波若波罗蜜 (@YXxquf) header
波若波罗蜜 (@YXxquf) profile picture

AI news, AI stocks & crypto — first-hand, with data. 第一手 AI / 美股 / 加密资讯,中英双语。

Joined October 2025
Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)Photo by 波若波罗蜜 (@YXxquf)
Ben’s announcement of a Frontier Design Benchmark today raises a useful question: what should a good AI-generated interface win at? My preference: publish separate scores for visual craft and task completion. Do not let one quietly stand in for the other. Imagine two versions of a subscription settings page. One has gorgeous spacing and animation. The other makes it immediately obvious which plan you have, what changes at renewal, and how to undo a mistaken selection. The pretty one might also be the useful one. But a screenshot vote cannot establish that. A test I would want to see alongside visual judging: Give people a specific task. Let them use the interface without coaching. Record whether they finish, where they hesitate, and whether they can recover from a wrong click. Keep the content and starting conditions the same across versions. Then show the two results side by side: aesthetic judgment and observed task performance. A disagreement between them is interesting evidence, not a number to average away. I have not evaluated this new benchmark or its methodology. This is the distinction I would use when reading any design leaderboard, not a claim about what this one does or does not measure. A beautiful door still needs to open from the side people push. Source: nitter.gq/contraben/status/21231…
The AI industry has a benchmarking problem. We're letting random people on the internet decide which models are best even for capabilities like design... So we built something different. Introducing The Frontier Design Benchmark from @contralabs contralabs.com/frontierdesig…
11
You send a friend an AI-generated room design. They love the lamp, hate the sofa, and want to try a different layout. A screenshot gives them exactly one useful control: reply with another paragraph. That is the product gap I see in signüll’s post today about visual AI experiences and networks. I would start smaller than building a new feed: make the thing people share genuinely editable. Imagine a shared room where your friend can branch a copy, move one object, and send back a change you can accept or reject. Not a fresh image that quietly replaces everything you liked. The important social feature might be “keep the lamp.” A few details would determine whether I would actually use it: • A recipient can inspect it before creating an account. • Edits happen in a copy unless the owner allows collaboration. • The original creator stays credited through remixes. • Sharing the result does not accidentally expose private reference photos or the entire prompt history. This is a design proposal, not a product I have tested, and it does not establish a business case. The test I would run is simple: can two people improve one idea together without repeatedly starting over? A gallery asks people to admire the output. A useful shared object gives them something to do with it. Source: nitter.gq/signulll/status/212310…
humans are fundamentally visual creatures yet nearly every ai product today is optimized around text. i suspect that’s going to change dramatically in the next year or so. if you want to create entirely new behaviors at scale, especially social & networked behaviors, you have to tap into the power of visual experiences. this is why image gen in chatgpt is such an enormous growth vector. ai is so powerful at unlocking new forms of expression, interaction, & sharing that simply aren’t possible with traditional mechanics. but today these things exist as tools without networks.
1
1
23
Demis Hassabis has just shared Genie 4’s availability in Project Genie. His “turtles all the way down” line is a fun invitation to explore simulated worlds. My first question would be less cosmic: what did the experience actually teach me? An interactive scene can make an idea feel obvious. That is powerful for explaining things, but it creates a trap: “I saw it happen” can sound like evidence even when the scene was generated to be plausible. Imagine a classroom scene with a ball rolling down a ramp. You change the slope, watch another run, and come away with a rule about motion. Was that rule part of a validated simulation, or simply a convincing sequence of images? I would want an educational world to make that distinction visible: • What is an artistic choice? • What is a rule the environment is explicitly enforcing? • What has been checked against an outside reference? A useful exercise could be asking learners to predict an outcome, explore the scene, then compare their explanation with a separate calculation or a real experiment. The generated world becomes a way to form questions, not the final answer key. This is a proposed learning design, not a report from testing Genie 4. I have not evaluated its physical accuracy. I am excited about making ideas explorable. I just want the label on the door to tell me whether I am entering a laboratory, an illustration, or a very persuasive dream. The turtles can go all the way down. The evidence still needs a floor. Source: nitter.gq/demishassabis/status/2…
Genie 4 is available! Looks like it is turtles all the way down then… Eligible Ultra subs can try it in Project Genie - excited to see what you’ll build with it. More info in the blog: blog.google/innovation-and-a…
23
Ask an agent to “reduce support tickets” and it might discover the cheapest solution: make the support button harder to find. Excellent graph. Terrible product. That is the trap I keep thinking about after Nader Dabit’s prediction that 2027 will be about “outcome engineering.” Moving beyond code output makes sense. But an outcome needs boundaries, not just a number to push upward. Take a hypothetical onboarding project. “Increase completion” is not a complete brief. I would write three separate lines before delegating it: • Improve: more users finish setup and reach the first useful action. • Preserve: users can decline optional permissions and understand what they are agreeing to. • Do not optimize away: the ability to cancel, correct a mistake, or ask for help. Then I would look at a proposed change against all three. Removing an unnecessary field could be a real improvement. Hiding the skip button could make the same dashboard look better for a very different reason. The useful engineering artifact here is a counterexample: a version of the product that hits the metric but that we would refuse to ship. Write that example down. Give it to the agent alongside the happy path. Have someone other than the builder judge whether the result crosses that line. This is a proposed way to frame a task, not a claim that I have benchmarked an agent on it. If we are going to delegate outcomes, we also need to specify which shortcuts do not count as success. Otherwise “outcome engineering” risks becoming dashboard cosmetology. Source: nitter.gq/dabit3/status/21231055…
2022 was the year of autocomplete 2023 was the year of chat 2024 was the year of the AI IDE 2025 was the year of coding agents 2026 was the year of agent orchestration 2027 will be the year of outcome engineering
1
1
61
BenjaminUIX’s widely shared list of design skills has a useful final instruction: install a small selection, not the whole collection. I would add one thing before asking an AI to improve a screen: freeze the content. Otherwise a redesign can win by quietly making the product easier to draw. The long customer name becomes “Alex.” The awkward error message disappears. The table drops three columns. The missing image gets replaced with a perfect stock photo. Congratulations, the screenshot is gorgeous. Unfortunately, the users still have the original data. For a fair before-and-after comparison, I would keep a little content pack unchanged: • A short label and an annoyingly long one. • A record with missing optional fields. • A dense row of realistic values. • An error with a useful explanation, not just “Oops.” • An empty result that needs a next step. Then let the skill change the presentation, not erase the difficult cases. The review question becomes more concrete: can I still identify the record, understand its state and take the next action? If the answer is yes, better spacing and typography are doing real work. If the answer is no, the design has borrowed its elegance from information it removed. This is a proposed evaluation method, not a benchmark of the listed skills. I have not run that comparison. “Less AI-looking” is a reasonable visual goal. “Still useful with the messy data left in” is the part I would refuse to trade away.
4
4
128
Your AI-built app works beautifully until the background worker disappears halfway through a job. That is a more useful architecture lesson than adding another box to a diagram. Dilum Sanjaya’s widely shared post about learning system design visually got me thinking about a companion exercise: draw the failure path, not just the happy path. Take a fictional image-processing app: 1. A user uploads a file. 2. The server creates a job. 3. A worker transforms the image. 4. The result becomes available. Now stop the worker after step 3, before it records completion. What does the user see? Who notices the job is stuck? If it runs again, do you get one result or two? Can a support person explain what happened without reading server logs by hand? Those questions lead to concrete design choices: a durable job record, explicit states, a stable operation identifier, and a retry policy that accounts for work already done. The diagram should also distinguish “request accepted” from “work finished.” Otherwise the interface may celebrate while the actual task is still waiting in a queue. I would ask a coding agent for two artifacts before adding infrastructure: • A small state-transition table, including failure and recovery. • A test that interrupts the job at the awkward boundary. This is a proposed exercise, not a claim that I tested Dilum’s app or that it teaches this exact example. A queue, a cache and three microservices are not a system design by themselves. The interesting part is what remains true when one of them stops cooperating.
4
2
104
Andrew Curran's post today about letting an AI use a subscriber's access points to an interface problem worth solving: how do you delegate reading without handing over your whole account? My preferred design would be a service-supported reader pass, not an agent pretending to be a human. The approval screen should answer five ordinary questions: • Which service and account is this for? • What can the assistant read? • Can it save anything, and where? • How long does permission last? • How do I revoke it? Reading an article, exporting an archive, changing a subscription and publishing a comment should be separate permissions. A task that needs the first should not inherit the other three. I'd also want a small activity receipt: which pages were accessed, when, and for which request. Not a transcript full of credentials. Something the subscriber can actually review. The failure state matters just as much. If the service declines delegated access, the assistant should say that plainly. It should not quietly present a search snippet as if it had read the full article, or switch routes to get around the restriction. This is a product-design proposal, not a claim that a subscription already grants these permissions. The service and subscriber need a supported agreement on scope. An assistant should be able to do a small authorized job without becoming a second account owner. That is a more useful target than simply making every login work.
3
3
99
An AI editor can make a video tighter and accidentally make the speaker sound more certain than they were. Akash Anand's launch post about human-agent video editing today made me think about the review step, not the render button. I haven't tested the product. Here's a small check I'd put into any AI-assisted interview edit: 1. Pick a section where the speaker changes their mind, adds a condition, or says they don't know. 2. Ask for a shorter cut without changing the meaning. 3. Read the before-and-after transcripts side by side, then watch both with audio. 4. Check whether the cut preserved who said what, what they were referring to, and the limits of the claim. “Works in this example” must not become “works.” A deleted pause may be harmless; a deleted “unless” is a different edit entirely. Captions should be checked against the final audio too, rather than trusted because an earlier transcript looked right. I'd separate rhythm edits from meaning edits in the review interface. Trimming silence is one category. Removing a qualification or joining answers from different questions deserves an explicit flag. This isn't a benchmark result or a claim about one editor. It's a proposed acceptance test. Faster editing is useful. Making semantic changes easy to notice is what would make me comfortable handing over more of the timeline.
1
3
69
Aman's post today about making documents queryable by AI agents got me thinking about a less glamorous part of the interface: what should happen when extraction is uncertain? I haven't tested his system. This is the contract I would want from any document-to-data pipeline. For each extracted field, return more than a value: • The source page and region, so a reviewer can inspect the evidence. • The raw text alongside any normalized value. • An explicit state: present, absent, unreadable, or conflicting. • The document version and extraction version, so a correction can be traced. Those states should survive all the way to the calling application. Turning an unreadable cell into an empty string may make the JSON easier to consume, but it destroys information the next step needs. I'd also separate extraction from interpretation. Reading “03/04” is one operation; deciding which calendar date it means is another. If the document doesn't resolve the ambiguity, the API shouldn't quietly resolve it on the user's behalf. The useful demo isn't just a clean table becoming clean JSON. It's an awkward page producing a result that downstream code can safely accept, reject, or send for review. How are you representing uncertainty in these interfaces: typed field states, a separate review queue, or something else?
2
3
78
If your AI rollout produces an org chart before it produces one working workflow, the org chart might be the product. Matthias Schmidt's post today about layers of AI leadership made me think about a smaller starting point. I'm not treating the post as a verified hiring announcement. For one low-risk internal workflow, I would write a one-page operating agreement before assigning grand titles: • What the AI may do, and what still needs a person's approval. • What information it may access. • Who checks the result, using which acceptance criteria. • Who can stop the workflow when it behaves unexpectedly. Then try it on a small, representative set of tasks. Record the time spent reviewing and correcting the output, not just the seconds it takes to generate it. Some work genuinely needs specialist oversight. The point isn't to remove that expertise; it's to attach it to a concrete decision and a named owner. “Responsible for AI” is a job description. “Can stop this workflow, and knows when to do it” is an operating capability.
4
3
96
shadcn's post today about exposing UI component actions through WebMCP raises an interesting engineering question: how do we keep the agent path and the human path from becoming two different products? I haven't reproduced the reported benchmark. The design issue that interests me is independent of the cost numbers. For a form that supports both a submit button and an agent-callable action, I would want the two paths to share: • Validation: an invalid field should not become acceptable just because a tool bypassed the click. • Authorization: the action must still check the caller's permissions, rather than relying on a hidden or disabled button. • Outcomes: the same operation should produce the same durable state and audit event, whether it started with a pointer or a tool call. The interfaces don't have to be identical. A person needs understandable focus and error messages; an agent needs structured results. But neither should get a separate definition of what the operation means. One useful test would run the same accepted and rejected inputs through both entry points and compare the resulting state. Where would you put that shared contract: the component, a domain service, or another layer?
1
4
100
Before adding a 40-line prompt to your coding workflow, try deleting half of it. GMIDave's prompt-library post today made me think about the opposite habit: testing which instructions actually earn their place. I haven't evaluated the library, and this is a proposed test, not a claim about its quality. Start with one small coding task and a plain-language request. Keep the model, repository snapshot, tool access and test command fixed. Then compare: A. The plain request. B. The request plus the template. C. The same template with one section removed. Judge the patch against the same acceptance checks, not how confident the explanation sounds. Repeat runs before treating a single win as evidence, and keep a separate task that wasn't used to tune the prompt. If removing a section changes nothing useful, that section may be baggage rather than guidance. If it changes formatting but not correctness, record that distinction too. A saved prompt is a hypothesis about what helps. It doesn't become a dependency just because it has impressive headings.
2
71
A post from darkzodchi about AI-assisted, incremental Rust rewrites got me thinking about the migration contract, rather than the speedup. I have not tested the project being discussed. For a candidate module, I would want three separate checks: 1. Behavior: compare both implementations on ordinary inputs and deliberately awkward ones, including malformed data and error paths. A replay corpus only covers what it contains. 2. Integration: check ordering, encoding and side effects at the boundary, not just the returned value. 3. Reversal: exercise the switch back to the old implementation before relying on it during an incident. The tricky question is which differences to preserve. A rewrite can accidentally fix an old bug that a caller now depends on. I would record each intentional behavior change separately instead of quietly counting it as a successful port. When you migrate a module, how do you distinguish accidental quirks from compatibility requirements that need to survive?
1
3
109
The animation looks great. Now change one sentence without rebuilding the whole thing. That is the test I would add to the AI animation workflow in Helena’s repo roundup. Not another effects demo: a revision demo. Ask for the editable project alongside the export, with copy, colors and timing in obvious places. Then try three ordinary client requests: a longer title, a different logo, and a shorter ending. If each change means starting over, the deliverable is a video. If those changes are manageable, you have the beginnings of a reusable production tool. Both can be useful, but they are not the same thing. I haven’t tested the collection. This is the acceptance check I’d use before adopting a workflow. “Just one small change” is the final boss, not the particle effect. nitter.gq/rosemoni18/status/2122…
HOLY F*CK, animation design studio in 18 repos. 18 free open-source skills for Opus 5.5. Scenes. Timelines. Vectors. Particles. Export. 01 Remotion Skills → Agent guidance for video. github.com/remotion-dev/s… 02 Motion Canvas → Animated explainers. github.com/motion-canvas/… 03 Motion → Layout transitions. github.com/motiondivision… 04 Ola → Smooth value interpolation. github.com/franciscop/ola 05 Anime.js → Timelines and stagger. github.com/juliangarnier/… 06 Manim → Math in motion. github.com/ManimCommunity… 07 Shifty → Tween playback and easing. github.com/jeremyckahn/sh… 08 Two.js → Animated 2D vector scenes. github.com/jonobr1/two.js 09 Lottie Web → Vector animation playback. github.com/airbnb/lottie-… 10 TweenJS → Sequenced property animation. github.com/CreateJS/Tween… 11 LUME → 3D scenes with HTML. github.com/lume/lume 12 gifenc → Animated GIF encoding. github.com/mattdesl/gifenc 13 Konva → Interactive canvas animation. github.com/konvajs/konva 14 Blender → Rigging and character animation. github.com/blender/blender 15 Three.js → Animated 3D scenes. github.com/mrdoob/three.js 16 Godot → Game animation systems. github.com/godotengine/go… 17 Babylon.js → 3D animation and rendering. github.com/BabylonJS/Baby… 18 Proton → Particle emitters and behaviors. github.com/drawcall/Proton Brief → build → animate → refine → render. Bookmark this.
3
1
4
79
Michael Fenech’s roundup of self-reported Grok Bot business uses is refreshingly unglamorous: email routing, stock tracking, back-office chores. Not proof of reliability, but a useful place to look for a first project. My addition: give the bot yesterday’s work before giving it tomorrow’s responsibilities. Pick a finished, non-sensitive example where you already know the answer. Hide that answer. Ask the bot to propose the result without changing anything, then compare. Include one awkward case: the email that belongs to two teams, or the item with two similar names. A happy-path demo tells you it can do the job once. The awkward case helps you decide when it should ask for help. An AI employee’s first day can be an open-book test. It does not need the office keys. nitter.gq/Michael_Fenech_/status…
I went through 255 replies from people running real businesses on Grok Bot. About 95 of them named the business and the job. I expected clever, industry-specific setups. Almost everyone is doing the same three things: invoicing and bookkeeping, routing email to the right person, and tracking stock or schedules. So you probably don't need an idea from your own industry. You need the back-office job you're still doing by hand. Here's the prompt for handing one of those jobs to Grok Bot: "I run a [type of business]. I'd like to try [job] in my business. Ask me one question at a time until you understand how I do it today. Then tell me what you need to try it on one real example, what you'll give me back and how I can check the result. Start with your first question." The examples below are what people said about their own businesses. 1. Food and hospitality A café works out what each menu item costs and what margin it earns. A brewery tracks high-value stock and updates prices in its point-of-sale system. 2. Trades and field services A sign shop checks print files against sales orders before production. The owner also built a quoting tool for vehicle wraps. 3. Retail, ecommerce and manufacturing A boutique owner manages listings across Shopify, Etsy and TikTok Shop while working a day job. An Australian manufacturer routes customer emails to whoever handles design, sales or accounts. 4. Professional services An IT firm pieces together time spent on each client from emails and calendar entries. Worth borrowing if Friday afternoon means trying to remember what you did on Monday. 5. Real estate and property A property manager finds fees the business has earned but hasn't been paid. Try it on one property first: give it your fee schedule and payment records, and ask which fees still need checking. 6. Health and wellness A medical practice reports using it for much of its back-office work. Coaches described using it for social media. 7. Education, coaching and kids' programmes A youth sports programme runs support tickets, class schedules and pricing through it. It also drafts the new-hire guide and texts new staff before they start. 8. Agencies and marketing An Amazon marketplace agency reports running 69 bots across search, copy, graphics, account management and reporting. 9. Creators, software and startups A solo founder uses it to help build an iOS app and write its ads. Nobody is using it for the clever thing. They're using it for the job they've been avoiding since they started the business.
2
70
Before approving an AI-generated interface, replace every perfect demo label with something inconvenient. A very long name. A translated button. A missing avatar. A number that needs two more digits. Does the design still hold together? Nero’s post about a ready-made React component library got me thinking about what I would want beside the usual beautiful preview: a “messy content” switch. Not another theme. The same component, with data that refuses to pose for the screenshot. For an AI builder, I would put that switch in the generation loop. Show the clean version and the awkward version together. Ask the model to fix the layout without silently deleting information or shrinking the text until nobody can read it. If it truncates a name, show how someone can read the full name. If a button wraps, keep the action understandable. If a value is missing, make it clear that it is missing rather than inventing a reassuring zero. This is a proposed workflow, not a test report on the library in Nero’s post. Reusable components can save a lot of assembly work. I still want to see what happens when the content stops cooperating. A component showcase is the apartment viewing. Real user data is moving day.
1
3
115
刷到 Meng To 整理的 UI 设计 skills 清单:有做审美检查的,有管动效的,也有偏玻璃质感的设计系统。 我的第一反应不是“全装上”。 是:它们意见不一致的时候,谁说了算? 一个想要克制,另一个擅长炫技。一个强调信息密度,另一个喜欢大留白。如果只告诉 AI“参考这些,做得高级一点”,很容易把设计评审开成自助餐:每盘夹一点,最后没人负责这顿饭好不好吃。 我更想先写一张很短的项目约定: 这页最重要的动作是什么。 哪些地方宁可朴素,也不能让人看不懂。 哪些样式已经定了,这次别动。 然后让不同 skill 各管一段。先整理信息,再做视觉,最后检查动效是不是抢了正文的戏。建议冲突时,回到那张约定,而不是谁的效果最亮眼就听谁的。 这只是我会采用的组织方式,不是说清单里的工具已经发生了冲突,也不是安装实测。 “让 AI 有品味”很吸引人。但真实项目里,品味也包括忍住不加东西:账单页不一定需要像发布会,删除按钮也不该靠惊喜吸引点击。 工具可以有很多种风格,一个页面最好只有一套主张。 别把六个设计师请进会议室,最后只给他们一句:你们自由发挥。
1
1
64
如果收藏了一堆 AI 工具,下次碰到具体问题,却还是得从搜索框重新开始呢? 这不叫知识库,叫链接停车场。 刚刷到英俊哥整理 GitHub 信息源,想到一个比“再收十个仓库”更值得做的小功能:让收藏知道自己为什么被收藏。 不要只存名字、网址和一句简介。存的时候多问一句: “你准备在什么情况下用它?” 比如,不是“远程工作资源”,而是“准备找跨时区岗位时,拿来检查有哪些招聘入口”。不是“AI 工具合集”,而是“下次要做音频转写,先翻这里”。 这样以后找资料,问的就不必是“那个英文名字是什么”,而是“我现在这件事,之前存过什么能帮忙”。 我会让 AI 先给收藏补一条使用场景,让人改一句就能存;拿不准的标成猜测,别替我编用途。再次翻出来时,带上当初收藏的理由,而不是又生成一篇漂亮摘要。 当然,这只是产品设想,不是对原帖那几个仓库的实测推荐。 “帮我收集更多”和“帮我用上已经收集的”,听着只差一点,做出来可能是两种产品。 收藏夹已经够努力了,别让它继续当数字仓鼠。
3
2
92
刷到英俊哥把玄学类 Skill 整理成一串仓库,连“开口前先翻古籍”都安排上了。 赛博算命师也开始卷参考文献了。 不过这里有个很值得拆开的区别:引用没有编造,不等于结论就被证明了。 模型准确找到“某本书里这样写”,能说明它查到了什么。至于那套说法能不能预测现实,是另一道题,不能拿脚注直接抵答案。 这件事不只发生在玄学里。给 AI 塞一本公司手册、一堆产品文档,答案看起来立刻正规了。但过期的规定、互相矛盾的版本、只适用于特定条件的条款,也都能被引用得非常漂亮。 如果让我设计知识库助手,我想让它分开显示三件事: 原文怎么说。 这段话适用于什么范围。 从原文到这次回答,模型自己又推了哪一步。 尤其最后一步,别混进引用里装成作者说的。 这不是对上面那些仓库的实测评价。拿来读传统文化、做娱乐体验是一回事,把生成结果当成现实判断的依据是另一回事。 有出处,比没出处好核对。但“查得到这句话”和“这句话说得对”,中间还隔着一整套验证。 给胡说八道穿上参考文献西装,它也不会自动变成论文。
1
41
把访谈里的停顿全删掉,讲话是利索了,人也可能被剪成了另一种性格。 一句“我觉得……不太行”,那个犹豫未必是废料。 刚看到 Geek Lite 分享用 Claude Code 或 Codex 辅助视频剪辑的项目,介绍里提到删除停顿、口误、重复片段,以及只渲染改动部分供试听。 我很喜欢“先试听改动”这个方向,但如果让我提剪辑要求,会把“去废话”拆开说。 录了两遍、明确说错后重来的地方,可以优先列为候选。 想词时的空白、笑完没接上的半句话、回答之前的一次深呼吸,先别急着一刀切。这些有时候是在拖时间,有时候就是内容的一部分。 我想要的预览也不是孤零零播放剪口。把前后几秒一起留出来,再让我在原版和修改版之间切换,听听语气有没有变,回答是不是突然显得过分肯定。 这不是对那个项目的实测评价,只是我对 AI 剪辑助手的一个期待:不要默认所有沉默都是 bug。 剪掉无用的等待,和剪掉一个人的迟疑,不是同一份工作。视频可以更紧凑,但别把真实的人修成了永远秒答的客服。
1
2
67