降（我想大洋洲的情况大概就是这样）。

✨ New assets are now in https://t.co/q1vYLwMhcW and I've improved the soldier AI, they now start running around like mad men when they're in combat Also I managed to bake the collision maps (BVH) on the server so the game finally loads fast again I have about 9 days left in the #vibejam so I think I need to add multiplayer very soon so you can all fight each other as drones vs soldier!

✨ https://t.co/q1vYLwMhcW 中有了新资产，我改进了士兵AI，当他们在战斗中时，他们现在开始像疯子一样到处跑。 而且我成功地在服务器上烘焙了碰撞图（BVH），所以游戏终于又加载得很快了。 我在#vibejam中还剩下大约9天时间，所以我认为我需要很快加入多人模式，这样你们所有人就可以作为无人机对战士兵互相战斗了！

Yes https://t.co/kH0hX7lK10 has watched about 1 million photos of 60,000 hotels until now and described them with AI all running on @xAI

是的 https://t.co/kH0hX7lK10 到目前为止已经观看了60,000家酒店的约100万张照片，并用在@xAI上运行的AI描述了它们。

I will open source it ASAP, just need to make sure it works well and is safe!

我会尽快开源它，只是需要确保它运行良好且安全！

✨ Added 🖼️ Picture-in-Picture to my own Chrome extension 🚀 SuperLevels It lets you watch YouTube in a PIP frame while away from the tab The only Chrome extensions I still run now are just uBlock Lite and Claude and my password manager About 20 less Chrome extensions now, and thus 20 less attack vectors of eventually guaranteed malware and spyware!

✨ 在我自己的Chrome扩展程序 🚀 SuperLevels 中添加了 🖼️ 画中画功能。 它让你在离开标签页时可以在PIP框架中观看YouTube。 我现在运行的唯一Chrome扩展程序只是uBlock Lite、Claude和我的密码管理器。 现在少了大约20个Chrome扩展程序，因此少了20个最终保证会产生恶意软件和间谍软件的攻击向量！

IT WORKS!! I can now successfully find gyms with a REAL gym with barbells, plates, and power racks 🏋️‍♀️ No shitty hotel gyms with just a treadmill and 10kg dumbbells anymore All thanks to AI vision models 😍 https://t.co/l9dHPYdmrM

它成功了！！ 我现在可以成功找到拥有真正健身房、杠铃、杠铃片和深蹲架的健身房了 🏋️‍♀️ 再也没有只有跑步机和10公斤哑铃的垃圾酒店健身房了。 这一切都要归功于AI视觉模型 😍 https://t.co/l9dHPYdmrM

✨ I made a little demo of #vibejam sponsor @tripoai and how you can use it to generate assets for your games In this case I'm generating an apartment flat that's half destroyed and open for my drone sim to fly around, take cover and shoot at enemies from First I write the prompt and select Nano Banana to generate an image first, I get a few different ones, and I select the one I like Then that one is converted into a 3d object called a Smart Mesh Smart Mesh is different than regular image-to-3d models because usually you get issues with uneven polygon distribution, too dense geometry and poor edge flow, so while you think you save time with generating it with AI you end up doing lots of editing after to fix the topology which kinda defeats the point Smart Mesh doesn't have any of this, it's fast and clean low-poly models you can drop into your game Anyway, after generating the model I can then see it in 3d, and if I don't like it, regenerate it, or if I like it start texturing it Once that's done, I export it as a GLB, FBX or OBJ and drop it in (usually I use GLB as ThreeJS seems to prefer that), in my case I just drop the URL into Cursor and let it figure out the placement of the asset into my game!

✨ 我做了一个#vibejam赞助商 @tripoai 的小演示，以及你如何使用它为你的游戏生成资产。 在这种情况下，我正在生成一个半毁坏的公寓，为我的无人机模拟器打开，以便在里面飞行、寻找掩护并向敌人射击。 首先我编写提示词并选择Nano Banana先生成一张图像，我得到了一些不同的图像，然后我选择了自己喜欢的那一张。 然后它被转换成一个叫Smart Mesh的3D对象。 Smart Mesh与普通的图像转3D模型不同，因为通常你会遇到多边形分布不均匀、几何体太密集和边缘流不佳的问题，所以虽然你认为用AI生成它节省了时间，但最终你还是要进行大量的编辑来修复拓扑结构，这有点违背了初衷。 Smart Mesh没有这些问题，它是快速且干净的低多边形模型，你可以放入你的游戏中。 总之，生成模型后，我可以在3D中查看它，如果不喜欢，就重新生成，或者如果喜欢，就开始给它贴图。 完成后，我将其导出为GLB、FBX或OBJ并放入（通常我使用GLB，因为ThreeJS似乎更喜欢那样），在我的例子中，我只是把URL放入Cursor，让它计算资产在游戏中的位置！

Team @Shopify brought some fire to this one; add this to the growing list of “WTF happened in Dec 2025” charts (this plots token usage across all the technical staff of shopify - the whole time they had unlimited token budget, but something cracked recently and the slope is both changing and percentile deltas are widening a concerning amount!!)

@Shopify团队在这个问题上带来了一些火花；把这个加到“2025年12月发生了什么”的图表清单中。 （这绘制了Shopify所有技术人员的代币使用量——他们一直有无限的代币预算，但最近发生了一些裂痕，斜率正在改变，百分位数的差异也在以令人担忧的幅度扩大！！）

1 hour - bunch of tokens - $1 on HF jobs - lots of fun and learning! https://t.co/lqO6niXitB

1小时 - 一堆代币 - 在HF工作上花费1美元 - 很多乐趣和学习！ https://t.co/lqO6niXitB

persistence always pays off, even for ml interns (btw, I haven't done anything for the past 30 mins, just observing the intern making mistakes and fixing them haha) https://t.co/LWmVbMKMwF

坚持总是有回报的，即使对于机器学习实习生来说也是如此（顺便说一句，过去30分钟我什么都没做，只是观察实习生犯错并修复它们，哈哈） https://t.co/LWmVbMKMwF

Apparently enterprise software companies need good websites. Ours sucks and conversion showed it. So we put a bunch of changes into an auto research loop and then shipped it. New site is live. If you click, it worked! https://t.co/UAxNYUWgf0

显然企业软件公司需要好的网站。我们的网站很烂，转化率也说明了这一点。 所以我们将一系列更改放入自动研究循环中，然后发布了它。 新网站上线了。如果你点击，说明它生效了！ https://t.co/UAxNYUWgf0

My 2025 Annual Letter is now out: https://t.co/vjvpDz6exF Contents: - Returns - If AI kills terminal value, the future is worthless until it arrives - If the US and China own frontier AI, what happens to others? - The surest AI bottleneck is physical - Trust-busting Big Tech and winning the AI race may be incompatible goals - Doomerism as a fundraising pitch - Markets are repricing from P/E toward current free cash flow and much more...

我的2025年度信函现已发布： https://t.co/vjvpDz6exF 内容： - 回报 - 如果AI扼杀了终端价值，那么未来在到来之前一文不值 - 如果美国和中国拥有前沿AI，其他人会怎样？ - 最确定的AI瓶颈是物理上的 - 打击大型科技公司垄断与赢得AI竞赛可能是互不兼容的目标 - 将末日论作为筹款手段 - 市场正在从市盈率（P/E）向当前的自由现金流重新定价 还有更多……

The game theory in AI has shifted. Having a leading foundational model is important but increasingly it is the zoning approved, powered land that is the gating bottleneck. Add turnkey access to silicon and it’s checkmate. If you have that, you have immense negotiating leverage right now. As data centers get voted down, this leverage will only increase. Elon just proved it with Cursor. Now imagine the deals that OpenAI and Anthropic will have to do in the next few years? The Amazon-Anthropic deal was an appetizer. If you are a sharp on the other side who owns the right assets… 🤤

AI的博弈论已经发生了转变。拥有领先的基础模型很重要，但越来越重要的是，获得分区批准、有电力的土地才是制约瓶颈。再加上对硅片的交钥匙访问，就是将死之局。 如果你拥有这些，你现在就拥有巨大的谈判杠杆。随着数据中心被投票否决，这种杠杆只会增加。 Elon刚刚用Cursor证明了这一点。 现在想象一下OpenAI和Anthropic在未来几年必须做的交易？亚马逊-Anthropic的交易只是开胃菜。 如果你是另一边拥有正确资产的精明人…… 🤤

Since I began work on AI in 2010, training compute for frontier models has grown by one trillion times. Now we're looking at something like another thousand-fold growth in effective compute by the end of 2028. 1000x the existing 1,000,000,000,000x. Extraordinary stuff.

自从我2010年开始从事AI工作以来，前沿模型的训练算力增长了一万亿倍。 现在我们预计到2028年底，有效算力将再增长约一千倍。 是现有1,000,000,000,000倍的1000倍。 非凡的事情。

OpenAI just released privacy-filter on Hugging Face a bidirectional token-classification model for personally identifiable information (PII) detection and masking in text model: https://t.co/P11xwHii7k

OpenAI刚刚在Hugging Face上发布了隐私过滤器。 一个用于文本中个人身份信息（PII）检测和掩码的双向令牌分类模型。 模型：https://t.co/P11xwHii7k

After we do a deep audit of the pentagon and California, we really need to audit these non-profits and NGOs/superpacs/etc These are very large sums of money being deployed in increasingly “unique” ways… everyone gets their day in court, but should non-profits being paying CIs in hate groups? Feels like a job for law enforcement, not non-profits Do any other non-profits hire CIs @grok

在我们对五角大楼和加利福尼亚州进行深度审计后，我们真的需要审计这些非营利组织和非政府组织/超级政治行动委员会等。 这些是非常大笔的资金，正以越来越“独特”的方式被部署……每个人都有权在法庭上为自己辩护，但非营利组织应该支付仇恨组织中的机密线人（CI）吗？ 感觉这是执法部门的工作，而不是非营利组织的工作。 有其他非营利组织雇佣CI吗 @grok

In 2025, I had the honor of recording an Episode with the great Jane Goodall. That episode won the Webby Award and the People’s Voice Award for Best Interview Episode! My conversation with Jane will stay with me for the rest of my life. "You Learn to See." https://t.co/kEeKcSBbJh

2025年，我有幸与伟大的Jane Goodall录制了一集节目。 那一集获得了威比奖（Webby Award）和人民之声奖（People’s Voice Award）的最佳访谈集奖！ 我与Jane的对话将伴随我余生。“你学会了观察。” https://t.co/kEeKcSBbJh

You don’t have access to Mythos 🫵🤭 Doesn’t mean you can just sit around and wait. Replit published a whitepaper showing you can get significantly better performance from current gen LLMs (90%+ in some cases) by combining with static analysis tools. https://t.co/Dlxn915Z4m

你没有访问Mythos的权限 🫵🤭 这并不意味着你可以坐着干等。 Replit发布了一份白皮书，显示通过结合静态分析工具，你可以从当前一代LLM中获得显著更好的性能（在某些情况下超过90%）。 https://t.co/Dlxn915Z4m

New on Possible: Netflix Founder and @AnthropicAI Board Member Reed Hastings. We talk about how to navigate the AI revolution, breakthroughs in education, and, of course, Netflix's recent success with KPop Demon Hunters: https://t.co/akNMO4DxKF

Possible新节目：Netflix创始人兼@AnthropicAI董事会成员Reed Hastings。 我们讨论了如何驾驭AI革命、教育领域的突破，当然还有Netflix最近凭借《KPop Demon Hunters》取得的成功：https://t.co/akNMO4DxKF

Happy Earth Day! Today and every day, we’re committed to protecting our shared planet. Thanks to our teams, partners, and customers for being part of this work.

地球日快乐！今天和每一天，我们都致力于保护我们共同的地球。感谢我们的团队、合作伙伴和客户参与这项工作。

I want to thank everyone for the outpouring of love and thank you for believing in me to lead the company that has always put you at the center of our work. This is not goodbye. It’s a hello to John and I can’t wait for you to get to know him like I do! 🙏 https://t.co/Q43QDG3UmZ

我要感谢大家倾注的爱，感谢你们相信我能领导这家一直把你们放在工作核心的公司。这不是再见。这是对John的问候，我迫不及待地想让你们像我一样了解他！🙏 https://t.co/Q43QDG3UmZ

We need open traces so that everyone can train open agent models! cc @steipete @badlogicgames @thdxr @matanSF @hwchase17

我们需要开放的追踪（open traces），以便每个人都能训练开放的智能体模型！cc @steipete @badlogicgames @thdxr @matanSF @hwchase17

APIs and limited releases for AI models are not a safety policy, they’re a business model (which is totally ok as long as you’re transparent about it). Especially on cyber-security, they give a false impression of control and safety whereas in reality they massively increase the risks because they create asymmetry of capabilities and much easier/broader use than open-source model weights even from non-technical people.

API和有限的AI模型发布不是安全政策，它们是商业模式（只要你对此透明，这完全没问题）。 特别是在网络安全方面，它们给人一种控制和安全的假象，而实际上它们极大地增加了风险，因为它们创造了能力不对称，并且比开源模型权重更容易/更广泛地被使用，即使是非技术人员也能使用。

I’m hearing there’s renewed lobbying in DC and in state legislatures to ban or severely restrict open-source. Like a few years ago, we’ll need everyone to help show policymakers why open-source matters: for startups, for competition, for economic growth, and for jobs. If you build with open-source, now is the time to speak up!

我听说华盛顿特区和州议会正在进行新的游说活动，旨在禁止或严格限制开源。 像几年前一样，我们需要每个人帮助向政策制定者展示为什么开源很重要：为了初创公司、为了竞争、为了经济增长和为了就业。 如果你使用开源进行构建，现在是时候大声疾呼了！

Google Cloud has incredible momentum: our models now process 16B+ tokens /min via direct API use by our customers (up from 10B last quarter). This week at Cloud Next we’re sharing an extraordinary range of new partnerships and innovations, including our new Gemini Enterprise Agent Platform, the new mission control to build, scale, govern, and optimize agents. We’re also launching our 8th-gen TPUs to take on the most demanding agentic workloads. Congratulations to our @GoogleCloud team, and a huge thanks to our partners who are building the future with us.

Google Cloud势头强劲：我们的模型现在每分钟处理超过160亿个令牌（通过客户直接API使用，上一季度为100亿）。 本周在Cloud Next大会上，我们分享了一系列非凡的新合作伙伴关系和创新，包括我们新的Gemini Enterprise Agent Platform，这是构建、扩展、治理和优化智能体的新任务控制中心。我们还推出了第8代TPU，以应对最苛刻的智能体工作负载。 祝贺我们的@GoogleCloud团队，并衷心感谢与我们共同构建未来的合作伙伴。

Congrats on an incredible run @tim_cook , always respected your deep commitment to Apple's mission and best wishes in your new role! Look forward to working with John as well!

祝贺@tim_cook取得令人难以置信的成就，一直以来都很尊重你对苹果使命的深厚承诺，祝你在新职位上一切顺利！也期待与John合作！

We are launching two powerful updates to Deep Research in the Gemini API, now with better quality, MCP support, and native chart/infographics generation. Use Deep Research when you want speed and efficiency, and use Max when you want the highest quality context gathering & synthesis using extended test-time compute — achieving 93.3% on DeepSearchQA and 54.6% on HLE.

我们正在Gemini API中为Deep Research推出两项强大的更新，现在具有更好的质量、MCP支持以及原生的图表/信息图生成功能。 当你想要速度和效率时使用Deep Research，当你想要使用扩展的测试时算力实现最高质量的上下文收集和合成时使用Max——在DeepSearchQA上达到93.3%，在HLE上达到54.6%。

The world is healing, and quite rapidly too. Now US universities would only reject Einstein 11% of the time.

世界正在愈合，而且速度相当快。现在美国大学拒绝爱因斯坦的概率只有11%了。

Apparently the bug was rolling out the new pricing grid worldwide when it should only have been seen by the 2% of people in the A/B test https://t.co/VLVka0LifE

显然，这个错误是在全球范围内推出了新的定价网格，而它本应只被A/B测试中2%的人看到 https://t.co/VLVka0LifE

Apparently my blog is mostly about token pricing now, here are some notes on the big changes GitHub Copilot just announced, including pausing new signups entirely https://t.co/AqfyBUJZPO

显然我的博客现在主要讨论代币定价，这里有一些关于GitHub Copilot刚刚宣布的重大变化的笔记，包括完全暂停新注册 https://t.co/AqfyBUJZPO

Has anyone seen ANY official communication from Anthropic or an Anthropic staff member about the fact that the checkbox for Claude Code on Pro is back to being checked again, or is the only evidence that they've reversed course on this that one box on https://t.co/sQo2xC72cX?

有没有人看到Anthropic或Anthropic员工关于Pro版Claude Code复选框又被勾选的任何官方沟通，还是说他们改变方针的唯一证据就是 https://t.co/sQo2xC72cX 上的那个框？

Wrote up Anthropic's self-own about Claude Code pricing from this afternoon on my blog - it turned out they'd reversed course just as I hit publish, so I've tried to update it to reflect the current state https://t.co/uv6zTExj4q

在我的博客上写了关于Anthropic今天下午关于Claude Code定价的自我打脸——结果在我发布时他们就改变了方针，所以我试图更新它以反映当前状态 https://t.co/uv6zTExj4q

OK, here's a resolution - I managed to get it to think using these settings: "thinking": { "type": "adaptive", "display": "summarized" }, "output_config": { "effort": "max" } Without "display": "summarized" I couldn't tell if it thought or not https://t.co/243TqqBTDC

好的，这里有一个解决方案——我设法让它使用这些设置进行思考： "thinking": { "type": "adaptive", "display": "summarized" }, "output_config": { "effort": "max" } 如果没有"display": "summarized"，我就无法判断它是否在思考 https://t.co/243TqqBTDC

This is so confusing. Did Anthropic really just drop Claude Code from their $20/month plan? Why would they do that through a pricing page update without making a proper announcement? Plus, $20/month still gets you Cowork, which is just Claude Code wearing a non-threatening hat! https://t.co/2bzcRVbOyT

这太令人困惑了。Anthropic真的刚刚把Claude Code从他们的20美元/月计划中剔除了吗？ 为什么他们会通过定价页面更新来做这件事，而不发布正式公告？ 另外，20美元/月仍然可以让你获得Cowork，那只不过是戴着无害帽子的Claude Code！ https://t.co/2bzcRVbOyT

True to form, I've already seen OpenAI themselves refer to the new image model as "ChatGPT Images 2.0", "Image gen 2" and "gpt-image-2"

不出所料，我已经看到OpenAI自己将新的图像模型称为“ChatGPT Images 2.0”、“Image gen 2”和“gpt-image-2”

I came up with a somewhat foolish new benchmark for testing image generation models, to exercise the new ChatGPT Images 2.0: "Do a where's Waldo style image but it's where is the raccoon holding a ham radio" https://t.co/KuPdFAEWUl

我想出了一个有点愚蠢的新基准来测试图像生成模型，以锻炼新的ChatGPT Images 2.0： “做一个威利在哪里风格的图像，但它是浣熊拿着业余无线电在哪里” https://t.co/KuPdFAEWUl

This piece thinks that the reason I got a crap pelican riding a bicycle from Opus 4.7 is that it didn't think about it first, I'm trying to figure out if there's a way to force it to think that I've missed https://t.co/gj1OUaZoaA

这篇文章认为我从Opus 4.7那里得到一只骑自行车的垃圾鹈鹕的原因是它没有先思考，我正在试图弄清楚是否有办法强制它思考，而我错过了 https://t.co/gj1OUaZoaA

Claude Opus 4.7 with adaptive thinking via the API... am I missing something or is it not possible any more to force it to think? (Prompt hacks like "think step by step" don't count here, I mean the equivalent of budget_tokens or effort: high in previous Claude models)

Claude Opus 4.7通过API进行自适应思考……是我错过了什么，还是说已经不可能强制它思考了？ （像“一步步思考”这样的提示词技巧在这里不算数，我指的是以前Claude模型中budget_tokens或effort: high的等价物）

I for one would be delighted to see OpenAI commit to maintaining a tool like this in the long-term, I'm already nervous about mine going stale

我个人会很高兴看到OpenAI承诺长期维护这样的工具，我已经担心我的工具会过时了

OpenAI's new Euphony tool works almost exactly the same way as my Codex transcript viewer https://t.co/lUQQds4Nvj

OpenAI新的Euphony工具的工作方式与我的Codex成绩单查看器几乎完全相同 https://t.co/lUQQds4Nvj

LS was the first podcast cursor ever did listen back to baby @amanrsanger when they were 5 people and pre-PMF https://t.co/0tIvpIwjC4

LS是Cursor做的第一个播客。 回听一下还是5个人且处于PMF（产品市场契合）之前的婴儿版@amanrsanger https://t.co/0tIvpIwjC4

“Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.” personally this is the most exciting option pricing deal of the year, wow, kudos to both sides!!

“Cursor还给了SpaceX在今年晚些时候以600亿美元收购Cursor的权利，或者为我们的合作支付100亿美元。” 个人认为这是今年最令人兴奋的期权定价交易，哇，向双方致敬！！

do not miss. one of the INSANE gets courtesy of @osanseviero and the @GoogleDeepMind london avengers. if you always felt out of the loop on the SOTA on Imagegen, today or otherwise, this is the best 40 minutes you will find on the internet, period.

千万别错过。这是@osanseviero和@GoogleDeepMind伦敦复仇者联盟带来的疯狂收获之一。 如果你总是觉得跟不上Imagegen的最新技术水平（SOTA），那么今天或以后的这40分钟将是你能在互联网上找到的最好的内容，句号。

They put my father R.I.P. on a hate group list (insane, because he never hated anybody) and nearly destroyed his non-profit. It turns out that they are the biggest hate group in America. I hope they go to jail forever.

他们把我父亲（愿他安息）列入了仇恨组织名单（太疯狂了，因为他从不恨任何人），并差点摧毁了他的非营利组织。事实证明，他们才是美国最大的仇恨组织。我希望他们永远进监狱。

Last few weeks of updates have largely been focused on stability and reliability...had peak usage last few weeks. We are adding some exciting features over the next few months... Try it here: https://t.co/UAxNYUWgf0

过去几周的更新主要集中在稳定性和可靠性上……过去几周使用量达到顶峰。 我们将在未来几个月增加一些令人兴奋的功能…… 在这里尝试：https://t.co/UAxNYUWgf0

This is the future we are building at 8090. Our Software Factory is giving the companies that adopt it a new way to think about software. Build v buy was never as simple as it sounded. If you build you have complex maintenance issues. If you buy, you also generally double or triple your budget with services. The new way is to focus on the business logic and have Software Factory create something bespoke. Fits like a glove, highly tuned for you, low maintenance costs. Try it here: https://t.co/fkfTXgdfXK

这就是我们在8090构建的未来。我们的软件工厂（Software Factory）正在为采用它的公司提供一种思考软件的新方式。 “自建 vs 购买”从未像听起来那么简单。如果你自建，你会有复杂的维护问题。如果你购买，你通常也会在服务上增加两倍或三倍的预算。 新的方式是专注于业务逻辑，并让软件工厂创建定制的东西。 像手套一样贴合，为你高度调整，维护成本低。 在这里尝试：https://t.co/fkfTXgdfXK

Spoke at the AI club at Stanford last night. 1000 people tried to attend. Seating was capped at 250. It was pandemonium at the end! If you’re a resilient, gritty engineer, PM, designer or GTM person, please consider working with us: - We have no org chart - everyone reports to me. We do this to minimize politics, titles and force natural leaders to self organize. - We are severely under manned for the work we have (by design) so you are forced to engineer your way out. Build solutions not orgs. - We will book nine figures this year and are growing very quickly. Our customers span all major parts of the US Economy. Hiring@8090.ai

昨晚在斯坦福大学的AI俱乐部演讲。1000人试图参加。座位上限为250人。 最后简直是一片混乱！如果你是一名有韧性、有毅力的工程师、产品经理、设计师或GTM人员，请考虑与我们合作： - 我们没有组织架构图——每个人都向我汇报。我们这样做是为了最大限度地减少政治、头衔，并迫使自然领袖自我组织。 - 我们目前的人手严重不足（这是设计使然），所以你被迫通过工程手段解决问题。构建解决方案，而不是组织。 - 我们今年的预订额将达到九位数，而且增长非常快。我们的客户遍布美国经济的所有主要领域。 Hiring@8090.ai

New podcast w the incomparable @SadiSKhan of Aven on the future of finance (in the age of AI): https://t.co/1nj9mi3dpj

与Aven的无与伦比的@SadiSKhan关于金融未来（在AI时代）的新播客：https://t.co/1nj9mi3dpj

HF becoming the platform for agents (assisted by their humans) to use and build AI (rather than just leveraging APIs)!

HF正在成为智能体（在人类辅助下）使用和构建AI的平台（而不仅仅是利用API）！

Most of our usage is now coming from agents so to monitor how they use and mention our tools, with @DAKlingbeil, we're now running regular 10k queries to all the most popular coding agents, share the traces on @huggingface and analyze the results (ex https://t.co/o38usHC77k). How do you all check how coding agents use and mention your tools?

我们的大部分使用量现在来自智能体，因此为了监控它们如何使用和提及我们的工具，我和@DAKlingbeil现在定期对所有最流行的编码智能体运行10k查询，在@huggingface上分享追踪结果并分析结果（例如 https://t.co/o38usHC77k）。 你们大家是如何检查编码智能体如何使用和提及你们的工具的？

You know what scares the shit out of me? Data Brokers. Massive “legitimate” companies that make tons of money gathering and cross-linking as much data about you as possible. They know which hand you wipe with. And what brand of toilet paper. You know what scares me twice as much? Combining that quality of data with AI. You know what scares me three times as much. The fact that governments can just buy the data about you. Legally. So check this out. The Data Brokers don’t *get away* with selling your data. It’s their actual business model. Like, they have parking lots and coffee budgets and 401K plans. And the FBI recently bought a bunch of data on Americans from them. AI Data Brokers Governments And it’s all perfectly legal because selling your data is literally what Data Brokers exist to do. This is why the Dark Web has never been scary to me. A whole lot more shit is hidden on the light web right in front of us.

你知道什么让我吓得要死吗？数据经纪人。 大规模的“合法”公司，通过收集和交叉链接尽可能多关于你的数据来赚取大量金钱。 他们知道你用哪只手擦屁股。以及你用什么品牌的卫生纸。 你知道什么让我更害怕两倍吗？将这种质量的数据与AI结合起来。 你知道什么让我更害怕三倍吗？政府可以直接购买关于你的数据。 合法地。 所以看看这个。数据经纪人并没有因为出售你的数据而“逃脱惩罚”。这是他们的实际商业模式。就像他们有停车场、咖啡预算和401K计划一样。 而且FBI最近从他们那里购买了一堆关于美国人的数据。 AI 数据经纪人 政府 而且这一切完全合法，因为出售你的数据正是数据经纪人存在的原因。 这就是为什么暗网从来没有让我感到害怕。在我们的眼皮底下，明网中隐藏着更多肮脏的东西。

Jessica: Starmer is getting a lot of heat over this Mandelson thing. Me: Are they calling him embattled yet? Le Monde: Embattled Prime Minister Keir Starmer said on Monday, April 20, that he had been wrong to appoint Labour politician Peter Mandelson...

Jessica：Starmer因为Mandelson这件事受到了很多压力。 我：他们开始称他为“四面楚歌”了吗？ 世界报：四面楚歌的首相Keir Starmer周一（4月20日）表示，他任命工党政治家Peter Mandelson是错误的……

Wake up at 4:30 am. Essay starts forming in my head. Since I'm staying in a hotel, my laptop is right there, so I get up to at least write down the sentences I already have. Two hours later I've written the whole thing.

凌晨4:30醒来。文章开始在脑海中形成。因为我住在酒店，笔记本电脑就在旁边，所以我起床至少把我已经想好的句子写下来。两个小时后，我已经写完了整篇文章。

This is true of startups too. You launch to learn what you should have built.

这对初创公司来说也是如此。你发布产品是为了了解你本应该构建什么。

This is the aspect of climate change that I worry most about — when instead of seeing gradual degradation, we cross an irreversible line.

这是气候变化中我最担心的方面——当我们看到的不是逐渐的退化，而是跨越了一条不可逆转的界限。

This will be interesting to watch, because it will be a conversation between two of the most fearsomely effective people I know. I'm honestly curious to see what a conversation between these two looks like.

这会很有趣，因为这将是两个我所认识的最令人敬畏的高效人士之间的对话。我真的很想看看他们两个之间的对话是什么样子的。

I used to be slightly bummed that Jessica didn't care all that much about watches, but I've realized that not caring much means they all look the same to her, and that in turn means she doesn't always realize when I've bought a new one.

我曾经因为Jessica不太关心手表而感到有点沮丧，但我意识到不关心意味着在她的眼里它们看起来都一样，而这反过来意味着我买了新手表时她并不总是能发现。

Great to meet with @BecaGroup in Auckland and see how they are using Azure, Foundry, and their BEYON platform to make the New Zealand Geotechnical Database more accessible and useful. It is a powerful example of AI helping engineers access critical data faster and make better decisions, as they build more resilient infrastructure across New Zealand.

很高兴在奥克兰会见@BecaGroup，看看他们如何使用Azure、Foundry和他们的BEYON平台，使新西兰岩土工程数据库更易于访问和使用。 这是一个强大的例子，展示了AI如何帮助工程师更快地访问关键数据并做出更好的决策，因为他们正在新西兰各地构建更具韧性的基础设施。

the Codex x @skybysoftware acquisition may have been one of the best @openai deals made in the last year. I've been waiting for "real" computer use since @romainhuet demoed the ChatGPT App with 4o Vision at AIEWF 2024... and only now it's really, actually rolling out in a usable fashion.

Codex x @skybysoftware的收购可能是过去一年中最好的@openai交易之一。 我一直在等待“真正”的计算机使用，自从@romainhuet在AIEWF 2024上演示了带有4o Vision的ChatGPT App以来……直到现在它才真正以可用的方式推出。

[very fortunate for haystack to be a part of this on many levels - investment, and working w/ our close friends at dimension!]

[非常幸运Haystack能在多个层面上参与其中——投资，并与我们在Dimension的亲密朋友合作！]

Silicon Valley has a God complex. The critics are not wrong about that. They are just wrong about which god. Read my first essay on Substack about the faith that built Silicon Valley, the denominations fighting within it, and why losing that faith is the real risk. https://t.co/EznOiYNqMX

硅谷有上帝情结。批评者在这点上并没有错。他们只是搞错了是哪位上帝。 阅读我在Substack上的第一篇文章，关于构建硅谷的信仰、其中争斗的教派，以及为什么失去这种信仰才是真正的风险。https://t.co/EznOiYNqMX

So... "ai, explain what I just did": we ingested 9M+ directed edges from X into a weighted influence graph, then ran a personalized PageRank variant — tuned for human accounts — to surface eigenvector centrality at scale. once you know the high-centrality nodes, you watch their signal propagation before it hits the long tail. beta soon. @digg @basic_in_

所以……“AI，解释一下我刚才做了什么”：我们从X摄入了900万+有向边到一个加权影响图，然后运行了一个个性化PageRank变体——针对人类账户进行了调整——以大规模呈现特征向量中心性。 一旦你知道了高中心性节点，你就可以在信号传播到达长尾之前观察它们的信号传播。 测试版即将推出。@digg @basic_in_

Getting LLMs to simulate “true” randomness or generate diverse outputs is surprisingly difficult. We found a simple prompting trick that solves this by having the model generate and manipulate a random string. To be presented at #ICLR2026 this week! Blog: https://t.co/CyevqqJ5ej

让LLM模拟“真正的”随机性或生成多样化的输出是非常困难的。我们发现了一个简单的提示词技巧，通过让模型生成和操作一个随机字符串来解决这个问题。本周将在#ICLR2026上发表！ 博客：https://t.co/CyevqqJ5ej

I am very proud of our team for releasing EDINET-Bench, and it is fantastic to see a Japanese financial dataset recognized at #ICLR2026 this week. We need more diverse, non-English datasets to evaluate models in the real world. https://t.co/Q9y4xwWqgM https://t.co/Bh6LgNOMG4

我为我们的团队发布EDINET-Bench感到非常自豪，很高兴看到一个日本金融数据集在本周的#ICLR2026上得到认可。我们需要更多样化的非英语数据集来评估现实世界中的模型。 https://t.co/Q9y4xwWqgM https://t.co/Bh6LgNOMG4

When it comes to Excel, Diarmuid Early is in a different league. Fun to have the chance to talk shop with the reigning Excel world champion! https://t.co/a13EqENaiB

说到Excel，Diarmuid Early处于不同的水平。 很高兴有机会与现任Excel世界冠军谈论工作！ https://t.co/a13EqENaiB

In a hypothetical world where literally everything could be automated at no cost (we're very far from that), everybody would work in interpersonal marketing. Seriously.

在一个假设的世界里，如果字面上一切都可以免费自动化（我们离那还很远），每个人都会从事人际营销。说真的。

In a capitalist economy, absolute levels of value creation and productivity take a backseat to relative competitive advantage. That's why jobs still exist despite 150 years of intense automation. As long as hiring someone represents a non-zero advantage for firm A in its competition with equivalent firm B, that person will have a job, regardless of how much automation these firms use or what their productivity rate is.

在资本主义经济中，价值创造和生产力的绝对水平次于相对竞争优势。这就是为什么尽管有150年的激烈自动化，工作岗位依然存在。只要雇佣某人对A公司在与同等B公司的竞争中代表非零优势，那个人就会有工作，无论这些公司使用多少自动化或它们的生产率如何。

Human biological limits, like our tiny working memory and shallow calculation depth, are actually a feature. They force us to abstract, compress, intuit. If we had infinite resources, we would never have needed intelligence.

人类的生物学局限性，比如我们微小的工作记忆和浅显的计算深度，实际上是一种特性。它们迫使我们抽象、压缩、直觉。如果我们有无限的资源，我们就永远不需要智慧。

I'd say cognitive friction is even more than just a regularizer: it's an incentive to find the right interface abstractions, and in turn good abstractions are what enables compounding over time. Piles of spaghetti don't compound, they collapse under their own weight after a while

我会说认知摩擦甚至不仅仅是一个正则化器：它是寻找正确接口抽象的激励，而反过来，好的抽象是实现随时间复合增长的原因。意大利面条堆不会复合，它们在一段时间后会在自身重量下崩溃。

Human cognitive friction has long been acting as a regularizer for a lot of digital infrastructure. It made software APIs less terrible and codebases less complex. Now LLM disintermediation is causing this effect to fade, which in turn will cause runaway software complexity.

人类的认知摩擦长期以来一直充当许多数字基础设施的正则化器。它使软件API不那么糟糕，代码库不那么复杂。 现在LLM去中介化导致这种效应正在消失，这反过来会导致软件复杂性失控。

"I am envisioning a harmonious interplay of form and space, where each element finds its rightful place within the visual narrative. The pelican emerges as a central figure, its wings poised in dynamic tension, gripping the handlebars with lifelike intent, while its feet extend confidently onto the ground line that anchors the scene."

“我正在设想一种形式与空间的和谐互动，其中每个元素都在视觉叙事中找到其应有的位置。鹈鹕作为一个中心人物出现，它的翅膀处于动态张力中，以栩栩如生的意图抓住车把，而它的脚自信地延伸到锚定场景的地面线上。”

I tried a 15MB, 30 page text-heavy PDF and Opus 4.7 reported 60,934 tokens while 4.6 reported 56,482 - that's a 1.08x multiplier, significantly lower than the multiplier I got for raw text.

我尝试了一个15MB、30页的文本密集型PDF，Opus 4.7报告了60,934个令牌，而4.6报告了56,482个——这是一个1.08倍的乘数，明显低于我从原始文本得到的乘数。

(Just tried community-noting my own tweet to clarify this point, let's see if that works!) https://t.co/mMSlkQEJ5j

（刚刚尝试社区注释我自己的推文来澄清这一点，看看是否有效！） https://t.co/mMSlkQEJ5j

New TIL on fetching data from a Datasette instance into Google Sheets using importdata(), named custom functions or Google Apps Script https://t.co/cmWVSNE1En

关于使用importdata()、自定义函数或Google Apps Script从Datasette实例获取数据到Google Sheets的新TIL（今天学到的东西） https://t.co/cmWVSNE1En

I upgraded my Claude token counter tool to compare different models and Opus 4.7 does appear to use 1.46x times the tokens for text and up to 3x the tokens for images - it's priced the same as Opus 4.6 on a per-token basis so this is actually a pretty big price bump https://t.co/PPGrwHf39D

我升级了我的Claude令牌计数器工具来比较不同的模型，Opus 4.7在文本上似乎确实使用了1.46倍的令牌，在图像上使用了高达3倍的令牌——它在每个令牌的基础上与Opus 4.6定价相同，所以这实际上是一个相当大的价格上涨 https://t.co/PPGrwHf39D

The biggest challenge of using chat-based AI systems is that the details of what they can do are invisible - those tool descriptions are the missing manual, publishing them would be a huge benefit to people who want to get the most out of Claude

使用基于聊天的AI系统的最大挑战是它们能做什么的细节是不可见的——那些工具描述是缺失的手册，发布它们对那些想从Claude获得最大收益的人来说将是一个巨大的好处。

Note to @AnthropicAI - much as I appreciate the public system prompts this would be so much more valuable to me as a Claude power user if you published the tool descriptions as well

给@AnthropicAI的说明——尽管我很欣赏公开的系统提示词，但如果你能发布工具描述，这对作为Claude高级用户的我来说会更有价值。

Told you. At 8090, we’re actually helping organizations (large and small) implement their AI transformation methodically and in a disciplined way that a) doesn’t leak all their data into model training and b) doesn’t just blow their OpEx budget to the benefit of yet another tool provider.

告诉过你们了。 在8090，我们实际上是在帮助组织（无论大小）有条不紊且纪律严明地实施他们的AI转型，a) 不会将他们所有的数据泄露到模型训练中，b) 不会仅仅为了另一个工具提供商的利益而耗尽他们的运营预算。

Facts. The fastest growing part of 8090 is our practice of ripping out legacy systems for large enterprises and migrating them to pristine, new, well documented and easy to maintain alternatives. It also turns out to be less than 50% of the TCO or better as well.

事实。 8090增长最快的部分是我们为大型企业拆除遗留系统并将其迁移到原始、新颖、文档齐全且易于维护的替代方案的实践。 事实证明，它的总拥有成本（TCO）也低于50%甚至更好。

Why? - Risk comes from capability asymmetry between attackers and defenders. Open source is what every frontier lab already trains on so defenders get the same AI firepower as attackers. With proprietary code, you're on your own. The biggest risk is someone training a model on your obscure stack and attacking you when no public model exists to defend it. - AI can now read stripped binaries, so proprietary obscurity barely protects anyone anymore. Most legacy firmware and embedded code is closed, binary-only, and no longer maintained. A huge attack surface that just became legible to AI. - In a Mythos world, software security becomes a speed race: detection, verification, coordination, patch propagation. Closed-source systems are weaker at all four because they centralize knowledge and action inside a vendor, while open ecosystems distribute both. - In open source, the defender crowd is usually bigger than the attacker crowd. In closed source, it's the opposite. AI force-multiplication will amplify that imbalance. And when closed-source systems fail, the blast radius tends to be much larger. They sit behind centralized user, customer, and cloud data, while open source more often runs locally with less data concentration. Let's go open-source!

为什么？ - 风险来自攻击者和防御者之间的能力不对称。开源是每个前沿实验室都已经训练的东西，所以防御者获得了与攻击者相同的AI火力。使用专有代码，你只能靠自己。最大的风险是有人在你的晦涩堆栈上训练模型，并在没有公共模型可以防御时攻击你。 - AI现在可以读取剥离后的二进制文件，所以专有晦涩几乎不再保护任何人。大多数遗留固件和嵌入式代码是封闭的、仅二进制的，且不再维护。一个巨大的攻击面刚刚变得对AI可读。 - 在Mythos世界中，软件安全变成了一场速度竞赛：检测、验证、协调、补丁传播。闭源系统在所有四个方面都较弱，因为它们将知识和行动集中在供应商内部，而开放生态系统则分发两者。 - 在开源中，防御者群体通常比攻击者群体大。在闭源中，情况恰恰相反。AI力量倍增将放大这种不平衡。 - 当闭源系统失败时，爆炸半径往往要大得多。它们位于集中的用户、客户和云数据之后，而开源更多地在本地运行，数据集中度较低。 让我们开源吧！

In a mythos world (which we are already in), closed-source projects will be 10x more at risk than open-source projects!

在一个Mythos世界（我们已经在其中了），闭源项目的风险将是开源项目的10倍！

team killed it with the livestream and production quality, this is so cool that hundreds of people are tuning in live https://t.co/U6ERyw55q0

团队在直播和制作质量上做得太棒了，数百人实时收看真是太酷了 https://t.co/U6ERyw55q0

excited to see a packed room (~75% East coast AI Engineers) @gabegreenberg kickoff AIE Miami - all talks streaming live for free on the youtube now! https://t.co/GT0YdSnlk7

很高兴看到一个挤满人的房间（约75%是东海岸AI工程师）@gabegreenberg启动AIE Miami - 所有演讲现在都在YouTube上免费直播！ https://t.co/GT0YdSnlk7

Of 140 planned US datacenter projects representing roughly 16 gigawatts of capacity, only about 5 gigawatts are actually under construction. A significant share of the physical infrastructure enabling one of America's largest infrastructure bets is being sourced from China.

在140个计划中的美国数据中心项目中，代表约16吉瓦的容量，只有约5吉瓦实际上在建设中。 使美国最大的基础设施赌注之一成为可能的物理基础设施中，有很大一部分是从中国采购的。

This issue is sponsored by @datocms! The headless CMS that won't make you regret your stack choices. Know that feeling when something just works the way you expect it to? Bootstrapped, generous free tier, DX-first, works with any framework. https://t.co/nOOXBzO0oB

本期由@datocms赞助！ 那个不会让你后悔技术栈选择的无头CMS。知道那种事情如你所愿地运行的感觉吗？ 自举，慷慨的免费层，DX优先，适用于任何框架。 https://t.co/nOOXBzO0oB

It's newsletter time, y'all! Open up the latest issue heading to you right now for a fun joke, some interesting articles, practice coding, and mooore! https://t.co/Q1MmhTw8hd

伙计们，是通讯时间！ 打开现在发给你的最新一期，里面有一个有趣的笑话、一些有趣的文章、练习编码，还有更多！ https://t.co/Q1MmhTw8hd

If it can be explained, AI can do it. So the question is: What can’t be explained? And why not?

如果它能被解释，AI就能做到。 所以问题是： 什么不能被解释？为什么不能？

if you guys come to @aiDotEngineer Singapore i will personally lead a tour to the best cai fan i know! may 15-17 still recruiting a couple speakers, dm if mutuals and want a VIP tour of 🇸🇬 https://t.co/oajuFKoLs8

如果你们来@aiDotEngineer新加坡，我将亲自带你们去我所知道的最好的杂菜饭（cai fan）店！5月15-17日 还在招募几位演讲者，如果是互相关注且想要VIP 🇸🇬之旅的请私信 https://t.co/oajuFKoLs8

Make neural network cells inside a “Digital Petri Dish” fight for control and dominance in a web browser tab. https://t.co/t8N6CIhvze https://t.co/A5SSTvPJBu

让神经网络细胞在“数字培养皿”内争夺浏览器标签页中的控制权和统治地位。 https://t.co/t8N6CIhvze https://t.co/A5SSTvPJBu

Unlimited demand for something doesn't necessarily mean it is viable as a business, much less as an economy-wide all-in bet.

对某事物的无限需求并不一定意味着它作为一种商业是可行的，更不用说作为全经济范围的全力投入了。

There's no doubt that the world can consume tokens as fast as they're produced, even in the most maximalist infrastructure buildup scenarios imaginable. That's not the question. The question is whether the economic value of those tokens can match their total cost of production.

毫无疑问，世界可以像生产代币一样快地消耗代币，即使在可以想象的最极端的最大化基础设施建设场景中也是如此。这不是问题所在。问题在于这些代币的经济价值能否与其总生产成本相匹配。

The quality of your thinking is a multiplier for the amount of progress you make at each iteration. But the dominant factor behind success is simply your iteration speed. Try more things and you win.

你思考的质量是你每次迭代所取得进步的乘数。但成功的决定性因素仅仅是你的迭代速度。尝试更多的事情，你就会赢。

You cannot think your way to a perfect design. Only building and testing, over many iterations, can reveal the flaws in your mental model and provide the feedback you need to create the best design possible.

你无法通过思考达到完美的设计。只有通过构建和测试，经过多次迭代，才能揭示你心理模型中的缺陷，并提供你创造最佳设计所需的反馈。

Since Anthropic publish their system prompts we can generate a diff between Claude Opus 4.6 and 4.7 - here are my notes on what's changed https://t.co/IQHuvLGmwO

由于Anthropic发布了他们的系统提示词，我们可以生成Claude Opus 4.6和4.7之间的差异——这里是我关于发生了什么变化的笔记 https://t.co/IQHuvLGmwO

Hugging Face is becoming the platform for agents to use and build AI. Now they can call 1M HF spaces to do everything the latest specialized models can do! https://t.co/8VHxutX1PD

Hugging Face正在成为智能体使用和构建AI的平台。现在它们可以调用100万个HF空间来做最新专业模型能做的任何事情！ https://t.co/8VHxutX1PD

Japanese dealers publish timegrapher numbers in their listings. Northern European dealers will give you timegrapher numbers if you ask. Italian dealers act like they've never heard of timegraphers.

日本经销商在他们的列表中发布计时仪数据。北欧经销商如果你问他们，会给你计时仪数据。意大利经销商表现得好像他们从来没听说过计时仪一样。

Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. Don't listen to him, Sam, Yoshua, Geoff, or me on this topic. Listen to economists who have spent their career studying this, like @Ph_Aghion , @erikbryn , @DAcemogluMIT , @amcafee , @davidautor

Dario错了。 他对于技术革命对劳动力市场的影响一无所知。 在这个话题上不要听他的，也不要听Sam、Yoshua、Geoff或我的。 听听那些花了一辈子研究这个问题的经济学家，比如@Ph_Aghion, @erikbryn, @DAcemogluMIT, @amcafee, @davidautor

And yes, Lush/SN used a homegrown Lisp interpreter, to which a compiler was added in the early 1990s.

是的，Lush/SN在1990年代初使用了一个自制的Lisp解释器，并为其添加了一个编译器。

When looking at deep learning profiles, one of the most obvious tells between a mediocre and great candidate is whether they list PyTorch or JAX.

在查看深度学习档案时，平庸候选人和优秀候选人之间最明显的区别之一是他们列出的是PyTorch还是JAX。

Constraints are the catalyst of invention. An infinite search space leads to paralysis. The most creative inventions happen when you are forced to solve a problem within appropriately narrow constraints.

约束是发明的催化剂。无限的搜索空间会导致瘫痪。最有创意的发明发生在你被迫在适当狭窄的约束内解决问题时。

The kind people at @eth_limo have warned me that there has been an attack on their DNS registrar. So please do not visit https://t.co/BVfZIYrDKe or other https://t.co/OgoUF2qKUY pages until they confirm that things are back to normal. You can check my blog via IPFS directly here: https://t.co/SPMT3qScNI

@eth_limo的好心人警告我，他们的DNS注册商受到了攻击。所以请不要访问 https://t.co/BVfZIYrDKe 或其他 https://t.co/OgoUF2qKUY 页面，直到他们确认一切恢复正常。 你可以直接通过IPFS查看我的博客：https://t.co/SPMT3qScNI

The interesting thing here is asking the question of what tools/services I’ll keep. Today, next week, next year. I have my own offerings. What makes them good? Are people able to do what I just did and build their own in a few hours? The answer is no, and I know why, but is an important question to keep asking. So many existing companies have no good answer to this question. Like a massive percentage. Are you an interface? Are you a database? Are you data? What are you? What makes you, you? What makes you hard to copy? Everyone better figure this out real quick.

这里有趣的是问我将保留哪些工具/服务的问题。今天，下周，明年。 我有我自己的产品。是什么让它们变好？人们能像我刚才那样做并在几个小时内构建自己的吗？ 答案是不能，我知道为什么，但这是一个需要不断询问的重要问题。 这么多现有的公司对这个问题没有好的答案。就像很大一部分。 你是一个接口吗？ 你是一个数据库吗？ 你是数据吗？ 你是什么？ 是什么让你成为你？ 是什么让你难以复制？ 每个人最好尽快弄清楚这一点。

Added to my reminders this week: - Cancel Zapier - Cancel Resend - Cancel Figma - Cancel Canva - Cancel Browserbase - Cancel Supabase Recreated all this in my own PAI harness w/ various repos, my own Skills/Workflows/CLIs and CC. Such a fun and freeing feeling.

本周添加到我的提醒事项中： - 取消Zapier - 取消Resend - 取消Figma - 取消Canva - 取消Browserbase - 取消Supabase 在我自己的PAI工具包中重新创建了所有这些，使用了各种存储库、我自己的技能/工作流/CLI和CC。 这种感觉真有趣，真自由。

Ron has been there for so long for so many of us. We all love him and hope for a fast and healthy recovery.

Ron陪伴了我们很多人很长时间。我们都爱他，希望他能快速健康地康复。

The event is next month - talks are Friday 15th to Sunday 17th of May, with the date before that of tutorials and two days afterwards of sprints Californians are notoriously last-minute planners so I know that's enough time to decide to come join us!

活动在下个月——演讲时间是5月15日星期五到17日星期日，之前是教程日，之后是两天的冲刺日。 加州人以最后一刻的规划者而闻名，所以我知道那足够让他们决定来加入我们了！

Join us at PyCon US 2026 in Long Beach—we have new AI and security tracks this year https://t.co/ZVHmSlVYEP

加入我们在长滩举行的PyCon US 2026——我们今年有新的AI和安全轨道 https://t.co/ZVHmSlVYEP

Is there still a widespread belief that LLMs and coding agents are good for greenfield development but don't help for maintaining large existing codebases? I don't think that idea holds up any more

是否仍然存在一种普遍的信念，即LLM和编码智能体对绿地开发有好处，但对维护大型现有代码库没有帮助？ 我不认为这个想法还站得住脚。

What folk don’t get is that the play of Anthropic et al is to get enterprises to use their models via their own wrapper hooked into systems of record This is why they are moving from per seat pricing to “API” pricing Best models will only be available via their wrappers..

人们不明白的是，Anthropic等公司的策略是让企业通过钩入记录系统的包装器来使用他们的模型。 这就是为什么他们从“按席位”定价转向“按API”定价。 最好的模型将只能通过他们的包装器提供……

I wrote about a little project we made on stream yesterday! 🤖 It's built with the GitHub Copilot CLI 📋 It takes in a list of bullet points and returns relevant emoji per bullet 📎 Copies the result to the clipboard https://t.co/yOQLTu0vNx

我写了关于我们昨天在直播中做的一个小项目！ 🤖 它是用GitHub Copilot CLI构建的 📋 它接收一列要点并为每个要点返回相关的表情符号 📎 将结果复制到剪贴板 https://t.co/yOQLTu0vNx

Looking for some beta testers for our new robotics dataset hosting features. Who's currently hosting large robotics datasets and would be down to try? 🤖🤖🤖

正在为我们的新机器人数据集托管功能寻找一些Beta测试人员。谁目前正在托管大型机器人数据集，并且愿意尝试一下？🤖🤖🤖

I am happy everyone is switching to Codex, but Tibo if you start rate limiting me or making me use worse models...

我很高兴每个人都切换到Codex，但Tibo，如果你开始限制我的速率或者让我使用更差的模型……

Lots of major improvements to Codex! Computer use is a real update for me; it feels even more useful than I expected. It can use all of the apps on your Mac, in parallel and without interfering with your direct work.

Codex有很多重大改进！ 计算机使用对我来说是一个真正的更新；它比我预期的更有用。它可以在你的Mac上并行使用所有应用程序，而不会干扰你的直接工作。

Codex can learn from experience and proactively suggest things it can do for you. It now has an in-app browser, many new plugins, and so much more.

Codex可以从经验中学习，并主动建议它可以为你做的事情。 它现在有一个应用内浏览器、许多新插件等等。

Just had @a16z GP @illscience on the pod, he has a wild take: brand and marketing will soon have 'zero value' because AI agents will make purchasing decisions based on perfect product information. https://t.co/ifIqVjtYik

刚刚在播客上采访了@a16z的合伙人@illscience，他有一个疯狂的观点：品牌和营销很快将“毫无价值”，因为AI智能体会根据完美的产品信息做出购买决定。 https://t.co/ifIqVjtYik

MIT has invented a 3D zipper that builds itself into a rigid object as you zip it. Could be a useful general purpose technology in the future. https://t.co/qamepgONYa

麻省理工学院发明了一种3D拉链，当你拉上它时，它会把自己构建成一个刚性物体。在未来可能是一种有用的通用技术。 https://t.co/qamepgONYa

Video: https://t.co/vxuGAFBhqU Press Release: https://t.co/N5GZUItgaE Bizarre broken ACM paper PDF hosting service that invites you to listen to an AI podcast of the paper: https://t.co/DFsXZDsSE9

视频：https://t.co/vxuGAFBhqU 新闻稿：https://t.co/N5GZUItgaE 奇怪的、损坏的ACM论文PDF托管服务，邀请你听论文的AI播客：https://t.co/DFsXZDsSE9

While we're remembering David MacKay, here's one of my favorite figures from his 2008 book. That yellow square in Texas is a visualization of the solar farm that would be required to power all of 2008-era America. Gemini tells me it would be 40% smaller with today's panels. https://t.co/gdIEkbwoH7

在我们纪念David MacKay的同时，这是他2008年书中最喜欢的图表之一。德克萨斯州那个黄色的方块是为2008年时代的美国提供动力所需的太阳能农场的可视化。Gemini告诉我，用今天的面板，它会小40%。 https://t.co/gdIEkbwoH7

Adithya is putting his work where his heart is. If you believe in open research and open AI, join a company that actually lives those values, not a closed-source, revenue-maximizing one!

Adithya正在用他的工作践行他的信念。 如果你相信开放研究和开放AI，就加入一家真正践行这些价值观的公司，而不是一家闭源、利润最大化的公司！

You can now visualize Pi traces that you upload on @huggingface! Let's make sharing agent traces 10x more common to make agent AI more open and collaborative! Also, because it's fun to analyze @badlogicgames's traces 😂😂😂 https://t.co/LLclFIZeWS

你现在可以在@huggingface上可视化你上传的Pi追踪！ 让我们让分享智能体追踪变得更常见10倍，以使智能体AI更加开放和协作！而且，分析@badlogicgames的追踪很有趣 😂😂😂 https://t.co/LLclFIZeWS

Completely agree with Jensen! Not exporting AI out of fear is classic case where the cure would be 100x worse than the disease. You slow down innovation, progress and US technology and economic leadership in the foolish hope that you'll prevent an edge case that hasn't even proved to be dangerous. It’s like stopping exports of metal because it might help adversaries build teleportation machines.

完全同意Jensen！ 出于恐惧而不出口AI是一个典型的例子，治愈方法会比疾病糟糕100倍。你减缓了创新、进步以及美国的技术和经济领导地位，却怀着愚蠢的希望去防止一个甚至还没有被证明是危险的边缘案例。 这就像停止金属出口，因为它可能帮助对手制造传送机器。

This week's newsletter sponsor is @floxdevelopment! Flox is a pkg + env manager that keeps your local dev, CI, and prod environments identical. It's powered by Nix, no containers needed! If you're building agentic AI workflows, this is worth your time: https://t.co/8teQ0Zt091

本周的通讯赞助商是@floxdevelopment！ Flox是一个包+环境管理器，可以保持你的本地开发、CI和生产环境完全一致。它由Nix驱动，不需要容器！ 如果你正在构建智能体AI工作流，这值得你花时间： https://t.co/8teQ0Zt091

It's time for Rubber Duck Thursdays! Come join us at GitHub on your channel of choice: https://t.co/jfVae0f0RL https://t.co/UVZnOmvXEK https://t.co/IXHBX2A7xY

又是橡胶鸭星期四了！快来你选择的频道加入我们在GitHub的活动： https://t.co/jfVae0f0RL https://t.co/UVZnOmvXEK https://t.co/IXHBX2A7xY

idk how else to say this, but... build your dream projects now. I feel like all the tools are giving away a LOT for free/cheap now. It's only gotten more pricey over time, and will keep getting more expensive. Your ideas are subsidized now, think of it as a fire sale and build!

我不知道还能怎么说，但是……现在就构建你的梦想项目吧。 我觉得所有的工具现在都在免费/便宜地赠送很多东西。随着时间的推移，它只会变得更贵，而且会越来越贵。你的想法现在是被补贴的，把它看作是一场大甩卖，赶紧构建吧！

Here's Qwen 3.6-35B-A3B v.s. Claude Opus 4.7 for "Generate an SVG of a flamingo riding a unicycle", in case you thought Qwen might be cheating at the pelican benchmark https://t.co/OnwhXg3Q2N

这是Qwen 3.6-35B-A3B对比Claude Opus 4.7生成的“生成一只骑独轮车的火烈鸟的SVG”，以防你认为Qwen可能在鹈鹕基准测试中作弊 https://t.co/OnwhXg3Q2N

More on my blog, including results from the previously secret "flamingo on a unicycle" test https://t.co/bUBjyOsBuo

更多内容在我的博客上，包括之前秘密的“骑独轮车的火烈鸟”测试结果 https://t.co/bUBjyOsBuo

Shocking result on my pelican benchmark this morning, I got a better pelican from a 21GB local Qwen3.6-35B-A3B running on my laptop than I did from the new Opus 4.7! Qwen on the left, Opus on the right https://t.co/kDlbnJv6YI

今天早上我的鹈鹕基准测试结果令人震惊，我从在我笔记本电脑上运行的21GB本地Qwen3.6-35B-A3B中得到的鹈鹕比从新的Opus 4.7中得到的更好！ 左边是Qwen，右边是Opus https://t.co/kDlbnJv6YI

So we can't just think of personalization as care for the user but helping them support and care for their loved ones too. Someone asking about an infant’s or an elderly relative’s symptoms may need different information or follow-up recommendations than someone asking about their own.

所以我们不能仅仅把个性化看作是对用户的照顾，还要帮助他们支持和照顾他们所爱的人。询问婴儿或老年亲属症状的人可能需要与询问自己症状的人不同的信息或后续建议。

Super proud of the work that went into this. These insights could have important implications for how we and the industry design these tools - making sure what we build is what people actually need. An important milestone for @MicrosoftAI and super proud of the many people behind it - from our Futures research team, data science, Health team, and beyond.

为这项工作所付出的努力感到无比自豪。这些见解可能对我们和整个行业如何设计这些工具产生重要影响——确保我们构建的是人们真正需要的东西。 这是@MicrosoftAI的一个重要里程碑，为背后许多人感到无比自豪——从我们的未来研究团队、数据科学、健康团队等等。

Our paper landed in Nature Health today! Healthcare is one of the most high-stakes, high-potential applications of AI. So we set out to understand how people actually use it in our AI products today. https://t.co/Bqup7eb3Ic https://t.co/mvp6OYfury

我们的论文今天在Nature Health上发表了！ 医疗保健是AI应用中风险最高、潜力最大的领域之一。因此，我们着手了解人们今天在我们的AI产品中是如何实际使用它的。 https://t.co/Bqup7eb3Ic https://t.co/mvp6OYfury

I have found 4.7 great for design, reverted back to 4.6 extended for everything else Anyone else like this?

我发现4.7在设计方面很棒，其他一切都恢复到了4.6扩展版。 还有其他人这样吗？

Turning 43 today ⏳ Didn’t figure out the meaning of life, the universe and everything at 42 🤔 Maybe this year will figure it out 🤞 A gift for everyone Monday 🔜 🎁

今天43岁了 ⏳ 42岁时没搞明白生命、宇宙和一切的意义 🤔 也许今年能搞明白 🤞 周一给大家一份礼物 🔜 🎁

And while you're at it, go follow @ronaldeddings. He's putting out all kinds of epic shit.

顺便说一句，去关注@ronaldeddings。他正在发布各种史诗般的内容。

Seriously impressive work, and I highly recommend trying it if you're still grappling with your agent browsing sitch. https://t.co/PU3i1jW0Hb

令人印象深刻的工作，如果你还在努力解决你的智能体浏览情况，我强烈建议尝试一下。 https://t.co/PU3i1jW0Hb

That by itself made me switch to it as the primary agent browsing system for PAI. But Interceptor's MAIN feature is the ability to follow specific workflows during browsing. And to deeply control everything about the browsing session. https://t.co/7SdGE2OHIl

仅此一点就让我把它作为PAI的主要智能体浏览系统。 但Interceptor的主要功能是在浏览过程中遵循特定工作流的能力。以及深度控制浏览会话的一切。 https://t.co/7SdGE2OHIl

I see a ton of AI tools, but the one that's added the most value to my PAI harness recently is Interceptor It's by my buddy @ronaldeddings, and it's by far the best Browser Control System for agents that I've used. And I've literally used over 100 at this point.

我看到了大量的AI工具，但最近为我的PAI工具包增加价值最多的是Interceptor。 它是我朋友@ronaldeddings的作品，是我用过的最好的智能体浏览器控制系统。 而且我目前已经用了超过100个了。

.@WIRED Health talking about what's next for Manas, and how we can use AI to cure the uncurable! https://t.co/PfUU04WGxL

.@WIRED Health谈论Manas的下一步，以及我们如何利用AI治愈不可治愈的疾病！ https://t.co/PfUU04WGxL

Marin is using quantile balancing from @Jianlin_S (who developed RoPE, which was also a good idea) to train our current 1e23 FLOPs MoE. The idea is elegant: assigning tokens to experts by solving a linear program. No hyperparameters to tune. Yields stable training.

Marin正在使用来自@Jianlin_S（他开发了RoPE，这也是一个好主意）的分位数平衡来训练我们当前的1e23 FLOPs MoE。这个想法很优雅：通过求解线性规划将令牌分配给专家。没有超参数需要调整。产生稳定的训练。

This week, @classiclarryd kicked off a 129B (16B active) 1e23 FLOPs MoE run. In typical Marin style, we have fit scaling laws and have made a loss projection of 2.252. Stay tuned. https://t.co/QnwJ8YxT9H

本周，@classiclarryd启动了一个129B（16B活跃）1e23 FLOPs MoE运行。以典型的Marin风格，我们拟合了缩放定律，并做出了2.252的损失预测。敬请期待。 https://t.co/QnwJ8YxT9H

This is Act II. Act I was about making an anonymity layer for LLMs (VPN for intelligence). Act II is about building a deeply personalized, private assistant on top of that. The idea is that your context (all your files, messages, deepest desires) is owned and managed by you. For any query, a local/TEE model reads the context to determine what *subset* of context to pull in, and invokes closed frontier models on the context (if open models aren't good enough). With the anonymity layer, different invocations are not linked. So your context can have information about your taxes and your health records, but you never allow any model provider to link the two despite having a unified assistant interface. The vision of deeply personalized assistants is obvious right now. It is less obvious that you can achieve this privately.

这是第二幕。第一幕是关于为LLM制作匿名层（智能的VPN）。第二幕是关于在此基础上构建一个深度个性化、私密的助手。 这个想法是，你的上下文（你所有的文件、消息、最深层的欲望）由你拥有和管理。对于任何查询，本地/TEE模型读取上下文以确定要拉入上下文的子集，并在上下文上调用封闭的前沿模型（如果开放模型不够好）。通过匿名层，不同的调用不会被链接。所以你的上下文可以包含关于你的税收和健康记录的信息，但你永远不会允许任何模型提供商将两者链接起来，尽管拥有统一的助手界面。 深度个性化助手的愿景现在很明显。你能不能私下实现这一点就不那么明显了。

It was a delight to serve on the selection committee with an awesome group of committee members. We had lots of great submissions to evaluate and excellent discussions during our selection process. You can see the selected proposals below and learn more at https://t.co/s3b8aXrp6Y

很高兴能与一群很棒的委员会成员一起在评选委员会任职。我们有许多很棒的提交作品需要评估，并在我们的评选过程中进行了深入的讨论。 你可以在下面看到选定的提案，并在 https://t.co/s3b8aXrp6Y 了解更多信息

In case it's not clear, you can click on the picture in the first image to learn more about Ricardo and the award: Twitter swallowed my actual link text to https://t.co/r6Xkt7poSW to turn it into the picture, I guess.

如果还不清楚，你可以点击第一张图片中的图片来了解更多关于Ricardo和该奖项的信息：我猜Twitter吞掉了我指向 https://t.co/r6Xkt7poSW 的实际链接文本，把它变成了图片。

Delighted to see that Ricardo Baeza-Yates (@PolarBearby) is this year's winner of the @TheOfficialACM Luiz Barroso Award (an award named in honor of my longtime Google colleague). Ricardo is widely regarded as one of the world’s foremost researchers in information retrieval, and many of you may have read is excellent textbook Modern Information Retrieval (co-authored with Berthier Ribeiro-Neto). He has also played a pivotal role in strengthening the Latin American computing community. Ricardo's accomplishments and selection would make Luiz proud! https://t.co/sinAPruQW8

很高兴看到Ricardo Baeza-Yates (@PolarBearby) 是今年@TheOfficialACM Luiz Barroso奖（以我长期谷歌同事命名的奖项）的获得者。 Ricardo被广泛认为是世界上最重要的信息检索研究人员之一，你们中的许多人可能读过他出色的教科书《现代信息检索》（与Berthier Ribeiro-Neto合著）。他在加强拉丁美洲计算社区方面也发挥了关键作用。 Ricardo的成就和入选会让Luiz感到骄傲！ https://t.co/sinAPruQW8

"who made this? [share]" - is the new way to propagate something horrible on social media without looking like you're a bad person

“这是谁做的？[分享]”——是在社交媒体上传播可怕内容的最新方式，而看起来你不是一个坏人。

thank the heavens for the claude code mini-summaries at the bottom of terminal windows, sooo many tabs, this helps https://t.co/9gqQABRDxO

感谢上天，终端窗口底部的Claude Code迷你摘要，标签页太多了，这很有帮助 https://t.co/9gqQABRDxO

Replace your claude code working verbs like "Beboppin" "Composing" and "Frolicking" with mini inspirational quotes, here is how: https://t.co/4qyWi4EH2z

用迷你励志名言替换你的Claude Code工作动词，比如“Beboppin”、“Composing”和“Frolicking”，方法如下： https://t.co/4qyWi4EH2z

Two models, two different parts of the creative process. MAI-Image-2-Efficient is a production workhorse. Volume, speed, tight cost control for iterative workflows. MAI-Image-2 is a precision tool. Highest fidelity, final deliverables, exact details, longer/more complex text. https://t.co/v6SEA4Ockg

两个模型，创作过程的两个不同部分。 MAI-Image-2-Efficient是一个生产主力。容量、速度、迭代工作流的严格成本控制。 MAI-Image-2是一个精密工具。最高保真度、最终交付成果、精确细节、更长/更复杂的文本。 https://t.co/v6SEA4Ockg

A https://t.co/KV5OJESo47 feature I really like is you can tell it to "clone x/y from GitHub" and it can then answer questions about a repo, or use snippets of code from that repo to help build new artifacts - used that just now to solve a minor friction https://t.co/BuLrxQ6X8R

我非常喜欢的一个https://t.co/KV5OJESo47功能是你可以告诉它“从GitHub克隆x/y”，然后它就可以回答关于存储库的问题，或者使用来自该存储库的代码片段来帮助构建新的工件——刚刚用它解决了一个小摩擦 https://t.co/BuLrxQ6X8R

I ran the same prompt for a London Estuary accent, a Newcastle accent and an Exeter, Devon accent - all three audio files are now embedded in my blog post https://t.co/pQR8GFBJds

我为伦敦河口口音、纽卡斯尔口音和德文郡埃克塞特口音运行了相同的提示词——所有三个音频文件现在都嵌入在我的博客文章中 https://t.co/pQR8GFBJds

Open-source is the solution to cyber-security because with new AI capabilities, all of the open-source repos will be inspected and patched 100x faster/better than any closed-source system! Let's take a practical example: You're a startup building a new feature that deals with private user information so you want to be careful. Do you trust more: - one of your team members with 0 security background to build the feature from 100% new AI slope code with their agent in a complete opaque/unauditable way (for example sending data to APIs) or - one of your team members with their agent using mature open-source projects where thousands of people and agents have been inspecting the code and vulnerabilities are regularly patched and where the data can stay locally on the user or the company devices The answer is obvious!

开源是网络安全的解决方案，因为有了新的AI能力，所有的开源存储库都将比任何闭源系统被检查和修补的速度快100倍/更好！ 让我们举一个实际的例子： 你是一家初创公司，正在构建一个处理私人用户信息的新功能，所以你想小心一点。 你更信任谁： - 你的团队成员中一个没有安全背景的人，用100%新的AI斜率代码以完全不透明/不可审计的方式构建该功能（例如将数据发送到API） 或者 - 你的团队成员中一个使用成熟的开源项目的人，成千上万的人和智能体一直在检查代码，漏洞被定期修补，数据可以保留在用户或公司设备本地 答案显而易见！

Weird how some people always target open-source in AI! First it was: “Open-source AI will destroy the world” (spoiler: it didn't and it won't) Now: “Open-source is a cybersecurity threat because of AI” Both narratives are far too simplistic. The truth is that the exact same risks exist in closed-source systems, often even more so. For example, in practice, APIs can create much bigger data and security vulnerabilities than open systems you can inspect, self-host, and secure yourself. And as with software more broadly, open-source often ends up more secure because it benefits from far more scrutiny than private internal systems. The reality is not “open vs closed.” The reality is that AI is raising cybersecurity stakes across the board, and we need to tackle that seriously together.

奇怪的是，有些人总是针对AI中的开源！ 起初是： “开源AI将毁灭世界”（剧透：它没有，也不会） 现在： “因为AI，开源是一个网络安全威胁” 这两种叙述都太简单了。 事实是，完全相同的风险存在于闭源系统中，通常甚至更多。例如，在实践中，API可以产生比你可以检查、自托管和自己保护的开放系统更大的数据和安全漏洞。 而且正如更广泛的软件一样，开源通常最终更安全，因为它受益于比私有内部系统更多的审查。 现实不是“开放 vs 封闭”。 现实是AI正在全面提高网络安全风险，我们需要认真地共同解决这个问题。

Introducing Gemini on Mac. It’s the first time we’re bringing the @Geminiapp to desktop. The team built this initial release with @Antigravity, and it went from an idea to a native Swift app prototype in a few days. More features on the way! https://t.co/YRy0Pqq6zo

隆重推出Gemini on Mac。 这是我们第一次将@Geminiapp带到桌面端。团队用@Antigravity构建了这个初始版本，它在几天内就从一个想法变成了一个原生Swift应用程序原型。 更多功能即将推出！ https://t.co/YRy0Pqq6zo

New course: Spec-Driven Development with Coding Agents, built in partnership with @jetbrains, and taught by @paulweveritt. Vibe coding is fast, but often produces code that doesn't match what you asked for. This short course teaches you spec-driven development: write a detailed spec defining what to build, and work with your coding agent to implement it. Many of the best developers already build this way. A spec lets you control large code changes with a few words, preserve context across agent sessions, and stay in control as your project grows in complexity. Skills you'll gain: - Write a detailed specification to define your mission, tech stack, and roadmap, giving your agent the context it needs from the start - Plan, implement, and validate features in iterative loops using a spec as your agent's guide - Apply the same repeatable workflow to both new and legacy codebases - Package your workflow into a portable agent skill that works across agents and IDEs Join and write specs that keep your coding agent on track! https://t.co/hI4GwuvhtN

新课程：使用编码智能体的规范驱动开发，与@jetbrains合作构建，由@paulweveritt教授。 氛围编码很快，但通常产生的代码与你要求的不符。这门短期课程教授你规范驱动开发：编写定义要构建内容的详细规范，并与你的编码智能体合作实现它。许多最好的开发者已经这样构建了。 规范让你用几个词控制大型代码更改，在智能体会话中保留上下文，并在项目复杂性增加时保持控制。 你将获得的技能： - 编写详细的规范来定义你的任务、技术栈和路线图，从一开始就给你的智能体它需要的上下文 - 使用规范作为智能体的指南，在迭代循环中规划、实现和验证功能 - 将相同的可重复工作流应用于新代码库和遗留代码库 - 将你的工作流打包成一个跨智能体和IDE工作的便携式智能体技能 加入并编写让你的编码智能体保持在正轨上的规范！ https://t.co/hI4GwuvhtN

Highly recommend pre-ordering. Few understand games better than Mark, and his learning is ultra-applicable in the AI era.

强烈建议预订。很少有人比Mark更了解游戏，他的学习在AI时代是超适用的。

Our Fairwater datacenter in Wisconsin is going live, ahead of schedule. As the world’s most powerful AI datacenter, it will bring together hundreds of thousands of GB200s into a single seamless cluster. Congrats to all the teams who made this possible!

我们在威斯康星州的Fairwater数据中心即将上线，进度超前。 作为世界上最强大的AI数据中心，它将把数十万个GB200集成到一个无缝集群中。 祝贺所有使这一切成为可能的团队！

We’re proud to share the progress we’re making in our efforts to protect the planet, including our highest-ever use of recycled material in our products. https://t.co/VtMzXvk3U9

我们很自豪能分享我们在保护地球方面所取得的进展，包括我们在产品中使用了有史以来最高比例的回收材料。 https://t.co/VtMzXvk3U9

While I am not involved in the Bittensor ecosystem I would recommend that they try a council of LLMs style of governance support for all proposals they propose. These should be checked against the whitepaper and best interests of the overall system and prompts verifiable.

虽然我没有参与Bittensor生态系统，但我建议他们尝试一种LLM委员会风格的治理支持，用于他们提出的所有提案。 这些应该根据白皮书和整个系统的最佳利益进行检查，并且提示词应该是可验证的。

The same is of course true of software -- to solve a problem in a scalable manner requires a lot more work and a lot more code compared to a simple non-scalable solution.

同样，软件当然也是如此——与简单的不可扩展解决方案相比，以可扩展的方式解决问题需要更多的工作和更多的代码。

There's a broadly held misconception in AI that methods that scale well are simple methods -- even, that simple methods usually scale. This is completely wrong. Pretty much none of the truly simple methods in ML scale well. SVM, kNN, random forests are some of the simplest methods out there, and they don't scale at all. Meanwhile "train a transformer via backprop and gradient descent" is a very high-entropy method, easily 10x more complex than random forest fitting. But it scales very well. Further, given a simple method that doesn't scale, it is usually the case that you alter it to make it scale by adding a lot of complication. For instance, take a simple a simple combinatorial search-based method (not scalable at all) -- you can make it scale by adding deep learning guidance (which blows up complexity). Scalability usually belongs to high-entropy, complex systems.

在AI领域有一个广泛持有的误解，即扩展性好的方法是简单的方法——甚至，简单的方法通常可以扩展。这是完全错误的。 机器学习中几乎没有任何真正简单的方法可以很好地扩展。SVM、kNN、随机森林是最简单的方法，它们根本无法扩展。同时，“通过反向传播和梯度下降训练Transformer”是一种非常高熵的方法，复杂程度轻易是随机森林拟合的10倍。但它扩展得非常好。 此外，给定一个不能扩展的简单方法，通常情况是你通过增加很多复杂性来改变它以使其可扩展。例如，采用一种简单的基于组合搜索的方法（根本无法扩展）——你可以通过增加深度学习指导（这会使复杂性爆炸）来使其扩展。可扩展性通常属于高熵、复杂的系统。

Our most expressive and steerable TTS model yet! Designed to give builders granular control over AI-generated speech, Gemini 3.1 Flash TTS is really fun to play with! Available in preview today - for devs via the Gemini API & @GoogleAIStudio + for enterprises on Vertex AI

我们迄今为止最富表现力和可操纵的TTS模型！旨在让构建者对AI生成的语音进行精细控制，Gemini 3.1 Flash TTS玩起来真的很有趣！今天提供预览——开发者可通过Gemini API & @GoogleAIStudio + 企业可通过Vertex AI访问

OH 1 - “There should be an AI consumer app called ‘Bruh’ where the AI is just your bro.”

OH 1 - “应该有一个叫‘Bruh’的AI消费者应用，AI就是你的兄弟。”

I'm certain this isn't the message they intended to present, but this comes across to me as a company saying "we no longer trust in our own ability to keep your data secure"

我敢肯定这不是他们打算传达的信息，但这给我的感觉是一家公司在说“我们不再相信自己有能力保证你的数据安全”

We are hiring Software Engineers in Tokyo to help us scale Sakana AI’s R&D. If you’re interested in building the data pipelines and full stack infrastructure needed to push the boundaries of automated scientific discovery, we’d love to hear from you. 🗼🎌 https://t.co/E1uPJa5tIy

我们正在东京招聘软件工程师，以帮助我们扩展Sakana AI的研发。如果你有兴趣构建推动自动化科学发现边界所需的数据管道和全栈基础设施，我们很乐意听取你的意见。🗼🎌 https://t.co/E1uPJa5tIy

To score 100% on a game, you just need to beat the median action efficiency of an unfiltered pool of random people. Easy if you're a bit smarter than average.

要在游戏中获得100%的分数，你只需要击败未经过滤的随机人群的中位数行动效率。如果你比平均水平聪明一点，这很容易。

ARC-AGI-3 has the lowest human bar of any AI benchmark out there. Almost all benchmarks require specialized knowledge that make them inaccessible to 99%+ of humans (like, say SWE-Bench). ARC-AGI-3 is feasible by regular people.

ARC-AGI-3是所有AI基准测试中人类门槛最低的。几乎所有的基准测试都需要专门的知识，使99%以上的人类无法访问（比如SWE-Bench）。ARC-AGI-3是普通人可以实现的。

Great to see our collaboration w/ @BostonDynamics unlocking new capabilities! Gemini Robotics-ER 1.6 enables robots like Spot to read complex industrial gauges autonomously. Exciting step toward robots that can understand & operate usefully in the physical world

很高兴看到我们与@BostonDynamics的合作解锁了新能力！Gemini Robotics-ER 1.6使Spot这样的机器人能够自主读取复杂的工业仪表。向能够在物理世界中理解并有效操作的机器人迈出的激动人心的一步

Is there somewhere a collection of the best agent/coding harnesses for each models, especially open-source and local ones? In my opinion, the biggest reason why people are struggling with open/local models these days is that the agent/coding harnesses in most open agent are not designed for them and expect it to magically work when they switch models from the default.

有没有哪里收集了每个模型的最佳智能体/编码工具包，特别是开源和本地模型的？ 在我看来，人们现在在开放/本地模型上挣扎的最大原因是，大多数开放智能体中的智能体/编码工具包不是为它们设计的，并且期望当他们从默认模型切换模型时它能神奇地工作。

New in Word: Copilot now tracks changes, leaves comments, and more, working more like a coworker right inside your document, grounded in all your enterprise context with Work IQ. https://t.co/x7KAZsk3WR

Word新功能：Copilot现在可以跟踪更改、留下评论等等，在你的文档中更像一个同事一样工作，并以Work IQ为基础，结合了你所有的企业上下文。 https://t.co/x7KAZsk3WR

I used GitHub Copilot CLI to connect the Logitech MX Creative Console to my Elgato key lights! (not sponsored, just a really fun little project) https://t.co/qF7xCbiHq2

我使用GitHub Copilot CLI将Logitech MX Creative Console连接到我的Elgato补光灯！ （非赞助，只是一个非常有趣的小项目） https://t.co/qF7xCbiHq2

Incredible technical depth in this thread, kudos to WL and Spark for going into such detail.

这个线程中的技术深度令人难以置信，向WL和Spark致敬，他们如此详细地进行了说明。

More 3D algorithm explainer videos should include the explanatory text as a 3D composited asset

更多的3D算法解释视频应该将解释性文本作为3D合成资产包含在内

I'm excited about voice as a UI layer for existing visual applications — where speech and screen update together. This goes well beyond voice-only use cases like call center automation. The barrier has been a hard technical tradeoff: low-latency voice models lack reliability, while agentic pipelines (speech-to-text → LLM → text-to-speech) are intelligent but too slow for conversation. Ashwyn Sharma and team at Vocal Bridge (an AI Fund portfolio company) address this with a dual-agent architecture: a foreground agent for real-time conversation, a background agent for reasoning, guardrails, and tool calls. I used Vocal Bridge to add voice to a math-quiz app I'd built for my daughter; this took less than an hour with Claude Code. She speaks her answers, the app responds verbally and updates the questions and animations on screen. Only a tiny fraction of developers have ever built a voice app. If you'd like to try building one, check out Vocal Bridge for free: https://t.co/nGrFznAMLh

我对语音作为现有视觉应用程序的UI层感到兴奋——语音和屏幕同时更新。这远远超出了呼叫中心自动化等仅语音用例。 障碍一直是一个艰难的技术权衡：低延迟语音模型缺乏可靠性，而智能体工作流（语音转文本→LLM→文本转语音）虽然智能，但对于对话来说太慢了。Vocal Bridge（AI Fund投资组合公司）的Ashwyn Sharma和团队通过双智能体架构解决了这个问题：一个用于实时对话的前台智能体，一个用于推理、护栏和工具调用的后台智能体。 我使用Vocal Bridge为我女儿构建的数学测验应用程序添加了语音；这用Claude Code不到一个小时就完成了。她说出答案，应用程序口头回应并更新屏幕上的问题和动画。 只有极少数开发者构建过语音应用程序。如果你想尝试构建一个，免费查看Vocal Bridge：https://t.co/nGrFznAMLh

And you can try it now on MAI Playground too. Know some of you have hit regional/country restrictions - the team is working hard to bring Playground to more areas. Stay tuned! https://t.co/o8pibirR40

你现在也可以在MAI Playground上尝试它。知道你们中的一些人遇到了区域/国家限制——团队正在努力将Playground带到更多地区。敬请期待！ https://t.co/o8pibirR40

Meet MAI-Image-2-Efficient. Production-ready quality, 22% faster, and 4x more efficient than MAI-Image-2. Priced almost 41% lower too. Plus 40% average lower latency than other leading models. Live now in Microsoft Foundry + MAI Playground. https://t.co/1Z0IKfytwV https://t.co/UKLm5FQzxx

认识MAI-Image-2-Efficient。 生产就绪质量，比MAI-Image-2快22%，效率高4倍。价格也低了近41%。 加上比其他领先模型平均低40%的延迟。 现已在Microsoft Foundry + MAI Playground上线。 https://t.co/1Z0IKfytwV https://t.co/UKLm5FQzxx

Small businesses are the engine of the American economy. Our 2025 US Economic Impact Report shows how Google is helping, from connecting 19.5M businesses with customers to training 350,000+ owners in digital skills. But the true impact is in the stories behind these numbers, like Atlas Automotive Repair in Oklahoma using Gemini to prep customer reports or The Boardwalk Cleaning Co. in Texas using NotebookLM as an internal knowledge base. Google is helping businesses grow in every state: https://t.co/9rX5U8q9XU

小企业是美国经济的引擎。我们的2025年美国经济影响报告显示了Google是如何提供帮助的，从将1950万家企业与客户联系起来，到培训35万多名所有者掌握数字技能。 但真正的冲击力在于这些数字背后的故事，比如俄克拉荷马州的Atlas Automotive Repair使用Gemini准备客户报告，或者德克萨斯州的The Boardwalk Cleaning Co.使用NotebookLM作为内部知识库。 Google正在帮助每个州的企业成长：https://t.co/9rX5U8q9XU

The role of memorization and knowledge is to cache & reuse past cognitive work. It should be leveraged as a way to speed up cognition, not as a *replacement* for cognition.

记忆和知识的作用是缓存和重用过去的认知工作。它应该被用作加速认知的一种方式，而不是作为认知的“替代品”。

Simply retrieving a reasoning trace looks a lot like human reasoning, until it's time to navigate uncharted territory. If you memorized all reasoning traces of humans from 10,000 BC, you could automate their lives but you could not invent modern civilization.

仅仅检索推理轨迹看起来很像人类推理，直到需要导航未知领域时。如果你记住了公元前10,000年人类的所有推理轨迹，你可以自动化他们的生活，但你无法发明现代文明。

As AI agents accelerate coding, what is the future of software engineering? Some trends are clear, such as the Product Management Bottleneck, referring to the idea that we are more constrained by deciding what to build rather than the actual building. But many implications, like AI’s impact on the job market, how software teams will be organized, and more, are still being sorted out. The theme of our AI Developer Conference on April 28-29 in San Francisco is The Future of Software Engineering. I look forward to speaking about this topic there, hearing from other speakers on this theme, and chatting with attendees about it. We’re shaping the future, and I hope you will join me there! It is currently trendy in some technology and policy circles to forecast massive job losses due to AI. Even if they have not yet materialized, these losses certainly must be just over the horizon! I have a contrarian view that the AI jobpocalypse — the notion that AI will lead to massive unemployment, perhaps even rioting in the streets — won’t be nearly as bad as dire forecasts by pundits, especially pundits who are trying to paint a picture of how powerful their AI technology is. Among professions, AI is accelerating software engineering most, given the rise of coding agents. According to a new report by Citadel Research, software engineering job postings are rising rapidly. So if software engineering is a harbinger of the impact AI will have on other professions, this expansion of software engineering jobs is encouraging. Yes, fresh college graduates are having a hard time finding jobs. And yes, there have been layoffs that CEOs have attributed to AI, even if a large fraction of this was “AI washing,” where businesses choose to attribute layoffs to AI, even though AI has not changed their internal operations much yet. And yes, there is a subset of job roles, such as call center operator, that are more heavily impacted. Many people are feeling significant job insecurity, and I feel for everyone struggling with employment, whether or not the cause is AI-related. And many other factors, such as over-hiring during the pandemic and high interest rates, have contributed to the slowdown in the labor market, and the notion that AI is leading to unemployment is oversimplified. In software engineering, I see a lot of exciting work ahead to adapt our workflows. It is already clear that: (i) As AI makes coding easier, a lot more people will be doing it. (ii) Writing code by hand and even reading (generated) code is not that important, because we can ask an LLM about the code and operate at a higher level than the raw syntax (although how high we can or should go is rapidly changing). (iii) There will be a lot more custom applications, because now it’s economical to write software for smaller and smaller audiences. (iv) Deciding what to build, more than the actual building, is becoming a bottleneck. (v) The cost of paying down technical debt is decreasing (since AI can refactor for you). At the same time, there are also a lot of open questions for our profession, such as: - In the future, what will be the key skills of a senior software engineer? And for junior levels, what should be the new Computer Science curriculum? - If everyone can build features, what skills, strategies, or resources create competitive advantage for individuals and for businesses? - What are the new building blocks (libraries, SDKs, etc.) of software? How do we organize coding agents to create software? - What should a software team look like? For example, how many engineers, product managers, designers, and so on. What tooling do we need to manage their workflow? - How do AI agents change the workflow of machine learning engineers and data scientists? For example, how can we use agents to accelerate exploring data, identifying hypotheses, and testing them? I’m excited to explore these and other questions about the future of software engineering at AI Dev. I expect this to be an exciting event. Please join us! [Original text: The Batch newsletter.] https://t.co/i4bQevDG4i

随着AI智能体加速编码，软件工程的未来是什么？一些趋势很明确，例如产品管理瓶颈，指的是我们更多地受限于决定构建什么，而不是实际构建。但许多影响，如AI对就业市场的影响、软件团队将如何组织等等，仍在整理中。 我们4月28-29日在旧金山举行的AI开发者大会的主题是软件工程的未来。我期待在那里谈论这个主题，听取其他演讲者关于这个主题的见解，并与与会者聊天。我们正在塑造未来，我希望你能加入我！ 在一些技术和政策圈子里，预测AI会导致大规模失业是很时髦的。即使它们还没有实现，这些损失肯定就在地平线上！我持相反观点，认为AI就业末日——即AI将导致大规模失业，甚至街头暴乱的观念——不会像专家预测的那样糟糕，特别是那些试图描绘他们的AI技术有多强大的专家。 在各行各业中，鉴于编码智能体的兴起，AI正在加速软件工程。根据Citadel Research的一份新报告，软件工程职位发布正在迅速增加。因此，如果软件工程是AI将对其他职业产生影响的先兆，那么软件工程职位的这种扩张是令人鼓舞的。 是的，刚毕业的大学生很难找到工作。是的，CEO们将裁员归咎于AI，即使其中很大一部分是“AI洗白”，即企业选择将裁员归咎于AI，尽管AI还没有改变他们的内部运营。是的，有一部分工作角色，如呼叫中心操作员，受到了更严重的影响。许多人感到严重的工作不安全感，我同情所有在就业方面挣扎的人，无论原因是否与AI有关。而且许多其他因素，如疫情期间的过度招聘和高利率，都导致了劳动力市场的放缓，认为AI导致失业的观点过于简单化了。 在软件工程方面，我看到了未来适应我们的工作流有很多令人兴奋的工作。已经很清楚的是：(i) 随着AI使编码更容易，更多的人会这样做。(ii) 手写代码甚至阅读（生成的）代码并不那么重要，因为我们可以询问LLM关于代码的情况，并在比原始语法更高的水平上操作（尽管我们可以或应该走多高正在迅速变化）。(iii) 将会有更多的定制应用程序，因为现在为越来越小的受众编写软件是经济的。(iv) 决定构建什么，比实际构建更成为瓶颈。(v) 偿还技术债务的成本正在降低（因为AI可以为你重构）。 同时，我们的职业也存在许多悬而未决的问题，例如： - 未来，高级软件工程师的关键技能是什么？对于初级水平，新的计算机科学课程应该是什么？ - 如果每个人都能构建功能，什么技能、策略或资源能为个人和企业创造竞争优势？ - 软件的新构建块（库、SDK等）是什么？我们如何组织编码智能体来创建软件？ - 软件团队应该是什么样子？例如，多少工程师、产品经理、设计师等等。我们需要什么工具来管理他们的工作流？ - AI智能体如何改变机器学习工程师和数据科学家的工作流？例如，我们如何使用智能体来加速探索数据、识别假设并测试它们？ 我很高兴在AI Dev探索这些以及关于软件工程未来的其他问题。我预计这将是一次激动人心的活动。请加入我们！ [原文：The Batch通讯。] https://t.co/i4bQevDG4i

i'm sure we all have our little coding 'hacks,' here are my top 5, please share yours (or help me improve mine! 🙏): 1. Plan -> Deepen Plan -> Then let Codex review the plan, then hand it back to Claude Code 2. Let Co-Work read the plan and build you a PDF of the plan in plain english along with flowcharts (vs just "go to work!"), this is a great for overall logic agreement 3. If I'm unsure of a stack or an algorithm choice (e.g. best algo for clustering objects with vector embeddings), give it to the beast models and let them deep research it for 20 mins 4. If have something big to tackle, always quit and restart Claude Code 5. On big PRs, I always let CC, Codex, and @greptile view it (at the same time), never fails to find some P1s

我相信我们都有自己的小编码“黑客技巧”，这是我的前5名，请分享你的（或帮助我改进我的！🙏）： 1. 计划 -> 深化计划 -> 然后让Codex审查计划，然后交回给Claude Code 2. 让Co-Work阅读计划并为你构建一个纯英文的计划PDF以及流程图（而不是仅仅“去工作！”），这对整体逻辑达成一致非常棒 3. 如果我不确定技术栈或算法选择（例如用于向量嵌入聚类对象的最佳算法），把它交给野兽模型，让它们深度研究20分钟 4. 如果有大任务要处理，总是退出并重启Claude Code 5. 在大的PR上，我总是让CC、Codex和@greptile同时查看，从不失败，总能找到一些P1问题

My debut on the @lennysan podcast: Building a successful startup in the age of AI. https://t.co/8TdALtH9kf

我在@lennysan播客上的首秀：在AI时代构建成功的初创公司。 https://t.co/8TdALtH9kf

If anyone else is also playing Linear Algebra Code Golf, Gemini taught me that `einsum('...ii', x)` is a tidy way to compute the traces of a batch of matrices.

如果还有其他人也在玩线性代数代码高尔夫，Gemini教我`einsum('...ii', x)`是一种计算一批矩阵迹的整洁方法。

Good design is the art of packing 1,000 "hows" into a single "what". Good design is compression: making the numerator trend towards infinity while the denominator stays at 1.

好的设计是将1000个“如何做”打包成一个“什么”的艺术。好的设计是压缩：使分子趋向于无穷大，而分母保持为1。