现象与趋势

Codex now has more than 5M weekly active users. But the bigger story is what people are using it for: not just writing code, but getting more work done across research, analysis, content, and operations. Our new report on how Codex is becoming a productivity tool for knowledge work: openai.com/index/codex-for-k…

So I spent some time studying the new Twitter/X algorithm today since the latest version was published about a week ago on Github (github.com/xai-org/x-algorit…). My goal was to answer why so many people have seemingly seen such a dramatic drop in their posts' reach. The first answer, which is actually somewhat unrelated to the ranking algorithm on Github, is the auto-translate feature, rolled out worldwide on April 7, 2026 (nitter.net/nikitabier/status/2041…). Before that date, if you wrote in English about, say, the Trump-Xi Beijing summit, you were competing for attention with maybe 5,000 other English-language accounts writing on geopolitics. After that date, your post is competing for attention with other posts on the same topic IN EVERY LANGUAGE ON EARTH. For some topics that do command global attention like geopolitics, that's a very brutal multiplier: you used to be one of 5,000, you're suddenly one of 50,000 (something of that order): MUCH more difficult to stand out. Secondly, the number of followers you have matters far less than it used to: each post now has to earn its audience reader by reader, on the predicted engagement of the post, and how its topic matches what each reader has recently been engaging with. Here is how the algorithm works, in simple terms: when you, as a reader, open your feed, the algorithm doesn't load "posts from accounts you follow." Instead it runs a 2-stage prediction of what posts you're likely to engage with in that very moment. The first stage is the retrieval stage. The system narrows billions of posts on X/Twitter that day down to roughly 1,500 candidates by matching the semantic content of each post - what it's about - against what you as a reader have recently engaged with. Some candidate posts come from accounts you follow; others are pulled from across the platform by pure topic similarity to your recent interests. You can test this retrieval stage easily: start disproportionally engaging with - say - Brad Pitt videos and you'll bit by bit see your timeline flooded with Brad Pitt content, most of it from accounts you've never followed and never heard of. Then there's the ranking stage. Each of these candidate posts for your feed is fed through a Grok-based model that tries to understand if you'll engage with the post. It looks at 15 engagement metrics: 1) P(favorite) — the reader likes the post 2) P(reply) — the reader replies to it 3) P(repost) — the reader reposts it 4) P(quote) — the reader quote-tweets it 5) P(click) — the reader clicks a link in it 6) P(profile_click) — the reader taps through to your profile 7) P(video_view) — the reader watches the video 8) P(photo_expand) — the reader expands an image 9) P(share) — the reader shares it (DM, off-platform, etc.) 10) P(dwell) — the reader stops scrolling and lingers on the post 11) P(follow_author) — the reader follows you after seeing it 12) P(not_interested) — the reader marks "not interested" 13) P(block_author) — the reader blocks you 14) P(mute_author) — the reader mutes you 15) P(report) — the reader reports the post Fifteen predicted actions, each multiplied by a weight, summed: that sum is the score that determines in which priority a post will be seen among other candidates. Please note that posting something with a video or an image can give your post an advantage as 2 actions are specifically for these: video_view and photo_expand. No video or photo and you don't get a score for these. Also, naturally, having a video maximizes the chance that a user will "dwell" on your post to watch it. Also note that 4 of these actions carry negative weights (not_interested, block_author, mute_author and report): meaning that if the model expects a post to generate a lot of negativity, it'll get de-boosted quite dramatically. But note, first and foremost, what's NOT in there: none of the things that, naively, one might think a serious information platform would weigh. There is no P(this post is true and well-sourced). No P(the author actually knows what they're talking about). No P(this person has spent a decade building a body of work that has held up). No P(this account has earned the right to be taken seriously on this topic). No P(the author has a large following from credible people). The model does not seem to care - at all - about any of that. Every post starts from zero. You could have ten years of rigorous, well-sourced analysis behind you - or you could be just an uneducated rando who registered yesterday. To this algorithm, you're both just a bag of engagement probabilities. Now, sure, to be fair, there is a "brand" effect that's not covered by the algorithm: someone who has in fact built a brand will naturally have better engagement metrics because people recognize their account. But that's an indirect, second-order effect. And crucially, it's legacy: those "brands" were built under earlier versions of the algorithm that gave followers and reputation more weight. Lastly, several other features of the new algorithm compound the dilution, none of them visible from outside but all consequential. The May 15 update added an "impression bloom filter," tightening the rule that once a reader has been served a post, the system won't serve it to them again. Before, a strong post could marinate in someone's feed across multiple refreshes and accumulate engagement on the second or third pass. Now it basically gets one shot. Also, your own posts compete with each other. An "Author Diversity Scorer" inside the ranking stage attenuates the score of every subsequent post of yours that ends up in a reader's candidate pool. In plain terms: if multiple of your posts land in a reader's candidate pool, the system shows one at full strength and dampens the others. So don't post several times consecutively on the same topic. And, last but not least, another huge impact on reach is that, in the old algorithm, when someone reposted or quote-tweeted you, your post was broadcast to their followers' timelines - a repost from an account with 100,000 followers was a huge boost. In the new algorithm, that mechanism is vastly demoted: reposts - like every post - need to go through the retrieval and ranking stage mentioned above, so a repost from a big account is a long way from the boost it used to be. This is especially brutal for low-effort quote tweets, which used to function as cheap amplification: now they often can't even clear the retrieval stage - they simply don't contain enough novel semantic content for the system to match them to anyone's interests. So, putting it all together, the reach collapse comes from many forces stacking at once: - Auto-translate makes your posts compete for attention against an order of magnitude more content - The retrieval stage matches posts by topic, not by who follows you - The ranking stage scores purely on predicted engagement with no weight for credibility, expertise, or track record - The bloom filter narrows every post's window to one strong shot - The diversity scorer penalizes prolific posting - Reposts no longer carry much distribution power Each of these alone would dent your reach. Combined, they amount to a complete reset: your audience that you built painstakingly over years basically doesn't matter much anymore, and it's much - much - harder to stand out even if you're a big account. People structurally rewarded by this algorithm are folks who: - Post visually (videos/images) - Post on globally popular topics because they clear the retrieval stage easily - Provoke strong emotional reactions - likes, replies, reposts - Don't care about accuracy or seriousness because the algorithm doesn't measure it - Don't care about their existing audience because every post is judged in isolation anyway In short this new algorithm, like so many on social media, is all about maximizing whether people will engage with something - not about whether they should.

So I spent some time studying the new Twitter/X algorithm today since the latest version was published about a week ago on Github (github.com/xai-org/x-algorit…). My goal was to answer why so many people have seemingly seen such a dramatic drop in their posts' reach. The first answer, which is actually somewhat unrelated to the ranking algorithm on Github, is the auto-translate feature, rolled out worldwide on April 7, 2026 (nitter.net/nikitabier/status/2041…). Before that date, if you wrote in English about, say, the Trump-Xi Beijing summit, you were competing for attention with maybe 5,000 other English-language accounts writing on geopolitics. After that date, your post is competing for attention with other posts on the same topic IN EVERY LANGUAGE ON EARTH. For some topics that do command global attention like geopolitics, that's a very brutal multiplier: you used to be one of 5,000, you're suddenly one of 50,000 (something of that order): MUCH more difficult to stand out. Secondly, the number of followers you have matters far less than it used to: each post now has to earn its audience reader by reader, on the predicted engagement of the post, and how its topic matches what each reader has recently been engaging with. Here is how the algorithm works, in simple terms: when you, as a reader, open your feed, the algorithm doesn't load "posts from accounts you follow." Instead it runs a 2-stage prediction of what posts you're likely to engage with in that very moment. The first stage is the retrieval stage. The system narrows billions of posts on X/Twitter that day down to roughly 1,500 candidates by matching the semantic content of each post - what it's about - against what you as a reader have recently engaged with. Some candidate posts come from accounts you follow; others are pulled from across the platform by pure topic similarity to your recent interests. You can test this retrieval stage easily: start disproportionally engaging with - say - Brad Pitt videos and you'll bit by bit see your timeline flooded with Brad Pitt content, most of it from accounts you've never followed and never heard of. Then there's the ranking stage. Each of these candidate posts for your feed is fed through a Grok-based model that tries to understand if you'll engage with the post. It looks at 15 engagement metrics: 1) P(favorite) — the reader likes the post 2) P(reply) — the reader replies to it 3) P(repost) — the reader reposts it 4) P(quote) — the reader quote-tweets it 5) P(click) — the reader clicks a link in it 6) P(profile_click) — the reader taps through to your profile 7) P(video_view) — the reader watches the video 8) P(photo_expand) — the reader expands an image 9) P(share) — the reader shares it (DM, off-platform, etc.) 10) P(dwell) — the reader stops scrolling and lingers on the post 11) P(follow_author) — the reader follows you after seeing it 12) P(not_interested) — the reader marks "not interested" 13) P(block_author) — the reader blocks you 14) P(mute_author) — the reader mutes you 15) P(report) — the reader reports the post Fifteen predicted actions, each multiplied by a weight, summed: that sum is the score that determines in which priority a post will be seen among other candidates. Please note that posting something with a video or an image can give your post an advantage as 2 actions are specifically for these: video_view and photo_expand. No video or photo and you don't get a score for these. Also, naturally, having a video maximizes the chance that a user will "dwell" on your post to watch it. Also note that 4 of these actions carry negative weights (not_interested, block_author, mute_author and report): meaning that if the model expects a post to generate a lot of negativity, it'll get de-boosted quite dramatically. But note, first and foremost, what's NOT in there: none of the things that, naively, one might think a serious information platform would weigh. There is no P(this post is true and well-sourced). No P(the author actually knows what they're talking about). No P(this person has spent a decade building a body of work that has held up). No P(this account has earned the right to be taken seriously on this topic). No P(the author has a large following from credible people). The model does not seem to care - at all - about any of that. Every post starts from zero. You could have ten years of rigorous, well-sourced analysis behind you - or you could be just an uneducated rando who registered yesterday. To this algorithm, you're both just a bag of engagement probabilities. Now, sure, to be fair, there is a "brand" effect that's not covered by the algorithm: someone who has in fact built a brand will naturally have better engagement metrics because people recognize their account. But that's an indirect, second-order effect. And crucially, it's legacy: those "brands" were built under earlier versions of the algorithm that gave followers and reputation more weight. Lastly, several other features of the new algorithm compound the dilution, none of them visible from outside but all consequential. The May 15 update added an "impression bloom filter," tightening the rule that once a reader has been served a post, the system won't serve it to them again. Before, a strong post could marinate in someone's feed across multiple refreshes and accumulate engagement on the second or third pass. Now it basically gets one shot. Also, your own posts compete with each other. An "Author Diversity Scorer" inside the ranking stage attenuates the score of every subsequent post of yours that ends up in a reader's candidate pool. In plain terms: if multiple of your posts land in a reader's candidate pool, the system shows one at full strength and dampens the others. So don't post several times consecutively on the same topic. And, last but not least, another huge impact on reach is that, in the old algorithm, when someone reposted or quote-tweeted you, your post was broadcast to their followers' timelines - a repost from an account with 100,000 followers was a huge boost. In the new algorithm, that mechanism is vastly demoted: reposts - like every post - need to go through the retrieval and ranking stage mentioned above, so a repost from a big account is a long way from the boost it used to be. This is especially brutal for low-effort quote tweets, which used to function as cheap amplification: now they often can't even clear the retrieval stage - they simply don't contain enough novel semantic content for the system to match them to anyone's interests. So, putting it all together, the reach collapse comes from many forces stacking at once: - Auto-translate makes your posts compete for attention against an order of magnitude more content - The retrieval stage matches posts by topic, not by who follows you - The ranking stage scores purely on predicted engagement with no weight for credibility, expertise, or track record - The bloom filter narrows every post's window to one strong shot - The diversity scorer penalizes prolific posting - Reposts no longer carry much distribution power Each of these alone would dent your reach. Combined, they amount to a complete reset: your audience that you built painstakingly over years basically doesn't matter much anymore, and it's much - much - harder to stand out even if you're a big account. People structurally rewarded by this algorithm are folks who: - Post visually (videos/images) - Post on globally popular topics because they clear the retrieval stage easily - Provoke strong emotional reactions - likes, replies, reposts - Don't care about accuracy or seriousness because the algorithm doesn't measure it - Don't care about their existing audience because every post is judged in isolation anyway In short this new algorithm, like so many on social media, is all about maximizing whether people will engage with something - not about whether they should.

We are making our discount permanent! 🎉 Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀

Cursor's new Composer 2.5 takes third on the Artificial Analysis Coding Agent Index and is ~10-60x lower cost than the higher-effort Opus 4.7 and GPT-5.5 variants above it. This release puts Composer among the leading coding agent models, something that wasn’t clear for past releases @cursor_ai has released Composer 2.5, the latest model in its Composer line. Composer 2.5 scored 62 on our Coding Agent Index, a 14 point gain over Composer 2 (48). This puts it in third place of our tested agents, behind only Claude Opus 4.7 (max) in Claude Code (66) and GPT-5.5 (xhigh reasoning) in Codex (65). These cost $4.10 and $4.82 per task respectively, ~10x the cost of Composer 2.5 Fast ($0.44) and ~60x the cost of Composer 2.5 standard ($0.07). Key results for Composer 2.5 in Cursor CLI: ➤ Cost-quality Pareto frontier: At $0.07 (standard) and $0.44 (Fast) per task, Composer 2.5 is cheaper than every other agent scoring above 60 on the Index. Medium-effort peers cost $1.24–$2.21 per task; higher-effort variants land 3-4 points above at $4.10–$4.82 ➤ Per-benchmark gains vs Composer 2: +35 points on SWE-Bench-Pro-Hard-AA (12% → 47%), +2 points on Terminal-Bench v2 (64% → 66%), and +3 points on SWE-Atlas-QnA (69% → 72%). At 47%, Composer 2.5's score on SWE-Bench-Pro-Hard-AA is comparable to Claude Opus 4.7 (max) in Claude Code ➤ Among the fastest coding agents: Composer 2.5 Fast runs at an average wall time of 6.7 minutes per task, the third-fastest agent on the Artificial Analysis Coding Agent Index, behind only Claude Opus 4.7 (medium) in Claude Code (5.8m) and GPT-5.5 (medium) in Cursor CLI (6.2m) ➤ Fast mode enables better responsiveness at 6x pricing: Fast runs 30% faster than standard Composer 2.5, but is ~6x the cost per task ($0.44 vs $0.07). Token pricing is 6x higher for Fast: $3.00/$15.00 vs $0.50/$2.50 per million input/output tokens Model details: ➤ Base model: Continued training on @Kimi_Moonshot's open weights Kimi K2.5 as with Composer 2, with Cursor reporting ~85% of total compute from its own additional training and reinforcement learning ➤ Pricing: $0.50/$2.50 per million input/output tokens for the standard variant; $3.00/$15.00 for the Fast variant (the default in Cursor) ➤ Available exclusively in Cursor: both Cursor IDE and Cursor CLI, an externally accessible API is not available Congratulations @cursor_ai and @mntruell on the impressive release!

LEAKED AUDIO: In an all-hands meeting on April 30, Mark Zuckerberg tells employees that he's training AI on them ahead of mass layoffs. "The AI models learn from watching really smart people do things... The average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks. So if we're trying to teach the models coding, for example, then having people internally build tools or solve tasks that help teach the model how to code, we think is going to dramatically increase our model's coding ability faster than what others in the industry have the capability to do, who don't have thousands and thousands of extremely strong engineers at their company." Video

兄弟们！ 今天已经可以在ZenMux上免费体验Gemini 3.5 Flash 了！ 我第一时间用它跑了那个经典的「AI模型递归二叉树生长测试」. 同一个 Prompt ，不同模型画出的树形态完全不一样。（见视频-Prompt见评论区） Gemini 3.5 Flash 从输入提示词到生成完整 HTML 动画网页（树干慢慢长出、分支递归展开、最后随风摇摆），全程只用了 77.56 秒！ 整体效果非常惊艳：树形态自然优雅、生长动画丝滑、视频和内容呈现都顶级！ 熟悉的老朋友都知道，ZenMux 每次新模型都是 ZeroDelay 首发. Google I/O 2026 今天刚发布，现在立刻就能通过 API 调用！ 还有免费额度可以白嫖～ 速度是真的没话说，还完美保留了旗舰级模型的能力。 专为 Agent 设计，在 MCP Atlas、Toolathlon、Finance Agent 等多项榜单直接拿下第一！ 多模态理解也极强：MMMU-Pro 83.6%、CharXiv Reasoning 84.2%，全面超越上一代 Gemini 3.1 Pro。 完全兼容主流 API 格式，无需改动现有工具链。 支持按量计费 + Builder 套餐。 👇 直接体验 正式版 → zenmux.ai/google/gemini-3.5-… 免费试用 → zenmux.ai/google/gemini-3.5-… Video

Just off stage at #GoogleIO, some highlights from this morning 🧵 Gemini 3.5 Flash is available today for everyone in @antigravity and across our products and APIs. Compared to 3.1 Pro, 3.5 Flash is better across almost all benchmarks with huge progress in coding. It’s also comparable to the best models but very fast (4x faster tokens/ second than other frontier models). And when looking at the intelligence versus output speed, it’s in a league of its own in the top right quadrant.

Anthropic is acquiring @stainlessapi, an SDK and MCP server platform that has powered every Anthropic SDK since the earliest days of our API. Read more: anthropic.com/news/anthropic…

飞书 CLI 牛皮啊，发布一个月多点就达到 10000 Star 了！ 说明用户和市场相当认可这个动作 最近我们可以发现，越来越多的传统办公产品开始发布 CLI 和 Agent。 AI 时代的 SaaS 软件可能得换个做法了：UI 只是最基本的，接下来还要竞争对 Agent 的适配程度以及覆盖率。在这块，我觉得飞书走得相当靠前。 作为一个 IM 软件，飞书在 AI 时代去做这种开放自己所有能力的 CLI 工具，其实是一种非常不传统互联网的尝试。 这对于之前的互联网产品逻辑和经验来说，是一个非常不应该做的决定。 因为他们这个 CLI 几乎可以控制飞书的所有能力：你可以完全不跟飞书的传统 UI 去交互。只跟 CLI 交互，也可以完成飞书上所有的工作。 传统的 IM 办公软件通常非常复杂，入门门槛相对较高。无论从产品逻辑、UI 设计还是交互设计的角度来看，都没有办法太好地消解这种复杂性。 但是 CLI 工具交付给 Agent 以后，就可以快速消解这种复杂性。用户只需要进行对话，这是非常本能的行为，不需要在繁杂的层级列表 UI 里去寻找功能入口。 我拉了一下数据，他们迭代效率也非常恐怖，它们是 3 月 28 号开源的，一个多月发了 32 个版本、385 个提交。 这说明飞书对这块是非常重视的，投入的人力和精力也非常大。 他们在 CLI 本身的设计上也考虑得非常多，下了很多功夫。主要分为三层： 面向日常任务的快捷命令、开放平台对应的标准 API、兜底的 API 调用。 因为人和 Agent 都不喜欢从 2500 个 API 里去寻找参数，但又需要把这些能力暴露出来，所以他们采用了这种分层的形式。 即使做了分层设计，CLI 本身的内容和 API 依然非常多。所以他们把 CLI 作为工具本身，同时做了很多 Skills 用来充当 CLI 的说明书。 Agent 可以分层、分类型地了解应该如何调用这些 CLI 及其命令。 此外，他们在对 Agent 友好的命令包装上做了很多工作，例如： (a) 内置了 Dry Run (b) 结构化输出 (c) 身份选择、权限检查与风险等级评估 (d) 允许 Agent 在发消息前预览请求 (e) 建立了输出格式的“契约”：将成功或失败的结果、原因以及风险提示都放在结构化数据里。 这样如果出错了，AI 可以非常清楚地进行调试和修改，而不是盲目猜测。 其实现在你如果要创业或者做自己的 Agent，就不需要非得写一个界面。 飞书 CLI 加上 Agent 框架可以完成所有的 Agent 产品常见的操作： 你的聊天界面就是你的 Agent 聊天界面； 你的数据库就是飞书多维表格和文档； 你的用户就是把你拉到组织里的群成员；