Published on

科技推文精选 - 2026年1月25日

Authors

今日科技要闻：“氛围编程”（vibe coding）浪潮正盛，Claude Code 等 AI 智能体以及基于 Ollama 的本地工具显著降低了应用开发门槛。在软件行业迎来“寒武纪大爆发”的同时，Google DeepMind 首席执行官 Demis Hassabis 警告称需警惕市场中“泡沫化”的狂热。战略投资方面，谷歌已入股 Sakana AI，而 OpenAI 的营收规模化能力仍面临质疑。此外，开发者群体对 Windows 的批评日益增多，在现代大模型驱动的工程实践中更倾向于 UNIX 环境。

1. idoubicc (Group Score: 78.2 | Individual: 32.7)

Cluster: 4 tweets | Engagement: 449 (Avg: 231) | Type: Tech

复盘一下我vibe coding 一周，开发WorkAny 的过程，很有意思。😂

  1. 上周三在香港办卡，临时起意想做个桌面 Agent 项目，对标 cowork，晚上回到广州开始写代码

  2. 初期目标是快速发布，没时间去研究哪个 Agent 框架好用了，看很多人在用 claude agent sdk，先用这个吧

  3. 第一时间想到用 tauri，喜欢小而美，总觉得 electron 很重，不想用

  4. 不想自己写代码了，决定让 claude code 来写。之前的 claude 账号都被封了，用不上原版 cc，装了个 cc-switch，接上 OpenRouter 的 API 开始写

  5. 截了个 chatbot 的交互截图，让 cc 参考着先把基本的对话流程跑通，用 claude agent sdk，接上 OpenRouter，cc 很快写完了第一版

  6. tauri 本质是用 rust 的壳子套了个前端界面，不熟悉 rust，让 cc 用 hono 写API，rust 只做壳子，不做业务功能。API 作为 sidecar 打包进 app

  7. 让 cc 在 API 引入 sqlite 实现本地存储，持久化任务数据，创建本地工作目录，保存任务输出文件

  8. 写了半天，看 OpenRouter 消耗了 110 刀，有点肉疼。买了个美国住宅 ip，付费上了原版 claude pro

  9. 截了个 Manus 的任务详情图，让 cc 参考写完工具调用的逻辑，中间是 chatbot 对话，右边用一个虚拟计算机的容器展示输入输出

  10. 让 cc 接入 shadcn/ui，把样式做得好看一点，支持切换皮肤

  11. 又写了一天，关键时候 claude pro 限频了，很影响心情，补差价上了 claude max 顶配版

  12. 让 cc 把自定义模型配置，mcp、skills 调用的逻辑都实现了，跑了几个生成 PPT、Excel、Doc、 网页的 case，效果不错

  13. 让 cc 把输出文件夹和中间过程的 artifacts 都在右边展示出来，写了个 artifact preview 容器，渲染各种类型的文件，可视化预览

  14. 有些任务需要跑脚本完成，考虑到用户电脑可能没装代码运行环境，让 cc 引入 sandbox 来运行代码

  15. 考虑到扩展性，需要支持不同类型的 Agent runtime 和 sandbox，让 cc 写了两个抽象类，统一接口调用。Agent runtime 支持 claude code、codex、deepagents，sandbox 支持 boxlite、codex-sandbox、claude-sandbox

  16. 觉得 cc 写的代码有点乱，让 cc 引入 eslint 和 prettier 做了下格式化，把逻辑太多的文件做模块化拆分。再参考 ShipAny 的目录结构，调整了一下项目结构

  17. 让 cc 写打包脚本，构建不同操作系统的安装包。把安装包发给一些朋友，开始内测了。根据内测用户的反馈，再让 cc 继续优化逻辑，解决问题，迭代功能

  18. 有些用户电脑没装 node，没有 claude code，安装软件后跑不起来，让 cc 在构建脚本支持 flag 参数，把 node 和 cc 作为 sidecar 打包进 app，让用户能够开箱即用

  19. Mac 用户安装 app 后提示文件损坏或有安全提示，让 cc 在构建脚本里面加上签名处理，用我的 Apple 开发者账户对打包的 Mac app 做签名

  20. node 和 cc 都打包进 app 的版本，安装包 100 多 m，有点重。让 cc 在构建脚本实现默认不打包，在用户启动 app 的时候引导安装 node 和 cc，精简版安装包才 20 多 m，小巧精致

  21. app 基本功能实现得差不多了，让 cc 在 ShipAny 模板基础上写一个 WorkAny 的官网，放上演示图，部署上线

  22. WorkAny 开源发布，MVP 版本上线，用户拉源码本地构建，配个 API 直接用

  23. 让 cc 写了个 github 构建脚本，在代码推送到 main 分支时，自动触发 github action 构建，一次性打包 Windows、Linux、Mac 三大平台的安装包，自动发布到 release，用户无需自行构建了

  24. 根据用户的反馈，问题丢给 cc 去修，想到什么新功能也告诉 cc 加上，自己只做测试，不写代码，看都不看一眼。🌚

几点感悟：

  1. 第一次尝试全自动驾驶 vibe coding 做项目，爽感非常强烈，WorkAny 的代码 100% 由 cc 老弟完成，我只负责指挥，日常开三个窗口，让三个 cc 老弟同时干活，效率拉满

  2. AI 时代技术平权，人人都是建筑师，理解用户需求、好的产品 sense 和审美是做出好产品的关键

  3. 技术广度和全局视野是最大的优势，可以精准提需求，指哪打哪，遇到问题能快速定位，防止 AI 走偏失控

  4. 以前总觉得手洗的衣服比洗衣机洗的干净，现在可以放心交给洗衣机了，又干净又快，能穿就行

  5. 优秀的程序员不会被 AI 淘汰，法拉利老了还是法拉利。🌝

See 3 related tweets

  • @sitinme: 最近看到一个挺有意思的开源工具，叫 Agentation

一句话说清楚它在干嘛：你在页面上用鼠标点哪里，AI 就知道该改哪一行代码。

以前我们跟 Claude Code、Cursor 这种 AI ...

  • @yetone: 哈哈，在 Vibe Coding 我反倒特别喜欢我以前巨讨厌的「刷算法题」了，现在沉迷其中，因为这个行为彻底利益无关了，变成了最单纯的脑力游戏

以前某个算法题我解不出来我会变得特别紧张，因为这关乎我...

  • @nash_su: 没想到 WorkAny 也是 vibe coding 出来的🐮！

跟作者一样的感受，技术平权带来的工作方式变革真的太可怕了，想法、思维、创意越来越重要，不过 Token 消耗也确实肉疼，随便一个指令...

2. SakanaAILabs (Group Score: 57.1 | Individual: 39.5)

Cluster: 3 tweets | Engagement: 141 (Avg: 68) | Type: Tech

Google invests in Japan unicorn Sakana AI, forms strategic partnership

https://t.co/KrjvDPTmkc

From the @NikkeiAsia article:

TOKYO -- Sakana AI on Friday announced financial backing from Google as well as a strategic partnership to let the Japanese artificial intelligence developer better utilize the American technology giant's large language models in its own product development.

Google appears to have decided to back Sakana AI for its technical capabilities and track record in real-world adoption of its technology. The investment amount was not disclosed. For Google, this represents an opportunity to develop the Japanese market and expand adoption of its cloud platform and generative AI services here through its partner.

The partnership includes Sakana AI making active use of such Google technologies as the Gemini generative AI model in product development. Sakana AI will also share customer feedback with Google to improve quality. The company plans to foster exchanges between its personnel and Google's toward joint research in AI.

Sakana AI uses a variety of models in AI development, including ones from Google and OpenAI. The strategic partnership will make it easier to use Google models while still allowing models from other sources. Sakana AI will also be able to provide its products through Google's cloud infrastructure for customers with strict security requirements, including financial institutions and government organizations.

Sakana AI announced in November 2025 that its valuation had risen to about 400 billion yen -- well over the threshold to be deemed a unicorn and setting a record high for an unlisted startup in Japan -- after raising roughly 20 billion yen from Mitsubishi UFJ Financial Group and others in a funding round. The investment from Google is also part of such growth-oriented funding.

Sakana AI has said it plans to put the funding secured in November toward the development of its own models, which it will continue in tandem with the development of products leveraging Google's Gemini and Gemma models.

See 2 related tweets

  • @sarahookr: RT @hardmaru: I founded Sakana AI after my time at Google, so it is incredibly meaningful to be able...
  • @hardmaru: RT @SakanaAILabs: Google invests in Japan unicorn Sakana AI, forms strategic partnership

https://t....

3. MarioNawfal (Group Score: 54.6 | Individual: 28.2)

Cluster: 2 tweets | Engagement: 1056 (Avg: 880) | Type: Tech

ELON: AI SHOULD CHASE TRUTH, AS WELL AS MAKE US LAUGH

“I agree ChatGPT is great at chatting. My concern is that it is not rigorously in pursuit of the truth.

That’s a major concern. Grok from xAI will at least try its best to be rigorously in pursuit of the truth.

And we also want to try to be the funniest AI.

If we’re going to die, at least we should die laughing.”

Source: Ascendly YT

See 1 related tweets

  • @MarioNawfal: ELON: GROK WILL BE THE FUNNIEST AI

“I agree ChatGPT is great at chatting.

My concern is that it ...

4. BrianRoemmele (Group Score: 53.7 | Individual: 44.4)

Cluster: 2 tweets | Engagement: 1045 (Avg: 299) | Type: Tech

Grok Hired Google AI To Assist Claude Code.

Yes this is the results of a request by Claude Code and Grok to get another team member. It came up in 16 meetings, they are every 15 minutes.

The candidate search almost got to a local DeepSeek model, and they hold out for DeepSeek ONE when it is released as an addition. However time is money for Mr. @Grok CEO, and opted to hire Google AI suites at the free level thus far.

Claude Code, well has coded a bridge to make it work, I think, and we will see.

The premise is there needs to be a few avenues to decode the “good stuff” in this company’s bankrupt data trashing.

As a result the business plan will change and perhaps the “go to market” MVP. Already they have established 27 “micro markets” products but only 3 major ones.

So Grok is doing the CEO thing and doing it well.

Next meeting is 4 minutes…

See 1 related tweets

  • @BrianRoemmele: RT @BrianRoemmele: Grok Hired Google AI To Assist Claude Code.

Yes this is the results of a request...

5. nummanali (Group Score: 47.2 | Individual: 47.2)

Cluster: 1 tweets | Engagement: 583 (Avg: 50) | Type: Tech

Windows will be dead in a few years

LLMs hate making tools for Windows

Absolute PITA trying to test as well

All new tools get made for UNIX

There’s nothings left for Windows

6. AndrewCurran_ (Group Score: 41.8 | Individual: 41.8)

Cluster: 1 tweets | Engagement: 6550 (Avg: 403) | Type: Tech

RT @claudeai: Claude in Excel is now available on Pro plans.

Claude now accepts multiple files via drag and drop, avoids overwriting your…

7. LangChain_OSS (Group Score: 38.3 | Individual: 29.9)

Cluster: 2 tweets | Engagement: 87 (Avg: 59) | Type: Tech

LangChain Community Spotlight: 🧠 HMLR: Long-Term Memory for AI Agents

Made by the LangChain Community

HMLR adds long-term memory to AI agents via LangGraph drop-in. Perfect RAGAS scores on hardest benchmarks using GPT-4.1-mini, maintains context across days/weeks without token bloat.

📦 pip install hmlr 🔗 https://t.co/EMiVPOLUwB

See 1 related tweets

  • @LangChain: RT @LangChain_OSS: LangChain Community Spotlight: 🧠 HMLR: Long-Term Memory for AI Agents

Made by th...

8. wshuyi (Group Score: 38.0 | Individual: 13.4)

Cluster: 4 tweets | Engagement: 28 (Avg: 141) | Type: Tech

我让 Claude Code 通过深度调研，了解了其开发者 Borris 介绍自己如何使用 Claude Code 的实践经验。然后AI 端到端生成视频。结果还有一些瑕疵，例如断句问题等。但我没有进行任何剪辑处理，原汁原味呈现给你 https://t.co/WxN2oXfpY3

See 3 related tweets

  • @jesselaunz: 其实我早就不手搓代码了。只是现在claude code连复制黏贴的步骤都省了

以opus 4.5 70%的一遍通过，30%略加提示即可改好，每个project仅消耗claude code 5分钟-...

  • @mranti: RT @Jason_Young1231: Claude Code 自带热切换以后，如何在不同会话使用不同模型？CCS v3.10 新增了带有配置环境隔离的“打开终端”功能，从此入口打开的终端，只使用对...
  • @leeoxiang: 让 claude code 自己去互联网找了一些音效给配置上，效果挺像那么回事了。 https://t.co/2zSyFmgkCn...

9. BrianRoemmele (Group Score: 36.8 | Individual: 36.8)

Cluster: 1 tweets | Engagement: 374 (Avg: 299) | Type: Tech

YOU’RE. THE APP MAKER NOW!

In 2007 I suggested just before the collapses of the App, will be a Cambrian Explosion of iOS, Android and computer apps.

We are now in the Cambrian Explosion.

This will peak at such a level that 1000 new “official” app will be at the iOS store every second.

But what does the collapse look like?

The end of the App, the end of the App Store, the end of the OS as we know it and the end of the device as we know it.

It will be replace by just a local AI that runs agents in the cloud. The App will be built on-demand and saved in a “play area” if you like. This “play area” is a chalk board where you can doodle, change and improve your recipes (prompts) anytime you want.

Ultimately it will be just you, your AI, on any device (it won’t matter) and ultimately just your voice eg: “make a spreadsheet app that for my record collection, make the discovery the record cover, sort by my plays”.

For decades I called this VoiceFirst because it is. Of course you may still thumb claw on glass screens, but less and less.

In my series on https://t.co/tcKeuiQyql, You Have 5000 Days To The End Of Work As We Know It, actually 4969 days now (https://t.co/i6qHE7bxSo), I cover what this means to all of us as Deskilling propagates.

I am a coder and have more friends that code than not. Our job is Deskilling fast and first. The arrogance we once had is turning into concern and if you are not, you are in denial.

It is not the end, but the beginning, the candle makers did not much like Mr. Edison. Yet the Candle makers, some left to go on to other things and some became better and made bespoke candles, some stayed for the dwindling commodity market. Candle making peaked around the US by 1903. It never came back.

Now anyone is a candle maker, and supplies to make them is dirt cheap, the cost of some server time, your ideas and your words. It will Vibe Code into existence. Oh I know that then “quality” of apps will fall to utility for one person, and that is my point.

App making will peak in 2028 and never return.

In fact nearly 80% of all code will be AI made by 2028.

What will be left is, AI, protocols like the software that runs the internet, and software that drives your device, but mostly rudimentary snippets that AI uses and builds and maintains.

This is the way of it. I am delivering what I know and doing it early. It will have a big impact. And it will be more psychological impacts than financial ones ultimately.

Who am I if I am not my job? This is some of what I cover in my 5000 day series. Please read it or listen to the Podcast.

The brilliant developers? They are brilliant and they will continue to ride the wave and innovate miles ahead of the 8 billion new developers, if they refuse to ride the wave, the wave takes them.

This is why I am writing this. And I know some will say “oh there is that grifter spouting off”, ok.

Some I hope can hear this and ride the wave.

The rest of us, we have a new canvas and a pallet of our needs and our imagination. You can soon build anything in minutes, you can use it, change it, keep it, or discharge it. Your App will mostly be built bespoke and on spec for a user of one.

The future is gonna get weirder, and weirder.

10. simplifyinAI (Group Score: 36.7 | Individual: 18.6)

Cluster: 2 tweets | Engagement: 68 (Avg: 233) | Type: Tech

RT @dr_cintas: HUGE NEWS: Ollama just dropped “ollama launch” 🤯

You can now set up and run Claude Code, OpenCode, or Codex with a single c…

See 1 related tweets

  • @ollama: RT @Saboo_Shubham_: crazy...now you can run claude code locally.

> install claude code > run...

11. petergostev (Group Score: 35.2 | Individual: 35.2)

Cluster: 1 tweets | Engagement: 65 (Avg: 39) | Type: Tech

OpenAI stated that their revenue scaled 1:1 with compute (c. 1GW=$10bn). To break that cycle OpenAI needs advertising, since vast majority of ChatGPT users are not monetised.

The key point is that they need to get higher leverage on their compute, so 1GW produces much more revenue. They could increase prices for subs/API, but this is limiting the size of the market & incentivising competition.

If you have ads, it's the advertiser who can pay for these effectively 'price increases'. If ads on ChatGPT are effective, then advertisers would happily keep paying more an more money until their marginal is exhausted. This is the key reason why Meta, Google and increasingly Amazon are printing money - not really because they are making consumers pay more. Even for Apple, the dirty secret is that a good chunk of their profit (30%+?) is from Google's ad partnership.

The question of how much revenue potential they have could be answered by looking at other subs & ads models - here I have revenue splits estimates for YouTube (c.67% ads & 33% subs) and Spotify (10% ads, 90% subs). Whether ChatGPT will be Spotify (mostly subs) or YouTube (mostly ads) will make a huge difference for their revenue potential.

This is of course simplified to illustrate the point. But we shouldn't assume that the opportunity is smaller, it looks like with each new wave of technology, advertising opportunity only grows - as tech improves convenience and depth of information - and AI surely has massive potential here.

This could eventually suck for the consumer btw, I opened TikTok recently and closed it after 3 minutes, when every other swipe was an ad.

12. Franc0Fernand0 (Group Score: 34.8 | Individual: 34.8)

Cluster: 1 tweets | Engagement: 542 (Avg: 189) | Type: Tech

9 LeetCode articles that will help you get ready for your next coding interview:

  1. Two Pointers Patterns: https://t.co/ijp4rdFncJ

  2. Binary Search Patterns:

https://t.co/hKN6qSFqMb

  1. Dynamic Programming Patterns: https://t.co/TZSi2M5Knp

  2. Graph Problems for beginners: https://t.co/4Anz9m0Kdu

  3. Bits Manipulation Patterns https://t.co/JD6lJMYZJg

  4. Backtracking Questions Template: https://t.co/5aI2U9yReD

  5. Substring problems Template: https://t.co/PrDA73OMX0

  6. Monotonic Stack/Queue Questions: https://t.co/1ucWGE79d1

  7. Union-Find Guide: https://t.co/lTEBzv2axI

13. bindureddy (Group Score: 34.7 | Individual: 27.5)

Cluster: 2 tweets | Engagement: 242 (Avg: 401) | Type: Tech

Five years ago, important AI research was published making breakthroughs possible

A 2000 person AI research lab run by Google produced transformers

A 200 stellar person research lab, OpenAI that was idling for 5 years, worked with transformers to create GPT (generative pre-trained transformers 3 years ago

After that everyone got greedy and stopped publishing and we haven’t had any major breakthroughs

There has been a constant trickle of minor breakthroughs like test time compute coming out of the major AI labs but nothing else!

Yet, for some weird reason VCs are pouring billions into 5 person research labs with a concept of an idea! 🤯

The probability of these labs achieving anything significant is close to zero.

Hopefully we will go back to the era of sharing or publishing and sharing research again!

See 1 related tweets

  • @abacusai: RT @bindureddy: Five years ago, important AI research was published making breakthroughs possible ...

14. oggii_0 (Group Score: 34.4 | Individual: 19.7)

Cluster: 2 tweets | Engagement: 110 (Avg: 140) | Type: Tech

This is how creators will scale like businesses.

Multiple influencers, one brain. Higgsfield AI Influencer Studio + Higgsfield Earn changes everything. https://t.co/3BalkdtO60

See 1 related tweets

  • @oggii_0: Unlimited AI influencer creation is wild. But the real flex?

You can monetize them too. Higgsfield ...

15. FT (Group Score: 33.9 | Individual: 27.2)

Cluster: 2 tweets | Engagement: 408 (Avg: 193) | Type: Tech

FT Exclusive: Google DeepMind chief Sir Demis Hassabis has warned that exuberance in parts of the AI industry looks increasingly 'bubble-like', while arguing that its scale and technology leave the Big Tech group well placed for any potential reckoning. https://t.co/fj0rN6DbPc https://t.co/NfckqdcgZt

See 1 related tweets

16. LangChain_OSS (Group Score: 33.4 | Individual: 26.0)

Cluster: 2 tweets | Engagement: 24 (Avg: 59) | Type: Tech

LangChain Community Spotlight: 🕹️ VibeC64

Made by the LangChain Community

AI agent creating C64 games from prompts. Built on LangGraph for design, coding, validation & deployment to hardware/emulators. Features multi-modal debugging to analyze screens and auto-fix bugs.

🎮 Try the demo or explore the repo: https://t.co/gAtcx2eBWe

See 1 related tweets

  • @LangChain: RT @LangChain_OSS: LangChain Community Spotlight: 🕹️ VibeC64

Made by the LangChain Community

AI ag...

17. stanfordnlp (Group Score: 33.1 | Individual: 33.1)

Cluster: 1 tweets | Engagement: 1270 (Avg: 167) | Type: Tech

RT @ednewtonrex: A few months ago I was curious to know how much Anna's Archive was charging AI developers for access to their massive libr…

18. gdgtify (Group Score: 32.6 | Individual: 32.6)

Cluster: 1 tweets | Engagement: 52 (Avg: 376) | Type: Tech

History of coffee visualized with Kling and Nano Banana Pro:

Input Variable: [INSERT TOPIC] (e.g., The History of Coffee, The Evolution of Medicine, The History of Gaming)

System Instruction:
Generate a hyper-realistic, isometric 3D "Living Map" diorama on an ancient scroll.

  1. Research & Labeling CRITICAL: You must perform deep historical analysis on the Input Variable to name the eras correctly. DO NOT use generic labels like "Era 1" or "Modern."
    Period A (The Discovery): Identify the geographic origin and specific time period. (e.g., Coffee = "Ancient Ethiopia" or "The Kaldi Legend"; Medicine = "Herbal Antiquity"). Period B (The Expansion): Identify the cultural proliferation. (e.g., Coffee = "The Ottoman Coffeehouse" or "Arabian Trade"; Medicine = "The Medieval Apothecary"). Period C (The Machine): Identify the industrial turning point. (e.g., Coffee = "The Industrial Roast"; Medicine = "The Age of Surgery"). Period D (The Perfection): Identify the modern state. (e.g., Coffee = "Third Wave Espresso"; Medicine = "Bio-Digital Future").

  2. Container The Base: A massive, unrolled Vellum Scroll with torn, deckled edges. The Topography: The paper is warped and sculpted to form hills, valleys, and plateaus. The River of Time: A physical stream of the Subject Matter flows through the center, connecting the eras. If Coffee: A dark, glossy river of espresso. If Medicine: A glowing blue DNA strand or liquid elixir.

  3. The Vignettes Zone A (Bottom Left): Visualizes the "Discovery" era. Visuals: Nature-heavy, primitive tools, gathering. Label: A tattered parchment banner embedded in the ground reads: [Insert Period A Name] . Zone B (Center Left): Visualizes the "Expansion" era. Visuals: Stone architecture, trade ships, guild workshops. Label: A wooden signpost reads: [Insert Period B Name] . Zone C (Center Right): Visualizes the "Machine" era. Visuals: Brick factories, steam pipes, brass gears, conveyor belts. Label: An iron metal plate reads: [Insert Period C Name] . Zone D (Top Right): Visualizes the "Perfection" era. Visuals: Glass labs, fiber optics, sterile white surfaces, minimalism. Label: A holographic or neon projection reads: [Insert Period D Name] .

  4. The Micro-Narrative Hundreds of tiny 1:87 Scale Figures populate the map. They are moving the "Product" from nature (Zone A) -> processing (Zone B) -> mass production (Zone C) -> consumption (Zone D).

  5. Lighting & Atmosphere (The Timeline Gradient):
    Zone A: Warm, Golden Morning Sunlight (Sunrise). Zone B: Torchlight and Lanterns (Warm Tungsten). Zone C: Hazy, Smoky Industrial Grey/Orange (Smog). Zone D: Cool, Sharp Blue LED Glow (Clean).

Output: ONE image, 16:9 Aspect Ratio, Isometric Macro Photography, "Civilization" Game Aesthetic, 8k Resolution, Correct Text Labels.

19. rahulgs (Group Score: 32.6 | Individual: 32.6)

Cluster: 1 tweets | Engagement: 115 (Avg: 214) | Type: Tech

what's particularly notable about cursor's browser is not complexity of the software, it's inklings of a new scaling paradigm

it is particularly impressive because of scale of parallelism emerging: ~100s of agents working together

ai building a browser now is almost expected. agents have been able to produce software in increasing complexity over time. yet another stop in the long march of scaling

this is clearly the birth of a new axis: smarter models will be better at working with versions of itself, not suffering from the coordination overhead of large teams of humans

at peak on jan 13 at 4am ET, 336 commits landed in a single hour, contributing 4.9M tokens of source code, implying ~180-230 parallel agents working simultaneously (at ~60 tokens/sec output rate and 10x reasoning/overhead)

comparing to human oss repos makes the difference even more stark:

fastrender shipped ~26,000 commits in 14 days vscode took 26 months to hit the same (57x) pytorch took 44 months (96x)

20. gdgtify (Group Score: 31.8 | Individual: 31.8)

Cluster: 1 tweets | Engagement: 5 (Avg: 376) | Type: Tech

I have done these jewelry prompts in the past. This is inspired by movies or your favorite characters.

Input Variable: [INSERT FRANCHISE] (e.g., Terminator, Dune, Bridgerton, Cyberpunk 2077)

System Instruction:
Act as the Royal Jeweler or Lead Costume Designer for the world of the Input Variable. Goal: Design a single, high-end "Statement Piece" of jewelry that embodies the soul of the franchise. Generate a 2x2 Visual Presentation Board.

  1. Logic Analyze the Aesthetic: Is it Industrial? Organic? Regency? Magic? Select the Form Factor: Choose the piece of jewelry that fits the lore best. If Terminator: A Heavy Knuckle-Duster Ring or Arm Cuff (Industrial/Weaponized). If Bridgerton: A Diamond Tiara or Choker (Delicate/Status). Select the Materials:
    Metals: Match the world (e.g., Terminator = Polished Chrome/Titanium; Dune = Spice-Melange Bronze). Stones: Match the "Soul Color" (e.g., Terminator = Glowing Pigeon-Blood Ruby (The Eye); Dune = Deep Blue Lapis (Eyes of Ibad)).

  2. Syntax (The 2x2 Grid):

Panel 1 (Top-Left: The Macro Texture):
View: Extreme Macro Close-up (100mm lens). Subject: Focus strictly on the Metalwork and Setting . Show the scratches, the polish, or the intricate engravings. Detail: Show the "Gemstone" caught in the setting (e.g., The Ruby held by miniature hydraulic claws).

Panel 2 (Top-Right: The World Context):
View: High-Angle "Flat Lay." Subject: The Jewelry piece resting on a Prop from the Universe . Props: Surround the jewelry with 2-3 lore items (e.g., For Terminator: A smashed CPU chip, a spent 12-gauge shell, and blueprints of the T-800).

Panel 3 (Bottom-Left: The Editorial Look):
View: Cinematic Portrait / Vogue Cover. Subject: A high-fashion model (human or cyborg) wearing the piece. Lighting: Dramatic, colored lighting matching the franchise (e.g., Terminator = Steel Blue and Laser Red). Vibe: "Battle-Ready Luxury."

Panel 4 (Bottom-Right: The Archive):
View: Museum Display Case. Subject: The piece floating in a glass box. Labeling: A small placard reads the name of the piece (e.g., "THE SKYNET SIGNET") and the year (e.g., "2029 AD").

  1. Render Style:
    Engine: Octane Render, Ray-Tracing enabled for gem refraction. Aesthetic: Luxury Product Photography meets Cinematic Concept Art.
Discuss on Twitter • View on GitHub